Strategies to improve dairy cow productivity and welfare in Vietnam

Ngoc Bang Nguyen

MSc. / BSc.

A thesis submitted for the degree of Doctor of Philosophy at

The University of Queensland in 2021

School of Veterinary Science Abstract

In Vietnam, smallholder dairy farms (SDFs) account for more than 97% of the total national dairy herd, and the number of SDFs has increased rapidly recently due to the high demand for fresh milk. While studies aiming to improve the productivity and welfare of cows in SDFs could help increase national milk production, such studies are currently extremely limited due to the lack of available data about dairy cow husbandry management, breed, pedigree, and performance. The aims of this project were to describe the productivity of Vietnamese SDF cows and their welfare in typical dairying regions and to determine and prioritise potential nutrition, breeding, and management strategies to improve the productivity and welfare of those SDF cows.

A total of 32 SDFs from four key contrasting dairy regions (8 farms per region) of Vietnam were chosen. The regions were those with highland/cooler climate: a southern highland region and a northern highland region; and those with lowland/hotter climate: a southern lowland region and a northern lowland region. Each farm was visited over an afternoon and during the following morning in autumn 2017 to collect necessary farming data and single-day data of all lactating cows. The farming data, including altitude, number of cattle, cow diets, feeding regimes, shed dimensions, heat stress abatement methods, and shed microclimate conditions, were obtained directly on the farms. Feed ingredients used for cow diets were sampled for chemical composition analysis. Single-day data of all lactating cows (n = 345), including milk yield, milk compositions, energy corrected milk yield (ECM), milk electrical resistance, body weight (BW), body condition score (BCS), number of inseminations per conception, panting score (PS, 4.5-point scale, 0 = normal, 4.5 = extremely heat- stressed) were measured directly on farms to be used as indicators of cow productivity and welfare. Genomic data of lactating cows were obtained by taking tail hair samples, extracting DNA, and genotyping with GGP bovine 50K chips. Infrared temperatures (IRTs) of lactating cows were measured at the inner vulval lip using an infrared thermometer, and measured at another ten different external areas, including: the outer vulval surface, inner tail base surface, ocular area, muzzle, armpit area, paralumbar fossa area, fore udder, rear udder, forehoof, and hind hoof, using an infrared camera. Data from all 32 farms were used for seven studies. The first study described herd structure and evaluated the productivity and welfare status of the cows. The second used genomic data to identify genetic diversity and breed composition. The third and fourth studies described and compared feeding regimes, diets, shed designs, heat stress abatement methods, and shed microclimate of the 32 farms in the four visited regions to evaluate the strengths and limitations of the cows’ nutrition and the heat stress management of each farm. The fifth study built multivariate models to identify nutrition, management and animal factors that were most associated with the productivity and welfare of the

I cows. The sixth study tested the capability of IRTs relative to PS to assess the level of heat stress and the reduction in the cows’ milk yield under heat stress conditions. The seventh and last study tested the application of genomic selection and genome-wide association studies on milk production, body conformation (BW and BCS), and heat tolerance straits (PS and IRTs) of the SDF herds.

The average herd size across regions was 29 animals comprising 11 lactating cows, four dry cows, six heifers, seven female calves, one male calf, one beef cattle, and no reproductively active bulls. All the cows were genetically very close to pure Holstein and retained a similar level of genetic diversity as the reference Holstein populations used in the USA, New Zealand, and France. The genetic proportionality of Holstein, Jersey, Brown Swiss, and Zebu, averaged across all herds, were 85.0%, 6.0%, 5.3%, and 4.5%, respectively. The most critical production and welfare issues were the high level of heat stress (96% of cows were moderate to highly heat-stressed in the afternoon), relatively low ECM (15.7 kg/cow/d), low BCS (2.8, 5-point scales), a high number of inseminations per conception in the lowland SDFs (1.9 to 3.2 times per conception), insufficient supply of water, insufficient floor allowance per cow (8 m2/cow), and excessive use of “tie” rather than “loose” stalls or pens for cows in sheds. From among the four regions, the cows in the southern lowland were least productive and had the poorest welfare conditions. The lactating cow diets in all regions were deficient in net energy and excessive in crude protein and fibre concentrations, which were associated with low milk yield and low feed efficiency in all regions. Increases in farm altitude, shed roof height, floor allowance for cows, percentages of cowshed sides open, and the use of roof soakers and fans were all associated with improved cowshed microclimate, reduced PS, decreases in the number of inseminations per conception, and improved milk production. All infrared temperatures showed potential to assess the level of heat stress in the cows. The infrared temperatures of the inner vulval lip, outer vulval surface, and inner tail base surface showed the best potential to predict ECM reduction during heat stress. Under SDF conditions, Holstein_BrownSwiss and Holstein_Jersey first crosses tended to be more productive and less heat-stressed than backcrosses of Holstein and Zebu with 15/16 Holstein purer. The moderate to high heritabilities (0.26 to 0.58) and moderate accuracies (0.23 to 0.50) of genomic estimated breeding values (GEBVs) for some milk production traits (ECM adjusted for BW and milk fat and protein concentrations), body conformation traits (BCS and BW), and heat tolerance traits (PS and IRTs at the rear udder, fore hoof and outer vulval surface) suggested the potential to apply genomic selection using single-test day phenotypic measurements for these traits. However, a larger sample size is required. Novel single nucleotide polymorphisms and genes for PS and infrared temperatures were located on chromosome (BTA) 5, BTA8, BTA11, BTA14, BTA19, and BTA20.

II

Overall, the most important short-term strategies to improve the productivity of SDF cows are heat stress amelioration and the measurement thereof using new technologies such as IRT. Next are nutritional strategies which include supplying ad libitum drinking water and the reformulation of cow diets by increasing available energy whilst decreasing fibre concentrations to meet the requirements of the cows in heat stress conditions. Long-term strategies could include the application of genomic selection to identify bulls and heifers with high genetic merit for milk production, heat tolerance and other key traits, from within the current Vietnamese SDF herd.

III

Declaration by author

This thesis is composed of my original work, and contains no material previously published or written by another person except where due reference has been made in the text. I have clearly stated the contribution by others to jointly-authored works that I have included in my thesis.

I have clearly stated the contribution of others to my thesis as a whole, including statistical assistance, survey design, data analysis, significant technical procedures, professional editorial advice, financial support and any other original research work used or reported in my thesis. The content of my thesis is the result of work I have carried out since the commencement of my higher degree by research candidature and does not include a substantial part of work that has been submitted to qualify for the award of any other degree or diploma in any university or other tertiary institution. I have clearly stated which parts of my thesis, if any, have been submitted to qualify for another award.

I acknowledge that an electronic copy of my thesis must be lodged with the University Library and, subject to the policy and procedures of The University of Queensland, the thesis be made available for research and study in accordance with the Copyright Act 1968 unless a period of embargo has been approved by the Dean of the Graduate School.

I acknowledge that copyright of all material contained in my thesis resides with the copyright holder(s) of that material. Where appropriate I have obtained copyright permission from the copyright holder to reproduce material in this thesis and have sought permission from co-authors for any jointly authored works included in the thesis.

IV

Publications included in this thesis

Since thesis submission, the following papers have been published:

Bang NN, Chanh NV, Trach NX, Khang DN, Hayes BJ, Gaughan JB, Lyons RE, Hai NT and McNeill DM (2021). Assessment of performance and some welfare indicators of cows in Vietnamese smallholder dairy farms. Animals, 11 (3), 674. https://doi.org/10.3390/ani11030674 (Chapter 3).

Bang NN, Chanh NV, Trach NX, Khang DN, Hayes BJ, Gaughan JB, Lyons RE, Hai NT and McNeill DM (2021). Issues of feeding strategy for lactating cows in Vietnamese smallholder dairy farms. Animals, 11 (3), 729. https://doi.org/10.3390/ani11030729 (Chapter 5).

Bang NN, Gaughan JB, Hayes BJ, Lyons RE, Chanh NV, Trach NX, Khang DN and McNeill DM (2021). Characteristics of cowsheds in Vietnamese smallholder dairy farms and their associations with microclimate - A preliminary study. Animals, 11 (2), 351. https://doi.org/10.3390/ani11020351 (Chapter 6).

Submitted manuscripts included in this thesis

Since thesis submission, the following manuscript has been submitted for publication:

Bang NN, Hayes BJ, Lyons RE, Randhawa IAS, Gaughan JB and McNeill DM (2021). Genomic diversity and breed composition of Vietnamese smallholder dairy cows. Submitted for publication at Journal of Animal Genetics and Breeding (Chapter 4).

Other publications during candidature

Peer-reviewed papers

Ngo TT, Bang NN, Dart P, Callaghan M, Klieve A, Hayes B and McNeill DM (2021). Feed preference response of weaner bull calves to Bacillus amyloliquefaciens H57 probiotic and associated volatile organic compounds in high concentrate feed pellets. Animals, 11 (1), 51. https://doi.org/10.3390/ani11010051.

Bang NN, Hayes BJ, Randhawa IAS, Lyons RE, Gaughan JB, Nguyen CV, Nguyen TX, Nguyen KD and McNeill DM (2019). Application of genomic selection in Vietnamese dairy herd. Proc. Assoc. Advmt. Anim. Breed. Genet. 23, 294–297.

Conference abstracts

Bang NN, Gaughan JB, Nguyen CV, Nguyen HT, Nguyen TX, Nguyen KD, Hayes BJ, Lyons RE and Mcneill DM (2019). Heat stress and productivity in lactating Vietnamese household dairy cows.

V

In Book of Abstracts of the 70th Annual Meeting of the European Federation of Animal Science, p. 611. Wageningen Academic Publishers.

Bang NN, Nguyen C V, Nguyen HT, Nguyen TX, Nguyen KD, Nguyen SH, Nguyen HT, Gaughan JB, Lyons RE, Hayes BJ and Mcneill DM (2019). Feeding value of lactating cow diets popularly used by Vietnamese household dairy farms. In Book of Abstracts of the 70th Annual Meeting of the European Federation of Animal Science, p. 557. Wageningen Academic Publishers.

Ngo TT, Bang NN, Dart PJ, Klieve AV, Callaghan MJ and McNeill DM (2018). Diet preference and ruminal pH effects associated with irradiated versus live probiotic spores of Bacillus amyloliquefaciens H57 in pellets for steers. In Proceedings of the 18th Asian-Australian animal production congress 1-5 Aug 2018, p. 170.

Ngo TT, Bang NN, Dart PJ, Klieve AV, Callaghan MJ and McNeill DM (2018). Ruminal pH and diet preference response to Bacillus amyloliquefaciens H57 probiotic in steers. In Proceedings of the 18th Asian-Australian animal production congress 1-5 Aug 2018, p. 329.

Ngo TT, Bang NN, Lisle AT, Dart PJ, Klieve AV, Callaghan MJ and McNeill DM (2018). One-page abstract: Bacillus amyloliquefaciens H57 alters diet preference and ruminal pH in steers. Animal Production Science 58, XXXIX.

Contributions by others to the thesis

Primary advisor Dr David McNeill and associate advisors Prof Benjamin Hayes, Dr Russell Lyons, Assoc Prof John Gaughan contributed to the concept and design of the project. Dr David McNeill and Prof Nguyen Xuan Trach acquired financial support for this research project before the commencement of the studies. A critical review of the thesis was provided by all advisors, especially Dr David McNeill. The adoption and validation of data-collecting procedures were contributed by Dr David McNeill. Professional advice on analysing the data of Chapter 4 and 9 were significantly contributed by Prof Benjamin Hayes and Dr Imtiaz Randhawa. Prof Nguyen Xuan Trach, Prof Nguyen Duong Khang, and Assoc Prof Bui Thi Nga assisted in getting permissions from local authorities to collect data in regions and establishing relationships with smallholder dairy farmers. Ms Nguyen Thi Thanh Thuy, Dr Nguyen Van Chanh, Mr Nguyen Thanh Hai contributed to data collection. Dr Nguyen Thi Huyen assisted in analysing dry matter of feed samples and sending feed samples to DairyOne Lab for feed chemical composition analysis. Ms Vu Thi Mai Lien contributed significantly to importing data into Excel files. UQ and MARD funded scholarships, and ACIAR funded the research project.

VI

Statement of parts of the thesis submitted to qualify for the award of another degree

No works submitted towards another degree have been included in this thesis.

Research Involving Human or Animal Subjects

All research involving human involvements was approved by the University of Queensland Human Research Ethics Committee A, approval number 2016001815 (Appendix 1).

All research involving animals was approved by the University of Queensland Animal Ethics Unit, approval numbers ANRFA/SVS/565/16/VIETNAM (Appendix 2 a) and SVS/010/18/VIETNAM (Appendix 2 b).

VII

Acknowledgements

I would like to express my utmost appreciation to my principal advisor Dr David McNeill and my associate advisors Prof Benjamin Hayes, Dr Russell Lyons, and Assoc Prof John Gaughan, for their mentorship, continual support, constructive feedback, and friendship during my doctoral candidature. I would especially like to thank Dr David for firstly providing me with the opportunity to be a postgraduate student at the Veterinary Science School and for being my principal advisor during my PhD study. For me, Dr David is not only an advisor but also a friend who patiently listened to, constantly advised, and continually encouraged me during my candidature. I especially thank Prof Ben and Dr Russell, who inspired me to find my love for genetics and statistics, and Prof John, who has always been a close and supportive advisor to me throughout my Masters and PhD studies. Also, I thank all advisors for allowing me to conduct this research with a level of independence that has helped me realise and develop my potential. I was fortunate and proud to be your student.

I am deeply grateful to the Australian Centre for International Agricultural Research (ACIAR) for funding this research project; and UQ-MARD for granting me the scholarships. I could not have finished this thesis without those funds. I am very grateful to the smallholder dairy farmers in Ho Chi Minh, Lam Dong, Ha Nam, and Son La provinces, who warmly welcomed me and provided the best conditions for me to collect necessary data. I gratefully acknowledge the technical and administrative staff at the School of Veterinary Science (UQ), Faculty of Animal Science (Vietnam National University of Agriculture-VNUA), Faculty of Animal Science and Veterinary Sciences (Nong Lam University-NLU), and Moc Chau Dairy Cattle Breeding JSC, especially Vice director Mr Nguyen Hai Nam and staff members Mr Bui Duy Quang and Mr Nguyen Quyet. Many thanks to Neogen Australasia Lab for providing training and support throughout the genetic sample analysis in the laboratory. Also, many thanks to the US Department of Agriculture and the University of California Davis for allowing me the right to access the Vietnamese version of the PCDairy software, which is used in Chapter 5.

I would also like to take the opportunity to thank my milestone review panel Prof Jenny Seddon, Assoc Prof Peter Murray, and Dr Di Mayberry, for providing constructive feedback and direction for every milestone of my doctoral candidature; I sincerely appreciate your support. I am deeply thankful to Dr Imtiaz Randhawa for his valuable detailed advice on analysing data, especially in Chapters 4 and 9; Mr Allan Lisle for an initial discussion about statistics; Ms Ngo Thi Thuy for technical and administrative discussions; Mr Jashim Uddin for the introduction about thermal cameras; and Mrs Diane Josey for editorial advice.

VIII

I would like to express my profound thanks to Prof Nguyen Xuan Trach, Prof Duong Nguyen Khang, and Assoc Prof Bui Thi Nga for their academic support and assistance in establishing connections with local authorities and smallholder farmers. I am deeply thankful to Ms Nguyen Thi Thanh Thuy for organizing accommodation, travel, and the materials necessary for data collection, and also for her direct involvement in collecting data and samples; and to Mr Nguyen Van Chanh, Dr Nguyen Thi Huyen, and Mr Nguyen Thanh Hai for their technical support and help with data/sample collection; and students Mr Nguyen Cong Trinh, Mr Tran Ba Phi, Mr Nguyen Van Khanh, Mr Nguyen Van Toan, Mr Huynh Xuan Anh, Mr Nguyen Minh Khanh, Mr Tran Cong Tien Dat, and Mr Vu Duc Quan for their help with the data collection. I would like to acknowledge Dr Chu Tuan Thinh for reading and giving comments on Chapters 4 and 9.

I want to express my gratitude and fondness for my father, Mr Nguyen Ngoc Bien, who passed away when I started my PhD in 2016. He had always given me immeasurable love, and he had always been an excellent example for me of filial piety, perseverance, and determination. I want to express my gratitude to my gentle mother, Mrs Dinh Thi Huan, who has always taught me righteousness, altruism, hard work, determination, and strong will to overcome difficulties. I am grateful to my parents-in- law, Mr Vu Van Toan and Mrs Nguyen Thi Man, who sincerely love my small family and wholeheartedly take care of my children so that I can focus entirely on my study. I am grateful to my wife, Vu Thi Mai Lien, my son Nguyen Ngoc Anh Minh, and my daughter Nguyen Ha Phuong, who has always been a solid spiritual support and who has always accompanied in overcoming the difficulties of the PhD candidature. Their unconditional love and encouragement helped me get over the stressful times.

I give thanks to my other family members, my colleagues at VNUA and UQ, and my friends for being supportive and highly understanding of the long journey of my doctoral candidature. To all of my fellow postgraduates, in particular Mrs Nguyen To Loan, Mr Dam Tuan Tu, and Mrs Mai Thi Phuong Thuy, who have become wonderful friends and for helping me in my journey, I offer my gratitude. I would like to thank everyone else who has provided friendship, guidance and encouragement and especially had faith in me over the last couple of years.

IX

Financial support

This research was supported by:

- UQ research Higher Degree Scholarship for living allowances.

- UQ Research Training Tuition Fee Offset Scholarship for tuition fees.

- MARD scholarship for partial tuition fees.

- Australian Centre for International Agricultural Research (ACIAR), project title “Improving dairy cattle health and production in Vietnam - AH/2016/02” for conducting the research.

Keywords

Smallholder dairy farms, feed types, feeding regime, housing, breed composition, genomic selection, GWAS, heat stress, infrared temperature, panting score.

Australian and New Zealand Standard Research Classifications (ANZSRC)

ANZSRC code: 070201, Animal Breeding, 50%

ANZSRC code: 070204, Animal Nutrition, 25%

ANZSRC code: 070203, Animal Management, 25%

Fields of Research (FoR) Classification

FoR code: 0604, Genetics, 50%

FoR code: 0702, Animal Production, 50%

X

Table of Contents

Abstract ...... I

Declaration by author ...... IV

Acknowledgements...... VIII

Keywords ...... X

Table of Contents ...... XI

List of Tables ...... XVIII

List of Figures ...... XX

List of Abbreviations ...... XXIII

Chapter 1 Introduction ...... 28

Chapter 2 Literature review ...... 31

2.1 The dairy sector in Vietnam ...... 31

2.2 Smallholder dairy farms in Vietnam ...... 32

2.2.1 Difficulties and challenges ...... 32

2.2.2 Productivity and welfare...... 33

2.2.3 Genetics and breeding ...... 34

2.2.4 Nutritional strategies ...... 38

2.2.5 Land and housing management ...... 40

2.3 Dairy regions ...... 40

2.4 Assessment of farm productivity and cow welfare...... 42

2.4.1 Key performance indicators ...... 42

2.4.2 Indicators of cow welfare ...... 43

2.5 Heat stress ...... 45

2.5.1 Heat stress and thermoregulatory in cattle ...... 45

2.5.2 Method to assess levels of heat stress ...... 46

2.5.3 Effects of heat stress on production and welfare ...... 49

2.5.4 Strategies to ameliorate heat stress in dairy cattle ...... 50 2.6 Advanced technologies with the potential to improve the productivity and welfare of smallholder dairy cows in Vietnam ...... 52

XI

2.6.1 Computer-based dairy nutrition models ...... 52

2.6.2 Infrared thermal technology ...... 53

2.7 Genomic technologies ...... 55

2.7.1 Identification of genetic diversity and breed composition using genomic data ...... 55

2.7.2 Genetic parameter estimate and genomic selection in dairy cattle using genomic data ...... 56

2.7.3 Genome-wide association studies ...... 63

2.7.4 Application of genomic technologies in developing countries ...... 67

2.8 Conclusions and identified research gaps ...... 67

Chapter 3 Performance and welfare status of Vietnamese smallholder dairy farms ...... 69

3.1 Introduction...... 70

3.2 Materials and methods ...... 71

3.2.1 Farm selection ...... 71

3.2.2 Selection of productivity and welfare indicators ...... 73

3.2.3 Farm visits and data collection ...... 74

3.2.4 Statistical analyses ...... 79

3.3 Results ...... 80

3.3.1 Herd characteristics ...... 80

3.3.2 Milk production ...... 84

3.3.3 The heat stress level of cows ...... 86

3.4 Discussion ...... 88

3.4.1 Herd structure and cow productivity ...... 88

3.4.2 Cow welfare ...... 90

3.4.3 Limitations ...... 92

3.5 Conclusions ...... 92

Chapter 4 Genomic diversity and breed composition of Vietnamese smallholder dairy cows . 94

4.1 Introduction...... 95

4.2 Materials and methods ...... 96

4.2.1 Tail hair samples ...... 96

4.2.2 DNA extraction and genotyping ...... 97

4.2.3 Merged genotype data and quality control ...... 97 XII

4.2.4 Genetic diversity ...... 98

4.2.5 Principal component and admixture analysis ...... 98

4.2.6 Classification of genotypes ...... 99

4.3 Results ...... 101

4.3.1 Genetic diversity within and between populations ...... 101

4.3.2 Principle component analysis to determine breed structure across populations ...... 103

4.3.3 Unsupervised hierarchical clustering of Vietnamese dairy cows...... 104

4.3.4 Association between genotype and coat colour ...... 107

4.4 Discussion ...... 110

4.4.1 Genetic diversity of Vietnamese dairy cows ...... 111

4.4.2 SNP-based estimates of breed compositions of Vietnamese dairy cows ...... 112

4.5 Conclusions ...... 113 Chapter 5 Issues of feeding strategies for lactating cows in Vietnamese smallholder dairy farms… ...... 114

5.1 Introduction...... 115

5.2 Materials and methods ...... 116

5.2.1 Farm selection and farm visits ...... 116

5.2.2 Feeding regime ...... 117

5.2.3 Diets ...... 117

5.2.4 Measurement of milk production ...... 119

5.2.5 Identification of dietary imbalance ...... 120

5.2.6 Statistical analysis ...... 121

5.3 Results ...... 123

5.3.1 Feeding regime ...... 123

5.3.2 Diets ...... 127

5.3.3 Nutrient composition of commonly used feeds ...... 131

5.3.4 Cow intake and nutrient concentrations of the diets ...... 134

5.3.5 Efficiencies of the diets ...... 136

5.4 Discussion ...... 137

5.4.1 Feeding regimes...... 137 XIII

5.4.2 Feeds and diets ...... 139

5.4.3 Limitations of the study ...... 143

5.5 Conclusion ...... 143 Chapter 6 Characteristics of cowsheds in Vietnamese smallholder dairy farms and their associations with microclimate ...... 144

6.1 Introduction...... 145

6.2 Materials and methods ...... 146

6.2.1 Farm visits and measurements of altitude, latitude, and microclimate data...... 146

6.2.2 Farm observation and barn measurements ...... 148

6.2.3 Data analysis ...... 149

6.3 Results ...... 151

6.3.1 Microclimate within the cowsheds ...... 151

6.3.2 Housing design ...... 154

6.3.3 Multivariate models identifying factors associated with cow shed microclimate ...... 160

6.4 Discussion ...... 161

6.4.1 Shed microclimate ...... 161

6.4.2 Associations between housing management and cowshed microclimate ...... 162

6.4.3 Tie-up, floor space and mat use ...... 163

6.4.4 Limitations ...... 164

6.5 Conclusion ...... 165 Chapter 7 Multivariate analysis identifying the main factors associated with cow productivity and welfare in tropical smallholder dairy farms in Vietnam ...... 166

7.1 Introduction...... 167

7.2 Materials and methods ...... 169

7.2.1 Study sites, time and data ...... 169

7.2.2 Building the models ...... 170

7.3 Results ...... 173

7.3.1 Variables associated with milk production...... 173

7.3.2 Variables associated with cow conformation ...... 176

XIV

7.3.3 Variables associated with the level of heat stress, reproduction, and udder health of the cows……...... 176

7.4 Discussion ...... 177

7.4.1 Housing variables ...... 178

7.4.2 Nutritional variables ...... 179

7.4.3 Animal variables ...... 181

7.4.4 Some limitations ...... 182

7.5 Conclusion ...... 183 Chapter 8 Application of infrared thermal technology to assess the level of heat stress and milk yield reduction of cows in tropical smallholder dairy farms ...... 184

8.1 Introduction...... 185

8.2 Materials and methods ...... 187

8.2.1 Farm visit ...... 187

8.2.2 Microclimate data ...... 187

8.2.3 Individua cow data ...... 188

8.2.4 Statistical analyses...... 190

8.3 Results ...... 192

8.3.1 Correlations between main variables ...... 192 8.3.2 Associations of heat load index with inner infrared vulva temperature, outer infrared vulva temperature, and panting score ...... 195 8.3.3 Associations of panting score, inner vulva infrared temperature, outer vulva infrared temperature, and inner tail base infrared temperature with energy corrected milk ...... 199

8.4 Discussion ...... 201 8.4.1 Pearson correlations between temperature-humidity index, panting score, infrared temperature measurements, and energy corrected milk yield ...... 201

8.4.2 Applicability of panting score and infrared technology in monitoring heat stress ...... 202

8.4.3 Suitable dairy crossbreeds for tropical smallholder dairy farms ...... 204

8.5 Conclusion ...... 205 Chapter 9 Genomic selection and genome-wide association studies for productivity and heat tolerance traits in smallholder dairy cows ...... 206

XV

9.1 Introduction...... 207

9.2 Materials and methods ...... 209

9.2.1 Animal and phenotype data ...... 209

9.2.2 Genotype data ...... 211

9.2.3 Genetic parameter estimates and genomic selection...... 211

9.2.4 Genome-wide association studies ...... 212

9.3 Results ...... 213

9.3.1 Genomic selection models ...... 213

9.3.2 Genome-wide association studies ...... 216

9.4 Discussion ...... 226

9.4.1 Genomic selection ...... 226

9.4.2 Genome-wide association studies ...... 229

9.4.3 Some limitations ...... 231

9.5 Conclusion ...... 232

Chapter 10 General discussion, implications of research, and conclusion ...... 233

10.1 Status of production and welfare ...... 233

10.2 Breeds, nutrition, and housing ...... 234

10.3 Multivariate analysis of strategies to improve productivity and welfare ...... 236

10.4 Infrared technology for assessing heat stress...... 239

10.5 Breeding and selection strategies...... 240

10.6 Limitations and future directions ...... 242

10.6.1 Limitations ...... 242

10.6.2 Future directions ...... 243

10.7 Conclusions ...... 243

List of References ...... 245

Appendices ...... 292

Appendix 1: Human research ethics approval certificate ...... 292

Appendix 2 a: Animal research ethics approval certificate ...... 293

Appendix 2 b: Animal research ethics approval certificate ...... 294

XVI

Appendix 3: Nutrient composition of feeds commonly used for dairy cows in regions (on dry matters basis, otherwise stated) A ...... 295 Appendix 4: Multivariate linear mixed effect models identifying the variables significantly and suggestively associated with infrared temperatures of rear udder (RUdT, °C), ocular area (EyeT, °C), muzzle (MuzT, °C); armpit (ArmT, °C), paralumbar fossa area (ParT, °C), fore udder (FUdT, °C), fore hoof (FHoT, °C), and hind hoof (HHoT, °C) ...... 296

Appendix 5: Significant and suggestive SNPs associated with the studied traits ...... 297

XVII

List of Tables

Table 2.1 Average monthly rainfall (mm), temperature (AT, oC), humidity (RH, %), and temperature- humidity index (THI, unit) during a period from 2002 to 2016 at weather stations nearest to the study sites A ...... 42 Table 2.2 Assessment of panting score (PS) based on breathing condition and associated respiration rate (RR; breaths/min)...... 48

Table 2.3 Some common statistical approaches for genomic selections ...... 59 Table 3.1 Comparisons of lactating herd characteristics, locomotion scores, and reasons for culling cows across four dairy regions ...... 82 Table 3.2 Comparisons of actual milk yields (kg/cow/d), milk yields adjusted for cow body weight (kg/100kg BW/d), and farmers’ milk yield target (kg/cow/d) across four dairy regions ...... 84 Table 3.3 Comparisons of farmers’ targets for milk concentrations (%), measured milk concentrations (%), and milk electrical resistance across four dairy regions ...... 85

Table 3.4 Comparisons of panting score of cows across four main dairy regions ...... 86 Table 4.1 Sampling information on Vietnamese dairy cattle populations and reference populations ...... 99

Table 4.2 Mean (SD) of observed heterozygosity (HO), expected heterozygosity (HE), and inbreeding

coefficient (FIS) of Vietnamese dairy cattle in each of four geographically contrasting regions compared to reference breeds populations A ...... 102

A Table 4.3 Pairwise fixation index (FST) between Vietnamese dairy cattle in each of four geographically contrasting regions compared to reference breed populations ...... 103

Table 5.1 Comparisons of feeding regime for dairy cows across four main dairy regions ...... 124 Table 5.2 Feed ingredients used across smallholder dairy farms (n) and mean dry matter intake (kg DM/cow/d) of each feed ingredient for lactating cows in four major dairying regions A ...... 128 Table 5.3 Nutrient concentration of feeds commonly used for dairy cows in regions (on dry matters basis, otherwise stated) ...... 133 Table 5.4 Comparisons of average nutrient composition (DM basis; % unless otherwise noted) of the lactating cow diets between four dairy regions and between these regions with aspirational targets ...... 135 Table 5.5 Diet intakes (kg/cow/d), diet cost, prediction of milk yield (kg/cow/d), and predicted methane emissions from the diets of cows in each region ...... 136

XVIII

Table 6.1 Comparisons of altitude and microclimate parameters (averaging from between 6000 h and 1800 h) inside the cowsheds across four dairy regions...... 152

Table 6.2 Housing management parameters of smallholder dairy farms in four dairy regions ...... 155

Table 6.3 Most significant variables characterizing each housing management cluster ...... 159 Table 6.4 Multivariate models identifying the factors associated with the temperature (AT, °C), humidity (RH, %), wind speed (WS, m/s), heat load index (HLI) and temperature-humidity index (THI) inside the cowsheds A ...... 160

Table 7.1 Summary of the studied variables A ...... 169 Table 7.2 Summary of independent variables with variance inflation factor less than ten that were included in the initial models A ...... 171 Table 7.3 Pearson correlations between variables included in the models (rows) and variables not included in the models due to variance inflation factor  10 (columns) A, B ...... 172 Table 7.4 Multivariate regression models identifying the variables affecting cow milk yield (MILK, kg/cow/d), milk fat (mFA, %), milk protein (mPR, %), milk dry matter (mDM, %) energy corrected milk (ECM, kg/cow/d), and energy corrected milk adjusted for body weight (ECMbw, kg/100kg BW/d)...... 175 Table 7.5 Multivariate regression models identifying the variables affecting cow heart girth (HG, cm), body weight (BW, kg), and body condition score (BCS) ...... 176 Table 7.6 Multivariate regression analysis identifying the variables affecting panting score (PS), times of artificial inseminations per conception (tAT, times), and milk electrical resistance (mRE, units) ...... 177 Table 8.1 Comparing means of the temperature-humidity index, heat load index, panting score, and infrared thermal temperatures across four main dairy regions A ...... 193 Table 8.2 Matrix of Pearson correlations between temperature-humidity index, heat load index, panting score, and infrared thermal temperatures A ...... 194 Table 8.3 Multivariate linear mixed effect models identifying the variables significantly or suggestively associated with panting score (PS); and infrared temperatures of cow inner vulva (IVuT, °C), outer vulva (OVuT, °C), and inner tail base (ITBT, °C) A ...... 195

Table 8.4 Thresholds to assess environmental head load and heat stress levels of cows A ...... 197 Table 8.5. Multivariate linear mixed effect models identifying the associations of panting score (PS); and infrared temperatures for cow inner vulva (IVuT, °C), outer vulva (OVuT, °C), and inner tail base (ITBT, °C) with energy corrected milk yield (ECM, kg/cow/d) A ...... 200 XIX

Table 9.1 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for milk production traits A ...... 214 Table 9.2 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for cow body conformation traits A ...... 215 Table 9.3 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for heat tolerance traits A ...... 215 Table 9.4 Single nucleotide polymorphisms (SNPs) and related genes most associated with milk production traits ...... 218 Table 9.5 Single nucleotide polymorphisms (SNPs) and related genes most associated with cow conformation traits ...... 222 Table 9.6 Single nucleotide polymorphisms (SNPs) and related genes significantly associated with heat tolerance traits ...... 225

List of Figures

Figure 2.1 Formation of the Vietnamese smallholder dairy herd in Vietnam ...... 35

Figure 2.2 Topographic map of Vietnam mainland and distribution of dairy cows by district ...... 41 Figure 2.3 Suggested body condition score profiles (10-, 8- and 5-point scales) for a lactating cow ...... 44

Figure 2.4 Description of panting scores from 1 to 4.5 ...... 49

Figure 2.5 Principle of genomic selection method ...... 58 Figure 2.6 Some prior assumptions for distributions of SNP effects in statistical approaches used in genomic selection...... 60

Figure 2.7 Common steps of genome-wide association studies (GWAS) ...... 64

Figure 3.1 Topographic map of Vietnam mainland and study sites ...... 73

Figure 3.2 Typical interiors of smallholder dairy farms in each study region ...... 75

Figure 3.3 Measurements of cow heart girth and body weight; and body condition score ...... 76

Figure 3.4 Measurement of daily milk yield and milk sampling ...... 78

Figure 3.5 Herd structure across the four contrasting regions ...... 80 Figure 3.6 Mosaic plots showing associations between regions and body condition score categories ...... 83

XX

Figure 3.7 Mosaic plots showing associations between regions and panting score categories with (a) morning panting score (mPS), (b) afternoon panting score (aPS), and (c) average day panting score (PS) ...... 87 Figure 4.1 Breed structure of Vietnamese dairy cow populations and reference population revealed by principal component analysis...... 104 Figure 4.2 Admixture plot for Vietnamese dairy cows and reference breeds analysed according to different numbers of assumed ancestors (K = 2 to 8) ...... 105 Figure 4.3 Cross-validation plot for the different number of assumed ancestors (K = 1 to 10) indicating the choice of the appropriate K = 4...... 106 Figure 4.4 Comparisons of genomic breed proportions between Vietnamese dairy cow populations when K = 4 ...... 107 Figure 4.5 Mosaic plots showing the relationship between regions and cow genotype identified by genomic data (Breed_G; a) and cow genotype reported by farmers (Breed_F; b) ...... 108 Figure 4.6 Agreement plot for the genotype reported by farmers and genotype identified by genomic data...... 109

Figure 4.7 Relationship between cow genotypes identified by genomics and coat colour ...... 110

Figure 5.1 Measurements of feed offered, feed refused, and appearance of PCDairy software ...... 119 Figure 5.2 Results of factor analysis (FAMD) and hierarchical clustering on principal components (HCPC) for feeding regime data ...... 127 Figure 5.3 Cluster dendrogram (a) depicting nine optimum diet clusters from C1 to C9; and biplot (b) a 2-dimensional view of the first two principal components ...... 130

Figure 6.1 Measurement of microclimate data inside cowsheds ...... 147 Figure 6.2 Changes in means of some main microclimate parameters in four regions during day time ...... 153 Figure 6.3 Results of factor analysis (FAMD) and hierarchical clustering on principal components (HCPC) for housing management data ...... 158 Figure 8.1 Cow identification (a), scoring panting score (b), and measurements of infrared temperature (c to l) ...... 188 Figure 8.2 Effect plots predicting the effects (with confidence intervals) of heat load index (HLI) on the panting score (PS), infrared temperatures of inner vulva (IVuT), outer vulva (OVuT), and inner tail base (ITBT)...... 196

XXI

Figure 8.3 Plots comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of panting score (PS) (a), infrared temperatures of inner vulva (IVuT) (b), outer vulva (OVuT) (c), and inner tail base (ITBT) (d) of the cows across regions ...... 198 Figure 8.4 Plot comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of panting score between different dairy cow crossbreeds ...... 199 Figure 8.5 Effect plots comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of energy corrected milk yields (ECM) across regions (a) and between crossbreeds (b) ...... 201

Figure 9.1 Manhattan and Q-Q plots of association results for milk production traits...... 217

Figure 9.2 Manhattan and Q-Q plots of association results for cow conformation traits ...... 221 Figure 9.3 Manhattan and Q-Q plots of association results for the first group of cow heat tolerance traits...... 223 Figure 9.4 Manhattan and Q-Q plots of association results for the second group of cow heat tolerance traits...... 224

XXII

List of Abbreviations

Abbreviation Meaning Unit Acc Accuracy of genomic selection - ACIAR Australian Centre for International Agricultural Research - ADF Acid detergent fibre % DM AHLU Accumulate heat load units units ArmT Armpit temperature oC AT Ambient temperature oC AUC Area under the receiver operating curve - B1 First backcrosses - B2 Second backcrosses - B3 Third backcrosses - BCS Body condition score (5-point scale) - BLUP Best Linear Unbiased Prediction - BRM Brahman - BSW Brown Swiss - BTA Bos taurus autosome - BW Body weight estimated from heart girth using an equation kg Ca Calcium % DM Cor Correlation - CP Crude protein - Cu Copper % DM CV Cross-validation - CYE Chinese Yellow cattle - d Day - DEC Chinese Dengchuan cattle - DEH Chinese Dehong cattle - DM Dry matter concentration of feed % as fed DMI Dry matter intake kg/cow/d DNA Deoxyribonucleic acid - EBV Estimated breeding value - ECM Energy corrected milk kg/cow/d ECMbw Energy corrected milk adjusted for body weight kg/100kg BW/d

XXIII

EyeT Temperature of ocular area oC F1 First crosses - FAMD Factor analysis of mixed data - FCM Fat corrected milk yield, 3.5% fat kg/cow/d Fe Iron % DM FHoT Fore hoof temperature oC

FIS Inbreeding coefficients -

FST Pairwise fixation index - FudT Fore udder temperature oC G Genomic relationship matrix - GBLUP Genomic best linear unbiased prediction - GEBV(s) Genomic estimated breeding value - GO Gene Ontology - GS Genomic selection - GT Globe temperature oC GWAS Genome-wide association study - h Hour - h2 Heritability - HCPC Hierarchical Clustering on Principal Components -

HE Expected heterozygosity - HG Heart girth cm HhoT Hind hoof temperature oC HLI Heat load index units

HO Observed heterozygosity - HOH Honghe cattle - HOL Holstein - HR Heat rate beats/min ID Identification - IRT(s) Infrared thermal temperature(s) oC ITBT Inner tail base temperature oC IVuT Inner vulval lip temperature oC JER Jersey K Potassium % DM

XXIV

K Number of distinct assumed ancestors - KEGG Kyoto Encyclopaedia of Genes and Genomes - KPI(s) Key performance indicator(s) - LD Linkage disequilibrium - MAF Minor allele frequency - Max Maximum - Mb Megabase Mb mDM Dry matter concentration of milk % mFA Fat concentration of milk % Mg Magnesium % DM MILK Milk yield kg/cow/d min Minute - Mn Manganese % DM mPR Protein concentration of milk % mRE Milk electrical resistance units MuzT Muzzle temperature oC Na Sodium % DM NDF Neutral detergent fibre % DM NEL Net energy for lactation MCal/kg DM NFC Non-fibre carbohydrates % DM NH North highland region - NL North lowland region - NRC National Research Council - ns Not significant - OT Other cattle breeds - OVuT Outer vulval temperature oC P Phosphorus % DM ParT Temperature of paralumbar fossa area oC

PC Agreement expected by chance - PC1 First principal component - PC2 Second principal component - PCA Principal component analysis - PMR Partial mixed ration -

XXV

PO Observed agreement - PS Day panting score (0 to 4.5) - Q-Q Quantile-quantile - QTL(s) Quantitative trait loci - R R software - REML Restricted maximum likelihood - RH Relative humidity % RR Respiration rate breaths/min RSI Red Sindhi - RT Rectal temperature oC RUdT Rear udder temperature oC S Sulphur % DM s Second - SAH Sahiwal - SD Standard deviation - SDF(s) Smallholder dairy farm(s) - SE Standard error - SEA South East Asia - SEM Standard error of the mean - SH South highland region - SL South lowland region - SNP(s) Single nucleotide polymorphism(s) - tAI Times of artificial insemination per conception - TBV True breeding value - TDN Total digestible nutrients % DM Tdp Dew point temperature oC THI Temperature-humidity index units TMR Total mixed ration - Tnawb Natural aspirated wet bulb temperature oC Twb Wet bulb temperature oC Twbg Wet bulb globe temperature oC UQ University of Queensland - USA United States of America -

XXVI

USD United States dollar $ USDA United States Department of Agriculture - V-NH Vietnamese dairy cows in north highland - V-NL Vietnamese dairy cows in north lowland - V-SH Vietnamese dairy cows in south highland - V-SL Vietnamese dairy cows in south lowland - VDC Vietnamese smallholder dairy cows - VIF Variance inflation factor - WS Wind speed m/s yDM Milk dry matter yield kg/cow/d yFA Milk fat yield kg/cow/d yFP Milk fat and protein yield kg/cow/d yPR Milk protein yield kg/cow/d ZEB Zebu breeds - Zn Zinc % DM 2 σ a Additive genetic variance - 2 σ c Random farm variance - 2 σ e Residual variance -

XXVII

Chapter 1 Introduction

Dairy production systems in tropical environments are constrained by many challenges, including hot and humid weather causing heat stress in animals, low productivity genotypes, shortage of land, small farm size, low forage quality and availability, a diversity of infectious and parasitic diseases, and inadequate research and extension services (Bonsma, 1949; Aminah and Chen, 1989; Owen et al., 2012; Moran, 2013; Bernard, 2015). These issues need to be studied and addressed, but currently, studies on tropical dairy production remain very limited (Hernández-Castellano et al., 2019; Ramírez- Rivera et al., 2019). This limitation is because the tropics are often associated with developing countries where both economic and scientific capacity are insufficient for high-quality research (Preston, 1989; Morton, 2007; Moran, 2013). Moreover, the most common type of dairy farm in tropical developing countries is of the smallholder type (Small holder dairy farms: SDFs), where cow numbers are low, often less than 20, and the input and output data are not systematically recorded commercially or for research (Morton, 2007; General Statistics Office of Vietnam, 2016; Nguyen et al., 2016b; Vinamilk, 2017). Compared to larger commercial farms, SDFs in the tropics are expected to be severely affected by climate change (Morton, 2007). Such climatic constraints, their small scale and relatively poor quality inputs mean that dairy cow performance in tropical developing countries is generally lower than that obtained in North America, Western Europe and Oceania (Renaudeau et al., 2012).

Dairy production is neither a strength nor a tradition of Vietnam (Trach et al., 2007). However, this sector is developing rapidly due to increasing domestic demands for fresh milk and dairy products (Cesaro et al., 2014). The demand for dairy products in Vietnam is increasing exponentially due to rapid growth in population and living standards; so improving the performance of SDFs may be the most practical way to meet that demand (Khanh, 2014; General Statistics Office of Vietnam, 2017a).

Commercial and SDF systems co-exist in Vietnam. Recent agricultural developments include the emergence of Asia’s largest individual dairy farm in central Vietnam, comprising approximately 44,000 cows (TH true milk Company) (Duteurtre et al., 2015). However, the overwhelming majority of fresh milk in Vietnam is produced by SDFs. Vietnamese SDFs are often owned and managed by single households and have varying herd sizes of less than five cows (36.7% of SDFs) up to 50 cows (2.2% of SDFs) (Nguyen et al., 2016b). There were approximately 28,695 SDFs in 2016 (General Statistics Office of Vietnam, 2016; An, 2019), accounting for more than 97% of the total national dairy herd (Trach, 2017a). These SDFs produced more than 80% of Vietnam’s fresh milk (Vinamilk, 2017). Whilst the government has facilitated the development of the commercial dairy production sector, it is the enhancement of SDFs, through the incorporation of modern dairy production 28 technologies and by increasing the number of cows per SDF, that appears to be the government preferred model for future development (Liem 2016; Dong 2017). Thanks to the strong support from the government, and apart from the regions with a long history of dairy farming, recently, smallholders in many regions that have only traditionally farmed rice, have turned to dairy farming (General Statistics Office of Vietnam, 2017a; Manh and Bich, 2017). Hanam province is an example where nearly 200 smallholders have begun raising dairy cows in the last five years (Manh and Bich, 2017). However, the rapid pace of technological development and its advancement in dairy farm management in terms of infrastructure and the purchase of cows have far exceeded the development of farmers’ knowledge and skills, particularly concerning general husbandry, including cow nutrition, genetics, and heat stress management. Consequently, the cows' productivity, health and welfare and the farm’s profitability remain compromised (Cai et al., 2005). Such problems can be even more severe in SDFs than in commercial dairy farms. Whereas commercial dairy farms have the financial capacity to import the latest and most complete dairy technology systems as well as expert advice for their farms, SDFs often do not have that capability. Also, most highly effective new technologies for dairying have been optimized in developed countries specifically to support commercial dairying with relatively large herds and mostly with temperate environments (Moran, 2013; Ban, 2014). Such technologies may be difficult or financially unviable when applied to SDFs in a tropical and developing country like Vietnam.

In Vietnam, the support for SDFs through targeted research has a history of being inadequate and disjointed. Currently, it is difficult to find basic objective production and husbandry information on these SDFs such as cow genotype, typical dietary regimes, management practices, milk yield, milk concentrations, or welfare issues in their herds such as heat stress, or the causes for their culling of cows (Vang et al., 2003; Chu et al., 2005; Vu et al., 2016). The farmers and even local extension staff and veterinarians have often not been taught, or do not fully understand, appropriate practices that promote cow productivity and welfare. Moreover, compared to the availability of technical support for dairy farmers in developed countries, local dairy extension staff and veterinarians are scarce in Vietnam, especially in regional areas. Necessary support includes help in genetic selection, diet formulation, housing modifications that facilitate cow comfort and hygiene, and the evaluation of animal welfare concerns such as nutritional condition and the heat stress status of the cows.

Therefore, the priorities of this thesis are to investigate suitable strategies to monitor and improve the productivity and welfare of Vietnamese SDFs. The thesis includes seven studies that aim firstly to characterise and understand the main constraints of these SDFs in terms of breeding, nutrition, management, productivity, and welfare; secondly, to test the application of genomic technologies,

29 diet balancing software, and infrared thermal technologies as potential farm management improvement technologies. Four typical but contrasting SDF regions of Vietnam were chosen as a focus for the thesis: a southern lowland (hot and humid, well-established dairying but minimal extension support), a southern highland (cool, well-established dairying but minimal extension support), a northern lowland (hot and humid, recently established dairying with developing extension support system), and a northern highland (cool, well-established dairy extension support system) (Cesaro et al., 2014; General Statistics Office of Vietnam, 2017a).

30

Chapter 2 Literature review

2.1 The dairy sector in Vietnam

The rapid development

Dairy farming is not a traditional farming pursuit in Vietnam; however, it started to appear in the 1930s, and in recent decades, it has developed rapidly (Trach et al., 2007). In 1996, there were approximately 22,000 cows producing 27,900 tons of milk per year, but in the ten years to 2006, that had grown to 113,000 cows and 215,900 tons of milk (Dairy Vietnam, 2017). In 2017, the total dairy herd was 301,649 cows, representing increases of 267% and 408% for cow numbers and milk produced, respectively (General Statistics Office of Vietnam, 2017a).

Companies involved

According to the Vietnamese government’s Department of Livestock Production (2017), Vinamilk, Friesland Campina Vietnam, Moc Chau Dairy Cattle Breeding JSC, and Ba Vi International Milk are the major milk processing companies that collect milk from SDFs (Khoi and Dung, 2014). Moc Chau Dairy Cattle Breeding JSC only collects milk in Son La province in the northern highlands. Ba Vi International Milk collects milk in several northern provinces, whereas Vinamilk and Friesland Campina Vietnam collect milk from SDFs all over the country, including Ho Chi Minh City and Lam Dong in the south and Ha Nam in the north. Of these companies, the most powerful is Vinamilk (Khoi and Dung, 2014). Another fast developing company is TH True Milk, but it focuses only on the large commercial farms it owns, not on SDFs (Hayes, 2012). Moc Chau Dairy Cattle Breeding JSC has an especially close relationship with SDFs; it allocates land to member farmers, supplies and manages inputs such as genetics, feed and veterinary services, and ensures the sale of all milk produced (Cuong and Nga, 2014).

Dependence on imports

Despite the rapid development of the dairy sector, it is unable to meet even the relatively low milk needs of Vietnamese consumers. In 2016, the total amount of domestic milk production was 795,144 tons of milk (Department of Livestock Production, 2017), which was less than 40% of the total demand of the 90 million-strong population of Vietnam. The average amount of milk consumption per person in Vietnam in 2014 was about 15 kg of fresh milk/year (Khanh, 2014) compared to about 35 kg/person/year in Thailand, 45 kg/person/year in Singapore (Vinamilk, 2017), and more than 100 kg/person/year in the developed world (Thong, 2017). Milk consumption in Vietnam has increased over time, in part due to economic growth and in part to promotional programs by the Government that encouraged milk to be included in the daily diet for all ages, but especially for school children 31

(Khanh, 2014). Milk consumption in Vietnam is projected to rise to 27 kg/person/year in 2020 (Khanh, 2014). To satisfy the remaining 60% of total demand, Vietnam imports milk products from New Zealand, Singapore, Germany, America, and Australia (Nguyen, 2014; General Department of Vietnam Customs, 2017). As presented in the import reports of the General Department of Vietnam Customs (2017), in 2017, Vietnam spent about 827 million USD on importing fresh milk. Additional to milk imports, the Vietnamese dairy sector also depends heavily on the importation of bulls, semen straws, heifers, and stock feed (Luong et al., 2006; Khoi and Dung, 2014; Nguyen, 2014).

Strong government support

The Vietnamese Government strongly supports the dairy sector, and Decision No. 124/QD-TTg of the Vietnamese Prime Minister (2012) approved the master plan for agricultural production development through 2020 with a vision toward 2030. The aim was to increase the dairy herd from just over 300,000 in 2017 to 500,000 dairy cows by 2020, producing one million tons of milk to supply an average of more than 10 kg of milk/person/year (Hai, 2008). It is now apparent that this target was not met in 2020. However, it highlights the Government’s ambition for the sector.

2.2 Smallholder dairy farms in Vietnam 2.2.1 Difficulties and challenges

Ban (2014) and Nguyen et al. (2016b) summarized the difficulties and challenges that the dairy sector, especially SDFs, must overcome to ensure sustainable development. Those include: (1) the unfavourable conditions for high yield dairy cows caused by the tropical climate, (2) limited land area for both grazing and the production of forage, (3) the high cost of importing pure Holstein genetics, especially heifers, (4) lack of suitable quality breeding stock and the regulation of the quality of stock within Vietnam, (5) the inadequate financial support to the dairy industry, and (6) poor extension services and the limited skills of farmers in the management and feeding of dairy cattle (Ban, 2014; Nguyen et al., 2016b).

To overcome the challenges mentioned above, commercial dairy farms owned by milk processing companies such as Vinamilk or TH True Milk, thanks to their abundant financial resources, usually import genetics, feed and technologies, and they hire international dairy experts (Hayes, 2012). TH True Milk is an example of a company that does all this. This company only started trading in 2010, and it did so by importing pure Holstein heifers from dairy-developed countries, importing complete dairy technologies, and hiring experts from the Afimilk Group (Israel) (Nguyen, 2014; Duteurtre et al., 2015). By 2015, this company had a high producing herd of 44,000 dairy cows with average milk production of 20 litres/cow/d (Duteurtre et al., 2015). This production level is reasonably high for

32 purebred Holstein cows in a tropical environment (Duteurtre et al., 2015). However, for SDFs, overcoming similar challenges in a similar way to TH True Milk is not easy because their land, finance, and expertise are all limited (Moran, 2013; Nguyen, 2014). Also, not all dairy technologies are applicable to SDFs (Moran, 2013). Thus, technical supports for SDFs to improve their productivity and reduce their production costs while simultaneously increasing the value of their end dairy product are necessary if government targets for national milk production are to be met (Ban, 2014; Moran, 2015a).

2.2.2 Productivity and welfare

The very limited number of published studies on SDFs in Vietnam suggests per cow productivity is relatively low (Ashbaugh, 2010; Lam et al., 2010; Vu et al., 2016). A few surveys reported average daily milk yields of 14 to 15 kg/cow/d in southern (Vu et al., 2016) and northern (Ashbaugh, 2010) provinces. Nationally, during 2001-2013, the average milk yield of cows was reported to have increased at a rate of only 2.56% per year (Dũng, 2014). Lam et al. (2010) estimated by survey that 14.2% of lactating cows also only produced milk for 7 to 8 months as compared to the targeted ten months in developed countries.

In Vietnam, the concept of animal welfare is still novel (Trach, 2017b). Article 21 of The Veterinary Act (2015) was the first legislation requiring organisations and individuals to minimize pain and fear, and to humanely treat animals through improved practices in animal husbandry, transportation, slaughter, destruction, disease prevention disease treatment, and scientific studies. Apart from that, Vietnam has no specific Animal Welfare Act and no specific regulations or guidelines on the welfare of domestic animals (Trach, 2017b). As a result, Vietnamese farmers are not fully aware of the welfare needs of animals.

Modern concepts of managing according to temperature-humidity index (THI), body condition score, locomotion score, and panting score, which are widely applied in dairy production around the developed world, are little known or applied in Vietnamese SDFs (Welfare Quality® Consortium, 2009; Ban, 2014; Trach, 2017b; RSPCA, 2018). A few studies have reported on aspects of welfare in dairy cows in Ho Chi Minh city, but little information is presented as animal welfare was not their focus (Lam et al., 2010; Vo, 2011; Östensson et al., 2013), and even fewer studies have attempted to measure heat stress among dairy cows objectively. Ngo and Trinh (2006) reported that the temperature-humidity index (THI) in Hai Duong and Hanoi city (low-land, hot and humid provinces in the North) from April to November was higher than THI = 68, a heat stress threshold suggested by Zimbleman et al. (2009). Nguyen et al. (2017b) reported that the THI in Ho Chi Minh City (a lowland, hot and humid province in the South) from February to December was also higher than 68. This 33 indicated that dairy cows in lowland provinces in both the north and the south of Vietnam could suffer from heat stress for most months of the year. A study by Lam et al. (2010) is one of the few studies that involved the measurement of indicators of heat stress in animals . In one of the hottest and most humid dairying areas in Vietnam and in the hottest months there (Ho Chi Minh City, May-June), the THI averaged 81 in the mornings and 85 by afternoon. Cow respiration rates averaged 54 (breaths/min) in the morning and 70 in the afternoon and cow rectal temperature averaged 38.8oC and 39.3oC in the morning and afternoon, respectively. It has not been possible to find published studies that have assessed the level of heat stress among dairy cows raised in the cooler highland provinces such as Son La (south Vietnam) and Lam Dong (north Vietnam), but SDF farmers there commonly consider that their cows do not suffer from heat stress.

Fertility is another likely problem in SDF cows as Lam et al. (2010) reported that 47.5% of lactating cows needed three to four inseminations per conception, and 43.4% of lactating cows needed five to seven inseminations per conception. Consequently, farmers had to sell their cows after seven failures with artificial insemination, and farmers had to milk 27.5% of their cows for more than 12 months rather than ten months as usual (Lam et al., 2010).

In addition, some studies reported that some SDF cows were not supplied with adequate amounts of drinking water (Suzuki et al., 2006; Lam et al., 2010). Suzuki et al. (2006) reported that only 29% of SDFs in Hanoi supplied cows water ad libitum, and Lam et al. (2010) in the South of Vietnam reported that 51% of SDFs provided less than 30 litres of water for a cow per day and only 35% provided fresh water ad libitum for the cows. In the tropics, farmers are advised to provide lactating dairy cows 60 to 70 litres of water per day for maintenance, plus an extra 4 to 5 litres for each litre of milk production (Moran and Chamberlain, 2017).

2.2.3 Genetics and breeding

Breeds

Cross-breeding of local cattle breeds with European dairy breeds has been widely practised in Vietnamese SDFs to improve the milk production potential by combining the good attributes of both local and exotic breeds and taking advantage of heterosis (Syrstad, 1989; Vang et al., 2003). Cows can be first crosses (F1), first (B1), second (B2), or third (B3) backcrosses, usually generated by back- crossbreeding between the local cattle breeds and imported dairy cattle breeds (Vang et al., 2003; Department of Livestock Production, 2009; Lam et al., 2010) (Figure 2.1). The local breeds are usually used as dams in crossbreeding programs and include Yellow Cattle, Lai Sind Cattle (a cross bred between Yellow Cattle and Zebu cattle such as Red Sindhi or Sahiwal), and other crossbreeds

34 between Yellow or Lai Sind Cattle and Brahman (Vang et al., 2003; Department of Livestock Production, 2009). The main exotic breed used is the Holstein, then come the Jersey and Brown Swiss, and occasionally French Brown (Brune), Montbeliarde, Bretonne Pie Noir, or Australian Friesian Sahiwal (Cai and Long, 2002; Vang et al., 2003; Department of Livestock Production, 2009; Ashbaugh, 2010). There are no official statistics for the exact numbers of each dairy crossbreed, but according to Tuyen (2009), 80% of the Vietnamese dairy herd are Holstein crosses, 15% are pure Holstein, and 5% are crossbreeds with non-Holstein dairy breeds such as Jersey and Brown Swiss. In recent years, a large number of pure Holstein heifers have been imported into Vietnam from Australia, America, New Zealand, Israel and some European countries, which is likely to have increased the percentage of pure Holstein in the national dairy herds (Vang et al., 2003; Luong et al., 2006). However, the majority of these imported cows went to the large commercial farms of the milk processing companies previously mentioned. Only a small proportion of these pure-bred imported cows were sold to SDFs (Luong et al., 2006; Dao Phuong, 2013). Therefore, the dominant genotypes in SDFs seem still to be Holstein crossbreeds, but it is difficult to be sure, given the lack of pedigree records.

Figure 2.1 Formation of the Vietnamese smallholder dairy herd in Vietnam

35

Genetic selection

Currently, there is a heavy dependence on imported semen straws, and virtually all SDF farmers rely on inseminators to artificially inseminate (AI) their cows (Vang et al., 2003; Chu et al., 2005; Suzuki et al., 2006). Natural mating is only used for cows with repeat breeding problems or in regions with no inseminators (Cai and Long, 2002). Imported semen is mainly from Holstein, Jersey, and Brown Swiss breeds but has included French Brown (Brune), Montbeliarde, Bretonne Pie Noir and Australian Friesian Sahiwal (Cai and Long, 2002; Vang et al., 2003). Holstein semen straws are usually sourced from Canada, France, USA, Australia, Spain, Cuba, New Zealand, Japan, Korea, Mexico and Hungary; Jersey straws from Belgium, New Zealand and the USA; and Brown Swiss straws from Hungary, Cuba and Mexico (Cai and Long, 2002). The relative numbers of straws imported each year were not published.

The selection and use of bulls for breeding are rare in Vietnam. Moncada is the only bull breeding centre, and it produces a limited number of semen straws, mainly from imported bulls (National Institute of Animal Sciences, 2017). Apart from imported bulls, traditional pedigree-based selection appears to not have been practised for dairy cattle. Male calves are usually sold at an early age for beef (Suzuki et al., 2006). For a few farmers who use bulls in some regions in the south of Vietnam where there is no inseminator, the bulls are selected based only on their physical appearances (Cai and Long, 2002). Tiem (2014) seems to be a landmark study in the application of the traditional performance-based best linear unbiased prediction (BLUP) selection method in selecting the three best Holstein bulls from a population of 35 male calves. As far as the author is aware, genomic selection methodologies have yet to be applied to the Vietnamese dairy herd.

Since the number of dairy cows relative to demand remains low in Vietnam, almost all female calves and heifers in SDFs are kept for reproduction and milk production and selection for keeping based on performance is not considered (Cai and Long, 2002; Chu et al., 2005; Lam et al., 2010). Farmers usually purchase female calves and heifers from their relatives, friends, neighbours, middlemen, and sometimes cooperatives and government breeding centres (Chu et al., 2005).

Pedigree and phenotype records

Despite the variety of dairy genetics that has been imported, the records of the pedigrees of imported semen lines are currently poorly maintained, if at all. In SDFs, almost all farmers fail to routinely record cow performance, health history or pedigree (Vang et al., 2003; Chu et al., 2005; Vu et al., 2016). A survey by Cai and Long (2002) in the south of Vietnam showed that only 33% of SDFs had

36 farm recording books, 25% had individual cow records, 21% kept semen straw records, and only 6% kept cow insemination records. For farms that had recording books, the format of the recording sheets and the recorded information were usually inconsistent and incomplete (Cai and Long, 2002). The bulls used by farmers also had no service recording books and no clear pedigree or breed composition (Cai and Long, 2002). The inseminators also kept poor records. Although all inseminators have certificates to practise, only 25% of them recorded the semen that they used. Even when they recorded, the information was mainly on the number of cows inseminated and farmer contact details. The main reason given by farmers and inseminators for the low rate of recording was that they felt recording was annoying and time-consuming (Cai and Long, 2002). Consequently, farmers do not appear to recognize the importance of recording activities yet (Chu et al., 2005). Similar to those in the south, most farmers in the north fail to maintain cow pedigree and phenotype records. The exception is those in Son La province, where record keeping is an expectation of the Moc Chau Dairy Cattle Breeding JSC and is actively facilitated by the cooperative that most farmers there belong to(Cuong and Nga, 2014). Because of the poor recording of cow pedigree, the risk of inbreeding and breed degeneration in Vietnamese dairy herds could be high.

Basic individual cow production trait details such as milk yield or quality are difficult to collect on most SDFs (Vang et al., 2003; Vu et al., 2016). Cows usually do not have ear tags or identification numbers. The farmers usually milk their cows by hand or by simple milking machines, which milk one to two cows at a time and have no weight or volume scale or meter (Lam et al., 2010). Milk from the cows is usually bulked on-farm and then transferred to the milk collecting centres where the technicians weigh the milk of each farm daily and periodically collect milk sample per farm for milk concentration analysis (Lam et al., 2010; Bui et al., 2013). Consequently, no performance data is commonly available for individual cows. Therefore, a significant challenge for future breeding and genomic selection studies and applications in Vietnam is how to derive phenotype and pedigree information for individual cows.

Performance of relevant dairy crossbreeds

The majority of dairy genetic studies in Vietnam have been limited to evaluating and comparing the performances of F1 (50% Holstein), B1 (75% Holstein), B2 (87.5% Holstein), and pure imported Holstein cows (Trach, 2004; Cuong et al., 2006b; Gioi et al., 2006; Luong et al., 2006; Noi et al., 2006; Tue et al., 2010). However, it appears none have occurred under SDF conditions; all quoted studies were performed under controlled experimental conditions. The F1 cows commonly rank as having the lowest milk yield, but there is no agreement between studies on the relative milk yields of B1 versus B2 cows. While the study of Phong and Thu (2016) conducted in Can Tho province (in the

37 south) concluded that the 305-day milk yield of B2 cows (4,179 kg) was higher than that of B1 cows (4,057 kg) and F1 cows (3,477 kg), the study of Tue et al. (2010) conducted in Bac Ninh province (in the north) indicated that the milk yield of B1 cows (4,231 kg) was higher than that of B2 cows (4,134 kg) and F1 cows (3,449 kg). The study by Trach (2004), in Son La province and in rural provinces adjacent to Hanoi (all in north Vietnam), also concluded that B1 cows outperformed F1 and B2 cows.

In terms of body weight, Phong and Thu (2016), in the south, showed that the weight of crossbred cows at the first and fourth parities was highest in B2 (432 kg at first parity and 484 kg at fourth parity), followed by B1 (424 and 468 kg), and the lowest in F1 (418 and 446 kg). In terms of fertility, the first oestrus age of F1, B1, and B2 crossbred cows were 18.5, 17.6 and 17.6 months, respectively (Phong and Thu, 2016). The first oestruses after calving of F1, B1, and B2 crossbred cows were 45.5, 45.2 and 44.1 days, respectively. Inseminations per conception F1, B1, and B2 were 2.0, 2.4 and 2.8 times, respectively (Phong and Thu, 2016).

The inconsistency between these study results suggests that the milk yield of different crossbred cow groups might depend more on general husbandry and nutritional conditions than genetics alone.

2.2.4 Nutritional strategies

Although Vietnamese scientists and farmers complain about the low productivity, few studies have documented the variety of dietary recipes or nutrient amounts and balance used in cow diets in SDFs. Most research stops at surveying the feed types commonly used by farmers and often only in a limited number of regions (Chu et al., 2005; Lam et al., 2010; Phong and Thu, 2016). What is generally assumed is that fresh Napier grass (Pennisetum purpureum, also known as Elephant grass, Uganda grass, or King grass) is the main source of roughage (Cuong et al., 2006a; Phong and Thu, 2016). Napier grass is promoted due to its high biomass, constant regrowth after cutting, and capacity to be harvested many times throughout the year. This helps to overcome the forage shortage experienced by many farmers (Chu et al., 2005). However, despite its high biomass, Napier grass is of inferior quality compared to other grasses, expressed in low dry matter digestibility and protein concentrations, but high fibre and water content (Dung et al., 2007; Ngo et al., 2009). Additional to Napier grass, other cultivated tropical grasses such as Ruzi grass (Brachiaria Ruziziensis), Long Tay grass (Brachiaria mutica), Mulato (Brachiaria ruziziensis x Brachiaria decumbens x Brachiaria brizantha), and some naturally available forage such as natural grasses at river banks or in fallow areas, banana stalks, and sometimes water hyacinth stalks (Eichhornia crassipes) collected at the ponds and rivers are also used as a forage source for dairy cows (Chu et al., 2005; Suzuki et al., 2006; Phong and Thu, 2016).

38

The main type of concentrate used is composite pellets, commonly called “concentrate pellets”. These are based on by-products fortified with key minerals and sometimes vitamins, with crude protein concentrations ranging from 16-18% of dry matter (Chu et al., 2005; Ho Chi Minh City Extension Center, 2015). Suzuki et al. (2006) reported that 94% of the SDFs around Hanoi offer cows commercial concentrate pellets. Farmers usually buy these concentrate pellets from feed companies which are often linked to the milk processing company that the farmer supplies their milk to (Phong and Thu, 2016). Only a small proportion of other types of concentrates such as corn powder, rice powder, and rice bran are used in cow diets (Phong and Thu, 2016). Some agricultural and industrial by-products are also used, such as rice straw, pineapple residual silage, brewers grain, cassava residual, or molasses (Chu et al., 2005; Suzuki et al., 2006; Ashbaugh, 2010; Phong and Thu, 2016).

Lam et al. (2010) reported that diets for SDF cows in the south might comprise, on a dry matter basis, 48.6% roughage, 19.5% by-products, and 33.7% concentrate. They claimed that, depending on the availability of grasses and agricultural by-products, farmers offer cows between 20 and 40 kg (as-fed basis) of roughage per cow per day, but no detail was given on the type of roughage. They also noted that regarding concentrate, farmers usually offer an amount of concentrate pellets to a cow in accordance with its milk yield (Loan et al., 2004). In the south, farmers commonly feed their cows 4 to 6 kg of concentrate pellets per day, depending on milk yield (Lam et al., 2010). In Son La and some other provinces in the north, farmers reportedly offer 0.5 kg of concentrate pellets per 1 kg of milk for a given cow (Cuong et al., 2006a). For example, if a cow produces 18 kg of milk per day, it could be given 9 kg of concentrate pellets in addition to forage (Cuong et al., 2006a). Such a simplistic rule could easily put a high performing lactating cow at great risk of ruminal acidosis due to the ingestion of excessive concentrate relative to roughage (Humer et al., 2018). It is quite obvious that there may be even many more inadequacies in the diets of SDF cows.

To optimise rumen function, a relatively recent extension booklet on the nutritional management of SDF cows in Vietnam recommends roughage level in the diets of lactating cows should be > 45% at early lactation, > 50% at mid-lactation, and > 55% at late lactation (dry matter basis) (Ngo et al., 2009). Some farmers have also had recent access to commercially available, bagged, total mixed ration (TMR) (Mai et al., 2011). Whatever the dietary ingredients, SDF farmers are commonly advised to feed their cows twice a day (Ngo et al., 2009). Roughage and concentrate are often offered to cows separately (Lam et al., 2010). At a given feeding period, roughages such as Napier grass (either chopped or not) may be given to cows first and then concentrates, usually during milking time (Mai et al., 2011). Similar to the supplying of excessively high concentrate diets to cows, feeding them concentrates separately from the roughage can also put them at risk of ruminal acidosis (Humer

39 et al., 2018). A study conducted to compare traditional and TMR feeding methods for SDF cows in Moc Chau (Son La province, north Vietnam) found that TMR improved dry matter intake, milk yield, and milk quality (Mai et al., 2011). However, this study also showed that in SDFs, feeding TMR did not bring economic benefits due to the high investment required, such as the need to invest in a TMR mixer. This may be at least one reason why Vietnamese SDF farmers prefer feeding concentrates separately to roughage rather than feeding them TMR.

A thorough analysis of SDF cow diets and how they are offered is required to confirm dietary regime inadequacy as a major cause of low productivity, health and welfare.

2.2.5 Land and housing management

The SDF farmers usually own relatively small land areas, ranging from 0.03–2.76 ha per farm (Suzuki et al., 2006; Lam et al., 2010). Only some have pasture land (Suzuki et al., 2006; Lam et al., 2010). Although barn designs and facilities are the main factors that affect microclimate conditions inside the barns, information on cowshed design is limited. Farmers usually build the cow barns by themselves using locally available materials. One report indicates barn areas average 29.4 m² with a median of 20 m², and a range from 7-200 m². The farmer's house is usually adjacent (Ashbaugh, 2010). Fifty-seven per cent of farmers in the survey reported using newly built facilities for their dairy farms, while 36.7% reported using previously built facilities and 6.7% of farmers reported using both new and previously built facilities. Previously built facilities were initially built as chicken, pig, or storage sheds (Ashbaugh, 2010). Typically, the barns are open on one or more sides, with open spaces in the walls or between the top of the wall and ceiling, for ventilation (Ashbaugh, 2010). Cows are usually kept inside the barns all day (Ashbaugh, 2010). A study by Suzuki et al. (2006) on 99 SDFs in Hanoi reported that 90% of SDFs kept cows inside cowsheds permanently, 87% of cows were on concrete flooring, and 67% of cowsheds were relatively closed in, with limited sides open.

2.3 Dairy regions

Dairy herds, especially SDF herds, are distributed widely in contrasting environments but are mainly concentrated close to the main milk demand areas of Ho Chi Minh City (the largest city in the South) and Hanoi (the capital and also the largest city in the North) (Cesaro et al., 2014; General Statistics Office of Vietnam, 2017a) (Figure 2.2). In 2016, 33 out of 63 provinces in Vietnam had a dairy farming industry, although about a half of the provinces had less than 2,000 cows and some provinces had virtually none (General Statistics Office of Vietnam, 2017a).

40

a) Topographic map of Vietnam mainland b) Distribution of smallholder dairy cows Map from https://en.wikipedia.org/wiki/Geography_of_Vietnam by district in 2011 Cartography from Cesaro et al. (2014)

Figure 2.2 Topographic map of Vietnam mainland and distribution of dairy cows by district Smallholder dairy farms distributed across the country, but the highest populations were in Ho Chi Minh and Hanoi, the two biggest cities in the north and south of Vietnam.

To highlight the environmental contrasts between provinces, the weather conditions of four provinces representing highland areas with cool climates (Lam Dong, Son La) and lowland areas with hot climates (Ho Chi Minh, Ha Nam), are described in Table 2.1.

41

Table 2.1 Average monthly rainfall (mm), temperature (AT, oC), humidity (RH, %), and temperature- humidity index (THI, unit) during a period from 2002 to 2016 at weather stations nearest to the study sites A

Parameter Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov De Average Station at Vung Tau, near Ho Chi Minh City Rainfall (mm) 2.3 10.8 11.1 63.5 171.9 181.7 211.3 184.0 189.3 215.4 52.9 22.5 129.6 RH (%) 75.4 75.8 75.9 76.0 78.4 80.2 81.7 81.3 82.2 81.5 78.1 76.8 78.6 AT (oC) 25.9 26.3 27.7 29.1 29.2 28.7 28.0 28.2 27.9 27.8 27.8 26.8 27.8 THI B 75.6 76.2 78.1 80.1 80.3 79.6 78.8 79.0 78.6 78.5 78.3 76.9 78.3 Station at Da Lat, Lam Dong province Rainfall (mm) 8.1 24.0 77.6 167.8 225.2 201.8 225.7 244.7 306.4 235.7 93.1 33.4 156.8 RH (%) 82.1 77.9 80.2 83.7 87.3 88.5 89.3 90.1 90.3 88.0 85.4 84.4 85.6 AT (oC) 15.9 16.9 18.1 19.2 19.7 19.4 18.9 18.8 18.7 18.3 17.9 16.7 18.2 THI 62.4 63.5 65.3 66.8 67.6 67.3 66.7 66.6 66.4 65.8 65.2 63.5 65.6 Station at Nam Dinh, near Ha Nam province Rainfall (mm) 34.9 23.5 41.2 68.1 178.6 148.6 251.5 288.0 291.8 110.6 58.8 21.5 126.4 RH (%) 82.9 86.7 87.7 86.5 82.1 79.4 80.8 84.7 84.5 80.9 80.2 79.3 83.0 AT (oC) 16.3 18.3 20.4 24.4 27.9 29.9 29.6 28.7 27.6 25.7 22.5 18.3 24.1 THI 62.9 65.7 68.6 74.0 78.6 81.2 80.9 79.8 78.3 75.6 71.2 65.6 73.5 Station at Son La Rainfall (mm) 37.8 20.9 46.8 120.0 170.4 200.0 276.9 264.7 146.7 53.0 37.7 34.1 119.4 RH (%) 79.7 75.5 72.9 74.9 77.7 83.1 85.4 85.5 83.5 80.9 80.7 79.7 80.0 AT (oC) 14.7 17.6 20.5 23.7 25.1 25.7 25.3 25.1 24.3 22.3 19.1 15.7 21.6 THI 60.7 64.4 68.3 72.6 74.7 75.6 75.2 74.9 73.8 70.9 66.7 61.9 70.0

A Data were derived from the General Statistics Office of Vietnam (2017), https://www.gso.gov.vn/SLTK/. B THI is calculated using equation: THI = AT + 0.36  Tdp + 41.2 (Yousef, 1985); with AT is dry bulb temperature (oC) and Tdp is dew point temperature (oC).

2.4 Assessment of farm productivity and cow welfare 2.4.1 Key performance indicators

Moran (2009) and Gonzalez-Mejia et al. (2018) outlined the key performance indicators (KPIs) which can be used to assess the performance and profitability of SDFs. These can be separated into two: firstly, feed-related KPIs, including stocking capacity, on-farm forage production, forage quality, concentrate feeding program, feed costs, milk income less feed costs; and secondly, herd- related KPIs, including herd proportion, per cent of productive cows, milk yield and quality, milk

42 production pattern, reproductive performance, and heifer management. These indicators can be used to identify the strengths and weaknesses of SDFs (Moran, 2009). Among them, milk yield and quality are the foremost indicators as they indicate production efficiency and turnover (Gonzalez-Mejia et al., 2018). These KPIs are also related to cow welfare assessment (Moran, 2015b). An animal will be in a good welfare condition if it is productive and healthy, and in return, if an animal is in good welfare condition, it will normally be healthy and productive (World Organisation for Animal Health, 2011).

2.4.2 Indicators of cow welfare

In developed countries, there are specific guidelines to monitor cow welfare (Australian Department of Agriculture; Welfare Quality® Consortium, 2009; de Vries et al., 2011; Popescu et al., 2014; BC SPCA, 2018; RSPCA, 2018) but few guidelines other than those recently proposed by Moran (Moran, 2015b) are available for SDFs.

Housing

Too high livestock density, bare concrete floor, tethering of cows, lack of feeding and watering troughs, and lack of cooling facilities for the cows during hot weather conditions all indicate poor cow welfare conditions (World Organisation for Animal Health, 2011; Moran, 2015b).

Feeding

Insufficient feed and water supply for the cows and inadequate feed intake indicate poor welfare conditions (World Organisation for Animal Health, 2011; Moran, 2015b).

Bodyweight and body condition score

Low body weight or body condition score indirectly indicates undernutrition (Agenäs et al., 2006). The body condition score (BCS) of a cow is an assessment of how fat it is. Although different countries might use different BCS score scales, high BCS values always reflect obesity, and low values reflect emaciation. In dairy cattle, both high and low BCS at calving is associated with reduced milk production, impaired reproduction, and reduced immune function. On the one hand, obesity may pose the risks of metabolic disorders, and on the other, thinness may pose the risk of discomfort in cold weather conditions (Roche et al., 2009). According to The British Columbia Society for the Prevention of Cruelty to Animals (BC SPCA, 2018), using a scale of one to five, the ideal BCS Range for growing heifers is 2.75 - 3.25; heifers at calving, 3.25 - 3.75; cows at calving, 3.25 - 3.75; early lactation, 2.50 - 3.25; mid-lactation, 2.75 - 3.25; late lactation, 3.00 - 3.50; and dry off, 3.25 - 3.75. Figure 2.3 shows the suggested BCS (between two bold lines) for a lactating cow that minimizes the

43 effect of energy balance and maximizes milk production while ensuring reproductivity, health, and welfare of the cow (Chagas et al., 2007; Roche et al., 2009).

Figure 2.3 Suggested body condition score profiles (10-, 8- and 5-point scales) for a lactating cow Adapted from Roche et al. (2009) and Chagas et al. (2007).

Udder health

Mastitis is the most costly disease of dairy farming due to economic losses from treatment costs, reduced milk production, increased labour costs and death and premature culling of dairy cows (Miller et al., 1993; Karimuribo et al., 2006; Almaw et al., 2008). Mastitis of the cow can be assessed by somatic cell count, using the California mastitis test (CMT), measuring milk electrical conductivity, or measuring milk electrical resistance (Miller et al., 1993; Van der Merwe et al., 2005; Hogeveen et al., 2010). Measuring milk electrical conductivity is commonly practised on commercial farms with automatic milking systems. However, for SDFs, using a portable mastitis detector such as the Draminski Mastitis Detector could be more convenient (Draminski, 2017). This device measures milk electrical resistance and classifies the cow's udder as a healthy udder if milk electrical resistance > 300 units; as a sub-clinically infected udder, 300 to 250 units; and as suffering from clinical mastitis, < 250 units (Draminski, 2017).

44

2.5 Heat stress 2.5.1 Heat stress and thermoregulatory in cattle

Heat stress in cattle has been defined as a situation when an animal’s body temperature is displaced from its resting or ground state due to the effects of the high temperature-related forces external to the cow’s body (Kadzere et al., 2002; West, 2003). It occurs when the animals cannot dissipate excess of endogenous or exogenous heat adequately to maintain their body temperature within a normal thermoneutral zone (Bernabucci et al., 2014). Cows gain heat from endogenous sources, which are metabolic heat productions, and exogenous sources, which are the heat gained from the environment through conduction, convection, and radiation (Kadzere et al., 2002). Cows dissipate heat into the environment through conduction, convection, and radiation, evaporation of water, and expired air (Kadzere et al., 2002). If the total amount of heat gain equals the total amount of heat loss, a cow is in homeothermy, as expressed by the equation: M  K  C  R – E – EA = 0, where M is the metabolic heat production; K, C, R, E, EA are the heat exchange by conduction, convection, radiation, evaporation, and expired air, respectively.

Both internal and external factors affect the thermoregulatory of the cattle. The factors that can increase metabolic heat production of the cow include: increase of heat increment due to increased production, shivering or exercise, tensing of muscles, increased metabolic rate, and fever due to disease (Kadzere et al., 2002). As a result, high producing cows are often more heat-stressed than the lower producing cow because the heat increment increases with increased milk production (West, 2003). For example, for cows producing more than 24 kg of milk per day, each kg increase in fat corrected milk was associated with an increase of 0.02°C in rectal temperature (Berman et al., 1985). On the other hand, heat loss can be enhanced by a cooler environment, panting, sweating, increased skin circulation (vasodilatation), shorter fur insulation, increased sensible water loss, increased radiating surface, and increased air movement or convection (Kadzere et al., 2002). In hot weather, cattle normally manage to balance their body temperature through avoiding exogenous heat sources, for example, by seeking cool places; decreasing endogenous heat production, for example by decreasing feed intake and production; and increasing heat dissipation– for example, by increasing respiration rate and sweating rate (Kadzere et al., 2002). A study in Mediterranean conditions by Bouraoui et al. (2002) showed that the daily THI correlated negatively to feed intake (r = -0.24) and milk yield (r = -0.76); and correlated positively with rectal temperature (RT) (r = 0.85), respiration rate (RR) (r = 0.89), heart rate (HR) (r = 0.88), and cortisol (0.31). When THI increased from 68 to 78, RT increased by 0.5°C, RR by five breaths/min, and HR increased by six beats/min.

45

Heat stress is predicted to become more serious in the future, especially in tropical countries, due to global warming (Morton, 2007; Henry et al., 2012; Sejian et al., 2015). Heat stress is more severe in tropical than temperate regions as both ambient temperature-humidity in the tropics are high. When the ambient temperature rises above the lower critical temperature, non-evaporative heat loss methods such as conduction, radiation, and convection become less efficient due to the decreased temperature gradient between the cow and the environment (Berman et al., 1985). Thus, in such circumstances, cows depend more on water evaporation and peripheral vasodilatation to enhance heat loss (Berman et al., 1985). However, when the humidity of the environment is high, the efficiency of heat loss through water evaporation decreases.

2.5.2 Method to assess levels of heat stress

The heat stress level of cattle can be assessed using environment-based indicators or animal-based indicators:

Environmental indicators

Four main climatic factors that define heat load in cattle are: ambient temperature, relative humidity, wind speed, and solar radiation. The most straightforward environmental indicator is ambient temperature. Some other more complicated indicators, which are calculated from environmental temperature and other climatic parameters, include the Temperature-humidity index (THI) developed by: Yousef (1985), Thom (1959), and the National Research Council (National Research Council, 1971); the Black globe-humidity index (BGHI) developed by Buffington et al. (1981); the Equivalent temperature index (ETI) developed by Baêta et al. (1987); the Heat load index (HLI) developed by Gaughan et al. (2008); the Adjusted temperature-humidity index (THIadj) developed by Mader et al. (2006); the Comprehensive Climate Index (CCI) developed by Mader et al. (2010); and the Index of thermal stress for cows (ITSC) developed by Da Silva et al. (2014). Comprehensive reviews by Herbut et al. (2018), Wang et al. (2018), and Rashamol et al. (2019) have summarized and compared the calculations and applications of those indicators. Among the environmental indicators, THI and HLI seem to be the most studied and applied in actual cattle production (Dunshea et al., 2013). Comparing THI and HLI, the advantage of HLI is that its calculation includes not only temperature and relative humidity as does THI, but also wind speed. Although environment-based indicators are used widely and can help predict the cow's heat stress level based on environmental parameters, they do not directly reflect the physiological changes in heat-stressed animals (Liu et al., 2019).

The thermoneutral zone is the range within which a change in a temperature-related parameter does not significantly induce physiological or behavioural changes amongst cows (West 2003). Based on

46 climatic indicators, the thermoneutral zone for lactating dairy cows is generally accepted as the temperature within a range of − 0.5 to 20.0°C (West, 2003), HLI < 70 (Gaughan et al., 2008), or THI < 68 (Zimbleman et al., 2009). When environmental conditions exceed the upper thresholds of the thermoneutral zone, feed intake decreases, cow performance is inhibited, and cooling cows becomes necessary (West, 2003; Könyves et al., 2017; Habeeb et al., 2018; Liu et al., 2019).

Animal-based indicators

Animal-based indicators to assess heat stress in cattle can be body temperature, respiration rate, heart rate, panting score (PS), sweating rate, dry matter intake, or behaviour (Gaughan et al., 2002; Wang et al., 2018; Galán et al., 2018; Rashamol et al., 2019). Many temperature measurements at different positions on the cow body can potentially be used to assess heat stress, such as body temperatures by direct contact measurements at rector, vagina, rumen, reticulorumen, skin (Liang et al., 2013; Dikmen et al., 2014; Lees, 2015). Alternatively, noncontact measurements using infrared technology can indicate temperatures of the cow’s body surface or milk (West et al., 2003). Reviews by Galán et al. (2018) and Liu et al. (2019) discussed the applications of these indicators.

Among the animal-based indicators, rectal temperature, respiration rate, heart rate, and PS (details below) are most commonly studied. Under thermoneutral conditions, rectal temperature, respiration rate, and heart rate of cattle are usually given as 38.5 ± 0.5ºC, 20 ± 10 breaths/min, and 70 ± 10 beats/min, respectively (Jackson and Cockcroft, 2002). When the ambient temperature (AT) increases from 4.4 to 37.8ºC, the rectal temperature remains quite constantly at 38.5ºC when AT < 26.7ºC, then it starts to increase considerably and reaches 40.6ºC when AT reaches 37.8ºC (Regan and Richardson, 1938). When AT increases from 4.4 to 37.8ºC, respiration rate increased rapidly from 12 to 124 breaths/min, but heart rate decreases slightly from 72 to 57 beats/min (Regan and Richardson, 1938).

Panting score

The PS is a well-established indicator of the extent to which cattle are heat-stressed (Gaughan et al., 2009). A PS of zero indicates a cow is breathing normally and is not panting, whereas a PS of 4.5 means a cow is excessively heat-stressed with rapid breathing from the flank, tongue fully extended, excessive drooling, neck extended, and head held down (Gaughan et al., 2009). Table 2.2 and Figure 2.4 describe how to score PS, and the correlation between PS and respiration rate (RR). Using PS, an animal is normal when PS = 0.0 to 0.4, slightly heat-stressed when PS = 0.4 to 0.8, moderately heat-stressed when PS = 0.8 to 1.2 and highly heat-stressed when PS > 1.2 (Gaughan et al., 2008; Lees et al., 2018). Assessment of heat stress by PS has been widely studied, and PS has been applied

47 to dairy and beef cattle under experimental and commercial farming conditions (Mader et al., 2006; Gaughan et al., 2009; Alfonzo et al., 2016; Unruh et al., 2017).

On comparing PS with other animal-based indicators, it is found that measurements of rectal temperature, respiration rate, and heart rate are quite complicated, invasive and time-consuming, and usually, there is a need to restrain the animal. These complications limit their application under actual production conditions. Measurement of PS, on the other hand, is simple and non-invasive, without the need to restrain the animal; and there is no need for accompanying equipment. Thus, PS seems to be very suitable for SDFs where restraint facilities are limited. Moreover, neither rectal temperature nor respiration rate responds directly to increased temperature; they usually lag behind ambient temperature changes by two to four hours (Hahn, 1995; Hahn et al., 1997; Gaughan et al., 2002). The disadvantages of PS are its subjectivity and dependence on the experience of the assessors.

Table 2.2 Assessment of panting score (PS) based on breathing condition and associated respiration rate (RR; breaths/min)

Adapted from Brown-Brandl et al. (2006b), Mader et al. (2006) and Gaughan et al. (2008).

48

Figure 2.4 Description of panting scores from 1 to 4.5 Photos from John Gaughan in the book, Tips and tools: Heat load in feedlot cattle (Meat and Livestock Australia, 2006).

2.5.3 Effects of heat stress on production and welfare

Several recent reviews have discussed the effects of heat stress on the productivity and welfare of cattle (Silanikove, 2000; West, 2003; Das et al., 2016; Polsky and von Keyserlingk, 2017). Briefly, heat stress affects all three key constructs of cow welfare, which are (1) the biological functioning and health, (2) the affective states, and (3) the natural living of the cows. Regarding functioning and health, heat stress causes reductions in feed intake, milk production, reproductive success, and function of the immune system, which ultimately causes economic loss (Kadzere et al., 2002; West, 2003; Hansen, 2007; Polsky and von Keyserlingk, 2017). Regarding affective states, heat stress induces hunger due to decreased feed intake, induces thirst due to increased water requirement, and causes strain, frustration, aggression, and lameness in cows (Polsky and von Keyserlingk, 2017). Finally, regarding naturalness, cows under high heat load conditions change their behaviour. They seek shade, bunch in tight groups, stand longer and ruminate less (Polsky and von Keyserlingk, 2017).

A decrease in feed intake, milk production, and milk yield are most often observed in heat-stressed cows. When THI increases from 68 to 78, dry matter intake decreases by 9.6%, and milk production increases by 21% (Bouraoui et al., 2002). When THI increases, a decline in daily milk yield of 0.08 to 0.26 kg per THI unit was reported in a study in Germany (Könyves et al., 2017), 0.41 kg per THI

49 unit in the Mediterranean (THI > 69) (Bouraoui et al., 2002), 0.2 kg per THI unit in the USA (THI > 72) (Ravagnolo et al., 2000), and 2.2 kg per THI unit in South Africa (THI > 65 to 73) (Preez et al., 1990). Cows’ milk fat (3.24 vs. 3.58%) and protein (2.88 vs. 2.96%) concentrations were also found in one study to be lower in summer compared to winter (Bouraoui et al., 2002). Although the mechanisms that mediate the effect of heat stress on milk yield can be multifactorial, at least half of the reduction is explained by the decrease in feed intake caused by heat stress (Tao et al., 2020).

2.5.4 Strategies to ameliorate heat stress in dairy cattle

Several recent reviews have discussed approaches to ameliorate heat stress in dairy cattle (West, 2003; Dunshea et al., 2013; Das et al., 2016; Polsky and von Keyserlingk, 2017; Lees et al., 2019). Briefly, there are three broad approaches: environmental, genetic and nutritional.

Environmental

Environmental approaches usually aim to minimize the heat that the cows gain from the environment and maximize heat dissipation from cows to the environment, based on the principles of convection, conduction, radiation, and evaporation (West, 2003). Firstly, cowsheds need to be designed and the building materials selected so as to minimize solar radiation exposure while maximizing air ventilation (Polsky and von Keyserlingk, 2017). Secondly, shade can be installed outside the cowsheds to reduce the cows’ exposure to direct solar radiation (West, 2003). Thirdly, some other cooling devices can be installed such as fans to increase air movement; foggers, misters, and sprinklers to enhance evaporative cooling; air conditioners; or tunnel ventilation systems to cool the air (West, 2003; Das et al., 2016; Polsky and von Keyserlingk, 2017).

In the case of SDFs, there are some difficulties in environmental approaches. Firstly, there are very few guidelines on how to design the SDF cowsheds. Commonly farmers build a cow shed on their own, based on their experience accumulated or learnt from other farmers (Ashbaugh, 2010). Secondly, it might not be possible for SDF farmers to choose the best materials and facilities due to their limited budgets. Also, evaporative cooling methods such as using foggers, misters, and sprinklers may be suitable in hot but low humidity environments, but in tropical conditions, they increase the humidity, which is already high. Thus, it is necessary to define affordable and suitable cowshed design parameters for SDFs in the tropics.

Nutritional

The formulation of diets for dairy cows during hot weather needs to take into account the reduced feed intake of cows caused by heat stress, the increased nutrient requirements of cows, especially for energy during hot weather, the dietary heat increment, and the energy cost of nutrient excesses (West, 50

2003). Thus, the common nutritional approaches to ameliorate heat stress are: (1) ensuring sufficient water for cows, (2) increasing dietary fat and starch concentrations to increase dietary energy density to compensate for reduced dry matter intake, (3) feeding lower fibre diets to decrease the extra dietary heat increment from the fermentation of fibre compared to starch or fat, (4) avoiding excesses of dietary crude protein to minimize metabolic heat production when excreting excessive nitrogen as urea, (5) supplementing vitamins C, E and A to minimize oxidative damage caused by heat stress; (6) supplementing with electrolytes to regulate acid-base balance in the blood, (7) supplementing with yeast by-products to decrease the production of ruminal ammonia and increase in ruminal microorganism population, (8) supplementing with betaine, and (9) altering feeding time to avoid the hottest hours of the day (West, 2003; Das et al., 2016; Lees et al., 2019).

Genetic

It is widely acknowledged that the dual-purpose Bos indicus breeds such as Gir, Red Sindhi, Sahiwal have greater heat tolerance compared to Bos taurus dairy breeds such as Holstein, Brown Swiss, or Jersey (Polsky and von Keyserlingk, 2017). Thus, the dual-purpose Bos indicus breeds or crosses thereof with Bos taurus dairy breeds are considered more suitable for tropical conditions if minimising heat stress is the main priority. However, the milk production of the Bos indicus breeds is low and does not meet the expectations of SDF farmers. Thus pure Bos taurus dairy breeds are still preferred for milk production in tropical SDFs. Among dairy cattle breeds, there is some evidence that Brown Swiss cows, and to a lesser extent, Jersey cows are more heat tolerant than Holstein cows (West et al., 2003; El-Tarabany et al., 2017), but this needs to be confirmed in SDF conditions in Vietnam.

Recently, some studies have indicated the potential for selecting heat-tolerant dairy cattle within Bos taurus dairy breeds. Olson et al. (2003) found the “slick” hair gene, which determines the character of the cow’s coat. Some studies have reported that the Holstein cows inheriting the slick hair gene associated with a short and sleek coat were more tolerant of heat than regular Holstein cows (Olson et al., 2003; Dikmen et al., 2008). A more long term approach is applying genomic selection to select heat-tolerant dairy cattle (Nguyen et al., 2016a). The effectiveness of this approach has been confirmed in a study by Garner et al. (2016), and this approach is currently being implemented in Australia (Nguyen et al., 2017a). This genetic approach might be suitable for the long-term development of dairy production in the tropics.

51

2.6 Advanced technologies with the potential to improve the productivity and welfare of smallholder dairy cows in Vietnam

Computer-based nutrition models, infrared thermal technologies, and genomic selection are advanced technologies that might be applicable for monitoring and improving the productivity and welfare of cows in Vietnamese SDFs. This section reviews computer-based nutrition models and infrared thermal technologies. Genomic technologies will be reviewed separately in the next section.

2.6.1 Computer-based dairy nutrition models

Empirical nutrition models, which include systems of mathematical equations, are commonly used to calculate the nutrition requirements of cattle, to estimate dry matter intake and nutrient concentrations of diets, and to predict the production of cows. Globally relevant models are based on equations derived from the feeding standards of the National Research Council (NRC, an American model), Agricultural Research Council (ARC, a British model), National Institute for Agricultural Research (INRA, a French model), and Commonwealth Scientific and Industrial Research Organisation (CSIRO, an Australian model) (Lean, 2011b; Tedeschi et al., 2014). Computer-based models developed from these feeding standards include PCDairy, Cornell Net Carbohydrate and Protein System (CNCPS), CPM-Dairy, CamDairy, Feed into Milk (FiM), Molly, Ruminant, and Rumen8 (Lean, 2011b; Robinson and Ahmadi, 2015). These computer-based nutrition models can provide a powerful means of assessing the nutrient balance of dairy cow diets and predicting animal responses to dietary nutrients or dietary manipulations. They can also formulate the least-cost or maximum- profit diets without the need for manual calculations (Kolver, 2003; Little et al., 2009; Lean, 2011b; Tedeschi et al., 2014).

As discussed in earlier sections, information required by such models on typical diets for Vietnamese SDF cows is minimal. Using computer-based nutrition models to analyse the current diet for SDF cows should help to systematically identify the limitations of the SDF diets for lactating cows. To enhance the balance of the SDF cows' diets whilst lowering the methane emissions of SDF cows, a cooperative program between the US Department of Agriculture (USDA), University of California Davis (UC-Davis), and the Vietnamese Government has initiated a training program for dairy farmers and local extension specialists (Mateo, 2016; Kebreab et al., 2019). Through this program, PCDairy, a piece of diet formulation software developed by scientists at UC-Davis, has been sponsored for use by farmers, nutritionists, and extensionists in the Vietnam dairy industry (Robinson and Ahmadi, 2015; Mateo, 2016). PCDairy was built based on the NRC nutrition system (2001) and equations to predict methane emissions, as suggested by Moraes et al. (2014). The advantage of PCDairy is that it

52 has been translated into Vietnamese. However, its limitation is that its current feed library does not include detailed nutritive value panels for locally relevant feedstuff.

2.6.2 Infrared thermal technology

As discussed in an earlier section, the rectal temperature measured by a thermometer is the most common way to measure the body temperature of cattle. However, it is time-consuming, potentially disruptive and stressful for the cows; and that stress may also lead to misleading results (Hoffmann et al., 2013). The rectal temperature of the cows is also difficult to measure under Vietnamese SDF conditions, as reliable cattle handling facilities are often not available. Infrared thermal temperature (IRT), measured by either infrared thermometers or thermal cameras, is a yet to be explored but potentially valuable option for monitoring the body temperature of cows in SDFs due to the minimal need for restraint.

Both infrared thermometers and thermal cameras work according to the same principle. They detect mid to long-wave infrared radiation emanating from an object and translate it into electronic signals, and these signals are then converted into a temperature reading (Usamentiaga et al., 2014; FLIR Systems, 2015; Tattersall, 2016). Infrared thermometers have only one infrared sensor, which enables the temperature measurement of only one spot (one pixel) and gives only one individual IRT reading. In contrast, thermal cameras have thousands of infrared sensors which enable the temperature measurement of thousands of spots (megapixels) on a surface to define maximum, average or minimum IRT for that surface (FLIR Systems, 2015). Infrared thermometers are a low-cost option and easy to buy and use. Thermal cameras are much more expensive, are supplied by few companies and are more complicated to use, although they can provide more sophisticated results. A review by Tattersall (2016) provides detailed background and guidelines for the proper use of thermal cameras by physiologists and biologists.

Advantages and disadvantages of IRT

Numerous advantages have resulted from the non-invasive and contactless approach of IRT technologies (Usamentiaga et al., 2014). These include: the ability to measure moving, physically inaccessible or distant animals; that there is no interference with the animal; that there is no risk of contamination; that repeated measurements can be taken with minimal preparation; that results are real-time; and that there is the potential to automate the process and continuously record the body temperature of cattle (Kastberger and Stachl, 2003; Hoffmann et al., 2013; Tattersall, 2016). When compared to the panting score, IRT measurements are found to be objective, while PS measurement is found to be subjective (Unruh et al., 2017).

53

However, the accuracy of IRT measurement is influenced by sometimes difficult to control variables including sunlight, wind, materials on the hair coat such as dirt and moisture, the distance of measurement, and the angle of measurement (Polat et al., 2010; Hoffmann et al., 2013; Church et al., 2014). When compared with rectal temperature, IRT measurement usually gives a lower temperature reading than the core body temperature of the cows because IRT measures the surface temperature of the animals. For example, a study by Church et al. (2014) reported that the mean IRT of the skin region immediately around the eyes, often considered to be the external surface region best correlated with core body temperature, was about 1.7°C lower than mean rectal temperature (37.55 oC vs 39.24oC).

The use of IRT technologies in dairy cow management

A review by Rekant et al. (2016) showed the veterinary applications of infrared thermography in cattle. Briefly, IRT has application for a hoof health assessment (LokeshBabu et al., 2018), mastitis detection (Berry et al., 2003b; Sathiyabarathi et al., 2016), oestrus detection (Talukder et al., 2014), and the prediction of heat stress in beef cattle (Unruh et al., 2017), purebred lactating Holsteins, lactating Girolando cows (crossbreeds of Holstein X Gir) (Daltro et al., 2017) and Jersey heifers (Salles et al., 2016). These studies have found positive correlations between the IRT measurements of cow body surface regions and: the panting score; physiological parameters including respiratory rate, heart rate, and rectal temperature; and environmental parameters such as THI. For example, Daltro et al. (2017) reported moderate positive correlations among IRT measurements taken at the eye, fore udder, and rear udder with: respiratory rate (r = 0.43, 0.57, and 0.50, respectively, P < 0.01); heart rate (r = 0.31, 0.47, and 0.37, respectively, P < 0.01); and with panting score (r = 0.37, 0.54, and 0.41, respectively, P < 0.01) in dairy cattle in Brazil. The authors concluded that the best region for the monitoring of heat stress in dairy cattle using IRT is the fore udder (Daltro et al., 2017). In beef cattle, Martello et al. (2016) found positive correlations between IRT measurement at the eye, cheek, flank, ribs, rump, and front feet regions and rectal temperature (r = 0.43 to 0.60, P < 0.01) and with respiration rate of the cattle (r = 0.37 to 0.47, P < 0.003). However, a study by Lees (2015) on feedlot cattle found no relationship between mean body IRT (head, shoulder, trunk, and rump) and core body temperature measured within the rumen. It is important to note that those studies (Martello et al., 2016; Daltro et al., 2017; Unruh et al., 2017) were all conducted under experimental conditions where confounding factors were carefully managed. It remains to be seen how reliable IRT is for the monitoring of SDF cows’ productivity and welfare given the diversity of climate, housing, nutrition and genetics in SDFs that could confound the

54 measurements (Polat et al., 2010; Church et al., 2014); but the potential for successful application is there.

2.7 Genomic technologies

A common breeding approach of dairy industries in tropical developing countries like Vietnam is that the high productive Bos taurus breeds, which commonly originated from temperate developed countries, are imported and used to cross with the locally adapted but low productive Bos indicus breeds to generate crossbreeds (Ducrocq et al., 2018). However, in tropical developing countries, the pedigree of the dairy cattle is not recorded properly; thus, the true genetic diversity of the population is hard to know (Ducrocq et al., 2018). Moreover, the lack of pedigree makes it difficult to apply traditional methods to select cattle of merit for breeding. Fortunately, the genomic technologies, which have been developing rapidly during recent decades, could help overcome these limitations (Fleming et al., 2018).

Genomic technologies are defined as those used to manipulate and analyse genomic data (Galas and McCormack, 2003). The development of genotyping technologies has generated a large quantity of genomic data (Fleming et al., 2018). The genomic data that are popularly exploited in animal breeding and genetics are single nucleotide polymorphisms (SNPs) data or sequencing data. Regardless of the availability of pedigree data, the development of mathematical algorithms and bioinformatics have enabled the exploitation of those genomic data to identify: the level of genetic diversity in a population; SNP-based estimates of breed composition of an animal; the genomic breeding value of animal traits of interest; or quantitative trait loci (QTLs) associated with quantitative traits (Meuwissen et al., 2016; He et al., 2018a). This section reviews some of the genomic technologies which might be applicable to SDFs in Vietnam.

2.7.1 Identification of genetic diversity and breed composition using genomic data

Genetic diversity and breed compositions

Knowing the genetic diversity of a population and knowing the breed composition of individual cattle are fundamentally important for understanding the evolution of genetic traits, for identifying the level of inbreeding and for designing breeding programs (VanRaden and Sanders, 2003; Bennewitz and Meuwissen, 2005; Stachowicz et al., 2011; Novembre, 2016).

Conventionally, the breed purity and composition are determined from pedigree records (Frkonja et al., 2012), but now it is possible to determine these based on SNP data (Kuehn et al., 2011; Melka and Schenkel, 2012; Sempéré et al., 2015; Visser et al., 2016; He et al., 2018a; b; Gobena et al., 2018; Lwin et al., 2018; Hulsegge et al., 2019). This could be extremely meaningful for Vietnamese SDF 55 herds, where pedigree records are often absent. Compared with the traditional pedigree-based method, the determination of breed composition using the SNP marker should be the more accurate method, as it is not affected by missing, incomplete, or inaccurate records (Frkonja et al., 2012; Dodds et al., 2014).

Statistical methods and software for genetic diversity analysis using SNP data

F-statistics (Wright, 1965), which include observed heterozygosity (HO), expected heterozygosity

(HE), inbreeding coefficients (FIS) of a population, and pairwise fixation index (FST) between populations, are the most common parameters used to assess the diversity level of populations. These parameters can be estimated using SNP data (Melka and Schenkel, 2012; Lwin et al., 2018). R packages such as ‘hierfstat’ (Goudet and Jombart, 2015), ‘dartr’ (Gruber et al., 2018), and ‘diveRsity’ (Keenan et al., 2013) can be used to estimate the F-statistics using SNP data.

Statistical methods and software for SNP-based estimates of breed composition

Genomically determined breed composition of an individual animal can be defined as the proportion of the genomic contribution of each ancestor breed to that animal (Frkonja et al., 2012; He et al., 2018a). SNP-based estimates of breed composition are usually estimated for mixed breed (admixed) animals and animals whose ancestors originated from different populations (He et al., 2018a). Using SNP data, the breed composition of an animal can be estimated using an admixture modelling method (Tang et al., 2005; Alexander et al., 2009; Frichot and François, 2015) or linear regression modelling method (Chiang et al., 2010; Kuehn et al., 2011). A study by He et al. (2018b) described and compared the two methods. The authors reported that the results of the admixture model were more consistent than those of the linear regression model (He et al., 2018b).

The paper of Liu et al. (2013) reviewed some methods and available software for studying population structure and identifying SNP-based estimates of breed composition using genomic data. Briefly, the publicly available software that enables identification of SNP-based estimates of breed composition of animals includes R package ‘LEA’ (Frichot and François, 2015), ADMIXTURE (Alexander et al., 2015), STRUCTURE (Porras-Hurtado et al., 2013), LAMP (Sankararaman et al., 2008), ADMIXMAP (Hoggart et al., 2004), and frappe (Tang et al., 2005).

2.7.2 Genetic parameter estimate and genomic selection in dairy cattle using genomic data

The main goal in animal breeding is to select as quickly as possible the animals with high breeding values for traits of interest to use as parents to produce subsequent generations (Dekkers, 2012). An essential prerequisite for designing animal breeding programs is knowing the heritability of the traits, knowing the genetic and phenotypic correlations between traits, and knowing the accuracy of 56 estimated breeding values for those traits. Traditionally, the best linear unbiased prediction (BLUP) method is used to calculate the heritability and estimated breeding values (EBVs) of the traits based on phenotypic records and the family relationship matrix, which are calculated from the pedigree records (Henderson, 1986; Mrode and Thompson, 2013; Weigel et al., 2017). Nowadays, many computing algorithms are available to estimate the heritability and breeding values of the traits using genomic data. The heritability of a trait estimated using genomic data is termed genomic heritability (de los Campos et al., 2015), and the selection of animals based on the genomic data is termed genomic selection (GS) (Meuwissen et al., 2001). Breeding values estimated using genomic data are termed genomic estimated breeding values (GEBVs).

Compared to traditional selection methods, which are based only on pedigree data, GS methods make the most of the available genetic information, including genomic data alone, or a combination of genomic data with pedigree data, when determining the animals with the most merit. Genomic selection showed increases in the accuracy of GEBVs when compared with traditional selection methods (Vanraden et al., 2009; Boichard et al., 2012; Thomasen et al., 2012).

Implementation of genomic selection

The implementation of GS generally includes five steps (Figure 2.5) (Kang et al., 2017): (1) a training/reference population is constructed, where all animals in this population are measured for phenotype data and genotyped by SNPs; (2) an appropriate statistical approach to obtain the equation estimating GEBVs from genotype and phenotype data of that reference population is applied; (3) all candidates are genotyped with the SNPs; (4) the GEBVs of each candidate are calculated by using the candidate’s genotype obtained from Step 3 and the equation obtained from Step 2; and (5) the candidates are ranked based on their GEBVs so that individual animals can be selected based on the desired level of genetic merit.

57

Figure 2.5 Principle of genomic selection method Adapted from Kang et al. (2017)

Statistical methods for GS

The challenge in GS is how to generate the prediction equation showing the effect of each SNP on the trait when the number of SNPs is much larger than the number of known-phenotype animals (Hayes et al., 2013). Many statistical methodologies and software programs have been developed to solve this problem by maximizing the use of prior information on the distributions of the SNP effects. Two main statistical approaches commonly applied to implementing GS are the frequentist method and the Bayesian method. Hayes et al. (2013) and de Los Campos (2013) have reviewed some of the approaches (Table 2.3). Some prior assumptions for distributions of SNP effects are presented in Figure 2.6.

58

Table 2.3 Some common statistical approaches for genomic selections

Method SNP effect assumption Software References rrBLUP (R package), Gaussian (Meuwissen et al., 2001; sommer (R package), GBLUP Covarrubias-Pazaran, 2016; AsREML, GenStat, Normal distribution of SNPs Endelman, 2017) with small effects Wombat, BGLR (R package), Scaled-t (Meuwissen et al., 2001; Abdel- gdmp (R package), Azim, 2016; Campos and BayesA BMTME (R Student's t distributions with a Rodriguez, 2016; Montesinos- package), small probability of large to López et al., 2019) moderate effects GenSel, BGLR (R package), Scaled-t mixture (Meuwissen et al., 2001; Abdel- gdmp (R package), Azim, 2016; Campos and BayesB BMTME (R A proportion of the SNPs with Rodriguez, 2016; Montesinos- package), zero effect, the remainder with López et al., 2019) Student's t distribution GenSel Gaussian mixture BGLR (R package), (Campos and Rodriguez, 2016; BayesC A proportion of the SNPs with BMTME (R package) Montesinos-López et al., 2019) zero effect, the remainder with Gaussian distribution Double exponential (Yi and Xu, 2008; de Los EBglmnet (R Bayesian Campos et al., 2013; Huang and Double-Exponential package), LASSO Liu, 2016; Montesinos-López et exponential distribution with a BMTME (R package) small probability of large to al., 2019) moderate effects A proportion of the SNPs with BayesCpi (Habier et al., 2011; Abdel-Azim, zero effect, the remainder with gdmp (R package) 2016) normal distribution A series of normal BayesR distributions, one with a mean - (Erbe et al., 2012) and variance of zero A proportion of the SNPs with almost zero effect, the BayesSSVS - (Verbyla et al., 2009) remainders with Student's t distribution Adapted from Hayes and Goddard (2010), Pérez and de los Campos (2014), and Ferrão et al. (2017).

59

Figure 2.6 Some prior assumptions for distributions of SNP effects in statistical approaches used in genomic selection Image from Pérez and de los Campos (2014).

Validation strategies and measures of prediction performance

In simulation studies, when true breeding values (TBVs) are available, the accuracy of prediction is calculated as the average correlation of GEBVs and TBVs of replicates (Meuwissen et al., 2001; Sun et al., 2012; Jiang et al., 2015). In empirical studies, there are two alternative approaches to assess the accuracy of GS. In the first situation, when some individuals have highly accurate EBVs estimated from pedigree and phenotype data, the training-testing approach can be used (Kang et al., 2017). In this approach, the highly accurate EBVs can be considered an alternative to TBVs (Kang et al., 2017). This situation is common in dairy cattle where some elite bulls have EBVs with accuracies up to 0.99 (Kang et al., 2017). In this case, animals are divided into a training population and a test population based on their ages (Vanraden et al., 2009; Luan et al., 2009; Li et al., 2014). Then, phenotypes of individuals in the test population are masked and only used when the GEBVs of these individuals have been estimated from a GS (Kang et al., 2017).

60

In the second situation, when the whole data set is small, the k-fold cross-validation approach is commonly applied (Luan et al., 2009; Sun et al., 2012; Fernandes Júnior et al., 2016). In this approach, individuals are divided into k same-sized subsets (Kang et al., 2017). One subset is used as a test population, and the others are used as training populations (Kang et al., 2017). Phenotypes of individuals in the test set are masked until their GEBVs are predicted (Kang et al., 2017). Then the correlation between their real phenotypes and GEBVs is the accuracy of prediction (Kang et al., 2017). Because there are k subsets, this process will be repeated k times so that each subset is treated as a test population. K is a critical number and is commonly five or ten (Daetwyler et al., 2013; Kang et al., 2017). To avoid inflated accuracy resulting from close relationships between training individuals and test individuals (Habier et al., 2007; Wientjes et al., 2013), subdivision into subsets should be based on the family, strains, and lines (Legarra et al., 2008; Sun et al., 2012). Besides k- fold cross-validation, repeated random sub-sampling validation can be used (Jiang et al., 2015). In this approach, the whole data set is divided into training sets and test sets randomly (Kang et al., 2017). Then the estimated accuracy of the statistics is calculated based on the repeated calculations (Kang et al., 2017).

Besides prediction accuracy, the area under the receiver operating curve (AUC), mean squared error (MSE) of prediction, and bias are other common measurements of prediction performance (Kang et al., 2017). AUC is usually applied for genomic prediction of binary/disease traits (Kang et al., 2017). The prediction with the highest value for the AUC is the best, and an AUC of one means the prediction is perfect (Kang et al., 2017). MSE evaluates the overall quality of prediction and is usually computed as the average square of the difference between TBV and GEBV (Kang et al., 2017). The smaller the MSE, the more precise and accurate the prediction performance will be (Kang et al., 2017). Bias is usually measured by the regression coefficient of GEBVs and TBVs (Daetwyler et al., 2013).

Factors affecting the prediction accuracy

Many factors can affect the accuracy of GEBVs estimated by GS.

Linkage disequilibrium between markers and quantitative trait loci: Some facts explain the effects of linkage disequilibrium (LD) between markers and quantitative trait loci (QTLs) on prediction accuracy. Firstly, higher accuracies can be gained with higher-density SNP chips or sequencing data (Solberg et al., 2008; Habier et al., 2009; Sun et al., 2012). This is because the LD between markers and QTLs is expected to be greater when a higher density of markers is applied. Thus, imputation of the low-density SNP data to high-density marker data is an effective solution to both reduce genotyping cost and improve prediction accuracy (Khatkar et al., 2012; Weng et al., 2013). The second fact is that the accuracy of both cross-population prediction and multi-population prediction 61 is relatively low. In a study by Hayes et al. (2009b), accuracies of GS for Jersey cattle using a training Holstein population and vice versa ranged from -0.06 to 0.23 for five traits. When including distant breeds in the training population, Moghaddar et al. (2014) saw either zero or negative effects on prediction accuracy. These low outcomes of the cross- or multi-population GS could be because the different populations have different LD patterns, allele frequencies, and QTLs (de Roos et al., 2009). The third fact is that the accuracy of GS declines when the number of generations between the training population and test population increases (Muir, 2007; Wolc et al., 2011; Sun et al., 2012; Akanno et al., 2014). The decrease in GS accuracy in this situation could be because, after many generations, the LD between markers and QTLs is broken, and the family relationships between training and test populations have become slighter (Kang et al., 2017).

Heritability of traits: More accurate results from GS can be derived for traits with high heritability (Luan et al., 2009; Guo et al., 2010; Hayes and Goddard, 2010). Heritability also dictates the required size of the training population, the accuracy of the response variables of the training population, and finally, the accuracy of predicted GEBVs of individuals in the test population (Hayes and Goddard, 2010; Kang et al., 2017).

Gene or quantitative trait loci effects: The common assumption of models used in GS is that the effects of genes are additive (Hayes et al., 2013). However, dominance and epistatic interaction effects also exist and contribute to the genetic variances of some traits (Sun et al., 2014; Guo et al., 2016). Therefore, the inclusion of these effects in the models may increase the GS accuracy. In actual data, some studies that included dominance effects in the models have shown improvement in GS accuracy (Nishio and Satoh, 2014; Sun et al., 2014; Guo et al., 2016), but others have not (Ertl et al., 2014; Sun et al., 2014; Santos et al., 2016).

Training population: The accuracy of GS increases when the size of the training population increases (Meuwissen et al., 2001; Zhong et al., 2009). The training population size should be large enough to accurately estimate the marker effects (Kang et al., 2017). Also, fewer relationships between the individuals in training sets result in greater accuracy in GS (Pszczola et al., 2012; Rincent et al., 2012). Relationships between the training set and the test set also affect the accuracy of GS. The closer the family relationships between the training and test sets, the higher the GS accuracy (Pszczola et al., 2012; Rincent et al., 2012; Habier et al., 2013; Wientjes et al., 2013).

The statistical methods: The accuracy of genomic predictions using the different methods listed in Table 2.3 is almost the same for most traits (Verbyla et al., 2009; Hayes et al., 2013). However, the BayesA, BayesB, and BayesR methods can perform better than GBLUP for traits with mutations of moderate to large effects (Verbyla et al., 2009; Habier et al., 2011; Erbe et al., 2012). Also, when the 62 number of SNPs is very large, for example, with sequence data, the methods that accept a large proportion of SNPs with zero effect can perform better than GBLUP (Erbe et al., 2012; Hayes et al., 2013).

Applications and achievements of genomic selection

Genomic selection has been successfully implemented in many countries, including the United States, Australia, England, Canada, New Zealand, Ireland, the Netherlands, France, Germany, and the Scandinavian countries (Silva et al., 2014). Genomic selection has been applied for almost all current traits used in dairy cattle selection programs such as milk production, body type, reproduction, health, workability and longevity traits; and has been applied in almost all dairy cattle breeds such as Holstein, Brown Swiss, Jersey, Ayrshire, and Guernsey Wiggans et al. (2017). A review by Miglior et al. (2017) reported heritably ranges of commonly selected traits in cattle from the Interbull project. Recently, GS has been applied for novel traits such as heat-tolerance (Garner et al., 2016; Nguyen et al., 2016a, 2017a).

García-Ruiza et al. (2016) reported that after nearly two generations of application of GS in the US Holstein cattle, the rates of annual genetic gain for yield traits increased from 50 to 100%. These rates for lower heritable traits such as herd life, female fertility, and somatic cell count increased three to four times (García-Ruiza et al., 2016). These achievements are derived because much younger bulls can be used as bull sires thanks to the application of GS (García-Ruiza et al., 2016). The sire-to-bull generation interval decreased from approximately seven years (before 2008) to just about 2.5 years in 2015 (García-Ruiza et al., 2016). Similarly, accuracy for selecting bull dams increased, and the average generation interval decreased from 4 years to 2.5 years (García-Ruiza et al., 2016).

2.7.3 Genome-wide association studies

Genome-wide association studies

Genome-wide association studies (GWAS) refers to studies in which numerous variants (most popularly SNPs) across the genome are genotyped and tested for associations with the traits of interest (Zeng et al., 2015; Carpio et al., 2018; Wang et al., 2019). The purposes of a GWAS can be to (1) identify risk alleles that can be used to predict the genetic predisposition of an individual to clinical diseases (Oetting et al., 2017), (2) discover the causal mutations controlling a trait (Henshall, 2013), (3) understand the quantitative trait loci that control quantitative traits (Liu et al., 2018), and (4) identify the genetic markers that can be incorporated into GS (Gebreyesus et al., 2019).

63

Steps in genome-wide association studies

Generally, a GWAS can be generalized into five main steps (Figure 2.7), including post-GWAS analysis: (1) identifying phenotypes and population of interest and considering the population structure of the interested population; (2) designing the trials to collect the phenotype data from a large number of individuals and taking samples from those individuals for genotyping with SNP platforms; (3) choosing a statistical model suitable for the collected phenotype and genotype data, taking into account the population structure or relatedness between samples; (4) using specialized software to perform quality control and conduct GWAS statistical analysis to identify the SNPs associated with the phenotype of interest; and (5) carrying post-GWAS analysis to identify biological evidence and functional annotations associated with the significant SNPs.

Figure 2.7 Common steps of genome-wide association studies (GWAS) Adapted from Carpio et al. (2018).

Statistical methods for genome-wide association studies

Several review papers have summarized the statistical method commonly used in GWAS (Zeng et al., 2015; Schmid and Bennewitz, 2017; Wang et al., 2019). Other papers and books have explained key concepts, illustrated the steps to conduct GWAS, and provided hands‐on practice with genomic data sets and scripts (Gondro et al., 2013; Carpio et al., 2018; Marees et al., 2018). Reed et al. (2015) outlined the main steps and necessary tools to conduct a complete GWAS analysis: reading and manipulating the genotype and phenotype data, conducting quality control of samples and SNPs, performing GWAS analysis, visualizing the GWAS results, and conducting the post-GWAS analyses.

64

Similar to GS studies, two main statistical approaches in GWAS are the frequentist approach and the Bayesian approach (Zeng et al., 2015; Schmid and Bennewitz, 2017; Wang et al., 2019).

Suggestive and significant thresholds for genome-wide association studies

In studies on humans, a common significant association threshold for SNP effects, which is Bonferroni corrected, is −log (5×10−8), and a less stringent suggestive association threshold is −log (5×10−6). In studies on animals, many thresholds have been reported. The sample size in a GWAS study can start from around 200 animals as in the study of Howard et al. (2014), to a huge number, for example, 294,079 Holsteins in the study of Jiang et al. (2019). In general, the larger the sample size, the smaller the significant association thresholds commonly used, and so more accurate results can be achieved.

Visualizing results of genome-wide association studies

After performing GWAS, some types of plots are usually used to visualize the GWAS results and perform quality control checks. The most common types of plots used for visualization of GWAS are Manhattan, quantile-quantile (Q–Q), heatmap and regional association plots (Reed et al., 2015). Manhattan plots are used to visualize the SNP effects and significance level by SNP locations on the chromosome and check if SNP effect signals are spurious or true (Schmid and Bennewitz, 2017). Q– Q plots are used to visualize the relationship between the observed and expected distributions of SNP- level test statistics to check if potential confounders such as population structure or sex have been adjusted (Schmid and Bennewitz, 2017). Heatmap and regional association plots are used in the context of GWAS to visualize the patterns of linkage disequilibrium between significant SNPs identified by GWAS and SNPs in adjacent regions that have been reported previously (Reed et al., 2015).

Software for genome-wide association studies

Almost all software used in GS also has functions for the conduct of GWAS. The common publicly available software that can be used to conduct GWAS are PLINK (Purcell et al., 2007) and R packages such as ‘sommer’ (Covarrubias-Pazaran, 2016), ‘rrBLUP’ (Endelman, 2011), ‘GenABEL’ (Aulchenko, 2015), ‘BGData’ (Grueneberg and Campos, 2019) and ‘BGLR’ (Pérez et al., 2014).

Validation strategies in genome-wide association studies

An ultimate goal of a GWAS is to identify the true association between an allele, or multiple alleles, of a single nucleotide polymorphism (SNP) with a phenotype or clinical outcome (Oetting et al., 2017). Genome-wide association studies have successfully identified the SNPs associated with

65 diseases and quantitative traits (Buniello et al., 2019). However, there have also been many GWAS based results that have found false positives for the latter (Oetting et al., 2017). Therefore, the validation of GWAS results is crucial. Also, when a GWAS identifies significant associations in a population, it is crucial to verify if those associations also exist in other populations (Henshall, 2013).

Two strategies commonly applied to validate the GWAS results are replication and meta-analysis (Oetting et al., 2017). Replication of GWAS in independent datasets is the most popular method of validating GWAS results (Konig, 2011). The validation dataset must not include samples that were already used in the discovery dataset (Henshall, 2013). Validation by replication can be internal validation or external validation. It is termed internal validation if both validation samples and discovery samples are drawn from the same population, and termed external validation if validation samples and discovery samples are drawn from different populations (Konig, 2011). Another common validation strategy is through meta-analysis studies (Zeggini and Ioannidis, 2009; Thompson et al., 2011).

Analysis after genome-wide association studies

Huang et al. (2009) discussed the underlying algorithms and pertinent details of approximately 68 bioinformatics enrichment tools that can be applied for the comprehensive functional analysis of large gene lists. These tools commonly provide a comprehensive set of functional annotation functions, which enable investigators to understand the biological meaning behind large gene lists (Huang et al., 2009). Functional annotations commonly applied to the gene list resulting from GWAS studies could be Gene Ontology (GO) term enrichment analysis, functional annotation clustering, BioCarta & Kyoto Encyclopaedia of Genes and Genomes (KEGG) pathway mapping of genes, and gene-disease association (Dennis et al., 2003). Some software and websites, which have been most widely used recently for gene set enrichment analysis, are DAVID (https://david- d.ncifcrf.gov/), Database for Annotation, Visualization, and Integrated Discovery (Dennis et al., 2003); WebGestalt (http://www.webgestalt.org), an integrated system for exploring gene sets in various biological contexts (Zhang et al., 2005); Enrichr (http://amp.pharm.mssm.edu/Enrichr/), a comprehensive gene set enrichment analysis web server (Kuleshov et al., 2016); PANTHER (http://pantherdb.org/), a large-scale gene function analysis system (Mi et al., 2013); and SNP2GO, an R package for Functional Analysis of Genome-wide association studies (Szkiba et al., 2014).

Application of genome-wide association studies in dairy cattle

A review of Sharma et al. (2015) provided an overview of successful stories and challenges to GWAS in livestock species, mainly in cattle. For example, GWAS has been conducted on milk production

66 traits such as milk yield, fat yield, protein yield, milk fat and protein concentrations (Streit et al., 2013; Nayeri et al., 2016; Jiang et al., 2019); conformation traits such as heart girth, hip height, body weight, and body condition score (Abo-Ismail et al., 2017; Zhang et al., 2017; Yin and König, 2019); female fertility traits such as heifer first service to calving interval, calving to first service interval, daughter fertility, days open, or calving performance (Nayeri et al., 2016; Abo-Ismail et al., 2017); beef cattle tympanic and vaginal body temperature (Howard et al., 2014); dairy cow rectal temperature, respiration rate and sweating rate (Dikmen et al., 2013, 2015); and slick hair coat (Huson et al., 2014). Through the application of GWAS in dairy production, many genomic markers or even candidate genes have been discovered and proven to have close associations with the traits (Ogorevc et al., 2009; Casas et al., 2016; Raven et al., 2016; Boichard et al., 2016).

2.7.4 Application of genomic technologies in developing countries

Although both traditional selection methods and modern genomic technologies have been applied popularly in developed countries, their application for animal breeding in developing countries has been minimal, especially in SDFs in tropical countries (Bortolussi et al., 2005; Boerner et al., 2014; Ducrocq et al., 2018; Mrode et al., 2019). The main reasons for the limited application are the lack of pedigree records, the limited facilities to record phenotype data accurately, the requirements for high levels of expertise in quantitative genetics and modelling to perform genomic analysis, and the high cost of genotyping and collecting phenotype data (Barendse, 2017; Ducrocq et al., 2018).

2.8 Conclusions and identified research gaps

In Vietnam, SDFs are the most popular type of dairy farms and account for most national milk production, but the cow productivity in SDFs remains relatively low, and the welfare of the SDF cows has not received much attention.

Dairy cows in SDFs in Vietnam are likely to suffer from unbalanced diets, heat stress, and poor genetics.

Basic information on the cow breeds, diets for the cows, and husbandry practices of SDFs in Vietnam is minimal and not uniform. There are no pedigree and individual data for SDF cows.

Vietnam has not yet had any unified national breeding program with specific goals for dairy cattle. Dairy sectors currently depend heavily on the importation of dairy breeds, including bulls, semen, and heifers.

Infrared thermal technology is a non-invasive technology with great potential for application in animal production. This technology has been studied widely under experimental farm conditions and

67 under commercial farm conditions, but limited studies have examined the applicability of this technology to SDF cows in the tropics.

Although modern molecular genomic technologies such as SNP-based estimates of breed composition study, GWAS and GS have been applied widely and successfully in developed countries, their application has been constrained in tropical developing countries, and they have never been applied in Vietnam.

68

Chapter 3 Performance and welfare status of Vietnamese smallholder dairy farms

Abstract

Smallholder dairy farms (SDFs) are the most popular type of dairy farms in Vietnam. They are distributed widely across lowland and highland regions, but data on the productivity and welfare status of the cows on these farms remains limited. This cross-sectional study aimed to describe and compare the productivity and welfare status of SDF cows across contrasting regions. The study was conducted in autumn 2017 on 32 SDFs randomly selected from four typical but contrasting dairy regions (8 SDFs per region); a southern lowland, a southern highland, a northern lowland, and a northern highland region. Each farm was visited over 24 hours (an afternoon followed by a morning of milking and adjacent husbandry activities) to collect the data of individual lactating cows (n = 345), which included: number of lactations, days in milk, inseminations per conception, milk yield and concentrations, body weight (BW), body condition score (BCS, 5-point scale, 5 = very fat), and level of heat stress experienced (Panting score, 4.5-point scale, 0 = no stress). Kruskal-Wallis, One- way ANOVA, and Chi-square independence tests were applied to compare not-normally distributed, normally distributed, and categorical variables, respectively, between regions. The high level of heat stress (96% of cows were moderate to highly heat-stressed in the afternoon), low energy corrected milk yield (15.7 kg/cow/d), low percentage of lactating cows (37.3% herd), low BW (498 kg), and low BCS (2.8) were the most critical productivity and welfare concerns determined, and these were most serious in the southern lowland. Cows in the northern lowland, a relatively hot but new dairying region, performed similarly to those in the southern highland, a region historically considered to be one of the most suitable for dairy cows in Vietnam due to its cooler environment. This indicates the potential to mitigate heat stress through new husbandry and genetic strategies. Cows in all regions were heat-stressed during the daytime, although less so in the highlands than in the lowlands. The region having the lowest milk yield was the southern lowland, which is a concern, given that it contains provinces with the highest numbers of dairy cows in Vietnam.

69

3.1 Introduction

Modern dairying demands the simultaneous improvement of cow productivity and welfare (Cardoso et al., 2016). Productivity, as indicated mainly by milk yield and quality, drives the biological and economic efficiency of the herd. Welfare, as indicated by the physical and mental health of the cow and provision of their basic needs such as freedom from pain, hunger and thirst, drives societal perceptions of dairying, which in turn can drive market demand and price received for dairy products (Rushen et al., 2008; Dairy Australia Limited, 2014; Moran, 2015a; Moran and Doyle, 2015b; Cardoso et al., 2016).

In Vietnam, the smallholder dairy farms (SDFs) with herd size ranging from two to 50 cows (Nguyen et al., 2016b) are the most common dairy farming systems. There were approximately 28,695 SDFs at the last census (General Statistics Office of Vietnam, 2016), and they account for 97% of the national dairy herd (Trach, 2017a) and 80% of the national fresh milk production (Vinamilk, 2017). Like many tropical countries in South East Asia, dairy industry professionals and farmers aim to reduce heat stress to obtain a higher milk yield. Consequently, the Vietnamese government initially prioritised developing SDFs in the highlands, which are regions such as Lam Dong province in the south and Son La province in the north, and considered most suitable for taurine dairy breeds. However, due to the increasingly high demand for fresh milk, SDFs expanded rapidly in the relatively hot lowland provinces with major populations, such as Ho Chi Minh city in the south and Hanoi city in the north. Recently they expanded to new lowland regions adjacent to the major cities, which are traditionally rice cultivating regions, such as Ha Nam province in the north or Tay Ninh in the south.

It is essential to record cow productivity and assess cow welfare regularly (Australian Department of Agriculture; Welfare Quality® Consortium, 2009; de Vries et al., 2011; Popescu et al., 2014; BC SPCA, 2018; RSPCA, 2018). However, in Vietnam, SDF farmers either do not have or rarely keep the records for the herd, cow pedigree, individual milk yield and composition. Also, there is no government regulation of dairy cow welfare in Vietnam and welfare considerations appear to have not been fully appreciated by Vietnamese farmers, dairy extension and technical advisors (Ban, 2014; Trach, 2017b). Consequently, the data on the production and welfare status of Vietnamese dairy cattle remain limited. A few reports, most likely based on verbal advice from the farmer rather than on direct measurement, indicate relatively low individual milk yields in the order of 14 kg/cow/d for cows in some provinces in the south (Vu et al., 2016) and 15 kg/cow/d for cows in Vinh Phuc province in the north (Ashbaugh, 2010). One of the few studies on heat stress is that of Lam et al. (2010), conducted in an exceptionally hot and humid lowland region near Ho Chi Minh city at the hottest 70 time of the year, from May to June, 2006. Lam et al. (2010) found an average temperature-humidity index (THI) of 85 units at 1400 h, a respiration rate of 70 breaths/min (versus an unstressed target of < 40 breaths/min at THI < 68; Pinto et al., 2019), and rectal temperatures of 39.3°C (versus an unstressed target of 38.6°C in Friesian cows: Andersson and Jóhasson, 1993; Ma and Du, 2010). Whilst it is reasonable to expect that cow productivity and welfare are improved by locating SDFs in the cooler highland provinces such as Lam Dong in the south or Son La in the north, there has not been a study that has systematically compared the productivity and welfare of SDF cows between the highland and lowland regions. In addition, the climatic records indicate that heat stress is still possible in all regions of Vietnam (General Statistics Office of Vietnam, 2017b).

Realizing the limitation of available data, we conducted this study to describe and compare the current stage of productivity and welfare of SDF herds across contrasting dairy regions of Vietnam to identify opportunities for future research. Four dairy regions were selected: a southern lowland (SL, Ho Chi Minh city), a southern highland (SH, Lam Dong province), a northern lowland (NL, Ha Nam province), and a northern highland region (NH, Son La province). Anecdotal evidence from industry specialists within Vietnam is that the productivity and welfare of Vietnamese SDF cows are impaired in some regions by factors including heat stress, simple diet formulations based mainly on high-fibre tropical forage such as Napier grass (Pennisetum purpureum) or rice straw plus concentrate at a ratio of 1 kg concentrate pellets to 2 kg of milk yield, high cow density per farm due to lack of land, simple husbandry techniques, lack of science-based advisory support and a lack in the application of animal welfare strategies (Cuong et al., 2006a; Moran, 2012; Hiep et al., 2016). We hypothesized that cows in the highland SDFs could be more productive and have better welfare than cows in the lowland SDFs in Vietnam.

3.2 Materials and methods 3.2.1 Farm selection

This cross-sectional study was conducted from 24 August to 7 October 2017 on 32 SDFs randomly selected from four main dairy regions of Vietnam with 8 SDFs per region and a total of 345 lactating cows. The four main dairy regions selected included: a southern lowland region (SL), which is the Cu Chi district in Chi Minh city (10.82oN, 106.63oE, 5-16 m above sea level); a southern highland region (SH), the Don Duong district in Lam Dong province (11.58oN, 108.14oE, 800-1000 m above sea level); a northern lowland region (NL), the Duy Tien and Ly Nhan districts in Ha Nam province (20.58oN and 105.92oE, 0.4-12 m above sea level); and a northern highland region (NH), the Moc Chau district in Son La province (21.33oN, 103.91oE, 600-700 m above sea level) (Figure 3.1). The longitude and latitude of the studied regions were derived from Google Earth Website 71

(https://www.google.com/earth/), and the altitudes were derived from the Website of Ministry of Planning and Management, 2017 (http://www.mpi.gov.vn/Pages/tinhthanh.aspx). These four regions were chosen due to their relatively long dairying history and current popularity as dairying areas (SL, SH, and NH regions); a short dairying history (NL region) but with recent government support to become a dairy region that embraces new technology (NL); and lastly, they were chosen for the contrast in their potential to produce heat stress. For a year, the mean monthly temperature, humidity, and rainfall figures obtained from the nearest weather station to each study area during the period from 2002 to 2016 were: 27.8°C, 78.6%, and 129.6 mm, respectively, in SL; 18.2°C, 85.6%, and 156.8 mm in SH; 24.1°C, 83.0%, and 126.4 mm in NL; and 21.6°C, 80.0%, and 119.4 mm in NH (General Statistics Office of Vietnam, 2017b). A temperature-humidity index, calculated from the temperature-humidity data and using the equation of Yousef (1985), showed that SL was the hottest region with THI = 78.3, followed by NL with THI = 73.5. The northern highland was a much cooler region than the lowland regions with THI = 70.0, and SH was the coolest region with THI = 65.6. There are usually four seasons in Vietnam, which include spring (January to March), summer (April to June), autumn (July to September), and winter (October to December). The current study was conducted in the autumn months, which were neither the hottest nor coolest months of the year.

The eight SDFs per region were randomly selected from 40 per region that had previously been included in a survey of SDF economics conducted in the same year as the current study by a collaborating scientist (Nga, 2017a; b). Briefly, in the economics survey, 160 SDFs with 40 per region were randomly selected from the lists of SDFs in each region that the local District Agriculture Departments supplied. Then, 8 SDFs per region that agreed to continue involvement in the current study were randomly selected from the SDFs in the economics survey. Contact was initially made by phone to inform the nominated farmer of the purpose of the study, assure them of the voluntary and confidential nature of it, and make an appointment for the visits, which occurred during one afternoon and the following morning. The visits were designed to coincide with milking and cow feed preparation and delivery times. All farmers agreed to be involved except for one in NH, so a new SDF there was randomly selected as a replacement.

72

Figure 3.1 Topographic map of Vietnam mainland and study sites Map adapted from https://www.google.com/earth/.

3.2.2 Selection of productivity and welfare indicators

It is essential to select indicators of production and welfare that are most simple, obtainable, and applicable (Rushen et al., 2008; Moran and Doyle, 2015b). The indicators selected for this study included: herd structure, cow breed, milk yield and quality, milk electrical resistance, reasons for culling cows, body weight, body condition score, number of inseminations per conception, and panting score. Herd structure, especially the percentage of lactating cows and replacement heifers in the herd, milk yield, and milk quality are the key performance indicators of a dairy farm (Moran, 2009; Gonzalez-Mejia et al., 2018). The breed of the cow defined the potential milk productivity and heat stress tolerance of the cows. Milk electrical resistance is an assessment of udder health, and mastitis is an indicator of poor welfare (Oltenacu and Broom, 2010). A high rate of involuntary culling is an indicator of poor welfare condition on farms (De Vries, 2020). Bodyweight and body condition score reflects the nutritional state of the animals, and low body weight or low body condition score indirectly indicates undernutrition (Agenäs et al., 2006). The number of inseminations per conception reflects the fertility state of the cows, and failure to conceive is also considered

73 reduced cow production and welfare (Oltenacu and Broom, 2010). The panting score is an indicator of the heat stress level of dairy cows (Gaughan et al., 2008).

3.2.3 Farm visits and data collection

Each SDF was visited in an afternoon and the next morning. All data were collected whilst the farmers applied themselves to their routine husbandry tasks. Examples of typical interiors of SDFs in each study region are presented in Figure 3.2.

At each SDF visit, the assessment team, the majority of whom visited every SDF and were trained by a visit to three practice SDFs immediately before the 32 study SDF visits, focused on collecting data by direct observation, subsampling and measurement using a standard protocol. At the beginning of each visit, to facilitate easy identification of cows, the cow’s identification (ID) number was written with a large black marker pen on an A5 size sheet of paper, and this was attached with a non-toxic, water-soluble glue to the left side of each cow.

Herd data and culling reasons

The numbers in each SDF livestock class – dairy bulls, lactating cows, dry cows, heifers, female dairy calves, male dairy calves, and beef cattle, were counted directly by the team. Male and female dairy calves were those animals under 12 months old; heifers were the female dairy cattle from 12 months to first calving; and beef cattle were the male or female beef cattle, often Zebu breeds, and raised for beef. Per lactating cow, estimated breed mix, age, lactation number, days in milk, pregnancy status, and time of artificial insemination (tAI) when the most recent pregnancy occurred were obtained by checking record books when available and asking the farmer. The most recent pregnancy was either the pregnancy leading to the current lactation, or pregnancy in the current lactation. Common reasons to cull cows and the targets farmers wished to obtain for milk yield, fat, protein, dry matter, solid non-fat were also requested from the farmers. Due to the insufficient record held by farmers, we were not able to estimate culling rates in the studied herds. It was only possible to ask farmers to estimate the percentages of cows being culled due to different reasons.

74

a) South lowland b) South highland

c) North lowland d) North highland

Figure 3.2 Typical interiors of smallholder dairy farms in each study region

Heart girth and body weight

The heart girth (HG, kg) of each cow was measured by using a commercially available tape measure with bodyweight relative to girth measurement indications on the actual tape (Asia Technology Service Company) and draping the tape around the girth closest to the heart (Figure 3.3 a). Cow body weight was estimated by reading body weight directly off the tape measure (BWt, kg) and by using a recently published algorithm for cattle (BWc, kg) (Goopy et al. 2017):

BWc0.3595 = 0.02451 + 0.04894  HG

Because BWt and BWc measurements were very close to each other, only the results of BWc are reported in this study (BW).

75

a) Measuring heart girth and body weight b) Scoring body condition

Figure 3.3 Measurements of cow heart girth and body weight; and body condition score

Body condition score

Body condition score (BCS) (Figure 3.3 b) was determined independently by two team members and averaged, using a 5-point system (1 for emaciated, 5 for obese with increments of 0.25; Edmonson et al. 1989). The team members were trained in how to score body condition using the Penn State Extension video (https://extension.psu.edu/learn-to-score-body-condition) (Heinrichs et al., 2017). Body condition score was classified as high when BCS ≥ 3.0; medium when BCS = 2.5 to 3.0; and low when BCS ≤ 2.50 (Polsky et al., 2017).

Milk yield and milk electrical resistance

As regular practices, on all SDFs, the cows were usually milked twice a day: in the early morning (0400 h to 0630 h) and the afternoon (1500 h to 1700 h), using either individual cow bucket milking machines or a vacuum pipeline system that allowed the milking of several cows simultaneously. Then all milk was bulked into 50 L milk cans and transported, often by motorbike, to the nearest milk collection centre of the relevant milk processing company. Hand milking was only applied to cows with colostrum or mastitis, and these cows were excluded from the lactating herd data.

In the current study, to measure an afternoon and the next morning milk yield of each cow, the individual bucket was emptied between each cow milked, rather than following the normal practice of letting the bucket fill from more than one cow. For those SDFs with a vacuum pipeline system, only one cow was attached per pipeline, and the milk was directed to a portable 20 L container rather than to the normal bulk vat. Afternoon and morning milk yield were weighed immediately after each cow was milked by pouring the milk into a separate bucket and weighing it using a digital hanging

76 scale that weighed to the nearest gram (Model OCS M 100, Vietnam Japan Digital Scale Company). An approximately 40 mL subsample was then taken for further analysis (detailed later). The afternoon and next morning yields were summed to give each cow’s milk yield per day (MILK, kg/cow/d) (Error! Reference source not found.).

Milk electrical resistance, an indicator of mastitis, was measured on each 40 mL sample of milk immediately it was taken, using a hand-held Draminski Mastitis Detector (DRAMINSKI U1. Owocowa 17 10-860 Olsztyn Poland) according to the manufacturer’s instructions (Draminski, 2017). The mRE was calculated as the reciprocal of milk electrical conductivity. Milk samples were classified as healthy udder, mRE > 300 units; sub-clinically infected udder, 300 to 250 units; clinical mastitis, < 250 units (Draminski, 2017).

Milk sampling and analysis

After each morning and afternoon milking of each cow, a milk sample was collected in 40 mL sterile test tubes (Error! Reference source not found.) and, immediately following the mRE test, a milk preservative tablet (18 mg tablet Broad Spectrum Microtabs II containing 8 mg Bronopol and 0.30 mg Natamycin (D & F control systems Inc) was added and mixed well into it. The sample was frozen within an hour of collection and kept at -18oC for later analysis for fat (mFA, %) by gravimetric method (TCVN 5504, 2010); for protein (mPR, %) by Kjeldahl method (TCVN 8099, 2015); and for dry matter (mDM, %) by the standard drying method (TCVN 8082, 2013) at the Nutrition Laboratory, Faculty of Food Science, Vietnam National University of Agriculture. Fat yield (yFA, kg/cow/d), protein yield (yPR, kg/cow/d), and milk dry matter yield (yDM, kg/cow/d) were calculated by multiplying MILK by mFA, mPR, and mDM, respectively.

Energy-corrected milk (ECM, 3138 KJ per kg ECM) was calculated using the equations of Tyrrell and Reid (1965):

MILK (kg/cow/d)  [376∗mFA (%) + 209  mPR (%) + 948] ECM (kg/cow/d) = 3138

Yields adjusted for BW (kg/100 kgBW/d) were calculated using the following equation, where yield was either ECM, yFA, yPR, or yDM:

Yield  100 Yield adjusted for BW (kg/100kg BW/d) = BW

77

a) Transferring milk to a bucket after milking b) Weighing milk

c) Sampling milk d) Adding preservative to milk

Figure 3.4 Measurement of daily milk yield and milk sampling

Panting score

Panting score (PS) was assessed twice a day, between 0500 h and 0600 h in the morning and between 1400 h to 1500 h in the afternoon, on a scale from 0 to 4.5 (0 indicates a normal cow breath, not panting; 4.5 indicates excessive panting with fast breath from the flank, tongue fully extended, excessive drooling, neck extended, and head held down (Gaughan et al., 2009)). Heat stress was classified as high when PS > 1.2, moderate when 0.8 to 1.2, slight when 0.4 to 0.8, and normal when 0 to 0.4 (Gaughan et al., 2008).

Locomotion score

For the SDFs where the cows could walk a distance long enough to assess the locomotion score (LS), the LS of the lactating and dry cows was assessed using a scale (0 to 3) from DairyNZ 78

(https://www.dairynz.co.nz/animal/cow-health/lameness/lameness-scoring/) (Chesterton et al., 2017). According to DairyNZ, a cow was classified as not lame or LS = 0 when she walked evenly and confidently; slightly lame or LS = 1, when she stood with straight backline but walked unevenly with possibly arched backline; lame or LS = 2, when she stood with arched backline, walked slowly and unevenly with often arched backline, and head bobbed up and down when walking; and very lame or LS = 3 when the cow stood with arched backline, walked very slowly and very unevenly with very arched backline; and head moved up and down widely when walking (Chesterton et al., 2017).

3.2.4 Statistical analyses

All statistics were performed using the base and additional packages of R software (R Core Team, 2018). Data were imported from Microsoft Excel 2016 (Microsoft, 2016) into R using the ‘readxl’ package (Wickham et al., 2016). The SDFs were the experimental units in all analyses.

Descriptive statistics for quantitative variables were calculated for each region using the ‘psych’ R package (Revelle, 2019). The results are presented as means for normally distributed quantitative variables, medians for not-normally distributed quantitative variables, and frequencies (percentages) for categorical variables.

The choice of suitable tests for comparisons of variables between regions was based on the guidelines of McDonald (2014). Before any statistical comparison, the normality of quantitative variables was tested using both the Shapiro-Wilk test and histogram. Variables that were found to be not-normally distributed were compared by Kruskal-Wallis tests followed by Dunn post-hoc tests (P < 0.05) using the R package ‘FSA’ (Ogle et al., 2019). Normally distributed variables were compared by One-way ANOVA tests followed by Tukey–Kramer tests (P < 0.05), using the R package ‘agricolae’ (Mendiburu, 2019).

To test for associations between categorical variables, two-way contingency tables between BCS categories and regions, and between PS categories and regions were generated using the ‘CrossTable’ function of the R package ‘gmodels’ (Warnes et al., 2018) and a chi-square independence test was applied. Mosaic plots generated by the R package ‘vcd’ (Meyer et al., 2017) were used to visualise the contingency tables, Pearson residuals, and independent test models that illustrate associations between categorical variables. The idea of Mosaic plots is that they recursively sub-divide a unit square into rectangular “tiles” for the cells of the contingency tables, such that the area of each tile is proportional to the cell frequency (Meyer et al., 2017). The tiles are shaded in various ways to reflect the Pearson residuals (resulting from deviations of observed frequencies from expected frequencies under a given log-linear model (Meyer et al., 2006). The ‘shading_max’ function of R package ‘vcd’

79

(Meyer et al., 2017) was used to shade the cells in the mosaic plots. The ‘Shading_max’ function applied the maximum statistics of the absolute Pearson residuals both to test the independence of the row and column variables in the contingency tables and to visualize significant cells which caused the rejection of the independence hypothesis (Meyer et al., 2006; Zeileis et al., 2007). By default, shading_max computed two cut-off points corresponding to confidence levels of 90% and 99% to shade the significant cells with less saturated colours (blue or red) and saturated colours, respectively (Meyer et al., 2017). Positive (blue) shades indicate the observed frequency is significantly greater than the expected frequency under the independence model, which is the row variable is independent of the column variable. The negative (red) shade indicates the observed frequency is significantly less than the expected frequency.

3.3 Results 3.3.1 Herd characteristics

Mean herd size was largest in the NH herds (45, ranging from 34 - 55) cattle, followed by the NL (27, 16 - 46 cattle) and SL herds (27, 11 - 44 cattle) and smallest in NH herds (17, 9 - 30 cattle), (P < 0.001) (Figure 3.5). Across regions, herd size averaged 29 cattle, consisting of 11 lactating cows, four dry cows, six heifers, seven female calves, one male calf, one beef cattle. There were no working bulls in any of the herds surveyed.

Herd structure was similar across regions. The % lactating cows averaged 37.3%, dry cows 13.7%, heifers 21.4%, female calves 21.8%, male calves 3.7%, and beef cattle 1.8%.

Figure 3.5 Herd structure across the four contrasting regions

80

The only region that maintained a collective record of cow genotype was the NH, and that was also the region with the greatest percentage (farmer reported) of pure Holsteins in the lactating herd (98% of cows, P < 0.001) (Table 3.1). Whereas, for all other regions, the farmers reported that less than 21% of their cows were pure Holsteins. Per lactating cow, the average number of lactations (2.2 lactations) and days in milk (181 d) were similar across regions. Cows in all regions were artificially inseminated with semen straws of dairy breeds, mainly Holstein. Whilst the mean number of artificial inseminations per conception was 2.1 across regions (P < 0.10), herds in the SL tended to require the most (3.2), whilst herds in other regions ranged in median value from 1.6 in NH to 1.9 in NL.

The mean BW ( SEM) of cows across regions was 498  18 kg. It was the lowest for the SL cows (450 kg), greatest for NL and NH cows and intermediate for the SH cows (P = 0.001). Mean BCS ( SE) of cows across regions was 2.8  0.1. It was greatest for NL cows (3.0), whereas all other regions were similarly lower (2.7 – 2.8, P = 0.007). Across regions, the overall percentages of cows with high, medium, and low BCS (as defined in Figure 3.6) were 34, 39 and 27%, respectively. Different BCS categories did not distribute evenly across regions (P = 0.016). A greater frequency of high BCS cows (coloured blue in Figure 3.6) and a lower frequency of low BCS cows (coloured red) were observed in the NL region than in other regions.

81

Table 3.1 Comparisons of lactating herd characteristics, locomotion scores, and reasons for culling cows across four dairy regions

B Region (n = cows), Mean or Median Overall, Parameter A P C SL SH NL NH Mean (SEM) Lactating herd n = 75 n = 44 n = 84 n = 142 Pure Holstein cows*, % 0b 21ab 8ab 100a 0.002 32 (23) Days in milk, days 176 172 200 177 0.767 181 (7) Lactation number 2.0 2.5 1.9 2.5 0.039 2.2 (0.2) Heart girth, cm 179b 185ab 187a 190a 0.002 186 (2.4) Body weight, kg 450b 496ab 513a 535a 0.001 498 (18) Body condition score 2.7b 2.7b 3.0a 2.8b 0.007 2.8 (0.10) Inseminations per conception* 3.2 1.7 1.9 1.6 0.061 2.1 (0.4) Lactating and dry herd n = 98 n = 64 n = 113 n = 193 Locomotion score (LS)D, % Not lame (LS = 0)* 79.00 71.00 - 53.00 0.078 68 (8) Slightly lame (LS = 1)* 14.00 22.00 - 15.00 0.780 17 (3) Lame (LS = 2)* 7.00 8.00 - 12.00 0.179 9 (2) Very lame (LS = 3)* 0.00 0.00 - 11.00 0.003 4 (4) Culling reason, % Lameness* 45 13 0 24 0.095 21 (10) Infertility* 20 21 17 0 0.432 15 (5) Old* 0 8 0 31 0.046 10 (7) Mastitis* 8 0 0 0 0.81 2 (2) Other* 8 13 0 0 0.752 5 (3) A Variables with * mark were not normally distributed; thus medians were presented. For other variables, means were presented. B Region: SL, South lowland (Ho Chi Minh city); SH, South highland (Lam Dong province); NL, North lowland (Ha Nam province); NH, North highland (Son La province). C P-values were given for either One-way ANOVA tests comparing means (superscript letters were given for post-hoc Tukey–Kramer test, P < 0.05) or Kruskal-Wallis tests comparing medians (superscript letters were given for post-hoc Wilcoxon rank sum test; P < 0.05). D Only locomotion score of lactating and dry cows in 2 SL SDFs, 3 SH SDFs, and 7 NH SDFs were obtained and used to calculate medians of locomotion score categories for those regions. a-b Means or medians with the different superscript letters within a row differ significantly from each other, P < 0.05.

82

Figure 3.6 Mosaic plots showing associations between regions and body condition score categories

Cow body condition (BCS) score was classified as low (L) when BCS ≤ 2.50, moderate (M) when BCS = 2.5 to 3 and high when BCS ≥ 3.00 (Polsky et al., 2017). Mosaic plots with residual-based shading based on the maximum statistic were constructed using R package ‘vcd’ (Meyer et al., 2017). In each mosaic plot, the size of each rectangle was proportional to its observed frequency (number displayed), and the rectangles were shaded according to Pearson residuals from the model of independence under the maximum statistic test. Two cut-off points were computed corresponding to confidence levels of 90% and 99%, respectively, to shade the rectangles. A positive (blue) shade indicates the observed frequency is significantly greater than the expected frequency under the null independence model, and a negative (red) shade indicates the observed frequency is significantly less, as shown in the legend. P-values presented in the legend were from the χ2 test of independence model based on the permutation distribution.

We were only able to score the locomotion score of cows in 2 SL and 3 SH where farmers agreed to let the cows walk outside the cowsheds, and in 7 NH SDFs where the cows walked a distance to the milking area. Based on the obtained data, the median percentage of lame and very lame cows (LS = 2 and 3) in NH, SH, and SL was 23%, 8%, and 7%, respectively. The median percentage of cows being culled for the range of reasons tested was similar across regions except for culling based on age (P < 0.05). The most popular reason for culling was lameness (21% of culled cows), followed by infertility (15%), age (10%), and mastitis (2%). Lameness was the major reason for culling in NL cows (45%), whilst infertility was a major reason in SH cows (21%) and NL cows (17%). Culling on age was the major reason in NH cows (31%).

83

3.3.2 Milk production

Mean ( SEM) MILK and FCM of cows across regions was 16.9  1.5 and 15.7  1.3 kg/cow/d, respectively (Table 3.2). The actual MILK amount was markedly lower than that targeted by farmers (23.6  2.1 kg/cow/d).

Table 3.2 Comparisons of actual milk yields (kg/cow/d), milk yields adjusted for cow body weight (kg/100kg BW/d), and farmers’ milk yield target (kg/cow/d) across four dairy regions

B Region (n = cows), Mean or Median Overall Parameter A SL SH NL NH P C Mean (SEM) (n = 75) (n = 44) (n = 84) (n = 142) Actual milk yields MILK 13.7b 16.4b 16.6b 21.0a < 0.001 16.9 (1.5) ECM 13.1b 15.1b 15.6b 19.2a < 0.001 15.7 (1.3) yFA 0.51b 0.56b 0.61ab 0.70a 0.002 0.60 (0.04) yPR 0.43c 0.50bc 0.56b 0.67a < 0.001 0.50 (0.05) yDM 1.75b 1.99b 2.05ab 2.42a 0.001 2.1 (0.14) yFA + yPR 0.94c 1.07bc 1.18ab 1.37a < 0.001 1.1 (0.09) Yields adjusted for BW ECM 2.95b 3.06ab 3.22ab 3.65a 0.024 3.3 (0.20) yFA 0.12 0.11 0.12 0.13 0.111 0.1 (0.00) yPR 0.10b 0.10b 0.11ab 0.13a 0.002 0.1 (0.01) yDM 0.40 0.41 0.42 0.46 0.190 0.4 (0.01) yFA + yPR 0.21b 0.22b 0.24ab 0.26a 0.012 0.2 (0.01) Farmers’ milk target MILK* 20b 25ab 20.5ab 29a 0.003 23.6 (2.1) A Abbreviations of yields per cow: MILK, raw milk yield; ECM, energy corrected milk; yFA, fat yield; yPR, protein yield; yDM, dry matter yield. Variables with * mark were not normally distributed; thus, medians were presented. For other variables, means were presented. B, C, a, b, c Other footnotes as in Table 3.1.

Across regions, when not adjusted for BW, all milk yield parameters in NH were significantly higher than those in the other regions (P  0.02, Table 3.2) and all milk yield parameters in SL tended to be lowest. All milk yield parameters in NL were similar to those in SH. Milk protein yield and yPR plus yFA were higher in the NL than SL region (P < 0.05).

84

When adjusted for BW, except for yFA and yDM (P > 0.111), milk production and quality parameters were greatest for NH cows and lowest for SL cows (P < 0.05). The yFA and yDM were similar across regions.

Milk component concentrations averaged across morning and afternoon and regions were: mFA, 3.66%; mPR, 3.27%; and mDM, 12.32% (Table 3.3). Cows in highland regions tended to produce milk with lower mFA, mPR, and mDM than the cows in the lowland regions (P < 0.05). NL cows had the greatest mFA and mPR, SL cows had the greatest mDM, NH cows had the lowest mFA and mDM, and SH cows had the lowest mPR (P < 0.05).

Table 3.3 Comparisons of farmers’ targets for milk concentrations (%), measured milk concentrations (%), and milk electrical resistance across four dairy regions

B Region (n = cows), Mean or Median Overall, Parameter A, % SL SH NL NH P C Mean (SEM) (n = 75) (n = 44) (n = 84) (n = 142) Actual milk concentrations mFA 3.92a 3.46ab 3.89a 3.38b 0.005 3.66 (0.14) Day: mPR 3.18b 3.10b 3.54a 3.24ab 0.004 3.27 (0.10) Day: mDM 12.96a 12.15bc 12.53ab 11.64c < 0.001 12.32 (0.28) Day: mRE 382c 431a 400bc 411ab < 0.001 406 (10) Farmers’ targetsD MornmFA* 4.0 3.8 3.7 3.7 0.010 3.8 (0.1) MornmPR* 3.2 - 3.3 - - - MornmDM* 12.1 - - 11.8 - - MornSolid non-fat* 8.7 8.8 8.7 - 0.305 8.5 (0.2) A Abbreviations: Mor, morning; Aft, afternoon; Day, an average of morning and afternoon values; mFA, milk fat concentration (%); mPR, milk protein concentration (%); mDM, milk dry matter concentration (%). Variables with * mark were not normally distributed; thus, medians were presented. For other variables, means were presented. B, C, a, b, c Other footnotes as in Table 3.1. D Only one farm in SL and one farm in NL reported milk protein targets, only one farm in SL reported milk dry matter target, and only one farm in NH reported milk solid non-fat target.

Milk electrical resistance, a measure of udder health, averaged across morning and afternoon and regions was 406 units. SH cows had the greatest mRE and SL the lowest (P < 0.01). Only one cow in the SL region had an mRE of less than 300 units (classified as subclinical mastitis), and one cow in the NL region, which was excluded from the study due to mastitis, had an mRE of less than 250 units.

85

All farmers across all regions nominated mFA targets, but only one farmer in NL and one farmer in NL nominated mPR targets. Farmer-nominated mFA targets differed across regions (P = 0.010). Targets for milk solid non-fat (mDM minus mFA) were nominated by all farmers in SL, SH, and NL but were not by farmers in NH. Across regions, the mean actual daily mFA (3.66%) was consistently lower than the mean target nominated (3.8%).

3.3.3 The heat stress level of cows

Means of mPS, aPS, and PS for SL cows, the hottest hot region, were highest, whilst those for SH, the coolest region, were lowest (P  0.007, Table 3.4). However, the means of mPS, aPS, and PS for cows in NH were similar to those in NL, despite NH being a cooler region.

Table 3.4 Comparisons of panting score of cows across four main dairy regions

B Region (n = cows), Mean Overall Panting score SL SH NL NH P C Mean (SEM) (n = 75) (n = 44) (n = 84) (n = 142) Morning (mPS) 1.4a 0.3c 0.9b 0.6bc < 0.001 0.8 (0.2) Afternoon (aPS) 2.2a 1.3b 1.9ab 2.0a 0.007 1.8 (0.2) Average day (PS) 1.8a 0.8c 1.4b 1.3b < 0.001 1.3 (0.2) B, C, a, b, c Abbreviations and footnotes as in Table 3.1.

In the morning, across regions, 12% of cows were highly heat-stressed, 39% were moderately heat- stressed, 35% were slightly heat-stressed, and 14% were un-stressed (Figure 3.7 a). All the cows in the SL region were slightly to highly heat-stressed. Only the SH region had no highly heat-stressed cows. In the SL, 100% of the cows were categorised as having slight to high heat-stress, 98% in NL, 80% in NH, and 61% in SH. Lowland regions (NL and SL) had more (P < 0.01, blue rectangles) moderately (NL) to highly heat-stressed (SL) cows, but less (P < 0.01, red rectangles) normally (both NL and SL) and slightly heat-stressed (SL) cows.

86

a) Morning panting score b) Afternoon panting score

c) Average daily panting score

Figure 3.7 Mosaic plots showing associations between regions and panting score categories with (a) morning panting score (mPS), (b) afternoon panting score (aPS), and (c) average day panting score (PS)

Based on panting score (PS), a cow was classified as normal (N) when PS = 0.0 to 0.4, slightly heat- stressed (S) when PS = 0.4 to 0.8, moderately heat-stressed (M) when PS = 0.8 to 1.2, and highly heat-stressed (H) when PS > 1.2 (Gaughan et al., 2008).

In the afternoon, across regions, 77% of cows were highly heat-stressed, 19% were moderately heat- stressed, 4% were slightly heat-stressed, and none were un-stressed (Figure 3.7 b). The SH region had a lower frequency of highly heat-stressed cows (41%, red rectangle, P < 0.01) and tended to have a higher frequency of slightly heat-stressed cows (34%, blue rectangle) (P < 0.10), whereas in all other regions, almost all the cows were moderately to highly heat-stressed in the afternoon. The frequency of highly heat-stressed cows was greatest in SL with 93%, then NH with 82% and NL with 73%, and lowest in SH.

During a day, across regions, 63% of cows were highly heat-stressed, 18% were moderately heat- stressed, 18% were slightly heat-stressed, and 1% were un-stressed (Figure 3.7 c). The SH region had 87 a lower frequency of highly heat-stressed cows (16%, blue rectangles) and higher frequencies of slightly heat-stressed (59%, P < 0.01) and un-stressed cows (9%, P < 0.01). The frequency of highly heat-stressed cows was greatest in SL with 89%, then NL with 66% and NH with 61%, and lowest in SH.

3.4 Discussion

As hypothesized, most cows in both the lowlands and highlands were moderately to highly heat- stressed during the day, and cows in the highlands were less heat-stressed than cows in the lowlands. However, despite expectations, a similar ranking of the region was not seen for milk yield. The NH milk yields were the greatest, but the SH milk yields were similar to the NL and SL regions. In addition to heat stress, the main productivity and welfare concerns were low milk yield and quality, low BW and low BCS of the SDF cows.

3.4.1 Herd structure and cow productivity

The mean herd size of Vietnamese SDFs in the current study (29 cattle) was greater than that determined in similar studies on SDFs in other Asian and African countries. In Thai SDFs, the mean herd size was 20 cows (Herawati et al., 2016). In Indonesian SDFs, 90% of SDFs had only three or fewer lactating cows; among Indian SDFs, 97% had only two dairy cows per farms (Moran and Chamberlain, 2017). In Zimbabwe SDFs, the mean herd size was ten cows (Paraffin et al., 2018).

The majority of the surveyed SDFs specialised in the production of dairy cattle for milk production. In such SDFs, the percentage of lactating cows and heifers in the whole herds were the key performance indicators of dairy farms since lactating cows generate income and replacement heifers define the sustainability of the dairy herds. Moran (2010) suggested that in tropical SDFs, the percentage of lactating cows should range from 40 – 48% and replacement heifers, 20 – 25%. Based on this guideline, the only region with an acceptable percentage of lactating cows was NL (41.1%) and possibly NH (39.1%), while SH (35.8%) and SL (34.8%) did not have acceptable percentages. The mean percentage of heifers in SL (22.5%), NL (29.9%), and NH (22.3%) were all within the desirable range of 20 – 25% suggested by Moran (2010) to ensure enough animals for replacements. Only the percentage of heifers in SH (18.8%) was lower than that suggested range.

The current study indicates the mean ECM of Vietnamese SDF cows across regions is 15.7 kg ECM/cow/d. This is greater than the milk yield reported by previous authors based on verbal advice from the farmer rather than direct measurement, which were 14 kg/cow/d for cows in some provinces in the south (Vu et al., 2016) and 15 kg/cow/d for cows in a province in the north (Ashbaugh, 2010). The mean ECM of Vietnamese SDF cows was also greater than the average yield of SDF cows in

88 some other tropical countries such as Thailand, whose multi-breed dairy cows yielded 12.4 to 14.1 kg/cow/d (Koonawootrittriron et al., 2009; Wongpom et al., 2017); Indonesia, whose Holstein cows yielded 14.4 kg/cow/d (Rahayu et al., 2018); and Ethiopia, whose Holstein cows yielded 11.49 kg/cow/d (Ayalew et al., 2017). However, the ECM of Vietnamese SDF cows was much lower than the milk yield of Holstein cows raised in commercial dairy farms also in Vietnam, such as on the farms of TH True milk company, with 20 kg/cow/d (Duteurtre et al., 2015), or commercial dairy farms in other developed countries such as Korea (28.6 kg/cow/d) (Shin et al., 2017); England (24.66 kg/cow/d) (Eaglen et al., 2013); or the USA (32.50 kg/cow/d) (Dikmen et al., 2012). Similarly to ECM, the average yFA (0.60 kg/cow/d) of Vietnamese SDF cows was greater than yFA of Thai multi- breed dairy cows (0.43 to 0.516 kg/cow/d) (Koonawootrittriron et al., 2009; Wongpom et al., 2017); but both yFA and yPR (0.50 kg/cow/d) of Vietnamese SDF cows were lower than the ranges of yFA (0.91 to 1.01 kg/cow/d) and yPR (0.79 to 0.90 kg/cow/d) of Holstein cows in commercial farms in Brazil, Korea, and England (Eaglen et al., 2013; Campos et al., 2015; Shin et al., 2017).

There may be factors other than tropical weather conditions that limit the productivity of Vietnamese SDF cows. The much lower milk yields in the SH herds than NH herds (both with relatively cooler weather) indicates that SH SDFs might not have had as good quality genetics, diet, or housing conditions as the NH SDFs had. Similarly, the similar milk yields from a hot region like NL and a cold region like SH along with the greater yPR plus yFA in NL compared to a similar hot region like SL, suggest that the disadvantage of lowland regions in terms of their hot and humid weather conditions could, perhaps, be overcome. Together, these results suggest that the weather conditions and other dairy farming factors such as cow genotypes, nutrition, and housing factors might also limit cow productivity. Thus, further studies are needed to evaluate this. The low productivity of cows in SL might be a reason why SL SDFs keep male calves and beef cattle for an extra income source.

The mean mFA of Vietnamese SDF cows (3.66%) was greater than that of dairy cows in some other tropical countries. It was greater than the Thai multi-breed dairy cows (3.19 to 3.59%) (Koonawootrittriron et al., 2009; Wongpom et al., 2017) and the Brazilian Holstein cows (3.45%) (Petrini et al., 2016), but much lower than that of Holstein cows in temperate countries such as Germany, whose Holstein cows ranged from 3.95 to 4.03% (Gieseke et al., 2018) or Holland, whose Dutch Holstein cows ranged had a mean mFA of 4.36% (Stoop et al., 2008). The mean mPR of Vietnamese SDF cows (3.27%) was also greater than that of Thai multi-breed dairy cows (3.01%) (Yaemkong et al., 2010) and Brazilian Holstein cows (3.05%) (Petrini et al., 2016), but lower than that of German Holstein cows (3.34 to 3.38%) (Gieseke et al., 2018) and Dutch Holstein cows (3.51%) (Stoop et al., 2008). Low mFA and mPR are indicators of unbalanced diets (Moran, 2012).

89

Overall, milk component concentrations of Vietnamese SDF cows seem to be comparable with that of cows in other tropical countries but still much lower than that of cows in temperate countries. Commercial dairy farms in developed countries did not commonly measure the parameter mDM, but in Vietnam and other South East Asia countries such as Thailand (Yaemkong et al., 2010), mDM and milk solid non-fat were commonly used by some milk collecting companies to define the milk price.

3.4.2 Cow welfare

Heat stress was indicated to be a serious welfare issue in Vietnamese SDF. This was especially the case in the afternoon compared to the morning, when 77% of cows across all regions were categorised as suffering from high heat stress. The high levels of heat-stressed cows in the lowlands in our studies were consistent with the study by Lam et al. (2010) which reported that SDF cows in the SL were highly heat-stressed during summer with rectal temperatures of 39.3°C, versus an unstressed target of 38.6°C in Friesian cows (Andersson and Jóhasson, 1993; Ma and Du, 2010). We are not aware of any previous study assessing the level of heat stress of SDF cows in the highland regions of Vietnam. As predicted from the historical weather data (General Statistics Office of Vietnam, 2017b), cows in all regions, including the coolest SH region, were slightly heat-stressed to highly heat-stressed from early morning to late afternoon. However, as expected, cows in highland regions suffered less heat stress than cows in lowland regions.

It is well documented that heat stress reduces feed intake and milk production in lactating cows (Kadzere et al., 2002; West, 2003; Das et al., 2016). Thus the high level of heat stress could be a reason for the low milk productivity of Vietnamese SDF cows. This suggests that the productivity and welfare of not only cows in lowlands but also cows in highlands might be improved if their heat stress is decreased. Also, interestingly, although the historical data showed that the weather conditions between the lowlands were quite similar, mPS and PS of cows in NL were significantly lower than that of cows in the SL. This indicated that NL SDFs might have better cowshed designs or apply more efficient heat stress abatement strategies than SL SDFs; thus, the heat stress of the cows decreased.

Low fertility was also identified as a productivity issue of Vietnamese SDF cows. The medians of number of AI per conception among Vietnamese SDF cows across regions, ranging from 1.6 in NH to 3.2 in SL, are quite high when compared to a median of 1.0 (a mean of 1.9) in Australian dairy herds, as reported by Talukder et al. (2015) and a mean of 1.5 (range from 1.4 to 1.9) in Danish dairy herds reported by Lehmann et al. (2017). Apart from the technical factors such as heat detection and AI skills of inseminators, the heat stress could be a reason for high tAI per conception in Vietnamese SDF cows, especially in SL regions, since it is well-documented that heat stress can significantly impair fertility and reproduction in dairy cattle (Jordan, 2003; Avendaño-Reyes et al., 2010; 90

Wolfenson and Roth, 2019). Avendaño-Reyes et al. (2010) reported that during hot months both primiparous and multiparous cows showed significantly greater tAI per conception and that primiparous cows showed significantly greater numbers of days open than in the cool months.

The average BW of Vietnamese SDF cows across four regions (498 kg) was much lower than the BW of Holstein cows in other countries such as Ireland (Berry et al., 2007), Brazil (Poncheki et al., 2015) or Israel (Van Straten et al., 2008). For example, the average BW of Holstein cows raised in humid subtropical farms in Brazil ranged from 522 kg at nadir (1st parity) to 670 kg at calving (> 3rd+ parity) (Poncheki et al., 2015). Assuming that the cow breeds reported by the farmer in the current study are true, the high percentage of crossbred cows as reported by farmers in SL, SH and NL SDFs (Table 3.1) is a likely reason for the low average BW of Vietnamese SDF cows. This is because the local Vietnamese Lai Sind cattle that SDF farmers used to cross with Holsteins has very small BW (249 - 281 kg) (Duy et al., 2013). Also, Holsteins usually have greater mature BW (590 - 680 kg) compared to other dairy cattle breeds such as Brown Swiss (509 – 537 kg) and Jersey (408 to 454 kg) (Capper and Cady, 2012; Piccand et al., 2013). Also, the low BW of Vietnamese SDF cows might reflect the poor nutrition status of the cows. However, this hypothesis can only be confirmed if the exact breed of the cows and the common diets for Vietnamese SDF cows, which are often not recorded systematically by SDF farmers, are known.

Similar to BW, the mean BCS of lactating cows across regions (2.8) in the current study was much lower than an average BCS of 3.18 (ranging from 2.85 to 3.54) (Vallimont et al., 2010) or an average BCS of 3.04 (ranging from 2.88 to 3.17) (Barberg et al., 2007) in US Holstein cows. Although it is widely accepted that BCS is a valuable indicator of cow welfare, guidelines to assess cow welfare based on BCS has remained limited (Ferguson and Matthews, 2011; Moran and Doyle, 2015b). A recent study suggested that a BCS of less than 2.5 may indicate that the cow could be experiencing hunger (Ferguson and Matthews, 2011). In the current study, 34% of the cows across all regions had BCS less than 2.5. According to Moran and Doyle (2015b), a low BCS in cows, similar to low BW, is an indicator of poor welfare and is usually due to poor feeding management. This seems to be reasonable because the available data showed that Vietnamese SDF cows are usually offered with mainly low-quality tropical forage such as Napier grass (Pennisetum purpureum) or rice straw topped up with concentrate pellets at a ratio of roughly 1 kg concentrate per 2 kg of milk yield (Chu et al., 2005; Cuong et al., 2006a; Moran, 2012; Hiep et al., 2016; Phong and Thu, 2016). However, further research is needed to confirm this.

Lameness and infertility ranked highest among the reasons for involuntary culling, and this was consistent with the high number of artificial inseminations per conception and the high percentage of

91 lame and very lame cows across regions. This suggests that improving the hoof health and reproductivity of the cows is essential to improve cow welfare in SDFs.

For udder health, according to the guidelines of the manufacturer of the Draminski mastitis detector, cows’ udders are classified into subclinical mastitis when mRE  300 units (Draminski, 2017). Based on this guideline, mastitis was not a big problem in the studied regions because the mean morning, afternoon and day mREs over each region were all higher than 300 units. This is not consistent with a study conducted in Dong Nai, a southern province of Vietnam, where it was reported that the prevalence of subclinical mastitis in cows was 88.6% (Östensson et al., 2013). However, it should be noted that the diagnosis of subclinical or clinical mastitis in the current study was purely based on mRE and the guidelines of the manufacturer of the mastitis detector, which is not a standard method. The accepted definition of mastitis is that clinical mastitis occurs when a cow produces milk with abnormal appearances and/or has swollen, red, or painful udder quarters (Pinzón-Sánchez and Ruegg, 2011). This might explain the inconsistency in the prevalence of subclinical mastitis between the current study and the previous study (Östensson et al., 2013).

3.4.3 Limitations

The current study has some limitations. Firstly, due to limited time and labour, only some selected performance and welfare indicators were assessed. Secondly, for the same reasons, only single-day data of productivity and welfare indicators were obtained. Thirdly, while the weather conditions of regions vary during a year and could affect the cows' productivity and heat stress levels, the current study was only conducted in an autumn period. These limitations should be taken into account in further studies.

3.5 Conclusions

It was as expected that cows in the highlands suffered less heat stress than cows in the lowlands. However, only cows in the NH, but not SH, were more productive than cows in the lowlands. Future studies or extension programmes should focus more on SL SDFs where the cows yielded the least milk and were in the poorest welfare conditions, and focus on SH SDFs where the weather conditions are most suitable for high yielding cows, but the cows were not most productive.

Across regions, the major productivity concerns of the Vietnamese SDFs were the low ECM, low percentage of lactating cows, relatively low mFA and mPR, and relatively high number of inseminations per conception. The major welfare concerns were the high heat stress levels, low BW, low BCS, and the culling of cows due to lameness and infertility. These concerns should be targeted in further studies and extension programmes. 92

Further studies are recommended to record the data for a more extended period, in different seasons, and across a broader range of welfare indicators to improve the results.

93

Chapter 4 Genomic diversity and breed composition of Vietnamese smallholder dairy cows

Abstract

Vietnamese smallholder dairy cows (VDC) have resulted from crossbreeding between local Yellow cattle breed, zebu breeds, and imported European dairy breeds through many undefined generations. Consequently, the predominant breed composition is currently unknown. High-performance dairy genetics, mainly from purebred Holstein, are imported regularly in the forms of semen straws, bulls, or heifers, but due to lack of pedigree records, farmers commonly estimate breed composition of individual VDC, at best, according to coat colour. The aim of this study was to use genomic data to evaluate the level of genetic diversity and breed composition of VDC. Genomic data from 345 lactating cows from 32 farms located across four typical but contrasting dairy regions of Vietnam were obtained from their tail hair DNA using a 50K SNP chip and merging it with genomic reference data of Zebu (ZEB) breeds: Red Sindhi (n=10), Sahiwal (n=17) and Brahman (n=25); taurine breeds: Holstein (HOL, n=93), Jersey (JER, n=49) and Brown Swiss (BSW, n=24); and a composite (Taurine x Indicine) breed: Chinese Yellow Cattle (CYE, three populations, n=59). The test herd was found to have a high level of similarity (Genetic proportion of 85%) to the HOL reference population and were not excessively inbred (inbreeding coefficient = -0.017 to 0.003, i.e., very close to zero). The genetic proportions of JER, BSW, and ZEB in the test herd were 6.0%, 5.3%, and 4.5%, respectively. Major genotype groupings and their corresponding proportions in the test herd were 48% pure HOL, 22% B3HOL_ZEB, 12% B2HOL_ZEB, 6% F1HOL_JER, 3% B1HOL_BSW, 2% B1HOL_JER, and 1% each of B1HOL_ZEB, F1HOL_BSW, and F1HOL_ZEB. No significant association was detected between coat colour and the genetic proportion of HOL in HOL_ZEB crosses. Since approximately 50% of tested VDC herds were pure HOL and 50% were crossbred HOL mainly with ZEB, the current VDC herd could be considered a good genetic base for the selection of either pure HOL cows or HOL genetics diluted with a certain fraction of tropically adapted ZEB.

94

4.1 Introduction

Vietnam is not a traditional dairy farming country; the current smallholder dairy herds are the result of crossbreeding European breeds with Zebu and native yellow cattle breeds through many generations (Vang et al., 2003; Trach et al., 2007). Initially, Vietnam had only Yellow cattle (Department of Livestock Production, 2009). Vietnamese Yellow cattle are believed to have originated from Chinese Yellow cattle (CYE) (Ly et al., 1999). Yellow cattle adapted well to the tropical conditions typical of Vietnam and were bred throughout the country as draught animals. However, Yellow cattle are small in stature and have low milk productivity (Duy et al., 2013). Dairy farms first appeared in Vietnam in the 1930s-40s using breeds including Jersey (JER), Ongole, Red Sindhi (RSI), Tharpara, Sahiwal (SAH) and Haryana (Trach et al., 2007). In the 1960s -70s, as domestic demand for milk developed, Vietnam simultaneously implemented two breeding programs to rapidly increase cow numbers (Vang et al., 2003; Trach et al., 2007). The first program aimed to improve the stature of Yellow cattle by encouraging farmers to cross them with Zebu breeds, including Red Sindhi, Sahiwal, and Brahman (BRM), from India and Pakistan, to create a hybrid breed called Lai Sind (Vang et al., 2003; Trach et al., 2007). The second program was designed to generate a more specialised dairy herd; farmers were encouraged to cross their Yellow and Lai Sind cattle with imported Holstein (HOL) genetics to create HOL crossbreeds, including first crosses (F1, 1/2 HOL), first backcrosses (B1, 3/4 HOL), second backcrosses (B2, 7/8 HOL), or third backcrosses (B3, 15/16 HOL) (Vang et al., 2003; Trach et al., 2007). During that period, JER and Brown Swiss (BSW) genetics were also imported to crossbreed with the Lai Sind and HOL crossbreeds (Trach et al., 2007). However, farmers were less impressed with these breeds as their coat colour was not black and white, and their stature was smaller than the much preferred purebred HOL (Trach et al., 2007). Since those initial programs, the demand for fresh milk in Vietnam has rapidly increased, and so the importation of bulls, semen straws and heifers of mainly HOL genetics and, to a much smaller extent, JER and BSW genetics has continued (Cai and Long, 2002; Trach et al., 2007). There have also been occasional imports of semen straws of other dairy breeds such as the French Brown (Brune), Montbeliarde, Bretonne Pie Noir, and the Australian Friesian Sahiwal (Cai and Long, 2002). Consequently, it can be assumed that the current VDC could include pure HOL and crossbreeds of HOL with JER, BSW, RSI, SAH, BRM, and CYE. However, due to a lack of pedigree records, the level of genetic diversity and predominant breed mixtures in VDC are unknown.

Knowledge of the level of genetic diversity is necessary to understand the evolution, level of inbreeding and degree of similarity between breeds. It is also necessary for consideration in the design of breeding programs (Weigel, 2001; Bennewitz and Meuwissen, 2005; Stachowicz et al., 2011;

95

Melka and Schenkel, 2012). Recent studies have revealed increasing rates of inbreeding in HOL, JER, and BSW populations worldwide (Weigel, 2001; Stachowicz et al., 2011).

Knowledge of breed purity and the breed compositions is also fundamentally crucial for breed societies, farmers, and scientists. For breed societies, this information is needed for breed registries (Hulsegge et al., 2019), identifying suitable selection or crossbreeding plans, and conservation (He et al., 2018a; Lwin et al., 2018). For farmers and nutritionists, this information is necessary for modifying diets and housing conditions because different breeds have different nutrient and environmental requirements (NRC, 2001). For scientists, the information about the breed is necessary to ensure quality control of experimental animals (He et al., 2018a) and to match genotype to the environment (Kuehn et al., 2011).

Conventionally, the breed purity and genetic merit of a cow is determined according to its pedigree records (Frkonja et al., 2012). This traditional method may be cheap and simple but is not applicable to VDC herds as pedigree records are rare. Currently, it is common practice for farmers to estimate cow genotype (breed composition) according to the cow’s coat colour. Black and white cows are considered pure HOL, and those with a higher percentage of black or other colours, usually brown or dark yellow, are considered crossbreeds. However, due to the recent development of genotyping technology and the availability of reference genotype databases, it has been possible to determine the genetic diversity, breed purity, and breed composition of a cow based on single nucleotide polymorphisms (SNPs, markers) instead of pedigree (Kuehn et al., 2011; Sempéré et al., 2015; Visser et al., 2016; He et al., 2018a; b; Gobena et al., 2018; Lwin et al., 2018; Hulsegge et al., 2019). These recent developments should help determine the genetic diversity and merit in Vietnamese dairy herds where pedigree records are generally unavailable.

Consequently, this study aimed to use genomic data to (1) evaluate the degree of genetic diversity of VDC, (2) identify the population structure and SNP-based estimates of breed composition of VDC, (3) compare the cow genotype inferred from genomic data with cow genotype reported by farmers, and (4) evaluate the relationship between cow genotype inferred from genomic data with observed coat colour which farmers commonly use to estimate cow genotype.

4.2 Materials and methods 4.2.1 Tail hair samples

From 24 August to 7 October 2017, tail hairs of all lactating cows (n = 345) from the 32 surveyed farms previously described in Chapter 3 were sampled, following the procedure of Neogen

96

Australasia Pty Limited. These farms are located in four main smallholder dairy regions of Vietnam, eight farms per region. The regions included a southern lowland (SL), a southern highland (SH), a northern lowland (NL), and a northern highland (NH). The regions were considered representative of the Vietnamese smallholder dairying areas in terms of their high population density, diverse weather conditions, and long dairy farming history except for the NL region, a new dairy region. Tail hair samples were collected from the tail switch of the cows by local veterinarians (approved by the University of Queensland Animal Ethics Unit, ethics approval number ANRFA/SVS/565/16/VIETNAM). The samples were preserved in dry and cool laboratory conditions in Vietnam and then exported, following Australian biosecurity regulations, to the Neogen Australasia laboratory at The University of Queensland, Gatton Campus, Australia.

4.2.2 DNA extraction and genotyping

DNA extraction and genotyping were conducted by Neogen Australasia Pty Limited (Neogen Australasia, 2018). Briefly, a DNA sample was extracted from each cow’s tail hair sample and purified using Sbeadex Livestock Kits (LGC Genomics GmbH, Germany). Procedures adapted for the kit by the Neogen Australasia Lab (SOP No. 325 - The University of Queensland) were followed. Then, the DNA samples were genotyped with a GeneSeek Genomic Profiler Bovine 50K chip (Neogen Corporation, Lincoln, NE), which assayed 48,268 SNP markers with an average SNP spacing of 59 kb. SNP chips were prepared using the producer’s assay protocol and scanned using the iScan System (Illumina Inc., San Diego, USA) to generate a genomic dataset of VDC.

4.2.3 Merged genotype data and quality control

Before statistical analysis, the genomic dataset of 345 VDC was merged with the reference genotype dataset of relevant cattle breeds (Table 4.1) using PLINK (Purcell et al., 2007). The reference genotype dataset was obtained from the WIDDE website (Web-Interfaced next generation Database dedicated to genetic Diversity Exploration): http://widde.toulouse.inra.fr/widde/ (Sempéré et al., 2015). The reference dataset comprised 277 cattle genotyped with 50,977 autosomal and unmapped SNPs, sampled from 11 populations of 7 reference cattle breeds, including Zebu (Bos indicus or indicine) breeds: Red Sindhi (coded as RSI, n=10), Sahiwal (SAH, n=17) and Brahman (BRM, n=25); taurine breeds (Bos taurus): Holstein (HOL, two populations, n=93), Jersey (JER, two populations, n=49) and Brown Swiss (BSW, n=24); and a composite (Taurine x Indicine) breed: Chinese Yellow Cattle (CYE, three populations, n=59). These breeds were included because they are assumed to be the ancestral breeds of VDC. In the reference genotype dataset, only RSI and SAH were genotyped with Illumina BovineSNP50 v2 (Decker et al., 2015) while all other breeds were genotyped with

97

Illumina BovineSNP50 v1 (Matukumalli et al., 2009; Flori et al., 2009; Gautier et al., 2010; Gao et al., 2017).

PLINK v1.9 (Purcell et al., 2007) was used for genomic data merging and quality control (Ahmad et al., 2020). Firstly, genotypes of 345 VDC were merged with genotypes of 277 reference animals using the command “--merge", resulting in a dataset of 623 animals with genotypes of 14,786 overlapping SNPs. Then, a quality control process was applied, which removed 4,079 SNPs with either call rate ≤ 95% or minor allele frequency (MAF) ≤ 5% using command “--geno 0.05” and “-- maf 0.05” (Lwin et al., 2018; Cheruiyot et al., 2018). An individual with a sample call rate of < 90% was also removed, using the command “--mind 0.1” (Lwin et al., 2018; Cheruiyot et al., 2018). After quality control, the final merged dataset comprised 621 samples typed with 10,707 common SNPs.

4.2.4 Genetic diversity

The final merged dataset was used to calculate F-statistics (Wright, 1965) for all breeds using the R package ‘hierfstat’ (Goudet and Jombart, 2015). Observed heterozygosity (HO), expected heterozygosity (HE), and inbreeding coefficients (FIS) were calculated to evaluate the levels of within‐ population genetic diversity (Nei, 1987; Weir and Goudet, 2017). Additionally, fixation index (FST) was calculated for each pair of populations to evaluate the level of population differentiation (Weir and Cockerham, 1984; Weir and Goudet, 2017).

4.2.5 Principal component and admixture analysis

Principal component analysis (PCA) and admixture analysis were applied to describe the population structure of the VDC population. PCA was performed by the PC-AiR method using the R package ‘GENESIS’ (Conomos et al., 2019). PCA results were then visualised using the R Package 'ggplot2' (Wickham et al., 2019).

The unsupervised model-based clustering methods implemented by the ADMIXTURE v1.3.0 program (Alexander et al., 2015) and the R package ‘LEA’ (Frichot and François, 2015) were used to estimate the breed composition of individual VDC using 10,707 markers. The analysis was run with different distinct assumed ancestors (K) ranging from 2 to 10, starting with the basic cross (indicine and taurine cross). Ten-fold cross-validation (CV = 10) was specified, and the error profile obtained after that was used to explore the most probable number of clusters (K), as described by Alexander et al. (2009). The results when using ADMIXTURE and LEA were very similar, so only the results derived from R package ‘LEA’ are presented. These admixture analysis outputs were visualised using R Package 'pophelper' v.2.3.0 (Francis, 2017). After identifying optimum K values, ANOVA was applied to compare breed compositions between the four VDC regions.

98

Table 4.1 Sampling information on Vietnamese dairy cattle populations and reference populations Population Breed Sampling No Breed name, Group Land of origin n Ref B name (Code) code country A Vietnam dairy populations 1 North highland (V-NH) Vietnamese dairy VDC Vietnam Various 141 1 2 North lowland (V-NL) Vietnamese dairy VDC Vietnam Various 84 1 3 South highland (V-SH) Vietnamese dairy VDC Vietnam Various 44 1 4 South lowland (V-SL) Vietnamese dairy VDC Vietnam Various 75 1 Reference populations 5 Holstein 1 (HOL1) Holstein, Taurine HOL USA, NZ Northern Europe 63 2 6 Holstein 2 (HOL2) Holstein, Taurine HOL France Northern Europe 30 3 7 Jersey 1 (JER1) Jersey, Taurine JER USA, NZ Channel Islands 28 2 8 Jersey 2 (JER2) Jersey, Taurine JER France Channel Islands 21 4 9 Brown Swiss (BSW) Brown Swiss, Taurine BSW USA Switzerland 24 2 10 Red Sindhi (RSI) Red Sindhi, Indicine RSI Pakistan Sindh, Pakistan 10 5 11 Sahiwal (SAH) Sahiwal, Indicine SAH Pakistan Punjab, Pakistan 17 5 12 Brahman (BRM) Brahman, Indicine BRM USA, AUS USA (India) 25 2 13 Dengchuan (DEC) Chinese Yellow CYE China Southwestern China 31 6 14 Honghe (HOH) Chinese Yellow CYE China Southwestern China 12 6 15 Dehong (DEH) Chinese Yellow CYE China Southwestern China 16 6 A Abbreviations of some countries name: USA, United States of America; NZ, New Zealand; AUS, Australia. B Sources of SNP genotype data: 1, This study; 2, Matukumalli et al. (2009); 3, Flori et al. (2009); 4, Gautier et al. (2010); 5, Decker et al. (2015); 6, Gao et al. (2017).

4.2.6 Classification of genotypes

Classification of a cow as F1HOL, B1HOL, B2HOL, B3HOL, or pure HOL was arrived at in two ways.

Firstly, genotype was derived from farmers using all available information at their farm, including their immediate judgment (most common), the memory of parentage (some), and their recording book (very few). Farmers commonly suggested genotypes based on cows’ coat colour (more white patches for more HOL percentage; other colours such as brown or dark yellow meaning genotypes other than HOL), based on cows’ dam and sire semen straws (if known), or based on what the cow seller told them. Basically, farmers classified their cows as F1HOL_ZEB (1/2 HOL + 1/2 ZEB), B1HOL_ZEB (3/4 HOL + 1/4 ZEB), B2HOL_ZEB (7/8 HOL+1/8 ZEB), B3HOL_ZEB and PureHOL (> 7/8 HOL + < 1/8 ZEB). Cows were classified as other genotypes (OT) if they included an additional breed other than HOL and ZEB, such as JER or BSW.

Secondly, genotype was identified based on the admixture analysis results of genomic data with K = 4. When K = 4, the admixture analysis results showed that all cows, the VDC and reference cows, were from four founder breeds (HOL, BSW, JER, and ZEB). After using the genomic data, a cow 99 was categorised as a first cross (F1HOL), first backcross (B1HOL), second backcross (B2HOL), third backcross (B3HOL), or pure HOL (PureHOL) if the genetic proportion of HOL in that cow was > 1/4 to <3/4,  3/4 to < 7/8,  7/8 to < 15/16,  15/16 to < 31/32, and  31/32, respectively. For example, a B1HOL_JER could have 75.0% to 87.5% of their genetics as HOL and the balance as JER. Consequently, VDC were categorised into nine genotypes: F1HOL_ZEB, B1HOL_ZEB, B2HOL_ZEB, B3HOL_ZEB, PureHOL, F1HOL_JER, B1HOL_JER, F1HOL_BSW, and B1HOL_BSW.

Mosaic plots were generated using the R package ‘vcd’ (Meyer et al., 2017) to visualize contingency tables, Pearson residuals, and independent test models that test the association between categorical variables (row and column variables of the contingency table). The idea of Mosaic plots is that they recursively sub-divide a unit square into rectangular “tiles” for the cells of the contingency tables, such that the area of each tile is proportional to the cell frequency (Meyer et al., 2017). The tiles are shaded in various ways to reflect the Pearson residuals (resulting from deviations of observed frequencies from expected frequencies under a given log-linear model (Meyer et al., 2006). The ‘shading_max’ function of R package ‘vcd’ (Meyer et al., 2017) was used to shade the cells in the mosaic plots. The ‘Shading_max’ function applied the maximum statistics of the absolute Pearson residuals both to test the independence of the row and column variables in the contingency tables and to visualize significant cells which caused the rejection of the independence hypothesis (Meyer et al., 2006; Zeileis et al., 2007). By default, shading_max computed two cut-off points corresponding to confidence levels of 90% and 99% to shade the significant cells with less saturated colours (blue or red) and saturated colours, respectively (Meyer et al., 2017). Positive (blue) shades indicate the observed frequency is significantly greater than the expected frequency under the independence model which is the row variable is independent of the column variable, and the negative (red) shade indicates the observed frequency is significantly less than the expected frequency.

The degree of agreement between cow genotype identified by genomic data and cow genotype reported by farmers was evaluated by some statistics and visualized by an agreement plot. The statistics, including unweighted Cohen’s Kappa (Cohen, 1960), unweighted Bangdiwala’s B-statistic (Munoz and Bangdiwala, 1997), and overall agreement Z-test, were calculated based on the methods described by Friendly and Meyer (2016) using the R package ‘vcd’ (Meyer et al., 2017). Kapa = (PO

– PC)/(1- PC), where PO is the observed agreement and PC is the agreement expected by chance if the two classifying methods were independent of one another (Cohen, 1960). Cohen’s Kappa statistics were calculated based on the equation of Cohen (1960). Kappa can range from a minus value to 1, and the Bangdiwala statistic ranges from 0 to 1 (Bangdiwala and Shankar, 2013). For both Kappa and

100

B statistics, a small value indicates a higher level of disagreement and 1 indicates perfect agreement (Bangdiwala and Shankar, 2013).

The agreement plot was a visual representation of a k × k square contingency table, where rows were admixture-based genotype classification, columns were farmer’s genotype classification, and k was the number of genotype categories (k = 5). The agreement plot was constructed based on the steps described meticulously in the paper of Bangdiwala and Shankar (2013) using the R package ‘vcd’ (Meyer et al., 2017). Firstly, an n × n square, where n was the total number of cows (n = 344), was drawn. Secondly, k white rectangles of dimensions based on the row and column marginal totals were drawn. The total area of the white rectangles represents the maximum possible agreement area. Thirdly, k black squares of dimensions based on the diagonal cell frequencies were drawn and placed within the white rectangles. The total area of the black squares represents the observed exact agreement area. Finally, k grey rectangles of dimensions based on the diagonal cell frequencies plus the closest off-diagonal cells (1 step from the main diagonal) were drawn and placed within the white rectangles. The total area of the grey rectangles minus the total area of black squares represents the partial area.

4.3 Results 4.3.1 Genetic diversity within and between populations

The FIS values for all populations, including the VDC per region, were very close to zero; they ranged from -0.037 to 0.056 (Table 4.2). For the VDC, FIS was closest to zero in V-NL (0.003) and furthest from zero in V-NH and V-SL (-0.017 for both). Mean HO was similar across the four VDC regions and similar to that of the reference HOL population. All Mean HO of VDC and HOL were relatively high compared to the non-HOL reference populations, especially RSI and SAH.

101

Table 4.2 Mean (SD) of observed heterozygosity (HO), expected heterozygosity (HE), and

inbreeding coefficient (FIS) of Vietnamese dairy cattle in each of four geographically contrasting regions compared to reference breeds populations A

Population/breed n HO HE FIS BRM 25 0.230 (0.175) 0.237 (0.172) 0.018 (0.184) BSW 24 0.370 (0.181) 0.356 (0.157) -0.037 (0.191) CYE 59 0.269 (0.153) 0.283 (0.155) 0.041 (0.138) HOL 93 0.409 (0.125) 0.406 (0.116) -0.008 (0.104) JER 49 0.355 (0.162) 0.358 (0.153) 0.010 (0.146) RSI 10 0.186 (0.194) 0.201 (0.191) 0.056 (0.291) SAH 17 0.190 (0.195) 0.193 (0.188) 0.009 (0.213) V-NH 141 0.419 (0.117) 0.412 (0.109) -0.017 (0.086) V-NL 84 0.417 (0.116) 0.418 (0.104) 0.003 (0.114) V-SH 44 0.416 (0.133) 0.410 (0.113) -0.015 (0.15) V-SL 75 0.425 (0.118) 0.417 (0.103) -0.017 (0.118) A Abbreviations: RSI, Red Sindhi; SAH, Sahiwal; BRM, Brahman; HOL, Holstein; JER, Jersey; BSW, Brown Swiss; CYE, Chinese Yellow Cattle; VDC, Vietnamese dairy cattle, V-NH, VDC in northern highland; V-NL, VDC in northern lowland; V-SH, VDC in southern highland; V-SL, VDC in southern lowland; FIS=1-HO/HE, the mean reduction in heterozygosity of an individual due to non-random mating within a subpopulation.

The FST between pairs of VDC regional populations and between pairs of the VDC and reference

HOL populations were the smallest, ranging from 0.006 to 0.010 (Table 4.3). The FST was greatest between VDC and SAH, RSI, BRM or CYE, and intermediate between VDC and JER or BSW.

102

A Table 4.3 Pairwise fixation index (FST) between Vietnamese dairy cattle in each of four geographically contrasting regions compared to reference breed populations Population B BRM BSW CYE HOL JER RSI SAH V-NH V-NL V-SH BSW 0.341

CYE 0.055 0.293

HOL 0.271 0.140 0.240

JER 0.327 0.174 0.290 0.144

RSI 0.036 0.352 0.046 0.285 0.340

SAH 0.060 0.376 0.066 0.299 0.358 0.032

V-NH 0.252 0.126 0.220 0.007 0.127 0.267 0.280

V-NL 0.245 0.123 0.215 0.006 0.117 0.258 0.274 0.003

V-SH 0.268 0.135 0.234 0.010 0.139 0.276 0.297 0.008 0.005

V-SL 0.232 0.124 0.200 0.010 0.123 0.243 0.260 0.008 0.002 0.007

A FST: the mean reduction in heterozygosity of a population (relative to the total population) due to genetic drift among populations. FST measures the extent of genetic differentiation among populations and ranges from 0.0 (not different) to 1.0 (completely different – populations fixed for different alleles). B Population/breed abbreviations as in Table 4.2.

4.3.2 Principle component analysis to determine breed structure across populations

The first principal component (PC1) accounted for 27.2% of the total variation and separated the Zebu breeds RSI, SAH, BRM from VDC and the Bos taurus breeds, including HOL, JER, and BSW (Figure 5.1). The PC1 also separated CYE from the Zebu and Bos taurus breeds. The second component (PC2) accounted for 8.9% of the total variation and separated the Bos taurus breeds from each other. The third component (PC3) accounted for 3.8% of the total variation and separated BSW from a group of CYE and other breeds. All three PCs failed to clearly separate VDC from HOL. Approximately 10% of the V-NL population were positioned closer to JER, 12% of the V-SL population were positioned closer to the Zebu breeds, and 9% of the V-SH population were positioned closer to BSW.

103

Figure 4.1 Breed structure of Vietnamese dairy cow populations and reference population revealed by principal component analysis. Population/breed abbreviations as in Table 4.2. All Vietnamese dairy cows (VDC) are presented in the dark blue colour, but different VDC populations sampled from different regions across Vietnam, namely V-NH, V-NL, V-SH, and V-SL, are represented by different shapes, as indicated below the graphs.

4.3.3 Unsupervised hierarchical clustering of Vietnamese dairy cows

Admixture analysis results for K = 2 to 8 are presented in Figure 4.2, and the cross-validation (CV) error plot is presented in Figure 4.3. VDC generally clustered and had the same colour patterns as HOL at all K values. Similarly, Zebu cattle breeds (or Bos indicus, hereafter abbreviated as ZEB), including RSI, SAH, and BRM, always clustered as a single group at all K values. Although VDC showed a slightly higher proportion of ZEB, they always shared very similar colour patterns with HOL at all K values.

The ZEB and Bos taurus cows were clearly distinguished at K = 2, similar to PC1 in principle component analysis. At K = 3 (similar to PC2 in principle component analysis), the separation between three groups, ZEB, HOL, and JER, was observed, although CYE and BSW showed an admixture between those three groups. At K = 4, ZEB, JER, BSW, and HOL clearly separated into four distinct groups. At K = 5 to 8, ZEB, JER, and BSW still clustered as distinct groups. Increasing K from 5 to 7 only showed that the HOL and VDC were admixtures. However, at K = 8, CYE populations began to cluster into a new group (illustrated by the yellow colour at K = 8, Figure 4.2) which was most evident for DEC and HOH populations. Also, at K = 8, VDC showed a difference

104 from HOL, indicated by the presence of yellow segments in the admixture plot, especially in V-SL. These yellow segments indicated the inclusion of DEC and HOH (which are CYE) in VDC.

Figure 4.2 Admixture plot for Vietnamese dairy cows and reference breeds analysed according to different numbers of assumed ancestors (K = 2 to 8) Population/breed abbreviations are as in Table 4.2. Each individual is represented as a single vertical line, and the proportion of the coloured segment represents their estimated ancestry derived from different populations.

A cross-validation error plot was constructed to choose the most appropriate value for K, which is the optimal number of populations in the whole dataset (Figure 4.3). The cross-validation errors

105 decreased most rapidly when K increased from 1 to somewhere between 3 and 4. Therefore, K = 4 was selected as the most appropriate number of populations to focus on in the study.

Figure 4.3 Cross-validation plot for the different number of assumed ancestors (K = 1 to 10) indicating the choice of the appropriate K = 4.

At K = 4, most VDCs were very similar to HOL. Only a small proportion of VDCs were identified as crossbreeds of HOL with JER, BSW and ZEB (Figure 4.2 at K = 4). ANOVA analysis comparing Admixture-based breed compositions between VDC populations showed that the genetic proportion of ZEB was highest in SL (8%) and lowest in NH (2%) (P < 0.001). The genetic proportions of HOL tended to be lowest in V-SL (80%) and highest in V-SH and V-NH (88%) (P = 0.069) (Figure 4.4). However, the average genetic proportions of JER and BSW were similar across regions. Averaged across regions, the genetic proportions of HOL, JER, BSW, and ZEB in VDC were 85.0%, 6.0%, 5.3%, and 4.5%, respectively.

106

Figure 4.4 Comparisons of genomic breed proportions between Vietnamese dairy cow populations when K = 4 Population/breed abbreviations are as in Table 4.2.

4.3.4 Association between genotype and coat colour

The classification of VDCs based on the genomic data resulted in nine genotypes (Breed_G), whereas classification according to cow coat colour by farmers resulted in five (Breed_F) (Figure 4.5 a and b). Farmers could not differentiate between B3HOL_ZEB and PureHOL or between F1HOL_JER, B1HOL_JER, F1HOL_BSW, and B1HOL_BSW. The Pearson correlation between the proportions of the cows reported by farmers as HOL and the proportions of the cows identified by Admixture analysis as HOL was significant but relatively small (r = 0.22; P < 0.001).

According to the genomic classification (Breed_G), out of the 344 VDC sampled, the overall percentage (frequency) of PureHOL cows was 48%; B3HOL_ZEB, 22%; B2HOL_ZEB, 12%; F1HOL_JER 6%; B1HOL_BSW 3%; B1HOL_JER, 2%; B1HOL_ZEB, 1%; F1HOL_BSW 1%; and F1HOL_ZEB 1% (Figure 4.5 a). All four of the F1HOL_BSW cows, six out of seven B1HOL_JER cows, and eight out of nine B1HOL_BSW cows were located in the NH region. All three of the F1HOL_ZEB were located in the SL region.

Also, based on the genomic classification (Breed_G, Figure 4.5 a), of the 166 PureHOL cows, 54% were in the NH region, 20% in the NL region, 14% in the SH region, and 11% in the SL region. The

107 maximum statistic test indicated that there was a tendency for a higher frequency of F1HOL_ZEB (corresponding blue rectangle in Figure 4.5 a, P < 0.10) and lower frequency of PureHOL (corresponding red rectangle, P < 0.10) in the SL region. There was also a lower frequency of B2HOL_ZEB in the NH region (Figure 4.5 a, P < 0.10). These exceptions aside, the cow genotypes were distributed quite evenly across regions. In contrast, (Breed_F, Figure 4.5 b) farmers in the NH region reported a higher frequency of B3HOL_ZEB and PureHOL (P < 0.01) while farmers in the NL region reported a lower frequency of these genotypes (P < 0.01). Farmers in the SL region reported higher frequencies of F1HOL_ZEB, B1HOL_ZEB, and B2HOL_ZEB (P < 0.01), farmers in the SH region reported a higher frequency of B1HOL_ZEB (P < 0.01), and farmers in the NL region reported a higher frequency of B2HOL_ZEB (P < 0.01).

a) b)

Figure 4.5 Mosaic plots showing the relationship between regions and cow genotype identified by genomic data (Breed_G; a) and cow genotype reported by farmers (Breed_F; b) For both Breed_F and Breed_G: 1, F1HOL_ZEB; 2, B1HOL_ZEB; 3, B2HOL_ZEB; 4, B3HOL_ZEB; 5, PureHOL; 6, F1HOL_JER; 7, B1HOL_JER; 8, F1HOL_BSW; 9, B1HOL_BSW. The size of each rectangle is proportional to its observed frequency (number displayed), and the rectangles are shaded according to residuals from the model of independence under maximum test statistic. Positive (blue) shade indicates the observed frequency is significantly greater than the expected frequency under independence, and the negative (red) shade indicates the observed frequency is significantly less, as shown in the legend. Two cut-off points are computed corresponding to confidence levels of 90% (light blue or red) and 99% (darker blue or red), respectively.

The groupings were further simplified to allow an Agreement Plot Analysis between the farmer (Breed_F) and genomic classifications (Breed_G) (Figure 4.6). The B3HOL_ZEB and PureHOL were combined as one group, and the F1HOL_JER, B1HOL_JER, F1HOL_BSW, and B1HOL_BSW as a second group. This analysis indicated agreement between farmer and genomic classifications (P < 0.001). The Cohen’s Kappa and Bangdiwala’s statistics, 0.195 and 0.396 respectively, indicated 108 fair to substantial agreement between the two methods. However, the farmers classified genotypes correctly for only 52.3% of all cows. The genotypes classified by the farmers had a bias towards more cows classified as B1HOL_ZEB and B2HOL_ZEB, but fewer cows classified as B3HOL_ZEB, PureHOL, and other genotypes, since the path of rectangles for these two smallholders is above the diagonal line of no bias.

Figure 4.6 Agreement plot for the genotype reported by farmers and genotype identified by genomic data. Breed_F, genotype reported by farmers; Breed_G, genotype identified by genomic data; For both Breed_F and Breed_G: 1, F1HOL_ZEB; 2, B1HOL_ZEB; 3, B2HOL_ZEB; 4, B3HOL_ZEB; 5, PureHOL; 6, F1HOL_JER; 7, B1HOL_JER; 8, F1HOL_BSW; 9, B1HOL_BSW. The total area of black squares and the total area of their enclosing white rectangles represent the observed exact agreement area and the maximum possible agreement area, respectively. The total area (including the total area of black squares) of the grey rectangles represents the total of exact and partial agreement areas. The extent to which the white rectangles are above or below the red line indicates the extent and direction of any disagreement. The overall strength of the unweighted agreement was measured by either Cohen’s Kappa or Bangdiwala’s B statistics (1960). The B statistics were the ratio of the total area of the black squares to the total area of their enclosing white rectangles. For both Kappa and B statistics, a small value indicates a higher level of disagreement and 1 indicates perfect agreement (Bangdiwala and Shankar, 2013). Kappa and B statistics are 1 when all the black squares exactly overlap the white rectangles.

109

Although there was an association between Admixture-based classifications of breed composition (Breed_G) and the coat colour of the cows (P < 0.001), this relationship was not clear in all subgroups (Figure 4.7). Associations were strongest for F1HOL_JER cows, which generally had brown-yellow hair (P < 0.01) and B1HOL_JER cows, which tended to have 76 to 100% black hair (P < 0.10). For all other crossbreeds, associations between their genotypes and hair coat colours were not significant (P>0.10).

Figure 4.7 Relationship between cow genotypes identified by genomics and coat colour Breed_G, genotype identified by genomic data; 1, F1HOL_ZEB; 2, B1HOL_ZEB; 3, B2HOL_ZEB; 4, B3HOL_ZEB; 5, PureHOL; 6, F1HOL_JER; 7, B1HOL_JER; 8, F1HOL_BSW; 9, B1HOL_BSW. Coat colour: Y, Brown yellow; 0-25, <25% black; 26-50, 26 to 50% black; 51-75, 51 to 75% black; 76-100, 76 to75% black.

4.4 Discussion

Knowledge of the status of diversity, inbreeding, and breed composition of VDC populations is crucial to the planning of a national dairy cattle selection and breeding program in Vietnam. However, the lack of cow pedigree records in the smallholder dairy farms examined in this study indicates the difficulty of reliance on traditional pedigree-based methods. The current study is the first, as far as we are aware, to use SNP chip and genomic methods to assess the level of genetic diversity and breed composition of VDC populations.

110

4.4.1 Genetic diversity of Vietnamese dairy cows

Inbreeding coefficient (FIS) is an important indicator of genetic diversity as it indicates the mean reduction in heterozygosity of individuals due to non-random mating within a population (Wright,

1965). A FIS close to zero implies a low degree of inbreeding (Lwin et al., 2018). In the current study, the close-to-zero FIS values of all VDC populations (from -0.017 to 0.003) suggest that inbreeding is not a concern.

Besides FIS, observed heterozygosity (HO) is an important indicator of genetic diversity. A higher Ho reflects a higher level of genetic diversity (Zhang et al., 2018). The high HO of VDC populations (from 0.416 in V-SH to 0.425 V-SL) compared to reference populations (from 0.186 in RSI to 0.409 in HOL) suggest that VDC populations have retained a quite similar level of genetic diversity to the reference HOL population, and that they are even more diverse than other reference populations included in the current study. Mean HO of VDC populations in the current study was also higher than the mean HO reported for HOL (0.31), JER (0.26), and BSW (0.27) in a study by Melka and Schenkel

(2012), and the mean HO reported for HOL (0.368) and JER (0.308) in the study of Cheruiyot et al. (2018). These results indicate that VDC populations are similar to or even more diverse than the reference HOL, JER, and BSW populations, which are the high genetic merit cows in developed countries.

The Pairwise fixation index (FST) measures the extent of the genetic relationship between two populations. Generally, a smaller FST indicates a closer genetic relationship (Wright, 1965). Wright

(1978) suggested the qualitative guidelines for the interpretation of FST: FST = 0.00 to 0.05 indicates little genetic differentiation, FST = 0.05 to 0.15 indicates moderate genetic differentiation, FST = 0.15 to 0.25 indicates high genetic differentiation, and FST above 0.25 indicates exceptional genetic differentiation. In the current study, the very low pairwise FST between VDC populations and reference HOL populations (ranging from 0.006 to 0.01) suggest that VDC are genetically similar to the reference HOL. The smaller pairwise FST between VDC and JER (0.117 to 0.139) compared to the pairwise FST between HOL and JER in the current study (0.144) indicates that the genetic relationship between VDC and JER is closer than that between HOL and JER. Similarly, the smaller pairwise FST between VDC and BSW (from 0.126 to 0.135) compared to a pairwise FST between HOL and BSW in the current study (0.140) indicates a genetic relationship between VDC and BSW closer than that between HOL and BSW. The closer genetic relationship between VDC with JER and BSW is because a certain per cent of VDC are the crossbreeds/composites of HOL with JER and BSW.

111

4.4.2 SNP-based estimates of breed compositions of Vietnamese dairy cows

Although the literature suggested that the main founder breeds of VDC included HOL, JER, BSW, RSI, SAH, BRM, and CYE (Cai and Long, 2002; Vang et al., 2003; Trach et al., 2007), the current study indicates that the major genetic proportions in VDC are from just three breeds: HOL (85.0%), JER (6.0%), and BSW (5.3%). The genetic proportions of all other proposed founder breeds (RSI, SAH, BRM, and CYE) could not be differentiated from each other. When combining these breeds into a ZEB proportion, they only accounted for 4.5% of the VDC genetics. While CYE genetics could exist in VDC, this may not be meaningful as they were only detected at K = 8. The variety of breeds in the VDCs appears similar to that in other tropical smallholder dairy herds, for example, in Thailand, where representative breeds have been defined as HOL, JER, BRM, RSI, SAH, Red Dane and Thai Native (Wongpom et al., 2017).

It has recently been reported that PureHOL cattle account for only 5-6% of VDCs while the rest are dairy crossbreeds (National Institute of Animal Science, 2016). Consequently, to drive the improvement of genetic merit in the domestic herd, Vietnam continues to import a large number of PureHOL cattle every year, including bulls, semen straws, and especially heifers, as it is believed that pure HOL cows are the most productive among dairy cow breeds (Trang, 2019; Tue, 2020). However, the current study indicates that far more VDCs than previously thought, up to 48% of them, are PureHOL, and 70% are greater than 7/8 HOL (B3 or better). Thus, if PureHOL is preferable, Vietnam can select PureHOL bulls and cows from the current VDC herd instead of relying on imports.

However, many studies have found that HOL crossbreeds appear to be more suitable for tropical dairy farming systems than PureHOL. In the tropics, HOL crossbreeds are considered more adaptable, fertile, and disease-resistant than PureHOL and have higher lifetime productivity (Garcia-Peniche et al., 2005; Effa et al., 2013; Dalcin et al., 2016; Alfonzo et al., 2016). Given that the HOL crossbreeds are more suitable for the tropics and with the result of the current study that VDC include mainly genetic proportions of dairy breeds (95%; HOL, BRW, and JER all together) but also a certain proportion of ZEB (4.5%), a suitable genetic resource as a base for selecting locally adaptable dairy cows for Vietnam is present.

Due to the lack of systematic recording of cows’ pedigrees, SDF farmers could not accurately classify their cows' genotype when they were asked to do so. Farmers were only able to identify correctly the genotypes of approximately 52% of VDC. Identifying cow genotypes on the basis of coat colour, as often practised by many farmers, is not an appropriate and accurate method of doing so. The current study showed that whilst F1HOL_JER cows are more likely to have brown-yellow hair coat (P < 0.01) and B1HOL_JER cows tended to have 75 to 100% black hair coat (P < 0.10), there were no 112 significant associations between SNP-based estimates of cow breed composition and coat colours for any other crossbreeds. The poor relationship between coat colour and breed composition in diverse crossbreds is understandable, as breed composition is a measure of the whole genome while coat colour is controlled by only a few genes (Seo et al., 2007; Dorshorst et al., 2015). Thus, the current results imply that a more rigorous recording of cow breed or breed mixtures is necessary for Vietnamese SDFs. Moreover, since knowledge of cow genotype is crucial for designing breeding programmes and planning dairy nutrition and management strategies, poor categorisation of genotype could lead to poor performance.

4.5 Conclusions

The majority of VDC tested were genomically determined to be very close to a pure HOL population, and they also retained a similar level of genetic diversity to that reference population. At least across the 32 herds tested, inbreeding appears not to be a problem.

The cow genotypes determined were distributed relatively evenly across dairy regions. Since approximately half of the VDC herd were PureHOL and half were crossbred HOL, mainly with ZEB, the current VDC herd could be considered a good genetic base for selecting either pure HOL cows or cows with a certain genetic fraction of tropically adapted ZEB combined with HOL genetics.

113

Chapter 5 Issues of feeding strategies for lactating cows in Vietnamese smallholder dairy farms

Abstract

Limited literature suggests relatively simple feeding regime and diet formulation strategies for dairy cows in Vietnamese smallholder dairy farms (SDFs). This study aimed to classify and compare feeding regimes and nutrient balance for lactating cows across four typical dairy regions in Vietnam and to evaluate the possibility of systematic dietary imbalance. Eight SDFs from each of four regions (a southern lowland, a southern highland, a northern lowland, and a northern highland) were randomly selected and visited for two adjacent milking periods per farm. For each visit, frequency and methods of feeding and supplying water to the lactating cows were recorded. Individual cow milk yields at each milking (pm and am) were weighed and sampled to assay and calculate fat corrected milk yield (FCM). Each diet ingredient offered and refused by each lactating group was weighed and sampled for calculation of dry matter intake per cow (DMI), and analysis of the nutrient composition in the component offered was undertaken. PCDairy, a diet formulation computer model, was used to calculate actual compared to recommended dietary nutrient concentrations, potential milk production, and predicted enteric methane emissions. Factor analysis, cluster analysis, and ANOVA were applied to determine grouping effects across as well as between regions. Feeding regimes were grouped into three clusters and diets into nine. Farmers in the same region tended to apply similar diets and feeding regimes. Across regions, 47% of all SDFs supplied water ad libitum to the cows, but all SDFs did not measure dry matter concentration of feeds, and only 9% of SDFs weighed feed ingredients when feeding. The most used roughages, including Napier grass, corn silage, fresh corn with cob, and rice straw, were all relatively high in ADF, NDF, and lignin. The diets in all regions were excessive in CP, ADF, NDF, lignin, and most minerals (Ca, P, Mg, K, Na, Fe, Zn, Cu, Mn, and S); but insufficient in net energy, non-fibre carbohydrate, and starch. Feed efficiency (0.98 kg ECM/kg DMI) and predicted methane emissions (21.0 g CH4/kg ECM) from the diets were sub-optimal. Feeding regimes and dietary nutrient balance of the southern lowland SDFs were most problematic. Increasing net energy concentration by increasing the use of concentrates and fat, and decreasing the fibre concentration of the diet by decreasing the use of Napier grass or rice straw to balance the diets, might help improve the milk production and thereby increase feed efficiency and reduce methane emission. Further studies aiming to identify suitable forage genotypes and optimal stage of harvest are recommended.

114

5.1 Introduction

As discussed in the previous chapter, Vietnam remains heavily reliant on smallholder dairy farms (SDFs) for its domestic supply of fresh milk. However, the average daily milk yields of these SDF cows, despite recent improvements, remain relatively low at 14 to 15 kg/cow/d in southern (Vu et al., 2016) and northern (Ashbaugh, 2010) provinces. This impacts not only national milk production but also the incomes of the SDFs. Although the reasons for low productivity can be multifactorial, poor nutrition is usually considered the most likely reason (Cuong et al., 2006a). High genetic-merit Holstein cows, such as those increasingly being imported into Vietnam, can only support high milk yield if cow nutritional requirements are met (NRC (2001).

A thorough analysis of lactating cow diets in the SDFs, relative to requirements, is required to define dietary limitations. To do that, the dietary ingredients and amounts of each that SDF farmers offer the cows and the nutrient composition of each must be determined. From that, the dietary supply of key nutrients or nutrient groupings (Crude protein, CP; Neutral or acid detergent fibre NDF, ADF; lignin; non-fibre carbohydrate, NFC; Starch; and Key minerals, such as Ca and P) can be calculated. Similarly, requirements can be calculated once key information on the cow is available (body weight, lactation number, days in milk, milk yield, milk fat and protein per cent, targeted body weight gain, days of pregnancy). From these data, the requirements can be estimated by using the nutritional guidelines of the National Research Council (NRC, America), Agricultural Research Council (ARC, United Kingdom), National Institute for Agricultural Research (INRA, France), or Commonwealth Scientific and Industrial Research Organisation (CSIRO, Australia) (Lean, 2011b; Tedeschi et al., 2014).

Although these calculations can be done manually, many computer-based nutrition models such as PCDairy, Cornell Net Carbohydrate and Protein System (CNCPS), CamDairy, Feed into Milk (FiM), Molly, and Rumen8 (Lean, 2011b; Tedeschi et al., 2014; Robinson and Ahmadi, 2015) are now available to assist. However, in the case of Vietnamese SDFs, there appears to have been no systematic published work on the required inputs for such models.

Only a few surveys have mentioned examples of locally relevant input data required for modelling the diets for the SDF cows in Vietnam (Loan et al., 2004; Chu et al., 2005; Cuong et al., 2006a; Lam et al., 2010; Phong and Thu, 2016). Fresh Napier grass (Pennisetum purpureum) appears to be the main type of forage used (Chu et al., 2005; Phong and Thu, 2016). This is followed by agricultural by-products such as rice straw, corn stalks (after the grain is removed), banana stalks; then other cultivated tropical grasses such as Ruzi grass (Brachiaria Ruziziensis), Guinea grass (Panicum maximum), Signal grass (Brachiaria Decumbens), Long Tay grass (Brachiaria mutica), and Mulato 115

(Brachiaria Ruziziensis x Brachiaria Decumbens x Brachiaria brizantha); and finally some naturally available forage collected from river banks or fallow areas (Chu et al., 2005; Phong and Thu, 2016). Commercial concentrate pellets, often called “complete pellets” in Vietnam and commonly purchased from milk processing companies or feed companies linked to the milk companies, are the main type of concentrate offered. SDF farmers do not commonly appear to mix their own concentrates or mix concentrates into the forage component, and concentrates are mainly offered to cows separately to the forage. In Ho Chi Minh City, in the south of Vietnam, Lam et al. (2010) reported lactating cow diets of 48.6% (dry matter basis) roughage such as Napier grass, Guinea grass or rice straw; 19.5% by-products such as brewers grain; and 33.7% concentrates. Cows were fed twice a day, and depending on the milk yield of the cows and the availability of the roughage, each cow was given (on a fresh basis) about 20 to 40 kg of roughage and 4 to 6 kg of commercial concentrates per day (Loan et al., 2004; Lam et al., 2010). The concentrate was offered before milking and roughage after (Lam et al., 2010). In Son La, a highland province in the north, Cuong et al. (2006a) reported that cows were given diets comprising 51.8% roughage and 48.2% concentrate. Napier grass and Signal grass were offered ad libitum, and commercial concentrates were available at a ratio of 0.5 kg concentrate per 1 kg of milk (Cuong et al., 2006a). Roughage was provided to cows first, then concentrate was offered to the cows during milking (Cuong et al., 2006a). Although the information from these studies is insufficient for reliable analysis of nutrient concentrations in the diets, it suggests the potential for diets to be regionally specific and imbalanced. The tropical roughage most commonly used are usually low in CP, high in PDF and ADF, and low in dry matter digestibility and net energy concentration (National Institue of Animal Husbandry, 2000; Dung et al., 2007). Similarly, the simple rules used to add roughage to concentrate indicate the risk of dietary imbalance and health concerns such as ruminal acidosis.

Realizing these concerns, this study was conducted to classify and compare the feeding regimes and diets for lactating cows in four typical but geographically contrasting dairying regions of Vietnam and to define the likely dietary imbalances within and across these regions. It was hypothesized that the feeding regimes and lactating cow diets are specific to each region, and that across regions, the lactating cow diets are excessive in fibre, with insufficient nutrient concentration for high yielding cows.

5.2 Materials and methods 5.2.1 Farm selection and farm visits

This study was conducted from 24 August to 7 October 2017 on 32 SDFs which were randomly selected from four main dairy regions (8 farms per region) of Vietnam, including a southern lowland 116 region (SL), a southern highland region (SH), a northern lowland region (NL), and a northern highland region (NH). The process of selecting and visiting the farms is detailed in Section 3.2 above in this thesis. Briefly, each SDF was visited for an afternoon and the next morning either side of and during milk times when feed was being offered and milk collected to allow the researcher to measure the inputs necessary to use the PCDairy model to calculate predicted against actual milk yield.

5.2.2 Feeding regime

A trained team of three to four observers visually recorded the feeding regime per SDF. The “Feeding regime” dataset included eight qualitative and four quantitative variables. The qualitative variables were: 1) quantity of water supplied (WaQuTi), water quantity was classified as ad libitum (AdWaQuTi) when water trough was always full and accessible for the cows, moderately (MoWaQuTi) when water trough was not always full or full but too small for the whole herds, and insufficient (InWaQuTi) when water trough was empty for an hour or more and the cows were observed waiting near the water trough; 2) Quality of water supplied (WaQuLi), was classified as poor (PoWaQuLi), when water was clouded and had a strange colour or smell, medium (MeWaQuLi) when water was clouded with visual contamination but had no strange colours or smells, or good (GoWaQuLi) when water was clear with no smell; 3) Water and concentrate troughs (SaCoWa): yes (YesSaCoWa) if the same trough was used for both concentrates and water, or no if not (NoSaCoWa); 4) Weighing feed (WeiFe): yes, if feed was weighed before feeding, (YesWeiFe) or no, if not (NoWeiFe); 5) Measuring feed dry matter (MeaDM): yes, if farmers measured feed dry matter (YesMeaDM) or no, if not (NoMeaDM); 6) Feeding partial mixed ration (PMR): yes, if partial mixed ration was used (YesPMR) or no, if not (NoPMR); 7) Feeding roughage before concentrate (CoBeRo): yes, if concentrates were offered before roughage (YesCoBeRo) or no, if not (NoCoBeRo); and 8) Mixing concentrates and roughage during feeding time (MixDu): yes, if concentrates and roughage were mixed during feeding time (YesMixDu) or no, if not (NoMixDu). The partial mixed ration was a mixture of corn silage, concentrates, and minerals that farmers bought from feed processing companies. The quantitative variables were: 1) Number of feed types used (FeTyp); 2) Times of feeding roughage per day (FeTim); 3) Times of feeding concentrates per day (CoTim); and 4) Times of cleaning feed trough per week (FeCle). Two feeding times were counted as two if one feeding occurred at least 3 hours from the other.

5.2.3 Diets

Each feed type offered to the lactating herd was weighed at feeding times using a digital hanging scale Model OCS M 100 (Vietnam Japan Digital Scale Company, 2017) (Figure 5.1 a to c). The total

117 amount of each type of feed offered to a cow per day was calculated as the total amount of that feed offered for all lactating cows summed across the day and divided by the total number of cows eating that feed. In all farms, concentrates were eaten completely, and only a small amount of roughage (< 3% offered) was leftover. The refusal roughage was collected, separated, and weighed to calculate intakes of each type of feed. Across regions, 19 feed types and some supplementations were observed, and amounts consumed were measured. The prices of the feed types were obtained by asking the farmers, and averaged per feed type.

Approximately 1 kg of each type of roughage and 400 g of each type of concentrate used by each SDF were sampled and stored in 30x30 cm plastic sealable bags and frozen in a fridge. After all feed samples had been collected, the same feed type samples within each region were counted. For the feed types with more than two samples per region, the samples were mixed, with the same weight ratio in the mix per initial sample, and a new sample was taken for nutritional panel analysis. There were two exceptions to this rule. There were more than two corn powder samples per region, but the nutrient composition of this feed was available and sufficient; thus, they were not analysed. In contrast, there was only one sample of passion fruit pulp, but the nutrient composition of this feed was not available; thus, it was analysed.

The samples were dried to determine dry matter concentration (DM) and transferred to dry samples at the Animal Nutrition Laboratory, Faculty of Animal Science, Vietnam National University of Agriculture. After that, the dry feed samples were ground and sent to Dairy One Forage Laboratory in America for analysis of other chemical components by traditional wet chemistry methods described in detail at Laboratory Analytical Procedures (Dairy One Forage Laboratory, 2015). Feeds were analysed for the following chemical compositions: crude protein (CP), acid detergent fibre (ADF), neutral detergent fibre (NDF), lignin, fat, starch, non-fibre carbohydrate (NFC), total digestible nutrients (TDN), net energy for lactation (NEL), calcium (Ca), phosphorus (P), magnesium (Mg), potassium (K), sodium (Na), iron (Fe), zinc (Zn), copper (Cu), manganese (Mn), and sulphur (S).

The chemical compositions of the feed types used by farmers but not analysed in the laboratory were derived from the available feed library of PCDairy (Robinson and Ahmadi, 2015) and available feed nutritive value books (National Institue of Animal Husbandry, 2000).

Using intakes “as fed” of 19 feed types and the DM of each feed, the DM intake (DMI) of each feed was calculated per cow per farm; and this became the “Diet dataset”.

118

a) Weighing concentrate b) Weighing roughage

c) Weigh refusal d) Run PCDairy

Figure 5.1 Measurements of feed offered, feed refused, and appearance of PCDairy software

5.2.4 Measurement of milk production

Milk yield (MILK, kg/cow/d) and milk fat concentration (mFA, %) were obtained from Chapter 3. Briefly, morning and afternoon milk yields were weighed per cow and summed to obtain single-day milk yield per cow (MILK, kg). Morning and afternoon milk were sampled per cow to analyse milk fat concentration (mFA, %) at the Nutrition Laboratory, Vietnam National University of Agriculture.

Different from Chapter 3, where milk yield was converted to energy corrected milk, in the current study, single-day MILK was converted to fat corrected milk (FCM, 3.5% fat) so that the actual FCM production would be comparable with the diet-allowable FCM (3.5% fat) predicted by PCDairy software. Specifically, FCM (3.5% fat) was calculated from MILK and mFA using the equation of Britt et al. (2003):

119

FCM (kg/cow/d) = MILK (kg/cow/d)  [16.23  mFA (%) + 0.432]

5.2.5 Identification of dietary imbalance

PCDairy

PCDairy version 2015 (Figure 5.1 d), a computer-based mechanistic nutrition model developed by the University of California Davis, was used. PCDairy estimates nutrients supplied by diets, from an inputted feed library, and nutrient requirements of animals based on the empirical nutrition models of NRC (2001) and estimated methane emission from the diets based on the equations of Moraes et al. (2014) (Ahmadi et al., 2013; Robinson and Ahmadi, 2015). PCDairy was chosen as it was made available and accessible for use by Vietnamese farmers, nutritionists, and extensionists through a cooperation program between the US Department of Agriculture (USDA), University of California Davis (UC-Davis) and the Ministry of Agriculture and Rural Development (MARD) (Mateo, 2016; Kebreab et al., 2019). Also, PCDairy was the only available nutrition model translated into Vietnamese to make it as user friendly as possible.

Background data of lactating herds

Besides data of MILK, and mFA, PCDairy required the background data of lactating herds, including the number of cows, the average number of lactations, days in milk, daily weight gain, and level of energy adjustment for activity. The majority of these data were obtained from Chapter 3. Specifically, at each visit, the total number of lactating cows was counted, and the farmers were asked to check their recording and memory to provide lactation number and days in milk for each cow. This information was used to calculate the percentage of the first and second lactation cows and the average days in milk of each herd. Mean ( SD) number of lactating cows in regions were 9  5 cows in SL, 6  2 cows in SH, 11 2 cows in SL, and 18  6 cows in SH. Means ( SE) of percentages of first and second lactation cows were similar across regions and were 36  22% and 31  20%, respectively.

The body weight (BW) of each cow was estimated from heart girth using an equation suggested for cattle by Goopy et al. (2017). The heart girth of each lactating cow was measured by draping a tape measure (Asia Technology Service Company) around the girth closest to the heart. Means ( SE) of average days in milk and BW of lactating cows across regions were 181 57 days and 498  49 kg, respectively.

120

Daily weight gain and level of energy adjustment for activity were set up as same as the default values of PCDairy, which were zero kg weight gain per day and 5% of energy requirement for maintenance of cows having limited space to move inside the cowsheds (Robinson and Ahmadi, 2015).

Analysing lactating cow diets by PCDairy

When using PCDairy, firstly, the FEEDLIST package was used to update the feed library with the price of all feed and the nutrient profile of new feeds. Then, the ANLSIS-L package was used to analyse the diets.

For each SDF, ANLSIS-L was used for four tasks that require slightly different input data. Task 1 was calculating dry matter intake, cost, and actual nutrient concentrations of the diets offered to cows. Task 2 was modelling the recommended nutrient concentrations to produce observed FCM (called PCDairy recommendation 1). Task 3 was modelling the recommended nutrient concentrations to produce a target FCM (3.5% fat) yield of 25 kg/cow/d (PCDairy recommendation 2). Finally, task 4 was predicting NEL-allowable FCM, CP-allowable FCM, and methane (CH4) emission per unit of FCM from the cows given a diet. For all tasks, the data including intakes “as fed” of dietary ingredients, percentage of the first and second lactation cows in each lactating herds, average days in milk, BW, daily body weight change, and NEL added for activity were entered into the ANLSIS-L package. The difference in input data ANLSIS-L between tasks 1, 2, 4 and task 3 was that the actual MILK and mFA were entered for tasks 1, 2 and 4, while a target milk yield of 25 kg FCM/cow/d and 3.5% milk fat was entered for task 3.

By comparing the actual nutrient concentrations with PCDairy recommendation 1, the particular nutrient deficiency in the actual diet for the production of the actual MILK can be determined (Robinson and Ahmadi, 2015). Similarly, by comparing the actual nutrient concentrations with PCDairy recommendation 2, the particular nutrient deficiency in the actual diet for the production of a target FCM of 25 kg/cow/d can be determined (Robinson and Ahmadi, 2015). Also, the balance between dietary NEL and CP of the actual diets can be determined by comparing NEL-allowable FCM and CP-allowable FCM.

5.2.6 Statistical analysis

Multivariate statistical analysis

The hierarchical Clustering on Principal Components (HCPC) analysis method was applied to cluster Feeding regime dataset (12 variables) and Dietary dataset (19 variables). For datasets that include many inter-correlated variables, as was the case in the current study, HCPC is commonly applied to categorize the system or regime (Gelasakis et al., 2012; Riveiro et al., 2013; Kuivanen et al., 2016; 121

Todde et al., 2016). Based on HCPC analysis, three standard methods, including principal component methods, hierarchical clustering method, and partitioning clustering method, were applied to cluster the farms into the groups so that the farms in the same group were more similar to each other than to those in other groups (Husson et al., 2010; Kassambara, 2010). Firstly, depending on the type of the dataset, either the Principal Component Analysis (PCA) method or Factor Analysis of Mixed Data (FAMD) method was applied to transform the dataset into non-correlated principal components (PCs). In the current study, PCA was applied to the Diet dataset as this dataset included only quantitative variables, while FAMD was applied to the Feeding regime dataset as this dataset included both quantitative and qualitative variables. After that, to reduce noise and increase cluster stability in the data, hierarchical cluster analysis was applied on only some first PCs to identify an initial number of clusters. The decision as to how many and which principal components (PCs) to keep was made based on Kaiser’s criterion; all PCs with an eigenvalue  1.00 were initially retained (Kaiser, 1961). Additionally, the cumulative percentage of variance explained by the retained PCs was cross-checked to make sure it was  70% (Kuivanen et al., 2016). Finally, the k-means clustering method was applied to identify an optimum number of clusters and assign farms to each cluster (Husson et al., 2010).

To further characterize each cluster in the final sets of clusters, V-tests (Husson and Josse, 2010) were used. For quantitative variables, V-tests compared the mean of each variable in each cluster to the mean of that variable in the whole data set (Kuivanen et al., 2016). For qualitative variables, V-tests compared the percentage of each category of each qualitative variable in each cluster to the percentage of that category in the whole data set (Kuivanen et al., 2016).

All the multivariate statistical analyses were performed using R package ‘FactoMineR’ (Husson et al., 2019), and the results of multivariate analyses were visualized using R package ‘factoextra’ (Kassambara and Mundt, 2019). The results of the HCPC analysis were visualised as dendrograms. Biplots of the first two principal components were drawn to visualize the correlations between the diets of 32 farms and the feed variables. Graphs were plotted to visualize the relationships between the feeding regimes of 32 farms and quantitative and qualitative feeding regime variables.

Statistical comparisons

All statistics were performed using the base and additional packages of R software (R Core Team, 2018). Farms were the experimental unit in all analyses. Descriptive statistics for quantitative variables were calculated for each region using the R package ‘psych’ (Revelle, 2019). Before any statistical comparison, the normality of quantitative variables was tested using both the Shapiro-Wilk

122 test and histograms. The results are presented as means for normally distributed quantitative variables, medians for not-normally distributed quantitative variables, and frequency for categorical variables.

The choice of suitable tests for comparisons of variables between regions was based on the guidelines of McDonald (2014). For variables that were found to be not-normally distributed, medians were compared by Kruskal-Wallis tests followed by Dunn post-hoc tests (P < 0.05) using the R package ‘FSA’ (Ogle et al., 2019). For normally distributed variables, means were compared by One-way ANOVA tests followed by Tukey–Kramer tests (P < 0.05), using R package ‘agricolae’ (Mendiburu, 2019). For categorical variables, frequencies of each sublevel of variables were compared by Fisher’s exact tests followed by Bonferroni-corrected pairwise Fisher's exact tests (P < 0.05), using the R package ‘rcompanion’ (Mangiafico, 2019).

5.3 Results 5.3.1 Feeding regime

The feeding regime dataset is presented, per region, in Table 5.1. In NL, both concentrates (4 times per day) and roughage (4 times per day) were offered more frequently than in all other regions, except for SL where roughage was offered at a similar level to NL. Across regions, concentrates were offered on average twice per day and roughage three times per day. Feed troughs were cleaned approximately twice as often in NL compared to NH and SH, with SL in between (P < 0.05). Across regions, feed troughs were cleaned an average of 10 times per week.

Only 15 out of all 32 SDFs supplied water ad libitum to the cows, and the majority of these SDFs were in NL (7 SDFs) and NH (6 SDFs) (Table 5.1). All SDFs in SL only supplied water to the cows after feeding concentrate. None of the SDFs in SL supplied water ad libitum to the cows. Similarly, only two out of 8 SDFs in SH supplied water ad libitum for the cows.

None of the SDFs took the dry matter of feeds into account when determining amounts to offer, and only three out of 32 SDFs across regions weighed feed ingredients before feeding (Table 5.1). None of the SDFs mixed concentrates and roughage before feeding time. One SDF in SH, and four in NL, mixed concentrates with roughage during feeding time. All SDFs in SL and all but one in NH offered concentrates separately to roughage.

The FAMD analysis, using the feeding regime dataset on the 13 variables (Table 5.1), defined the first five principal components (PCs), accounting for 78.0% of the total variance. HCPC, based on those first five PCs, defined three optimum housing management clusters (Figure 5.2 a). All NL SDFs

123 and one SH SDF grouped into the feeding regime Cluster 1 (coloured in red); all SL SDFs were in Cluster 2 (yellow), and all SH and NH SDFs were in Cluster 4 (purple).

Table 5.1 Comparisons of feeding regime for dairy cows across four main dairy regions

Region A, Median or n B Parameter P C Overall D SL SH NL NH Quantitative variables Mean  SEM Type of feeds 4.5 5 4 5 0.845 4.6  0.2 Roughage feeding times 3.5a 2.5b 4a 3b < 0.001 3.3  0.3 Concentrate feeding times 2b 2b 4a 2b < 0.001 2.5  0.5 Feed trough cleaning 10ab 7b 14a 7b 0.034 10.0  2.0 frequency (times/week Qualitative variables Frequency (%) Supply water ad libitum 0b 2ab 7a 6a < 0.001 15 (47) Same trough for water and 8a 0b 0b 0b < 0.001 8 (25) concentrate Using partial mixed ration 0b 0b 0b 8a < 0.001 8 (25) Water with visual contamination 3 3 0 5 0.079 11 (34) Weigh feeds before feeding 0 1 2 0 0.587 3 (9) Measure feed dry matter 0 0 0 0 1.000 0 (00) Feed concentrates and roughage 8a 0b 0b 7a < 0.001 15 (47) separately Mix concentrates and roughage 0 0 0 0 1.000 0 (0) before feeding Mix concentrates and roughage 0 1 4 0 0.034 5 (16) during feeding A SL, Southern lowland; SH, Southern highland; NL, Northern lowland; NH, Northern highland. B n, number of farms out of eight farms. C P-values are given for either Kruskal-Wallis tests (superscript letters are given for post-hoc Wilcoxon rank sum test; P < 0.05) or Fisher’s exact tests (superscript letters are given for post-hoc Bonferroni-corrected pairwise Fisher's exact test; P < 0.05). D Overall mean (SEM) of medians or overall frequency (percentage) of all farms. a, b, c Medians or percentages with the different superscript letters within a row differ significantly from each other, P < 0.05.

The directionality and the amount of variation in feeding regime variables and the associations of these variables with the feeding regime clusters are presented in a 2-dimensional view of the first two principal components (Figure 5.2 b for all variables and Figure 5.2 c for sublevels of qualitative 124 variables and feeding regime clusters). The qualitative variables that varied most (furthest from the original coordinates in Figure 5.2 b and c) and most meaningful in the partitions of the clusters were: “using the same trough for water and concentrate (SaCoWa), yes or no”, “concentrate feeding times (CoTim)”, “water quantity (WaQuTi), insufficient, moderate, or ad libitum”, “mixing concentrate and roughage during feeding (MixDu), yes or no”, and “using partial mixed ration (PMR), yes or no”.

The main characteristics of each feeding regime cluster are presented in Figure 5.2 c. Cluster 1 (all NL SDFs in red) was characterised by mixing feed during feeding and weighing feed before feeding (red SDFs are close to sublevels “YesMixDu” and “YesWeiFe” in Figure 5.2 c). However, Figure 5.2 c was only the visualized relationships of feeding regime cluster and feeding regimes variables in the view of the two first PCs. V-tests described better the main characteristics of each feeding regime cluster by statistically comparing the mean of each quantitative variable in each cluster with the mean of that quantitative variable in the whole dataset and comparing the percentage of categories of each qualitative variable in each cluster to the percentage of that category in the whole dataset. V-tests showed that the SDFs in Cluster 1 (all NL SDFs) offered cows concentrates and roughage more times per day and cleaning feed troughs more times per week than average. Also, more SDFs in this cluster than average supplied concentrates and roughage simultaneously and mixed them during feeding, supplied cows ad libitum good quality water, did not use the same trough for both water and concentrates, and did not use partial mixed ration. SDFs in Cluster 2 (all SL SDFs) offered cows concentrates fewer times per day than average. Also, more of the SDFs in this cluster than average used the same trough for both concentrates and water, offered cows concentrate before roughage, and supplied cows with moderate quality water than average, but fewer of the SDFs in this cluster than average supplied cows water ad libitum. SDFs in Cluster 3 (all NH and 7 SH SDFs) provided concentrates and roughage fewer times per day than average, used partial mixed ration more, did not use the same trough for both water and concentrates, and did not mix concentrates and roughage during feeding more than average.

125

a) HCPC – Cluster dendrogram

b) First two PC view of all variables

126

c) First two PC view of sublevels of qualitative variables and observations

Figure 5.2 Results of factor analysis (FAMD) and hierarchical clustering on principal components (HCPC) for feeding regime data Thirty-two farms: SL1 to SL8, SH1 to SH8, NL1 to NL8, NH1 to NH8 represent the farms numbered from 1 to 8 in southern lowland, southern highland, northern lowland, northern highland regions, respectively. Feeding regime variables: WaQuTi: if water supplied to cows ad libitum (AdWaQuTi), moderately (MoWaQuTi), or insufficiently (InWaQuTi); WaQuLi: if water quality is medium (MeWaQuLi) or good (GoWaQuLi); SaCoWa: if the same trough was used for both concentrate and water (YesSaCoWa) or not (NoSaCoWa); FeTyp: types of feed used, WeiFe: if feeds were weighed when feeding (YesWeiFe) or not (NoWeiFe); MeaDM: If farmers measured feed dry matter (YesMeaDM) or not (NoMeaDM); PMR: If partial mixed ration was used (YesPMR) or not (NoPMR); CoBeRo: if concentrates were offered before roughage (YesCoBeRo) of not (NoCoBeRo); MixDu: if concentrates and roughage were mixed during feeding time (YesMixDu) or not (NoMixDu); FeTim: times of feeding all types of feeds per day; CoTim: times of feeding concentrates per day, FeCle: times of cleaning roughage trough per week.

5.3.2 Diets

The number of SDFs using a given feed type and the average DM intake per cow for each feed are summarized in Table 5.2. A similar number of feed types was offered per region (4 to 5 types, P = 0.845, Table 5.1). Nineteen feed types were used by SDFs across regions. Eight SDFs reported using

127 sodium chloride in either the silage-making process or by spreading it over the feed at feeding time. Five SDFs reported occasionally using calcium supplements, two reported using mineral blocks, two reported using bypass fat, and one reported using sodium chloride, glucose, and whey. However, we either did not observe or could not measure the actual amounts of these feed additives; hence, they were not included in the dietary calculations in this study.

Table 5.2 Feed ingredients used across smallholder dairy farms (n) and mean dry matter intake (kg DM/cow/d) of each feed ingredient for lactating cows in four major dairying regions A

Feed intakes SL SH NL NH Overall P D (kgDM/cow/d) n B Mean (SD) C n Mean (SD) n Mean (SD) n Mean (SD) Mean  SEM Fresh Napier grass 6 2.6ab (2) 8 5.0a (1.6) 7 3.2ab (2.4) 5 1.2b (1.5) 0.009 3.0  0.8 Fresh tropical grass 2 0.7 (1.4) 0 - 1 0.5 (1.4) 2 0.5 (1.0) 0.513 0.4  0.1 Fresh corn with cob 0 - 2 1.1 (2.3) 2 0.7 (1.3) 1 0.2 (0.6) 0.430 0.5  0.2 Fresh corn leaves 0 - 0 - 0 - 1 0.3 (0.7) < 0.001 0.1  0.1 Corn silage 0 - 4 1.4bc (1.7) 7 3.2ab (2.7) 8 5.0a (1.0) < 0.001 2.4  1.1 Napier grass silage 0 - 0 - 1 0.2 (0.5) 0 - < 0.001 0.0  0.0 Fresh rice straw 1 0.6 (1.8) 0 - 0 - 0 - < 0.001 0.2  0.2 Dry rice straw 5 1.0 (0.9) 0 - 0 - 0 - < 0.001 0.3  0.3 Rice hay 1 0.3 (0.8) 0 - 0 - 0 - < 0.001 0.1  0.1 Partial mixed ration 0 - 0 - 0 - 8 2.0 (0.8) < 0.001 0.5  0.5 Concentrate pellets 8 6.1 (0.9) 8 6.4 (0.9) 8 6.6 (1.7) 8 7.3 (1.1) 0.112 6.6  0.2 Brewers grain 6 1.9a (1.5) 3 0.6b (0.8) 0 - 5 0.7ab (0.8) 0.007 0.8  0.4 Cassava residue 7 1.3 (0.8) 0 - 0 - 0 - < 0.001 0.3  0.3 Corn powder 0 - 5 0.8b (0.9) 8 1.6a (0.5) 0 - < 0.001 0.6  0.4 Whole soybean meal 0 - 1 0.1 (0.3) 2 0.1 (0.3) 0 - 0.272 0.1  0.0 Passion fruit pulp 0 - 1 0.3 (0.9) 0 - 0 - < 0.001 0.1  0.1 Sweet potato tuber 0 - 1 0.3 (1.0) 0 - 0 - < 0.001 0.1  0.1 Dried distillers grain 0 - 2 0.2 (0.4) 0 - 0 - < 0.001 0.1  0.1 Rice grain with husk 0 - 1 0.1 (0.3) 0 - 0 - < 0.001 0.0  0.0 A SL, Southern lowland; SH, Southern highland; NL, Northern lowland; NH, Northern highland. B n, number of farms using a given feed type out of 8 farms per region. C Mean of 8 farms; farms that did not use a feed were included as 0 kgDM/cow/d. D P-values were given for either One-way ANOVA tests; comparing means with superscript letters were given for post-hoc Tukey–Kramer test, P < 0.05. a-c Means with the different superscript letters within a row differ significantly from each other, P < 0.05.

128

The most popular roughage across regions was Napier grass (Table 5.2). Five to 8 SDFs per region offered Napier grass, and the average amount provided per cow was 3 kg DM/d. Next was corn silage which was popularly used in NH (8 SDFs), NL (7 SDFs) and SH (4 SDFs) but not at all in SL. Cows in NH were given the most corn silage (5 kg DM/d, P < 0.001) and the least Napier grass (1.5 kg DM/d, P = 0.009). Dry rice straw was used the most in SL (5 out of 8 SDFs) but not in any other regions. Fresh tropical grasses other than Napier were used by one to two SDFs in SL, NL, and NH but not at all in SH. Fresh corn with cob was used by one to two SDFs in SL, SH, and NH but not SL.

Concentrate pellets were the main concentrate source for cows in all SDFs in all regions (3.0 kg DM/cow/d, P = 0.112, Figure 5.2). When the mean amount of concentrate used across the region was divided by mean FCM production across regions (16.9 kg FCM/cow/d, Table 5.5), the ratio was 0.39 kg of concentrate pellets per kg of FCM. The next popularly used concentrate was brewers grain, which was used in SL (6 SDFs), SH (3 SDFs), and NH (5 SDFs) but not at all in NL. Corn powder and whole roasted soybean meal were used at a moderate to low extent only in SH and NL. The partial mixed ration was only used in NH and was used by all SDFs in this region at an amount averaging 2 kg DM/d. Other feeds such as passion fruit pulp, sweet potato tuber, dried distillers grain, and rice grain with husk were used by one to two SDFs, in SH only.

From the Diet data on the 19 feed types (Table 5.2), the FAMD analysis defined the first nine principal components (PCs), accounting for 79.9% of the total variance. HCPC, based on those first 9 PCs, defined nine optimum diet clusters (Figure 5.3 a). Similar to Feeding regime data, almost all diets in the same region clustered into the same group except for diets on the farms SH8, SL3, SL4, SL6, and NL7, which separated, mainly individually, from the main clusters. The largest clusters were Cluster 3 (6 NL and 3 NH SDFs coloured grey), Cluster 6 (8 NH and one NL SDFs in dark blue), and Cluster 9 (5 SL SDFs in dark red).

The directionality and variation of feed intake variables and the associations of these feed intake variables with the diet clusters are presented in the biplot in a 2-dimensional view of the first two principal components (Figure 5.3 b). Visually, the variables that varied most (longest arrows) and most meaningfully in the partitions of the clusters were cassava residue, brewers grain, and dry rice straw (close to Cluster 9); fresh Napier grass and sweet potato tuber (close to Clusters 1 and 3); and corn silage and partial mixed ration (Close to Cluster 6).

129

a) Cluster dendrogram

b) PCA – Biplot coloured by diet clusters Figure 5.3 Cluster dendrogram (a) depicting nine optimum diet clusters from C1 to C9; and biplot (b) a 2-dimensional view of the first two principal components Thirty-two diets (or observations): SL1 to SL8, SH1 to SH8, NL1 to NL8, NH1 to NH8 represent the diets for lactating cows on the farms from 1 to 8 in southern lowland, southern highland, northern lowland, northern highland regions, respectively. The diets clustered into the same cluster were coloured with the same colour. Nineteen feed intake variables (kg DM/d): NaGr, Fresh Napier grass; TrGr, Fresh tropical grass; CoCo, Fresh corn with cobs; CoLe, Fresh corn leaves; CoSi, corn silage; NaSi, Napier grass silage; RiStF, Fresh rice straw; RiStD, Dry rice straw; RiStH, Rice hay; PMR, Partial mixed ration; LaPel, concentrate pellets for lactating cows; BrGra, Brewer grain; CaRe, Cassava residue; CoPo, Corn powder; SoBe, Whole soybean meal; PaFr, Passion fruit pulp; PoTu, Sweet potato tuber; DiGr, Dried distillers grain; RiGra, Rice grain with husk.

130

Further characterization of each diet cluster by V-test showed that use of rice grain and sweet potato tuber was limited to C1 (one SDF only, in SH); dried distillers grain and fresh corn with cobs, to C2 (3 SDFs only, all in SH); fresh Napier grass, corn powder, and whole soybean meal, to C3 (mostly NL but some SH); fresh tropical grass, to C4 (2 SDFs only, SL and NL); passion fruit pulp, to C5 (one SDF only, in SH); partial mixed ration, corn silage, corn powder, to C6 (all 8 SDFs in NH and one in NL), fresh Napier grass, to C7 (One SDF only, in SL); fresh rice straw and cassava residue and rice hay, to C8 (One SDF only in SL); and dry rice straw, cassava residue, brewers grain, and corn silage, to C9 (5 SDF, all in SL). Figure 5.3 b shows these aspects in the first two PCs.

5.3.3 Nutrient composition of commonly used feeds

Eleven feed types (24 samples across regions) commonly used for dairy cows in each region were analysed for nutrient composition and are presented in Table 5.3. Another eight feed types, including Napier grass silage, fresh corn leaves, fresh rice straw, rice hay, corn powder, rice grain with husk, dried distillers grain and sweet potato tuber, were not analysed. The nutrient compositions of those feeds were obtained from the literature and are summarized in Appendix 3.

Nutrient compositions (DM on per cent as fed and other nutrients on DM basis) of fresh Napier grass, corn silage, fresh corn with cob, brewers grain, and whole soybean meal varied widely across regions (Table 5.3). For example, the DM, NEL, CP, ADF, fat, starch, and lignin concentrations of Napier grass, the starch concentration of fresh corn with cob, the lignin concentration of brewers grain, and the fat concentration of whole soybean meal varied widely across regions using that roughage. Differently from roughage, the main nutrient concentrations of concentrate pellets varied slightly across regions.

On comparing roughage types, as expected, concentrations of NEL (0.68 Mcal/kg), CP (7.2%), NFC (6.8%) of dry rice straw were found to be the lowest among the analysed roughage types, while concentrations of ADF (50.5%) and NDF (77.6%) of this feed were highest among the analysed roughage types. Also, as expected, DM concentrations of corn silage (23.8 to 26.3%), tropical grass (20.8 to 21.6%), fresh corn with cob (24.4 to 25.0%) were higher than the DM concentration of Napier grass. However, not as expected, concentrations of NEL, CP, ADF, NDF, fat, NFC, lignin, and minerals of corn silage samples in NL and NH were all within the ranges of corresponding nutrients of Napier grass. CP concentration of SH silage (10.4%) was also within the range of CP concentration in Napier grass. Only corn silage in SH had lower concentrations of ADF and NDF and higher concentrations of NEL, NFC and Starch than those of Napier grass. Concentrations of NEL, CP,

131

ADF, NDF, fat, NFC, lignin of tropical grass and fresh corn with cob were not much different from those of Napier grass.

Fibre and lignin concentrations of all roughage were high. Across roughage types, lignin concentration ranged from 5.1% in fresh corn with cob to 11.5% in NH fresh Napier grass. ADF and NDF concentrations ranged from 37.4% and 56.3, respectively, in SH corn silage to 50.5% and 77.6%, respectively, in dry rice straw.

132

Table 5.3 Nutrient concentration of feeds commonly used for dairy cows in regions (on dry matters basis, otherwise stated)

Price DM NEL TDN CP ADF NDF Fat NFC Starch Lignin Ca K Mg Na P S Cu Fe Mn Zn No Region-Feed name Mcal/ US/ton % AF % % % % % % % % % % % % % % ppm ppm ppm ppm kg 1 SL-Fresh Napier grass 38 19.8 1.08 58 15.8 41.9 65.2 2.2 10.0 0.2 5.3 0.27 3.19 0.17 0.02 0.32 0.14 7 234 56 55 2 SH-Fresh Napier grass 38 18.7 0.90 51 12.1 45.6 68.1 2.6 10.5 4.7 8.7 0.45 2.25 0.22 0.09 0.31 0.14 11 452 92 100 3 NL-Fresh Napier grass 38 19.1 0.77 49 10.1 47.9 72.5 2.7 7.9 1.0 10.1 0.51 1.74 0.36 0.09 0.34 0.18 8 605 91 33 4 NH-Fresh Napier grass 38 19.8 0.88 48 16.3 42.6 66.6 2.2 8.3 0.1 11.5 0.45 2.43 0.39 0.01 0.33 0.16 7 482 56 26 5 SH-Corn silage 104 23.8 1.28 61 10.4 37.4 56.3 3.0 25.1 17.8 6.4 0.22 1.23 0.18 0.90 0.23 0.13 7 735 84 27 6 NL-Corn silage 104 25.9 0.88 53 9.5 46.2 70.1 2.1 13.2 4.6 8.9 0.30 0.66 0.24 0.54 0.22 0.12 6 319 144 20 7 NH-Corn silage 104 26.3 0.90 52 13.0 49.6 69.0 2.0 10.8 0.7 9.5 0.41 1.73 0.50 0.01 0.23 0.14 6 317 33 25 8 SL-Fresh tropical grass 38 21.6 0.84 52 10.9 46.4 71.6 2.0 8.9 3.5 7.5 0.15 1.76 0.15 0.07 0.22 0.41 6 470 111 46 9 NL-NH-Fresh tropical grass 38 20.8 0.97 54 15.6 44.3 67.9 2.8 7.0 0.4 7.6 0.39 4.08 0.22 0.01 0.34 0.23 7 632 58 26 10 SH-Fresh corn with cob 75 25.0 1.19 61 12.4 38.2 61.4 2.3 18.7 15.1 5.1 0.27 1.60 0.13 0.04 0.22 0.13 6 688 114 37 11 NL-Fresh corn with cob 75 24.4 1.08 59 11.5 41.4 66.7 2.4 14.2 0.9 5.4 0.70 0.77 0.29 0.01 0.19 0.18 7 194 62 30 12 SL-Concentrate pellets 471 90.6 1.61 69 20.7 14.0 29.7 4.2 37.0 27.2 5.9 1.26 1.18 0.32 0.58 0.67 0.35 45 468 127 348 13 SH-Concentrate pellets 471 89.7 1.78 76 21.7 11.2 25.9 4.9 40.1 32.1 3.3 1.01 1.04 0.34 0.53 0.57 0.30 41 451 92 492 14 NL-Concentrate pellets 471 89.2 1.76 75 21.2 11.4 28.3 5.8 37.0 29.8 3.8 1.40 0.98 0.31 0.39 0.57 0.35 53 295 135 304 15 NH-Concentrate pellets 471 88.6 1.74 74 20.4 11.9 22.1 4.2 44.6 32.3 3.5 1.68 0.95 0.34 0.51 0.73 0.34 31 419 164 347 16 SL-Brewers grain 81 23.5 1.80 75 29.4 24.4 55.3 8.4 2.9 2.9 5.1 0.25 0.06 0.17 0.01 0.55 0.32 14 206 42 84 17 SH-Brewers grain 81 25.0 1.74 72 24.9 18.2 44.2 7.9 25.1 19.0 9.5 0.15 0.14 0.15 0.01 0.46 0.27 8 150 33 60 18 NH-Brewers grain 81 21.3 1.50 62 31.5 23.7 58.2 9.1 1.4 1.4 17.5 0.25 0.03 0.14 0.01 0.54 0.35 11 153 40 74 19 SH-Whole soybean meal 750 90.3 2.16 88 47.1 10.6 21.4 11.5 15.0 2.0 3.7 0.25 1.59 0.27 0.01 0.62 0.33 12 286 34 48 20 NL-Whole soybean meal 750 84.2 2.57 90 37.3 11.9 17.2 21.6 18.9 2.8 5.8 0.21 1.56 0.20 0.01 0.52 0.29 16 79 20 33 21 NH-Partial mixed ration 289 54.7 1.50 65 17.6 22.8 36.9 3.7 33.5 22.5 6.8 1.30 1.12 0.33 0.37 0.50 0.29 24 763 125 197 22 SL-Dry rice straw 173 84.2 0.68 53 7.2 50.5 77.6 1.4 6.8 0.5 5.4 0.37 1.96 0.16 0.15 0.13 0.18 4 532 287 35 23 SL-Cassava residue 54 19.8 1.63 72 2.5 25.6 31.2 0.1 63.2 48.6 4.8 0.50 0.23 0.10 0.02 0.02 0.02 1 190 39 12 24 SH-Passion fruit pulp 66 16.4 1.17 53 10.1 43.3 52.9 0.9 27.7 0.6 9.4 0.26 2.99 0.13 0.05 0.08 0.18 2 117 69 10 A Abbreviations: SL, southern lowland; SH, southern highland; NL, northern lowland; NH, northern highland areas; USD, United States Dollar; AF, as fed; DM, dry matter; NEL, net energy for lactation; TDN, total digestible nutrients; CP, crude protein; ADF, acid detergent fibre; NDF, neutral detergent fibre; NFC, nonfibre carbohydrate; Ca, calcium; P, phosphorus; K, potassium; Mg, magnesium; Na, sodium; S, sulphur; Cu, copper; Fe, iron; Mn, manganese; and Zn, zinc.

133

5.3.4 Cow intake and nutrient concentrations of the diets

Calculated DMI and dietary nutrient concentrations for average diets per region compared to nutrient concentration targets recommended by PCDairy are summarized in Table 5.4. Although DMI was highest (17.3 kgDM/d) in NH, similar in SH and NL and lowest in SL (14.4 kgDM/d) (P = 0.007), DMI calculated as per cent of cow BW was similar across regions (3.2  0.1% BW, P = 0.861). The dietary concentration of DM was highest in SH diets (39.2% as fed) and lowest in NH diets (32.3% as fed) (P = 0.002).

The dietary concentration of NEL was similar across regions, 1.4  0.02 MCal/kg (P = 0.176). However, this concentration was lower than the NEL concentration recommended by PCDairy for either observed milk production or a target production of 25 kg of FCM (1.50 MCal/kg and 1.59 MCal/kg, respectively). The dietary concentration of CP was highest in NH (17.5% DM) and SL (17.1% DM), followed by SH (16.4% DM), and lowest in NL (14.8% DM) (P = 0.003). The dietary CP concentrations in all regions were higher than the CP concentration (15.7%) recommended by PCDairy for actual MILK production and the target of 25 kg of FCM.

Mean concentrations of ADF (27.3  0.4% DM) and NDF (45.8  0.8% DM) in the diets were similar across regions (P > 0.192) and much higher than the target concentrations suggested by PCDairy, which were 21% DM and 28% DM for ADF and NDF, respectively. Dietary concentration of fat in all regions was higher than the recommended level of 3% DM. Dietary fat concentration was highest in NL (4.1% DM) and lowest in NH (3.4% DM) (P = 0.026). Mean concentrations ( SEM) of TDN (64.3  0.5% DM), NFC (27.4  0.9% DM) in the diets was similar across regions. Starch concentration was highest in SH (22.6% DM) and lowest in SL (16.7% DM) (P = 0.04), while lignin concentration was highest in NH (6.8% DM) and lowest in SL (5.5% DM) (P = 0.012).

Concentrations of all the measured minerals in all regions were higher than the recommended concentrations and lower than the maximum concentrations recommended by PCDairy. Concentrations of K (1.31  0.07% DM) and Na (0.30  0.01% DM) were similar across regions. Diets in SL and NH had the highest concentration of S (P < 0.001). Diets in SH had the highest concentrations of Fe and Zn but the lowest concentration of Ca and Ca:P ratio (P  0.003). Diets in NL had the highest concentrations of Cu but the lowest concentrations of P and Zn (P  0.003). Diets in NH had the highest Ca:P ratio, higest Ca, P, and Mg concentrations, but lowest Cu concentration (P < 0.001).

134

Table 5.4 Comparisons of average nutrient composition (DM basis; % unless otherwise noted) of the lactating cow diets between four dairy regions and between these regions with aspirational targets

Actual diets of regions B PCDairy recommendations D Nutrient A (Mean) P C Mean  SEM (Mean  SEM) (DM basis) SL SH NL NH Recomen_1 Recomen _2 Max Intake DMI, kg/cow/d 14.4b 16.4ab 16.2ab 17.3a 0.007 16.1  0.6 ------DMI, % BW 3.2 3.3 3.2 3.3 0.861 3.2  0.1 ------Concentration DM, % as fed 35.5ab 32.3b 36.6ab 39.2a 0.002 35.9  1.4 ------NEL, MCal/kg 1.40 1.44 1.36 1.38 0.176 1.40  0.02 1.50  0.03 1.59  0.02 -- TDN, % 64.7 63.7 65.4 63.4 0.481 64.3  0.5 65.4  2.0 69.9  1.8 -- CP, % 17.1a 16.4ab 14.8b 17.5a 0.003 16.5  0.6 13.8  1.5 15.7  0.5 -- ADF, % 27.4 26.2 27.5 28.2 0.507 27.3  0.4 > 21 > 21 -- NDF, % 47.4 43.9 46.9 44.9 0.192 45.8  0.8 > 28 > 28 -- Fat, % 3.6ab 3.9ab 4.1a 3.4b 0.026 3.8  0.2 > 3 > 3 -- NFC, % 24.9 29.3 27.9 27.4 0.163 27.4  0.9 ------Starch, % 16.7b 22.6a 20.3ab 16.8b 0.004 19.1  1.4 ------Lignin, % 5.8ab 5.5b 5.9ab 6.8a 0.012 6.0  0.3 ------Ca, % 0.71b 0.59c 0.78b 1.04a < 0.001 0.78  0.10 0.49  0.05 0.58  0.03 2.0 P, % 0.44b 0.41bc 0.40c 0.49a < 0.001 0.44  0.02 0.32  0.03 0.37  0.02 1.0 K, % 1.40 1.44 1.12 1.26 0.070 1.31  0.07 0.9 0.9 3.0 Mg, % 0.25b 0.25b 0.28b 0.37a < 0.001 0.29  0.03 0.20 0.20 0.5 Na, % 0.27 0.32 0.29 0.31 0.611 0.30  0.01 0.18 0.18 1.6 S, % 0.26a 0.20b 0.23ab 0.26a < 0.001 0.24  0.01 0.20 0.20 0.4 Cu, ppm 23ab 21bc 25a 18c < 0.001 22  1 10 10 100 Fe, ppm 346b 437a 341b 402ab 0.001 381  23 50 50 1000 Mn, ppm 101a 83b 108a 103a 0.003 99  6 40 40 1000 Zn, ppm 174b 232a 137c 181b < 0.001 181  19 40 40 500 Ca:P ratio 1.6b 1.5b 2.0a 2.1a < 0.001 1.8  0.2 ------A Abbreviations as in Table 5.3. B, C, a, b, c Other footnotes as in Table 5.2. D Recommended nutrient concentrations calculated internally by PCDairy for production of actual MILK (Reccomen_1), for production of a target 25 kg FCM (3.5% fat) per day (Reccomen_2), and the maximum concentrations of minerals (Max).

135

5.3.5 Efficiencies of the diets

Milk production, diet cost and diet efficiencies of regions are summarized in Table 5.5. NEL- allowable FCM was similar across regions (18.6 kg FCM/cow/d) while CP-allowable FCM was highest in NH (30.9 kg FCM/cow/d) and lowest in NL (23.9 kg FCM/cow/d) (P < 0.001). The ratios of CP-allowable FCM to NEL- allowable FCM were 1.4 in NL, 1.5 in SL and SH, and 1.6 in NH. Thus, the maximum predicted FCM in all regions was taken as NEL-allowable FCM.

Table 5.5 Diet intakes (kg/cow/d), diet cost, prediction of milk yield (kg/cow/d), and predicted methane emissions from the diets of cows in each region Region B, Mean Parameter A P C Mean  SEM SL SH NL NH

Predicted milk production NEL-allowable FCM, kg/cow/d 16.5 20.6 17.6 19.8 0.049 18.6  0.9 CP-allowable FCM, kg/cow/d 25.0b 27.6ab 23.9b 30.9a < 0.001 26.8  1.6 CP-allowable FCM: NEL-allowable FCM 1.5 1.5 1.4 1.6 0.034 1.5  0.1

Actual milk production FCM, kg/cow/d 14.2b 16.2b 16.9b 20.4a < 0.001 16.9  1.3 FCM, kg/kg DMI 0.99b 1.00b 1.04ab 1.19a 0.016 1.06  0.05

Diet costs Diet cost, USD/d 5.4c 6.4bc 6.5b 7.6a < 0.001 6.4  0.4 Roughage cost, USD/d 1.3c 2.0bc 2.3b 3.4a < 0.001 2.2  0.5 Concentrate cost, USD/d 4.2 4.4 4.2 4.0 0.765 4.2  0.1 Diet cost, USD/kg DMI 0.37b 0.39b 0.40ab 0.44a < 0.001 0.40  0.01 Diet cost, USD/kg FCM 0.39 0.39 0.39 0.37 0.846 0.39  0.01

Predicted methane emissions

CH4, % of gross energy 5.54 5.31 5.46 5.24 0.049 5.39  0.07

CH4, g/d 297 324 329 336 0.056 321  8

CH4, g/kg DMI 20.7 19.8 20.4 19.5 0.064 20.1  0.3

a ab ab b CH4, g/kg FCM 21.4 20.0 20.0 16.5 0.013 19.5  1.0 A All results were calculated per cow per day. Abbreviations: c, cow; d, day; DMI, dry matter intake; BW, body weight; MILK, raw milk yield; ECM, energy corrected milk; FCM, fat corrected milk; NEL, net energy for lactation; CP, crude protein; USD, United States dollar. B, C, a, b, c Other footnotes as in Table 5.2.

136

Means of actual FCM in SL, SH, NL were 2.3, 4.4, and 0.7 kg/cow/d lower than the means of predicted FCM for those regions, while the mean of actual FCM in NH was 0.6 kg/cow/d higher than predicted. The correlation coefficient (r) between predicted and actual FCM was 0.59 (P < 0.05).

The feed efficiencies (kg of ECM produced/kg of DMI) of the diets in SL and SH (0.99 and 1.00 kg of FCM/kg DMI, respectively) were both lower than that of diets in NH (1.19 kg of FCM/kg DMI) (P = 0.016). The mean of diet costs across regions was 6.4 USD/cow/d, of which the concentrate costs accounted for 52.6% (in NH) and 77.8% (in SL) of the total cost. Both diet cost calculated as USD/cow/d and diet cost calculated as USD/kg DMI were lowest in SL and highest in NH (P < 0.001). However, the diet cost for milk production, calculated as USD/kg FCM, was similar across regions (0.39 USD/kg FCM, P = 0.846).

Predicted methane emissions calculated as both per cent of dietary gross energy (5.39  0.07%) and as g/cow/d (321  8 g CH4/d) were similar across regions (P > 0.049). The environmental efficiency of the diets calculated per unit of DMI was also similar across regions (20.1  0.3 g CH4/kg DMI, P = 0.064). However, the environmental efficiency of the diets calculated per unit of FCM was poorest in SL (21.4 g CH4/kg FCM) and best in NH (16.5 g CH4/kg FCM) (P = 0.013).

5.4 Discussion

As hypothesized, feeding regimes and lactating cow diets were specific to regions, and SDFs across regions tended to apply the same feeding regimes and diets for cows. The quality of the main forage used across regions was low, and the diets across regions were unbalanced, showing excessiveness in ADF, NDF, CP, and key minerals but insufficiency in NEL, NFC, starch, and fat. Improving the quality of roughage and rebalancing the diets, particularly for net energy and fibre concentrations, should be promoted as key strategies for further research and practical assessment.

5.4.1 Feeding regimes

There were only three feeding regime clusters. The hypothesis was that SDFs in the same regions tended to apply similar feeding regimes. Through discussion with farmers, we found that SDFs in the same region tended to learn farming practices from their neighbours. On comparing feeding regime clusters, we found that Cluster 1, which comprised all the NL SDFs, employed the most “best practice” feeding regimes; they gave cows concentrates and roughage more times per day, cleaned feed troughs more often, supplied cows ad libitum good quality water, and mixed concentrate and roughage during feeding. In contrast, SDFs in feeding regime Cluster 2 (all SL SDFs) had the worst feeding regimes when they offered cow concentrate fewer times, used the same trough for both water

137 and concentrate and did not supply cows ad libitum good quality water. These results suggested that extension programs need to be specific to each region.

As expected, 29 out of 32 SDFs across regions did not weigh diet ingredients, and none of the SDFs measured dry feed matter. This indicates that farmers did not formulate diets based on cows’ requirements. Also, none of the SDFs mixed concentrates and roughage before feeding. This indicates that SDFs did not see the importance of mixing the feeds. Preparing total mixed rations based on cows’ requirements or at least mixing concentrate and roughage well when feeding can improve performance and decrease nutritional issues such as acidosis or laminitis (Pilachai et al., 2013; Humer et al., 2018). For example, a study by Pilachai et al. (Pilachai et al., 2013) in Thai SDFs showed that the prevalence of subclinical laminitis was associated with feeding concentrate and roughage separately. Thus, further studies or extension programmes should aim to change these feeding practices. If possible, formulating diets into a total mixed ration (TMR) would be the best feeding practice. In a study in Vietnam, Mai et al. (2011) showed that feeding cows total mixed ration improved dry matter intake, milk yield, and milk quality compared to traditional feeding methods usually practised by SDFs. However, the same study also reported that the high cost of mixing wagons made it uneconomical for the SDF farmers (Mai et al., 2011). Given this issue, at least farmers can do similar to what an SDF in SH and four SDFs in NL did: spread roughage into the feed trough first, then spread concentrate on the top of the roughage, and finally mix the concentrates with the roughage either by hand or by using a rake.

The limited supply of good quality water for cows in all SL and SH SDFs is a serious problem. These farmers used the same trough for both concentrate and water and only supplied water for cows after they had eaten all the concentrates offered, per feed. A study by Lam et al. (2010), also in SL, reported the same practice by SDF farmers, and so it appears that little has changed between then and the current study. This either reflects the conservatism of SDFs in these regions (SL and SH are among the regions with the most extended dairy farming history in Vietnam) or the limitations of the extension systems in these regions (Trach et al., 2007). Using the same trough for concentrate and water could limit the water supplied to the cows and promote fermentation, making the water less drinkable (Lam et al., 2010). Although different troughs were used for water and concentrate in the SH region, only two out of eight SDFs in this region supplied water ad libitum for the cows. This result is consistent with previous surveys, which reported that only 29% of SDFs in Hanoi (Suzuki et al., 2006) and 35% of SDFs in SL (Lam et al., 2010) supply fresh water ad libitum for cows. In the current study, when farmers were asked why they supplied limited water for cows, they answered that they wanted to ensure the high concentrations of milk fat and milk solid non-fat so that they can

138 get a good price for their milk per litre. The limited supply of water not only raised welfare concerns but it can also limit milk production. In tropical conditions, lactating dairy cows require 60 to 70 litres of water per day for maintenance, plus an extra 4 to 5 litres for each litre of milk production (Moran and Chamberlain, 2017). Insufficient water can also exacerbate the effects of high environmental heat load on the cows (Renaudeau et al., 2012). A study on the hot conditions of Pakistan showed that the provision of water ad libitum increased milk production by 1.5 L/cow/d (Moran and Doyle, 2015b).

In the current study, apart from NL farms, all other regions proffered cows concentrates twice per day and roughage two to three times per day. These results are in line with the results of previous studies (Loan et al., 2004; Lam et al., 2010). Again, this reflects the conservatism of the SDF farmers. Increasing feeding frequency, especially concentrate feeding frequency, is associated with increased feed intake, milk fat yield, and decreased severity of subacute ruminal acidosis in high producing cows (Moran, 2012; Macmillan et al., 2017; Humer et al., 2018). When cows are given food more frequently, they will eat more evenly, which helps prevent the rapid production of volatile fatty acid in the rumen caused by over-fermentation of starch (Krause and Oetzel, 2006; Abdela, 2016).

5.4.2 Feeds and diets

Diversity and quality of the feeds

Feed types used by SDFs were diverse (19 types). As reported by other authors, concentrate pellets, fresh Napier grass, dry rice straw, fresh tropical grass, brewers grain, the common roughages was the popular feeds used by SDFs (Loan et al., 2004; Chu et al., 2005; Cuong et al., 2006a; Lam et al., 2010; Phong and Thu, 2016). However, the ratio of 0.39 kg of concentrate pellets per kg of FCM in the current study was much lower than the 0.5 kg of concentrate pellets per kg of milk production reported by Cuong et al. (2006a). Besides those feeds reported by previous authors, the current study found the use of many other feeds across regions. Corn silage, corn powders, whole soybean meal, and dried distillers grain, which are the common feeds for commercial dairy farms globally (2001, 2016), were also used by Vietnamese SDFs. In addition, other industrialized feed (PMR), industrial by-products (cassava residue, passion fruit pulp), agriculture by-products (fresh corn leaves, fresh rice straw, rice hay), local feeds (rice grain with husk), silage (Napier grass silage), and some minerals and vitamin premixes were found to be used. The diversity of the feeds used by SDFs could provide opportunities for formulating least-cost or maximum-profit diets.

Compared with previous studies (Loan et al., 2004; Chu et al., 2005; Cuong et al., 2006a; Lam et al., 2010; Phong and Thu, 2016), the current study not only listed the feed types used by SDFs but reported the diets that SDFs in each region used for the cows and clustered the diets. Similar to feed

139 types, the diets for the cows were diverse and constituted nine diet clusters. The SDFs in the same regions tended to feed cows similar diets, just as they practised similar feeding regimes (Diet clusters 3, 6, 9), which again reflected Vietnamese SDF farmers’ habit of learning from neighbours. The diversity and localization of the diets implied that nutritionists or extensionists should formulate diets specific to each region. Also, the feed companies who sell concentrate pellets should balance them according to the background ingredients specific to each region rather than produce one pellet formulation to fit all regions. In addition, Diet Cluster 6 was applied by many SDFs across regions, which implies the potential for SDFs to learn from other regions. This could be the basis for the development of a more effective national extension programme.

Despite the diversity of the feed types found to be used across regions, the quality of the roughage was a major issue. As reported by other authors, Napier grass and dry rice straw are especially fibrous with high concentrations of ADF (41.9 to 47.9% in Napier grass, 50.5% in dry rice straw), NDF (62.5 to 75.2% in Napier grass, 77.6% in dry rice straw), and lignin (5.3 to 11.5% in Napier grass, 5.4% in dry rice straw) (National Institute of Animal Husbandry, 2000; Dung et al., 2007; Rusdy, 2016; Maaruf and Paputungan, 2017). However, unexpectedly, corn silage and fresh corn with cob did not show much higher nutrient concentrations of NEL, CP, NFC or Starch and did not show lower concentrations of ADF, NDF, and lignin than those of fresh Napier grass. In addition, the concentration of ADF (37.4 to 49.6%), NDF (56.3 to 70.1%), and lignin (6.4 to 9.5%) of corn silage across regions in the current study were higher than the normal means ( SD) of ADF (27.5  3.9%), NDF (44.5  4.9%), and lignin (4.0  1.3%) concentrations in mature corn silage (32 to 38% DM) and much higher than the means ( SD) of ADF (28.1  3.3%), NDF (44  5.3%), and lignin (4.3  1.0%) concentrations in mature corn silage (> 40% DM) presented by NRC (2001, 2016). The high concentrations of NEL, CP, NFC and relatively low concentrations of ADF, NDF, and lignin are often the reasons for the higher nutrient digestibility of corn silage compared to other roughage such as Napier grass, making it one of the most suitable and most used roughage types for dairy cows globally (Neal et al., 1935; NRC, 2001, 2016; Khaing et al., 2015). The high fibrous concentration of the roughage is also a common reason for decreased feed intake (NRC, 2001, 2016). Thus, these results suggest that improving the quality of the roughage by reducing fibrous concentrations of Napier grass, rice straw, corn silage, and corn with cobs is crucial for improving the quality of smallholder dairy cow diets in Vietnam. The concentrations of ADF, NDF, and lignin often depend on the genetics of the forage and the stage of harvest (NRC, 2001). The high fibrous concentration of the roughage in the current study suggests that SDFs might not have had suitable forage varieties or have harvested forage too late.

140

The current study highlighted the wide variations in nutrient concentrations of the same feeds across regions, suggesting the importance of having different feed nutrient libraries for different regions. Through this study, the nutritive values of 24 local feeds (11 feed types) were added to the feed library of PCDairy software for local nutritionists, extensionists and farmers to use. It serves as an important advance on what was previously available but could be further enhanced if more feed samples per region could be analysed and at various growth stages (e.g. season or harvesting age).

Imbalance and inefficiency of the diets

As predicted, the current study indicates that diets in all regions were insufficient in NEL concentration but excessive in CP, ADF, NDF, lignin and minerals. That SDF farmers did not weigh the feeds when feeding cows and the high fibre concentration of all used roughages as discussed previously could be reasons for this dietary imbalance. Excess protein and minerals usually are eliminated through faeces and urine (NRC, 2001). This causes both economic loss and environmental pollution (NRC, 2001). The solutions to solving the diet imbalance issues in Vietnamese SDFs could involve either decreasing CP and mineral concentrations in the diets, or increasing NEL concentrations. The latter appears to be more beneficial as it could both balance the diets and increase milk production. As indicated by the results, the CP concentrations in SL, SH, and NH, and mineral concentrations in all regions were even enough to produce 25 kg of FCM. Thus, if the NEL concentration can be increased to at least 1.59 MCal/kg DM, then milk production of 25 kg FCM/cow/d could be expected.

The solution to increasing NEL concentrations in the lactating cow diets might lie in increasing usage of starch and fat while decreasing usage of high fibre roughage such as Napier grass or rice straw. Also, improving starch and NFC but lowering the fibre concentration of corn silage is essential. Currently, the NFC concentrations in the diets (24.9 to 29.3% DM) are all lower than the range of 30 to 42% DM suggested by the Encyclopedia of Dairy Sciences (Lean, 2011a) and PCDairy (Robinson and Ahmadi, 2015). Starch concentrations in diets (16.7 to 20.3% DM) were all lower than the range of 22 to 26% DM suggested by PCDairy (Robinson and Ahmadi, 2015). Fat concentrations in diets (3.4 to 4.1% DM) were at the lower threshold of the range (3.0 to 6.0% DM) suggested by NRC (2001) and PCDairy (Robinson and Ahmadi, 2015). In contrast, the concentrations of ADF (26.2 to 28.2% DM), NDF (43.9 to 47.4% DM), and lignin (5.8 to 6.8% DM) in all regions were relatively high and higher than the range of 19 to 21% DM suggested for ADF, and the range of 27 to 32% DM suggested for NDF by Encyclopedia of Dairy Sciences (Lean, 2011a) and PCDairy (Robinson and Ahmadi, 2015). These indicate that there is still room for increasing starch and fat and decreasing fibre concentrations in the diets whilst still ensuring the cows' rumen health. Moreover, the risks when

141 feeding higher starch and fat could be further reduced by taking more care to mix forage with concentrates at each feeding.

The feed efficiencies (0.99 to 1.19 kg FCM/kg DMI) in the current study were slightly lower than those values (1.07 to 1.14 kg FCM/kg DMI) found in an experiment that was also conducted in the NH region of Vietnam (Hiep et al., 2016), but much lower than those values in the studies conducted in the USA (1.39 and 1.72 kg FCM/kg DMI) (Neylon and Kung, 2003; Schingoethe et al., 2004).

Methane emission per kg FCM in the current study (16.5 to 21.4 g/kg FCM) was slightly lower than that (22.5 to 27.4 g CH4/kg FCM) in an experiment by Hiep et al. (2014) which was conducted in

Vietnam, but is much higher than that (13.1 to 14.8 CH4/kg energy corrected milk) in an experiment by Kolling et al. (2018) in Brazil. Similarly, methane emission per kg of DMI in the current study

(19.5 to 20.7 g CH4/kg DMI) was slightly higher than the values ranging from 15.1 to 19.0 g CH4/kg DMI in an experiment by Hiep et al. (2016), which was also conducted on diets for Vietnamese SDF cows and higher than that found (15.3 to 19.7 g CH4/kg DMI) in an experiment by Kolling et al. (2018) in Brazil.

Although the diet costs, calculated as USD/kg DMI, in SL and SH regions (0.37 and 0.39 USD/kg DMI, respectively) were both significantly cheaper than the diet cost in the NH region (0.44 USD/kg DMI) (P < 0.001), the feed efficiency of the diet, calculated as kg FCM/kg DMI, in the sameSL and SH regions (0.99 and 1.00 kg FCM/kg DMI, respectively) was significantly lower than that in NH region (1.19 kg FCM/kg DMI) (P = 0.016). As a result, the diet costs calculated for each kg of FCM were similar between all regions (0.39 USD/kg FCM, P = 0.846). According to recent surveys conducted in the same regions and the same year as the current study, the mean price for the first- class milk in SL, SH, NL, and NH was 0.55 USD, 0.58 USD, 0.58 USD, and 0.53 USD, respectively. (Nga, 2017a; b). Therefore, diet costs could be considered relatively high as they accounted for approximately 70% of the milk price. Similarly, although methane emission calculated as g CH4/kg

DMI was similar across regions (20.1 g CH4/kg DMI, P = 0.064), the methane emission calculated as g CH4/kg FCM was highest in SL (21.4 g CH4/kg FCM) and lowest in NH (16.5 g CH4/kg FCM) (P = 0.013). This is because the feed efficiency of the diets in NH was higher than the feed efficiency of the diets in SL (1.10 vs 0.93 kg FCM/kg DMI, P = 0.014). This indicates that the more expensive diet in NH was of a higher quality, which enabled higher milk production and a lower estimated methane emission. Consequently, diet quality deserves further evaluation as a key driver of milk yield in SDF cows.

142

5.4.3 Limitations of the study

The greatest difficulty in the current study was how to measure accurately feed intake, especially roughage intake. Concentrates were often only offered to lactating cows; thus, they could be measured easily. However, farmers usually gave forage only to lactating cows and dry cows and at the same time. While lactating cows might consume more than dry cows, in the current study, intake of roughage was calculated by dividing the total amount of roughage supplied by the total number of lactating and dry cows. Another issue was that three farmers offered roughage to cows at night time when there was no observer. Thus, we could not directly weigh the amount that farmers gave the cows at night. We only asked farmers to bring roughly the amount they had offered the cows at night, and we weighed that amount. These factors might affect the accuracy of roughage intake measurements.

5.5 Conclusion

SDFs within a given region tended to practise similar feeding regimes and to feed cows similar diets. The most problematic region for each was SL. The feed types used by the SDFs were diverse, and the nutrient concentrations of the same feed varied widely across regions. Therefore, the formulation of the diets and the extension programs to improve feeding practice for lactating cows should be specific to each region. Among regions, SL might require the most support.

Feeding regimes, especially in the SL, SH, and NH regions, could be improved by supplying water ad libitum, increasing feeding frequency per day, and mixing feed before or during feeding.

Lactating cow diets in all regions were excessive in protein, NDF, ADF, lignin, and minerals but insufficient in NEL, NFC, and starch. These imbalances may have caused the estimated sub-optimal feed efficiency and excessive methane emission of the cows fed these diets. The diets should be balanced towards increased availability of starch and fat to replace high fibre roughage such as Napier grass or rice straw. Studies aimed at reducing the concentrations of NDF, ADF, and lignin concentrations in the most commonly used roughage by SDFs (Napier grass, corn silage, rice straw, and fresh corn with cob) are also required. Studies could include selecting more suitable forage genotypes and identifying the optimal stage of harvest.

143

Chapter 6 Characteristics of cowsheds in Vietnamese smallholder dairy farms and their associations with microclimate

Abstract

In smallholder dairy farms (SDFs), farmers often build cowsheds on whatever land is available, using local materials, and based on self-accumulated experience without considering reducing the risk of heat stress and improving welfare condition for the cows. This study aimed to characterise the heat stress abatement strategies and microclimate within SDF cowsheds from four typical dairy regions of Vietnam (southern lowland, southern highland, northern lowland, and northern highland); and identify the cow housing parameters most associated with the microclimate. This study was conducted on 32 SDFs (8 SDFs per regions) in Autumn 2017. Twelve housing management variables, illustrating cowshed design and heat stress abatement methods of each SDF, were collected. Six microclimate parameters collected within the cowshed were temperature (AT), humidity, wind speed (WS), heat load index (HLI), Temperature-humidity index (THI), and accumulated heat load units (AHLU) during a day (0600 h to 1800 h). Factor analysis and cluster analysis were applied to group SDF cowsheds into clusters where SDFs in a cluster had the same cowshed characteristics. Multivariable linear models were applied to define the parameters most likely to inform future research into heat stress mitigation on SDFs. Averaged from between 0800 h and 1800 h, the microclimate inside the cowsheds was considered hot (HLI > 79) in the highland and very hot (HLI > 86) in the lowland regions. Cows in the lowland regions accumulated high heat load (AHLU > 50) by 1800 h. Cowsheds of SDFs varied widely and grouped into seven, but no cluster was more effective than the others in reducing heat stress conditions within cowsheds. Using roof soakers and fans simultaneously decreased AT and HLI by 1.3°C and 3.2 units respectively at 1400 h compared to 1100 h. Each 100 m increase in altitude was associated with decreases of 0.4°C in AT, 1.3 unit in HLI, 0.8 units in THI (P < 0.001), but there was no association between altitude and WS within the cowsheds. Each metre increase in the eave height of the cowshed roof was associated with decreases of 0.87°C in AT, 3.31 units in HLI, 1.42 units in THI, and an increase of 0.14 m/s in WS (P < 0.05). Thus, the interventions that should be prioritised for future research into the amelioration of heat stress in SDF cows include using the roof soakers together with fans, eave roof height, and floor area per cow.

144

6.1 Introduction

Heat stress is an inherent difficulty associated with dairy farming in the tropics due to the hot and humid weather conditions that cows are likely to be exposed to in such regions (Lam et al., 2010; Moran, 2013). Such conditions reduce cow feed intake, milk production, reproduction and negatively impact welfare (Kadzere et al., 2002; Hansen, 2007; Polsky and von Keyserlingk, 2017). Parameters of the microclimate inside cowsheds that are often proposed as drivers of heat stress include mostly ambient temperature, humidity, wind speed, solar radiation, temperature-humidity index (THI, which combines temperature and humidity) and heat load index (HLI, which combines temperature, humidity, solar radiation, and wind speed) (Gaughan et al., 2008; Zimbleman et al., 2009; Herbut et al., 2018).

In the large scale dairies that predominate in developed countries, the risk of heat stress is managed by developing them in relatively cool regions (e.g. temperate or highland) using cowsheds designed to moderate THI or HLI within (Collier et al., 2006; Ambazamkandi et al., 2015; Fournel et al., 2017). Compared to developing countries, developed countries have a relative abundance of land, financial resources and ease of access to cowshed design standards that optimise cow welfare (Hayes, 2012; Flaba et al., 2014; Duteurtre et al., 2015; Moran and Chamberlain, 2017; Bewley et al., 2017). Increasingly, developed countries are also required by official regulations to meet specific criteria for cow welfare (Canadian National Farm Animal Care Council, 2009; Moran and Doyle, 2015; UK Red Tractor Assurance for Farms, 2017; British Columbia Society for the Prevention of Cruelty to Animals, 2018; UK Royal Society for the Prevention of Cruelty to Animals, 2018).

In contrast, cowsheds in smallholder dairy farms (SDFs: farms with < 20 lactating cows on average), which are the most popular type of dairy farm in tropical South-East Asian countries like Vietnam, vary significantly in style, size, design, construction material, and equipment (Moran, 2015a). Cowsheds on SDFs are often built on whatever available land there is, using locally available materials rather than those that might be more appropriate to minimise heat stress (Ashbaugh, 2010; Moran, 2015a; Phong and Thu, 2016). These farmers often design cowsheds based on personal experience or the accumulated experience of farmers they know, rather than in accordance with official regulations designed to optimise cow welfare (Chu et al., 2005; Phong and Thu, 2016). Currently, no such regulations exist in Vietnam.

Vietnam is a typical tropical country where dairy production is neither a strength nor a tradition (Trach et al., 2007). In the past, SDFs were mainly developed in highland regions of Vietnam to provide suitable cowshed microclimates for high yielding cows (Cai and Long, 2002). However, in recent decades, the development of SDFs has shifted toward the lowland regions where most of the 145 population resides, and hence where the demand for fresh milk is greatest. The much closer proximity of lowland regions to the market compared to highland regions means that fresh milk can also be more cost-effectively supplied to the consumer. However, since lowland regions are likely to be much hotter than highland areas, this shift needs to be matched with further research into strategies to manage the risk of heat stress for cows in SDFs. To the best of our knowledge, no published studies specific to Vietnamese SDFs are available to guide the targeting of research interventions on optimal cowshed design for the amelioration of heat stress. Thus, this study aimed to (1) classify and compare housing design relative to heat stress amelioration and the microclimate within cowsheds in typical highland compared to lowland regions of Vietnam; and (2) to define the housing parameters that are most associated with improved microclimate within the cowsheds.

6.2 Materials and methods 6.2.1 Farm visits and measurements of altitude, latitude, and microclimate data

Farm visits

This study was conducted from 24 August to 7 October 2017 on 32 SDFs which were randomly selected from four main dairy regions (8 SDFs per region) of Vietnam, including a southern lowland region (SL), a southern highland region (SH), a northern lowland region (NL), and a northern highland region (NH). The process of selecting SDFs and visiting the SDFs was described in Chapter 3. Each SDF was visited in an afternoon and the next morning. Specifically, the SL SDFs were visited during the period from 24 August to 1 September, the SH SDFs from 5 to 9 September and 3 to 7 October, the NL SDFs from 11 to 19 September, and the NH SDFs from 22 September to 1 October.

Altitude, latitude and microclimate data

The microclimate parameters inside the cowshed of each SDF were measured at 1400 h, 1600 h, 1800 h, 0600 h, 0800 h, 1000 h and 1100 h by holding a Kestrel 5400 Heat Stress Tracker (NIELSEN- KELLERMAN, USA) in a walkway as close as possible to the middle of the cowshed and at about 1.8 m above the floor (Figure 6.1). The measured microclimate parameters included: wind speed (WS, m/s), ambient temperature or dry bulb temperature (AT, °C), relative humidity (RH, %), black globe temperature (GT, °C), natural aspirated wet bulb temperature (Tnawb, °C), wet bulb globe temperature (Twbg, °C), dew point temperature (Tdp, °C), and wet bulb temperature (Twb, °C). Also, the Kestrel device was used to measure the altitude (m) of each SDF. The latitudes of the SDFs were simply recorded as north or south.

146

Figure 6.1 Measurement of microclimate data inside cowsheds

Based on AT, cows were predicted to be normal when AT < 20oC, at heat stress threshold when 20oC  AT < 27oC, and at mid-severe heat stress level (feed intake decreases and welfare is disturbed) when AT  27oC (Beede and Collier, 1986; West, 2003; Brouček et al., 2009).

Temperature-humidity index (THI, units) was calculated from AT (oC), Tdp (oC), and RH (%) using the equation of Yousef (1985):

THI = AT + (0.36  Tdp) + 41.2

Tdp = (237.3 x b)/(1.0-b)

b = [log(RH/100.0) + (17.27  AT)/(237.3 + AT)]/17.27

Based on THI, cows were predicted to be normal when THI < 68, at heat stress threshold when 68  THI < 72, at mild-moderate heat stress level when 72  THI < 80, and at moderate-severe heat stress level when THI  80 (Zimbleman et al., 2009).

The heat load index (HLI, units) was calculated from GT (oC), RH (%), WS (m/s), and base of the natural logarithm (e) using the equations of Gaughan et al. (2008):

(2.4 – WS) When BGT  25, HLI = 8.62 + 0.38  RH + 1.55  GT – 0.5  WS + e

When BGT < 25, HLI = 10.66 + 2.8  RH + 1.3  GT – WS

Based on HLI, the microclimate inside cowshed was categorized as cool when HLI < 70.0, moderate when 70  HLI < 77, hot when 77  HLI < 86, and very hot when HLI  86 (Gaughan et al., 2008).

Accumulated heat load units (AHLU, units), indicating the estimated amount of heat load accumulated by the cows, were also calculated using the equations suggested by Gaughan et al.

147

(2008). AHLU at a time point were calculated from AHLU at a previous time point, AHLU increment, and interval in hours between current and previous HLI measurements, using the equations:

AHLUCurrent = AHLUPrevious + AHLUIncrement*Interval

AHLU at 0600 h (the first time point of measurement) = AHLUIncrement at 0600 h

If an actual calculated AHLUCurrent was less than zero, it was set to zero, which indicates that the cow is in thermal balance.

AHLUIncrement was calculated from HLI at a time point, lower HLI threshold (HLI = 77) and upper HLI threshold (HLI = 86) as follows:

AHLUIncrement = HLICurrent – 77, if HLICurrent < 77

AHLUIncrement = 0, if 77  HLICurrent  86

AHLUIncrement = HLICurrent – 86, if HLICurrent > 86

Based on AHLU, the heat load that cows accumulated was categorized as low heat load when AHLU <10, moderate heat load when 10  AHLU < 25, high heat load when 25  AHLU < 50, and very high heat load when AHLU  50 (Gaughan et al., 2008).

6.2.2 Farm observation and barn measurements

The housing management dataset consisted of seven quantitative variables and five qualitative variables, which illustrated the design of the cowsheds, the facility used, and the heat stress abatement methods that farmers practised for the cows. All the variables of the housing management dataset were recorded pre afternoon milking time.

The seven measured quantitative variables were: 1) mat area (m2) per cow (abbreviated as MatCow), 2) floor area (m2) per cow (FloorCow), 3) roof height (m) at the highest point (RidgeHei), 4) roof height (m) at the lowest point (EaveHei), 5) per cent of shed sides open (SideOpen), 6) number of fans per cow (FanCow), and 7) frequency (times) of hosing cows and floors per day (HoseCoFlo). The dimensions of the cowshed, including length, width, highest point, and lowest point, and the dimensions of the open side areas of the cowshed were measured using a rolling tape. Floor area (m2) per cow was calculated as the total width × length of the cowshed (including stalls, alleys, and crossovers) divided by the number of cows present in the cowshed. The numbers of floor mats and fans used in each SDF were counted. Almost all SDFs used wall fans with fan diameters from 30 to 40 cm. The number of mats was counted, and the dimensions of each mat were measured to calculate the total mat area, then divided by the number of cows to achieve mat area per cow (m2/cow).

148

Percentage of cowshed sides which were open, an indication of potential ventilation in the cowshed, was estimated by the ratio of open shed side area over the total shed side area. The frequency of washing cows and the floor was obtained by both observation and asking farmers.

The five qualitative variables recorded were: 1) type of housing (Housing): housing was classified as tie-up housing (TieHousing) or loose housing (LooseHousing); 2) type of roof (RoofType) was classified as asbestos cement roof (AsbetosRoof) or sheet metal roof (MetalRoof); 3) roof ventilation (RoofVent): yes (YesRoofVent) if the roof had vent system or no, if not (NoRoofVent); 4) cool cows by sprinklers (Sprinkler): yes (YesSprinkler) if cows were cooled by sprinklers or no, if not (NoSprinkler); and 5) cool roofs by soakers (RoofCooler): yes (YesRoofCooler) if the cowshed has a soaker-cooling system above the roof or no, if not (NoRoofCooler). The roof cooling system is a soaker system fitted above the roof to cool the roof when it starts becoming hot, especially during noontime. Definitions of tie-up housing and loose housing were based on the description of Moran (2012). Tie-up housing (also called tie stalls) is the type of housing where the cows are tied up by a rope, whereas loose housing is where the cows are not tied up and can move freely around group pens within the cowshed. In loose housing, the lying area for the cows can be either shared open lounge or cubicles (also called free stalls). All of these qualitative management data were obtained by direct observations.

6.2.3 Data analysis

Statistical comparisons

All statistics were performed using the base and additional packages of R software (R Core Team, 2018). SDFs were the experimental unit in all analyses. Descriptive statistics for quantitative variables were calculated for each region using the R package ‘psych’ (Revelle, 2019). Before any statistical comparison, the normality of the quantitative variables was tested using both the Shapiro-Wilk test and histograms. The results are presented as means for normally distributed quantitative variables, medians for not-normally distributed quantitative variables, and frequency for categorical variables.

All variables were compared between regions. The choice of suitable tests for the comparisons of variables between regions was based on the guidelines of McDonald (2014). For variables that were found to be not-normally distributed, medians were compared by Kruskal-Wallis tests followed by Dunn post-hoc tests (P < 0.05), using R package ‘FSA’ (Ogle et al., 2019). For normally distributed variables, means were compared by One-way ANOVA tests followed by Tukey–Kramer tests (P < 0.05), using the R package ‘agricolae’ (Mendiburu, 2019). For categorical variables, frequencies of

149 each sublevel of variables were compared by Fisher’s exact tests followed by Bonferroni-corrected pairwise Fisher's exact tests (P < 0.05), using R package ‘rcompanion’ (Mangiafico, 2019).

Hierarchical clustering on principal components

Hierarchical clustering on principal components (HCPC) method was applied to partition SDFs into clusters where SDFs in the same cluster had more similarity in housing management than those SDFs in other clusters (Husson et al., 2010). The details of HCPC was described in Chapter 5. Briefly, factorial analysis of the mixed data method (FAMD) was applied first to transform the housing management dataset into non-correlated principal components (PCs). Then, some first PCs, which accounted for more than 70% of the total variance in the management dataset, were retained for hierarchical cluster analysis to identify an initial number of clusters (Kaiser, 1961; Kuivanen et al., 2016). Finally, the k-means clustering method was applied to identify an optimum number of clusters and assign SDFs to each cluster (Husson et al., 2010). The HCPC analysis results were visualised as the dendrograms. All the multivariate statistical analyses were performed using R package ‘FactoMineR’ (Husson et al., 2019), and the results of multivariate analyses were visualized using R package ‘factoextra’ (Kassambara and Mundt, 2019).

Characteristics of management clusters were further explored by V-tests statistics (Husson and Josse, 2010). V-tests compared the mean of each quantitative variable in each cluster with the mean of that variable in all clusters. For categorical variables, V-tests compared the percentage of each category of each qualitative variable in each cluster to the percentage of that category in the whole data set (Husson and Josse, 2010; Kuivanen et al., 2016). Through those comparisons, V-test statistics could point out the advantages and disadvantages of each management cluster, and thereby suggest the management clusters with the most advantages.

Although V-test statistics could point out the management clusters with the most advanced housing management characteristics, they did not prove if the most advanced clusters were more effective than the other clusters in improving shed microclimate. Thus, two-way ANOVA analysis was performed to compare AT, RH, WS, THI, and HLI between management clusters while accounting for the effects of altitude and latitude to assess if any management clusters were more effective than the others in improving the microclimate inside the cowsheds.

Multivariate linear regression

Multivariate linear regression was performed to determine the predictor variables significantly associated with AT, RH, WS, THI, and HLI inside the cowsheds. To eliminate multicollinearity among altitude, latitude, and all housing management variables (seven quantitative and five

150 qualitative variables), only the predictor variables with variance inflation factor (VIF) less than five were included in the initial multivariate models (Kock and Lynn, 2012). A manual backward elimination process was used to remove the variables one by one so that only the variables having P- values of regression coefficients  0.1 have remained in the final models. The final models were also evaluated by examining the standardized residuals and leverage to ensure model assumptions were met (Richards et al., 2014).

6.3 Results 6.3.1 Microclimate within the cowsheds

Mean altitudes and microclimate parameters within the cowsheds from 0600 h to 1800 h in four regions are presented in Table 6.1. Mean altitude was highest in SH (967 m), then NH (937 m), and was similarly low in SL and NL (47 m and 32 m, respectively) (P < 0.001). Mean RH (81.2%) and WS (0.40 m/s) were similar across regions. The means of AT, GT, Twbg, Tdp, Twb, Tnawb, THI, HLI, and AHLU in the highlands (SH and NH) were higher than those in the lowlands (SL and NL) (P < 0.001) (Table 6.1). However, these measurements were similar within the lowlands (SL and NL) and within the highlands (SH and NH).

Changes in the within-cowshed microclimate parameters during daylight hours are summarized in Figure 6.2. Wind speed remained steady throughout the day, and all regions showed a similar pattern. At all measurement times, mean WS was similar across regions (Figure 6.2c). The lowest mean WS was 0.12 m/s in NH SDFs at 0600 h, and the highest was 0.76 m/s in SH SDFs at 1400 h. Across regions, mean RH was consistently higher than 70% during the measurement hours, highest during the period from 0600 h to 0800 h (87 to 89%) and lowest during the period from 1100 h to 1600 h (Figure 6.2b). Mean RH was similar across regions at all measurement times except for RH at 1100 h when mean RH in NL SDFs (78.1%) was significantly higher than that in SH SDFs (70%) (P = 0.034).

The interior of the cowsheds in SL and NL were classified as very hot (HLI  86), and cows were predicted to be moderately to severely heat-stressed (THI  80) throughout the day (0600 h to 1800 h). In contrast, the interiors in NH and SH were classified as hot from 0800 h to 1800 h (77  HLI < 86), and cows in these regions were predicted to suffer mild-moderate heat stress (72  THI < 80) during this period. Mean AHLU in the highland regions increased steeply and similarly, exceeded the high heat load threshold (AHLU = 25) at approximately 1030 h, exceeded the very high heat load threshold (AHLU = 50) at approximately 1300 h, after which it continued to increase linearly at least until the last measurement at 1800 h (AHLU = 83.7 units in NL and 93.1 units in SL). In contrast to

151 the highland regions, mean AHLU in the lowlands increased only slightly, appearing to peak at approximately 1400 h, after which it plateaued at approximately 10 units until the last measurement (1800 h).

Table 6.1 Comparisons of altitude and microclimate parameters (averaging from between 6000 h and 1800 h) inside the cowsheds across four dairy regions Region B, Mean Parameter A P C Mean  SEM D SL SH NL NH Altitude, m 47c 967a 31c 937b < 0.001 496  264 AT, oC 29.5a 25.4b 29.7a 26b < 0.001 27.7  1.1 RH, % 81.8 80.5 82.0 80.6 0.887 81.2  0.4 WS, m/s 0.44 0.36 0.47 0.33 0.543 0.40  0.03 THI, units 82.5a 75.5b 82.9a 76.7b < 0.001 79.4  1.9 HLI, units 92.4a 80.0b 91.9a 81.2b < 0.001 86.4  3.3 AHLU, units 42.6a 6.8b 41.7a 5.1b < 0.001 24.1  10.5 GT, oC 30.0a 26.1b 29.9a 26.5b < 0.001 28.1  1.0 Twbg, oC 27.5a 23.5b 27.8a 24.0b < 0.001 25.7  1.1 Tdp, oC 26.4a 22.2b 26.7a 22.7b < 0.001 24.5  1.2 Twb, °C 27.2a 23.1b 27.4a 23.6b < 0.001 25.3  1.1 Tnawb, oC 26.6a 22.7b 27.0a 23.1b < 0.001 24.8  1.1 A Abbreviations: AT, temperature; RH, relative humidity; WS, wind speed; THI, Temperature-humidity index; HLI, heat load index; AHLU, accumulate heat load units; GT globe temperature; Twbg, wet bulb globe temperature; Tdp, dew point temperature; Twb, wet bulb temperature; Tnawb, natural aspirated wet-bulb temperature. B Regions: SL, Southern lowland; SH, Southern highland; NL, Northern lowland; NH, Northern highland. C P-values are given for One-way ANOVA tests comparing means. Superscript letters are given for post-hoc Tukey–Kramer test, P < 0.05. D SEM, standard error of mean. a-c Means or medians with the different superscript letters within a row differ significantly from each other, P < 0.05.

152

Figure 6.2 Changes in means of some main microclimate parameters in four regions during day time Abbreviations: AT, temperature (oC); RH, relative humidity (%); WS, wind speed (m/s); THI, Temperature- humidity index (units); HLI, heat load index (units); AHLU, accumulate heat load units (units). Error bars represent confidence intervals; Significant levels: *, P < 0.05; **, P < 0.01; ***, P < 0.001. AT: normal, AT < 20oC; heat stress threshold, 20oC  AT < 27oC; mid-severe heat stress, AT  27oC (Beede and Collier, 1986; West, 2003; Brouček et al., 2009). THI: normal, THI < 68; heat stress threshold, 68  THI < 72; mild-moderate heat stress, 72  THI < 80; moderate-severe heat stress, THI 80 (Zimbleman et al., 2009). HLI: cool, HLI < 70; moderate, 70  HLI < 77; hot, 77  HLI < 86; very hot, HLI  86 (Gaughan et al., 2008). AHLU: low, AHLU <10; moderate, 10  AHLU < 25; high, 25  AHLU < 50; very high heat load, AHLU  50 (Gaughan et al., 2008).

153

The changing patterns of AT (Figure 6.2 a), THI (Figure 6.2 d), HLI (Figure 6.2 e), and AHLU (Figure 6.2 f) were similar between the lowland regions and between the highland regions but were very different between the lowland and highland regions. Post-hoc Tukey–Kramer test showed that at 1400 h, mean AT in NL was similar to that in NH and SH, and mean HLI in NL was similar to that in SH; and at 0600 h, mean AT in SH (21.7oC) was lower than that in the NH (23.1oC) (P < 0.001). Apart from 0600 h and 1400 h, at all other measurement times, means of AT, THI, and HLI of SDFs in lowland regions (SL and NL) were similar to each other but significantly higher (P < 0.05) than those measurements of the SDFs in the highlands (NH and SH).

Smallholder dairy farms in NL stood out from those in other regions when showing that during the 1100 h to 1600 h period, AT, HLI, THI in NL tended to decrease and reached the lowest points at 1400 h (Figure 6.2 a, d, e). In this region, means of AT, HLI, and THI at 1400 h were 1.3°C, 3.2 units, and 2.5 units, respectively, lower than those at 1100 h. The reason for the decreases of AT, HLI, and THI of SDFs in NL during the hottest time of the day was the use of the soakers above the roof and the fan systems. We recorded that farmers in NL turned on the soakers and fan systems at around 1000 h and off at around 1600 h. Farmers reported that they turned on the cooling systems for the cows when they themselves felt hot.

6.3.2 Housing design

Summary of housing management variables

All cowsheds in NH and SH were loose housing whilst all in SL and 62% of cowsheds in NL were tie-up housing (P < 0.001) (Table 6.2). In all tie-up cowsheds, it was observed that cows were tied to the poles or bars adjacent to the feed and water troughs using a 1.2 to 2 m long rope threaded through a hole in the cows’ nasal septum. Farmers reported that cows were usually tied 24 hours per day for extended periods, and they were only moved to other places when they needed disease treatments or needed to be moved to dry herds. Floor areas per cow were largest in NH (12.5 m2/cow) and similar for SH (7.5 m2/cow), NL (6.7 m2/cow), and SL (5.2 m2/cow) (P < 0.001). Use of mats (mainly polyethylene foam mats) was similar across regions (0.9 m2/cow, P = 0.698). Sheet metal roofs were most popular in SL (all 8 SDFs), SH (7 out of 8 SDFs), and NH (5 out of 8 SDFs), whereas asbestos cement roofs were most popular in NL (7 out of 8 SDFs) (P < 0.001). Ridge roof heights were similar between regions (3.6 m, P = 0.118). However, eave roof height was highest in NL (3.4 m) and similar for NH (2.8 m), SL (2.6 m), and SH (2.3 m) (P = 0.008). All cowsheds in NL had roof vents, whereas only three in NH, one in SH, and none in SL out of 8 SDFs in each region had roof vents (< 0.001).

154

Table 6.2 Housing management parameters of smallholder dairy farms in four dairy regions

Region A, Median or n B Parameter P C Overall D SL SH NL NH Qualitative variables n (%) Housing: Loose 0b 8a 3b 8a < 0.001 19 (59) Housing: Tie-up 8a 0b 5a 0b 13 (41) Roof type: Asbestos cement 0b 1b 7a 3ab < 0.001 11 (34) Roof: Sheet metal 8a 7a 1b 5ab 21 (66) Cowshed has roof vents 0b 1b 8a 3ab < 0.001 12 (38) Cool cows with sprinklers 0 0 2 0 0.226 2 (6) Cool roof with soakers 0b 0b 7a 0b < 0.001 7 (22) Quantitative variables Mean  SE Floor area, m2/cow 5.2b 7.5b 6.7b 12.5a < 0.001 8.0  1.6 Mat area, m2/cow 0.6 0.0 1.4 1.4 0.698 0.9  0.3 Ridge roof height, m 3.3 3.3 4.1 3.6 0.118 3.6  0.2 Eave roof height, m 2.6b 2.3b 3.4a 2.8ab 0.008 2.8  0.2 Shed sides open, % 75 87 75 90 0.064 81.8  3.9 Fans per farm 1b 0b 8a 0b < 0.001 2.1  1.8 Fans per cow 0.1b 0.0b 0.8a 0.0b < 0.001 0.2  0.2 Hosing cows and floor, times/d 2 2 2 2 0.169 2  0 A Regions: SL, Southern lowland; SH, Southern highland; NL, Northern lowland; NH, Northern highland. B n, number of farms out of eight farms C P-values are given for either Kruskal-Wallis tests (superscript letters are given for post-hoc Wilcoxon rank sum test; P < 0.05) or Fisher’s exact tests (superscript letters are given for post-hoc Bonferroni-corrected pairwise Fisher's exact test; P < 0.05). D Overall mean (SEM) of medians or overall frequency (percentage) of all farms. a, b, c Medians or percentages with the different superscript letters within a row differ significantly from each other, P < 0.05. For cooling methods, all SDFs across regions used the hose to wash the cows and floors about twice per day, usually before milking time. While each SDF in NL had approximately 8 fans for cooling the cows, each SDF in SL had only about one fan. SDFs in NH and SH did not use fans at all (P < 0.001). Farmers in the highlands (SH and NH) reported that they did not use any cooling methods because they thought the weather there was already very cool. Comparing between the lowland regions, SDFs in NL put more effort into cooling the cows, shown by their supplying of approximately one fan per cow, cooling cows by sprinklers (two out of 8 SDFs), and especially cooling the roof by soakers fitted above the roof (7 out of 8 SDFs). In contrast, SDFs in SL used

155 neither sprinklers to cool the cows nor soakers to cool the roof, and fans were few (one fan for ten cows).

Factor analysis and clustering analysis

The FAMD analysis, using 12 housing management variables (Table 6.2), defined the first nine principal components (PCs), accounting for 79.9% of the total variance. HCPC, based on those first 9 PCs, defined seven optimum housing management clusters (Figure 6.3 a). SDFs in the same regions tended to group into the same clusters. Specifically, seven SDFs in SH and all SDFs in NH were quite similar to each other and came together into clusters C1 and C2. All SDFs in SL and one SDF in SH were in a single cluster (C3).

The directionality and the amount of variation in housing management variables, and the associations of these variables with the housing management clusters are presented in a 2-dimensional view of the first two principal components (Figure 6.3 b for all variables and Figure 6.3 c for sublevels of qualitative variables and housing management clusters). The qualitative variables that varied most (furthest from the original coordinates in Figure 6.3 b and c) and most meaningful in the partitions of the clusters were: “Roof type (RoTyp), asbestos cement or sheet metal”, “Cool cows with sprinklers (Sprinkler), yes or no”, “Cool roof with soakers (RoofCooler), yes or no”, “Housing, loose or tie-up”, and “Cowshed has roof vents (RoofVent), yes or no”. The quantitative variables that best characterized the partition were “fans per cow (FanCow)” and “eave roof height (EaveHei)”.

V-test results (Table 6.3) showed the main characteristics of each housing management cluster by comparing the mean of each quantitative variable in each cluster with the mean of that quantitative variable in the whole dataset and comparing the percentage of categories of each qualitative variable in each cluster to the percentage of that category in the whole the dataset. Cluster 1 SDFs (two NH and two SH) had more mat area per cow but lower ridge roof height than average. Cluster 2 SDFs (five SH and five NH) all had loose housing, 90% had sheet metal roof and no fans, and they had less mat area per cow but more percentage of sides open than average. In Cluster 3 SDFs (all eight SL and one SH), 87.5% were tie-up housing, all had sheet metal roof type, all lacked roof vents and had less floor area per cow than the average. In Cluster 7 SDFs (all 5 were NL), all had tie-up housing, all had asbestos roof type, all had roof vents, all cooled roofs with soakers, and all had higher eave roof height and more fans per cow than the average. Figure 6.3 a and c also show these aspects in the first two PCs.

156

a) HCPC – Cluster dendrogram

b) First two PC view of all variables

157

c) First two PC view of qualitative variables and observations Figure 6.3 Results of factor analysis (FAMD) and hierarchical clustering on principal components (HCPC) for housing management data Thirty-two housing conditions (or observations): SL1 to SL8, SH1 to SH8, NL1 to NL8, NH1 to NH8 represent the housing for lactating cows in the farms numbered from 1 to 8 in southern lowland, southern highland, northern lowland and northern highland areas, respectively. Twelve housing management variables: MatCow, m2 of mat per cow; FloorCow, m2 of floor per cow; RidgeHei, ridge roof height (m), EaveHei, eave roof height (m); SideOpen, per cent of shed sides open; FanCow, number of fans per cow; HoseCowFloor, times of hosing cow and floor per day; Housing, loose housing (LooseHousing) or tie-up housing (TieHousing); RoTyp, asbestos cement (AsbetosRoof) or sheet metal (MetalRoof); RoofVent, yes (YesRoofVent) if the roof has vent system or no, if not (NoRoofVent); Sprinkler, yes (YesSprinkler) if cows are cooled by sprinklers or no, if not (NoSprinkler); RoofCooler, yes (YesRoofCooler) if the roof is cooled by soakers or no, if not (NoRoofCooler).

158

Table 6.3 Most significant variables characterizing each housing management cluster Group Overall Cluster Most significant variables A V-test P B mean (SD) or % mean (SD) or % C1 Mat area, m2/cow 2.88 (0.82) 1.17 (1.09) 3.75 < 0.001 Ridge roof height, m 3.00 (0.25) 3.71 (0.74) -2.31 0.021 C2 Shed sides open, % 88.25 (8.54) 74.92 (20.12) 2.49 0.013 Mat area, m2/cow 0.57 (0.76) 1.17 (1.09) -2.08 0.038 Fans per cow 0.00 (0.00) 0.23 (0.34) -2.47 0.014 Housing = Loose 100 59.38 3.19 0.001 Roof type = Sheet metal 90.00 59.38 2.3 0.021 C3 Floor area, m2/cow 5.31 (1.01) 8.67 (4.24) -2.55 0.011 Housing = Tie-up 87.50 40.63 2.93 0.003 Roof type = Sheet metal 100 59.38 2.69 0.007 Cowshed has roof vents =No 100 62.5 2.51 0.012 Housing = Loose 12.5 59.38 -2.93 0.003 C4 Floor area, m2/cow 21.26 (0.00) 8.67 (4.24) 2.97 0.003 C5 Cool cows with sprinklers = Yes 100 6.25 3.09 0.002 Cool roof with soakers = Yes 100 21.88 2.03 0.042 C6 Hosing cows and floor 5.00 (0.00) 2.13 (0.6) 4.8 < 0.001 Ridge roof height, m 6.00 (0.00) 3.71 (0.74) 3.09 0.002 Fans per cow 1.00 (0.00) 0.23 (0.34) 2.25 0.024 C7 Fans per cow 0.79 (0.16) 0.23 (0.34) 3.96 < 0.001 Eave roof height, m 3.59 (0.31) 2.84 (0.61) 2.94 0.003 Cool roof with soakers = Yes 100 21.88 3.88 < 0.001 Roof type = Asbestos cement 100 34.38 3.05 0.002 Cowshed has roof vents =Yes 100 37.5 2.88 0.004 Housing = Tie-up 100 40.63 2.73 0.006 A Abbreviations as in Figure 6.3. B P-values were from V-tests, which compared the mean of each quantitative variable in each cluster with the mean of that variable in the whole dataset or compared the percentage of each category of each qualitative variable in each cluster with the percentage of that category in the whole dataset (Kuivanen et al., 2016).

As shown in Table 6.3, cowsheds in each cluster had different characteristics. When simply based on Table 6.3, cowsheds in Cluster 7 seem to be more advanced than the others because of having roof vents, having soakers to cool the roof, and having more fans and higher eave height. However, the results of two-way ANOVA, which compared means of WS, AT, RH, HLI or THI between housing management clusters in more than 3 SDFs (C1, C2, C3, and C7) while accounting for effects of the latitude and altitude, showed that no housing management clusters were more effective than the others in improving any microclimate parameter.

159

6.3.3 Multivariate models identifying factors associated with cow shed microclimate

The independent variables that were strongly correlated with other independent variables (VIF > 5) and therefore excluded from the initial models were: “housing, tie-up or loose”, “roof type”, “fans per cow”, and “cool roofs with soakers”. The independent variables that were included in the initial models but had no significant effect were “mat area per cow”, “frequency of hosing cows and floors”, “cool cows with sprinklers”, and “ridge roof height” (P > 0.1). A model was also fitted for RH, but no variable in that model reached significance; thus, that model was not presented.

Table 6.4 Multivariate models identifying the factors associated with the temperature (AT, °C), humidity (RH, %), wind speed (WS, m/s), heat load index (HLI) and temperature-humidity index (THI) inside the cowsheds A

AT WS HLI THI Variable Coef (SE) B P C Coef (SE) P Coef (SE) P Coef (SE) P Intercept 33.86 (1.56) < 0.001 0.02 (0.17) 0.916 107.1 (3.41) < 0.001 88.01 (1.81) < 0.001 Altitude, m -0.004 (0.001) < 0.001 -- -- -0.013 (0.001) < 0.001 -0.008 (0.001) < 0.001 Latitude: North Reference Reference Reference Latitude: South -1.43 (0.58) 0.019 -- -- -2.46 (1.08) 0.030 -1.57 (0.61) 0.016 Eave roof height, m -0.87 (0.41) 0.047 0.14 (0.04) 0.026 -3.31 (0.93) 0.001 -1.42 (0.52) 0.011 Floor area, m2/cow -0.12 (0.07) 0.094 ------Shed sides open, % ------0.05 (0.02) 0.052 -- -- R2, % 79 15 88 86 A In all models, the independent variables that were excluded due to VIF>5 were: “housing”, “roof type”, “fans per cow”, and “cool roof soakers”; B Coef (SE), Coefficient (Standard error). C The independent variables that were included in each model but had no significant effect (P > 0.1) were: “mat area”, “frequency of hosing cows and floors”, “cool cows with sprinklers”, “roof vents”, “ridge roof heigh”, and the variables with ‘--’ sign in the P column of each model.

The independent variables associated with AT, WS, HLI, and THI are presented in Table 6.4. Each 100 m increase in altitude was associated with decreases of 0.4°C in AT, 1.3 unit in HLI, and 0.8 unit in THI (P < 0.001). Cowsheds in the south were 1.41°C lower in AT (P = 0.019), 2.46 units lower in HLI (P = 0.030), and 1.57 units lower in THI (P = 0.016) compared to cowsheds in the north. Each metre increase in the eave height was associated with decreases of 0.78°C in AT (P = 0.047), 0.14 m/s in WS (P = 0.026), 3.31 units in HLI (P = 0.010), and 1.42 units in THI by (P = 0.011). Each m2 increase in floor area per cow tended to be associated with a decrease of 0.12°C in AT (P = 0.094).

160

Each 10% increase in cowshed sides open tended to be associated with a decrease of 0.5 unit in HLI (P = 0.052).

6.4 Discussion

As expected, cowshed microclimate measures of heat stress experienced by cows were dramatically higher in the lowland regions than in the highlands. However, opportunities were also identified for improvement in the highland regions. Heat stress abatement opportunities for future research were particularly indicated by some strategies employed by SDFs in the NL region.

6.4.1 Shed microclimate

The current study, to our knowledge, was the first to directly measure HLI and THI inside cowsheds across major contrasting SDF regions in Vietnam. Based on the guidelines for HLI (Gaughan et al., 2008) and THI (Zimbleman et al., 2009) to categorise the level of heat stress; the very hot microclimate (HLI  86, THI  80) in the cowsheds during the daytime from 0600 h to 1800 h in the lowlands (SL and NL) indicated that the cows in these lowlands need to be cooled from the early morning to late afternoon of a day. Although during the day HLI and THI in the cowsheds in the highlands always maintained about ten units lower than those in the lowlands, based on the guidelines for HLI (Gaughan et al., 2008) and THI (Zimbleman et al., 2009), the cowshed microclimate in the highlands was still considered moderate hot from 0700 h to 0800 h (70  HLI < 77), and hot from 0800 h to 1800 h (77  HLI <86, 72  THI < 80). Thus, cooling the cows in the highlands was also found to be necessary. Moreover, the risk of heat stress is likely to get even worse at other times of the year. The current study was performed in a relatively mild time of the year, autumn, whereas microclimate can be expected to be even more extreme in the summer (General Statistics Office Of Vietnam, 2017). For example, Lam et al. (2010) found that, in NL SDFs, THI measured during early summer (May-June) averaged 81 units in the morning and 85 in the afternoon, higher than the THI of 78.5 units at 0600 h and 83.2 units at 1800 h measured in the current study. These data indicate that heat stress abatement strategies need to be applied in highland as well as lowland SDFs. This contradicts the opinion of farmers in the SH and NH regions, who considered the highland region to be so cool as to not require heat stress abatement strategies.

The AHLU assesses heat load accumulation over time, and an AHLU higher than 50 predicts that cows will accumulate very high heat load (Gaughan et al., 2008). In the current study, AHLU of the cowsheds in the lowlands at 1800 h (93.1 units in SL and 83.7 units in NL) was considerably higher than the highest threshold (AHLU = 50) suggested by Gaughan et al. (2008). Therefore, these AHLU were extreme and indicated that cooling cows during the day time in the lowland regions was

161 inadequate. If cows cannot be sufficiently cooled during the daylight hours, they need to be cooled at night time to allow them to dissipate that daytime heat load to return them to their thermoneutral zone (AHLU = 0) as soon as possible (Meat and Livestock Australia, 2006).

The mean WS across regions (0.40 m/s) can be considered extremely low. Although few studies have defined the optimum wind speed in a cowshed, some extension websites have suggested targets of between 1 to 2 m/s (The Dairyland Initiative, 2020) or 2 to 3 m/s (Curt and Mcfarland, 2017). Increasing WS is important because WS is a key driver of convection and evaporation, which are the principal mechanisms for cooling cows in hot conditions (Blackshaw and Blackshaw, 1994; West, 2003). In addition, WS is a key component in the calculation of HLI (Gaughan et al., 2008). Low WS is also often associated with high AT and high RH (Mader et al., 2006; Dikmen and Hansen, 2009). Thus, the low WS could be a reason for the high RH, THI, and HLI in the current study. Furthermore, WS and airflow patterns also directly influence air quality parameters, including dust and concentrations of noxious gases such as ammonia, carbon dioxide, and methane (Fiedler et al., 2013). Therefore, further research into improving air movement in cowsheds in all regions is necessary. 6.4.2 Associations between housing management and cowshed microclimate

The current study showed that although housing management of Vietnamese SDFs varied widely to enable the definition of seven main clusters, the SDFs in the same regions were often in the same cluster. This is expected because SDF farmers from the same regions tend to learn cowshed design and construction from each other (Chu et al., 2005; Phong and Thu, 2016). Initially, the current study expected that some housing management clusters could be more effective than others in improving microclimate. For examples, SDFs in Cluster 7 appeared better than others since the sheds in this cluster had roof vents, soakers to cool the roof, higher eave height and more fans per cows. However, the Two-Way ANOVA analysis results showed that none of the housing management clusters was better than the others. This indicated that currently, no SDFs have cowsheds optimized for improving microclimate. Thus, at the current stage, the identified individual housing management variables best associated with microclimate should be relied on more than the housing management clusters to define future research directions for the abatement of heat stress in SDF cowsheds. However, in the long term, the identification of standard housing parameters optimized for SDF cowsheds in the tropics, similar to the standards currently applied in commercial dairy farms (Hayes, 2012; Flaba et al., 2014; Duteurtre et al., 2015; Moran and Chamberlain, 2017; Bewley et al., 2017), is very necessary.

Multivariate analysis identified that altitude, latitude, and eave height were the most important variables to focus on for the abatement of heat stress as they were all negatively associated with AT, 162

HLI, and THI. The identified decrease of 0.4°C in AT for every 100 m increase in altitude is consistent with the finding of Trewin (2014) that tropical temperatures decrease with altitude at the rate of approximately 0.6°C per 100 m. Since AT is the main component in the calculations of HLI and THI, each 100 m increase in altitude was also associated with decreases of 1.3 unit in HLI and 0.8 unit in THI in the current study. Consequently, high altitude regions should be preferred over low altitude regions to establish new SDFs where possible. However, high-altitude plateaus are few in Vietnam, and virtually all of the available land in the two provinces with the largest areas of such plateaus, namely Son La and Lam Dong, have already been selected for dairy development since the late 19th century (Cai and Long, 2002). The lower AT of the cowsheds in the south compared to the AT of those in the north was consistent with the historical weather data from the General Statistics Office of Vietnam (2017) (Table 2.1), which showed that during the studied period (mainly in September), the outdoor AT in SL (27.9oC) and NL (27.6oC) were similar, but the outdoor AT in SH (18.7oC) was much lower than that in NH (24.3oC).

Multivariate analysis also indicated that each metre increase in eave height was associated with decreases of 0.78°C in AT, 3.31 units in HLI, 1.42 units in THI, and was associated with an increase of 0.14 m/s in WS. These results are consistent with those of Hatem et al. (2004), who reported that in Egypt, increasing roof height of cowsheds from 5 to 8 m enhanced cowshed microclimate by increasing air velocities, which resulted in a decrease in maximum temperatures and an increase in milk production. Current results suggest that increasing roof height could be a potential intervention that would decrease AT within the cowshed and increase air movement through the shed. The median eave heights across regions were considerably low in the current study, ranging from 2.3 m in SH SDFs to 3.4 m in the NL SDFs. Although no studies have suggested optimum roof heights for SDFs, the suggested roof heights for large commercial dairy farms are recommended to be about 5 m for eave height and 9 m for ridge height to ensure sufficient ventilation and convenience for machinery (Moran and Chamberlain, 2017).

Apart from increasing eave roof height, the current study showed that the simultaneous use of roof soakers and fans was also associated with decreases in AT, THI, and HLI within the NL cowsheds during the hottest time of the day (1000 h to 1600 h). This could be an effective strategy for heat stress abatement within SDFs in all regions, especially SL. However, those cooling systems should be turned on earlier and run for a longer period than they currently are.

6.4.3 Tie-up, floor space and mat use

Optimizing the design of cowsheds is not only important for improving the microclimate but also crucial for ensuring the comfort of the cows (Canadian National Farm Animal Care Council, 2009; 163

Moran and Doyle, 2015; UK Red Tractor Assurance for Farms, 2017; British Columbia Society for the Prevention of Cruelty to Animals, 2018; UK Royal Society for the Prevention of Cruelty to Animals, 2018). The current study identified some cowshed parameters in Vietnamese SDFs that need to be improved to ensure the welfare of the cows. Firstly, the tie-up housing in all SDFs in NL and SL raised welfare concerns because the cows would not be comfortable when tied by a rope threaded through a hole in their nasal septum for extended periods. Compared with loose housing, tie-up housing is thought to compromise cow comfort by causing irritation and infection of the nose, causing knee and hock inflammations, reducing lying and resting time, and restricting self-grooming and social contact between cows (Veissier et al., 2008; Popescu et al., 2014; Moran and Chamberlain, 2017). Therefore, for Vietnamese SDFs, if possible, changing to loose housing is the best. However, if changing to loose housing is impossible, and the cows need to be tethered, halters should be used instead of nasal rope (Moran and Chamberlain, 2017).

Secondly, a small floor area also raised cow welfare concerns. Floor area per cow could be considered acceptable in NH (12.5 m2/cow), but in all other regions (5.2 to 7.5 m2/cow), it was too small to ensure normal cow activity. The lowest current recommended floor area per cow is 7.4 m2/cow, suggested in the United States since the 1980s (Bickert and Light, 1982). A larger area of around 8 to 11 m2/cow is needed to ensure cow comfort and animal welfare standards as recommended by global welfare regulations (NFACC, 2009; Moran and Doyle, 2015b; Red Tractor Assurance for Farms, 2017; RSPCA, 2018). In the present study, increased floor area per cow tended to be associated with decreased AT.

Thirdly, mat use across the SDFs was also determined to be inadequate, especially in SH. The cows rested mainly on bare concrete flooring. Guidelines for cow welfare developed by the New Zealand Ministry of Agriculture (2018) and the British Columbia Society for the Prevention of Cruelty to Animals (BC SPCA, 2018) suggest bare concrete is not suitable. Cows should be supplied with a sufficient and suitable resting surface for rest such as mats or bedding, or after standing on concrete surfaces for 12 hours per day for three consecutive days or more, cows should be given at least one day on a comfortable surface, where they can lie down and rest freely (National Animal Welfare Advisory Committee, 2018). According to this guideline, cows in NL and SL had inadequate lying conditions, which was made even worse by the cows mostly being tied rather than loose housed.

6.4.4 Limitations

The current study had some limitations. Firstly, the study was merely based on single-day data measurements of each SDF in autumn, whereas the microclimatic conditions change seasonally or even daily. Secondly, the microclimate data of all 32 SDFs should have been measured on the same 164 range of days, but the current study could not do that due to the lack of labour hours and the distance between the regions. Thirdly, while the microclimatic conditions within the cowsheds might be affected by variables such as cowshed orientation, the angle of the cowshed roofs, or roof colour (Angrecka and Herbut, 2016; Angrecka et al., 2017), these variables were not collected and analysed. Besides, the microclimate can vary at different positions within a given cowshed (Herbut, 2013). Thus microclimatic data should be measured at multiple points per cowshed. However, the current study measured microclimatic data at only one point per cowshed. Further studies need to take these limitations into account to improve the accuracy of the results.

6.5 Conclusion

During the daytime, wind speed inside the cowsheds across regions was very low, and microclimate inside the cowsheds across regions was hot (in highlands) to very hot (lowlands) for the cows. Thus, not only SDFs in the lowlands (SL and NL) but also SDFs in the highlands (SH and NH) need to consider heat stress abatement strategies for the cows.

Although the cowsheds of SDFs varied widely and clustered into seven groups, no group was more effective than the others in improving the microclimate inside the cowsheds. Choosing high-altitude regions to develop dairy farms, increasing the eave height of the cowshed and cooling the cowshed by using roof soakers and fans could be potential solutions for improving microclimates inside SDF cowsheds.

To improve welfare conditions for the cows, the tie-up housing in SL and NL should be minimized, the floor area for cows in SL, SH, and NH should be increased, and the cows across regions should be supplied bedding materials for resting.

165

Chapter 7 Multivariate analysis identifying the main factors associated with cow productivity and welfare in tropical smallholder dairy farms in Vietnam

Abstract

Dairy farming systems are multidimensional and complex, with many factors such as shed conditions, nutrition and genotype simultaneously influencing cow productivity and welfare, and some having stronger effects than others. This study aimed to rank potential drivers of cow productivity and welfare in the tropical smallholder dairy farms (SDFs), using Bayesian modelling. Forty-three variables were collected from 32 SDFs located in four geographically diverse dairy regions of Vietnam, with 8 SDFs per region. Twelve variables, including milk yield (MILK), percentages of milk fat (mFA), protein (mPR), dry matter (mDM), energy corrected milk yield (ECM), ECM per 100 kg of body weight (ECMbw), heart girth (HG), body weight (BW), body condition score (BCS, 1 = thin to 5 = obese), panting score (PS, 0 = normal to 4.5 = extremely heat stress), inseminations per conception (tAI), and milk electrical resistance (mRE, an indicator of udder health) of cows, were fitted as outcome variables in the models. Twenty-one other variables describing farm altitude, housing condition and diet for the cows, cow genotypes, and cow physiological stage were fitted as explanatory variables. Increased altitude was associated with increases in ECM and mRE and with decreases in PS and times in tAI (P < 0.05). Increases in ridge and eave heights and percentage of shed side open were associated with increases in ECM, mFA, and mDM (P < 0.05). Increased dry matter intake and dietary densities of dry matter and fat were associated with increased MILK, ECM, and ECMbw and decreased tAI (P < 0.05). Increased dietary densities of lignin and phosphorus were associated with increased PS. An increased genetic proportion of Brown Swiss in the herd was associated with increased MILK, ECM, and ECMbw, whereas an increased genetic proportion of Zebu was associated with decreased ECMbw, mPR, HG, BW and BCS (P < 0.05). To improve cow productivity and welfare in Vietnamese SDFs, the following interventions were identified for testing in future cause-effect experiments: increasing floor area per cow, roof heights, shed sides open, dry matter intake, dietary fat density, and the genetic proportion of Brown Swiss and decreasing dietary lignin concentration and the genetic proportion of Zebu in the herd.

166

7.1 Introduction

Smallholder dairy farms (SDFs) are the most popular type of dairy farming system in tropical developing countries in South East Asia (SEA), including Vietnam (Devendra, 2001; Moran, 2015c). For example, in Vietnam, approximately 28,695 SDFs were raising an average of 20 cows per farm or less and accounting for 80% of the national fresh milk production in 2016 (General Statistics Office of Vietnam, 2016; Nguyen et al., 2016b; Vinamilk, 2017). Currently, there are relatively limited studies on SDFs in SEA, and the majority of the available studies on SDFs are limited to describing basic characteristics such as cow numbers, herd structure, and farmer-reported milk yields (Moran and Brouwer, 2013). Common characteristics of SDFs across SEA countries are low milk production relative to both genetic potential and cows under similar environmental conditions in developed countries; poor welfare as shown by panting in response to heat stress; low fertility and high incidence of mastitis (Lam et al., 2010; Moran and Brouwer, 2013; Östensson et al., 2013; Moran and Morey, 2015). The few studies that have estimated average individual milk yields indicate it is 14 to 15 kg/cow/d for cows in some northern provinces (Ashbaugh, 2010; Vu et al., 2016) of Vietnam, 8.3 to 15.3 kg/cow/d in Indonesia (Moran and Doyle, 2015a) and 12.4 to 14.1 kg/cow/d in Thailand (Koonawootrittriron et al., 2009; Wongpom et al., 2017).

Finding the reasons for the low productivity and welfare of SDF cows is important but complicated because SDFs, like any agricultural system, are multidimensional and complex systems where many factors, including weather, shed conditions, nutrition and genotype, simultaneously influence cow productivity and welfare. It is reasonable to propose that the low productivity and welfare of the cows in SDFs in SEA prevail for the following reasons. Firstly, the hot and humid weather conditions in SEA are not conducive to raising high producing dairy cows. Secondly, the dairy cattle in SEA are commonly crossbreeds between internationally popular dairy breeds, mainly Holsteins (HOL), with local multipurpose Zebu breeds (ZEB). These crossbreeds might adapt well to the local conditions but be less productive than the pure dairy breeds such as HOL, Brown Swiss (BSW), or Jersey (JER) (Koonawootrittriron et al., 2009; Lam et al., 2010; Wongpom et al., 2017). Thirdly, the low quality of tropical roughage and by-product types commonly used by farmers, such as Napier grass (Pennisetum purpureum) or rice straw, and simple diet formulation such as roughage plus concentrate pellets added at a ratio of 1 kg per 2 kg of expected milk yield, does not supply enough or an appropriate balance of nutrients for the cows (Chu et al., 2005; Cuong et al., 2006a; Lam et al., 2010; Moran, 2012; Pilachai et al., 2013; Bernard, 2015; Hiep et al., 2016; Phong and Thu, 2016). Furthermore, the poor design of the cowsheds, including a lack of cow-cooling facilities, most likely 167 makes the cows uncomfortable and less productive (Chu et al., 2005; Moran and Brouwer, 2013; Phong and Thu, 2016).

While all the above-mentioned reasons can simultaneously serve to lower the productivity and welfare of cows, some might be more important than others. Identification of those that have the greatest impact on herd productivity and welfare is of practical importance as it can help to define the most needed short-term and long-term interventions. Mathematically, if all the input management and output data of the SDFs are available, the key drivers of cow productivity and welfare can be identified by building multivariate models which include only factors with significant effects. However, this type of study on SDFs is currently very limited because of the lack of accurate and complete dairy farming data. SDF farmers in SEA commonly do not record the production data properly. For example, a study in Thailand reported that 70% of all farms in the survey did not record production data (Yeamkong et al., 2010). Currently, the common strategies that farmers are applying to improve cow productivity appear to focus on genetic and nutritional aspects of herd management. Backcrossing current dairy herds with imported dairy sires of high genetic merit, mainly HOL, and directly importing HOL heifers, are common genetic strategies (Lam et al., 2010; Moran, 2015c). Buying high-quality commercial concentrated pellets rich in protein and energy, and offering them at least twice a day to complement the balance of nutrients in basal forage, is the common nutritional strategy (Lam et al., 2010). Choosing highland rather than lowland regions to develop dairy farms was also a common strategy in the past, but currently, this is limited as the demand for milk in the cities, all in hot lowland areas, is rapidly increasing, whilst the availability of affordable land in the highlands is rapidly decreasing. Dairying in the lowlands is also considered to be more attractive than in the highlands as fresh milk can be delivered to market more cheaply and more reliably in terms of quality assurance (Su and Binh, 2001; Cai and Long, 2002; Moran and Morey, 2015; Herawati et al., 2016).

In this study, we focused on the multidimensionality of Vietnamese SDFs. The aim was to combine the data from the different themes of previous chapters to build multivariate explanatory models to enable the identification and ranking of potential factors affecting the productivity and welfare of dairy cows in SDF. The factors include milk yield and quality, body weight and body condition score, heat stress level, fertility, and udder health of the cows. We hypothesized that current improvement strategies found to varying extents in SDF herds, such as increased altitude, infusion of HOL genetics, and increased dietary energy and protein concentrations, are all associated with the improvement of cow productivity and welfare.

168

7.2 Materials and methods 7.2.1 Study sites, time and data

The study was conducted from 24 August to 7 October 2017 on 32 SDFs selected randomly from four key dairy regions of Vietnam, including a southern lowland, a southern highland, a northern lowland, and a northern highland region, with eight farms per region. Each farm was visited over a 24-hour period to collect the necessary data. Microclimate data from 0600 h to 1800 h within the cowshed of each SDF was measured by a Kestrel 5400 Heat Stress Tracker (NIELSEN- KELLERMAN, USA). Mean ( SD) ambient temperature across farms was 27.7  2.3°C and ranged from 22.7 - 31.6oC; humidity, 81.2  4.6% and 70.2 to 91.0%; Wind speed, 0.40  0.03 m/s and 0.04 to 1.08 m/s; and temperature-humidity index (THI), 79.4  3.7 units and 71.9 to 85.6 units.

A total of 41 variables were used, as detailed in previous chapters (Table 7.1). In Table 7.1, to assist with the interpretation of the models and results, these variables were divided into three groups. Firstly, housing management; these were variables 1 to 7, from Chapter 6. Secondly, nutrition; these were variables 8 to 23, from Chapter 5. Thirdly, animal; these were variables 24 to 41, from Chapter 3. The animal variables were grouped into four subgroups, including genetic (variables 24 to 27, from Chapter 4), physiological stage (variables 28 and 29, from Chapter 3), and productivity and welfare (variables 30 to 41, obtained from Chapter 3). All cow-level variables were average per farm to obtain farm-level variables. Then, only farm-level variables were used to build the models.

Table 7.1 Summary of the studied variables A

No Variable, unit Explanation Mean SD Min Max Variable 1 Altitude, m Altitude 495 464 5 990 Predictor 2 FloorCow, m2/cow Floor area per cow 8.7 4.3 3.1 21.3 Predictor 3 MatCow, m2/cow Mat area per cow 1.2 1.1 0.0 4.0 Predictor 4 RidgeHei, m Ridge roof height 3.7 0.8 2.7 6.0 Predictor 5 EaveHei, m Eave roof height 2.8 0.6 2.1 4.2 Predictor 6 SideOpen, % Per cent of shed sides open 75 21 25 100 Predictor 7 FanCow, fans/cow Number of fans per cow 0.2 0.3 0.0 1.0 Predictor 8 DMIbw, % BW Dry matter intake per body weight 3.2 0.3 2.6 3.9 Predictor 9 DM, % Dietary dry matter concentration 35.9 4.0 27.5 44 Predictor 10 NEL, MCal/kg DM Dietary net energy for lactation 1.4 0.1 1.2 1.5 Predictor 11 CP, % DM Dietary crude protein 16.5 1.7 12.7 21.1 Predictor 12 ADF, % DM Dietary acid detergent fibre 27.3 2.7 22.1 35.2 Predictor 13 NDF, % DM Dietary neutral detergent fibre 45.8 3.7 37.4 56.6 Predictor 14 Fat, % DM Dietary fat 3.8 0.5 2.9 5.1 Predictor 15 Starch, % DM Dietary starch 19.1 4.1 10.4 30.4 Predictor 16 NFC, % DM Dietary nonfibre carbohydrate 27.4 4.0 16.6 37.6 Predictor 17 Lignin, % DM Dietary lignin 6.0 0.9 4.7 8.7 Predictor

169

18 Ca, % DM Dietary calcium 0.78 0.19 0.52 1.14 Predictor 19 K, % DM Dietary potassium 1.31 0.27 0.85 1.94 Predictor 20 Mg, % DM Dietary magnesium 0.29 0.06 0.22 0.42 Predictor 21 Na, % DM Dietary sodium 0.30 0.08 0.21 0.51 Predictor 22 P, % DM Dietary phosphorus 0.43 0.05 0.33 0.53 Predictor 23 S, % DM Dietary sulphur 0.24 0.03 0.18 0.32 Predictor 24 ZEB, % Average Zebu ancestry proportion 4.3 2.7 0.9 12.2 Predictor 25 HOL, % Average Holstein ancestry proportion 84.7 7.4 54.9 95.0 Predictor 26 BRW, % Average Brown Swiss ancestry proportion 4.9 1.6 1.7 7.9 Predictor 27 JER, % Average Jersey ancestry proportion 6.1 6.2 0.0 32.2 Predictor 28 DIM, days Days in milk 181 57 41 330 Predictor 29 Lactations Number of lactations 2.2 0.5 1.0 4.0 Predictor 30 HG, cm Heart girth 186 6 171 202 Outcome 31 BW, kg Body weight 498 49 392 623 Either 32 BCS, 1 to 5 Body condition score 2.8 0.2 2.4 3.4 Either 33 MILK, kg/cow/d Milk yield 16.8 3.6 9.2 23.7 Outcome 34 ECM, kg/cow/d Energy corrected milk 15.7 3.0 8.9 21.6 Outcome 35 ECMbw, kg/100 kg BW/d ECM per 100kg of body weight 3.2 0.5 2.0 4.3 Outcome 36 mFA, % Milk fat concentration 3.7 0.4 2.9 4.5 Outcome 37 mPR, % Milk protein concentration 3.3 0.3 2.7 4.2 Outcome 38 mDM, % Milk dry matter concentration 12.3 0.7 10.9 13.6 Outcome 39 mRE, units Milk electrical resistance 406 25 362 477 Outcome 40 PS, 0 to 4.5 Panting score 1.3 0.5 0.5 2.4 Either 41 tAI, times Inseminations per conception 2.3 1.2 1.0 6.0 Outcome A Body condition score: 1 = thin to 5 = obese; Panting score: 0 = normal to 4.5 = extreme heat-stressed. Variables 1 to 7 were from Chapter 6, 8 to 23 from Chapter 5, 24 to 27 from Chapter 4, and 28 to 41 from Chapter 1

7.2.2 Building the models

Models

Milk production, panting score (PS), number of artificial inseminations (tAI), and milk electrical resistance (nRE) are productivity and welfare indicators because they reflect the status of production, heat stress, fertility, and udder health of the cows, respectively. High productivity, low PS, low tAI, and high mRE are preferable because they indicate that a cow is productive, not suffering from heat stress, fertile, and has good udder health. Since the aim was to build explanatory models of cow productivity and welfare indices, these indicators were chosen as outcome variables (or dependent variables) and the variables that theoretically reflect the causal structure of the outcome variables were chosen as the predictor (independent) variables (Shmueli, 2010) (Table 7.1). The management, nutrition, genetic and physiological stage variables, all well-known as cow productivity and welfare

170 drivers, were chosen as candidate predictor variables (NRC, 2001; Food and Agriculture Organization, 2011; USDA, 2016). The matrix notation describing the models was:

y = X + e, where: y was the vector of dependent variables,  was the vector of fixed effects (independent 2 variables), and e was the vector of residual random effects [assumed e ~ N(0, Iσ e)], and X was the incidence matrices of the fixed effects.

Identification of initial independent variables

When building the models, firstly, multicollinearity was assessed using variance inflation factor (VIF) statistics. Predictor variables with VIF > 10 were excluded from the initial multivariate model (Kock and Lynn, 2012). Based on this procedure, 13 variables (Region, FanCow, NEL, CP, ADF, NDF, Starch, Ca, K, Mg, S, HOL, and HG) were excluded from all models. Table 7.2 summarises the included variables.

Table 7.2 Summary of independent variables with variance inflation factor less than ten that were included in the initial models A

Outcome variables (y) Predictor variables (or fixed effects ) with VIF < 10

MILK, ECM, ECMbw, Altitude + FloorCow + MatCow + RidgeHei + EaveHei + SideOpen + mFA, mPR, mDM, DMIbw + DM + Fat + NFC + Na + ZEB + BRW + JER + DIM + mRE, or tAI Lactations + BW + BCS + PS

HG, BW, or BCS Altitude + FloorCow + MatCow + RidgeHei + EaveHei + SideOpen + DMIbw + DM + Fat + NFC + Na + ZEB + BRW + JER + DIM + Lactations + Lignin + P + PS

PS Altitude + FloorCow + MatCow + RidgeHei + EaveHei + SideOpen +

DMIbw + DM + Fat + NFC + Na + ZEB + BRW + JER + DIM + Lactations + BW + BCS + Lignin + P

A Abbreviations, see Table 7.1 Since the excluded variables may still be correlated with explanatory and outcome variables included in the models, a Pearson correlation matrix was also determined (Shmueli, 2010) (Table 7.3). In Table 7.3, correlations were only considered significant if they passed a Bonferroni corrected P-value. Farms with high altitude were associated with fewer fans per cow. Farms with high NFC were

171 associated with low ADF, low NDF, and high Starch. Farms with high lignin concentration were associated with lower NEL and ADF. Farms with high P were associated with high CP and Ca. Farms with high JER were associated with low HOL. Magnesium was the only variable that was associated positively with MILK and ECM.

Table 7.3 Pearson correlations between variables included in the models (rows) and variables not included in the models due to variance inflation factor  10 (columns) A, B

Included Not included variables variables FanCow NEL CP ADF NDF Starch Ca K Mg S HOL Altitude -0.66* 0.20 0.31 -0.06 -0.38 0.15 0.17 0.17 0.34 -0.27 0.43 FloorCow -0.23 -0.41 0.07 0.43 0.10 -0.24 0.43 0.02 0.57 -0.03 0.31 MatCow 0.10 -0.31 -0.08 0.31 0.35 -0.2 0.02 -0.25 0.14 -0.05 0.18 RidgeHei 0.45 0.01 -0.27 -0.20 -0.24 0.27 0.28 -0.36 0.32 -0.09 0.03 EaveHei 0.59 -0.27 -0.38 0.08 0.04 0.06 0.33 -0.58 0.42 0.11 -0.14 SideOpen -0.03 -0.17 -0.31 0.12 -0.15 0.25 0.05 0.07 0.23 -0.42 0.18 DMIbw -0.02 0.10 -0.12 -0.15 -0.18 0.23 -0.01 -0.06 0.29 -0.15 -0.02 DM 0.06 0.32 0.12 -0.39 -0.47 0.09 0.80 -0.37 0.50 0.48 0.00 Fat 0.30 0.50 0.27 -0.47 -0.17 0.21 -0.25 -0.23 -0.31 0.09 -0.18 NFC 0.01 0.47 -0.36 -0.69* -0.87*** 0.90*** 0.18 -0.16 0.08 -0.18 0.05 Lignin -0.03 -0.65* -0.06 0.68* 0.47 -0.52 0.29 -0.24 0.61 0.07 0.06 Na -0.11 0.05 -0.14 0.08 -0.03 0.01 -0.01 -0.15 -0.03 -0.11 0.01 P -0.41 0.32 0.77*** -0.24 -0.31 -0.3 0.69* 0.15 0.39 0.64 -0.08 ZEB 0.09 0.16 0.18 -0.10 0.20 -0.19 -0.39 0.19 -0.41 0.16 -0.41 BRW -0.08 0.24 0.32 -0.19 -0.22 -0.03 0.23 -0.14 0.08 0.39 -0.48 JER 0.18 0.06 -0.08 -0.13 -0.03 0.07 0.04 -0.26 -0.15 0.48 -0.9*** DIM 0.19 -0.18 -0.07 0.08 0.09 -0.02 -0.03 0.06 -0.16 -0.17 0.20 Lactations -0.30 0.14 0.12 -0.07 -0.28 0.1 0.14 0.16 0.31 -0.11 0.21 BW 0.12 -0.31 -0.15 0.29 0.05 -0.09 0.38 -0.11 0.42 -0.18 0.50 BCS 0.60 -0.08 -0.24 0.00 0.09 0.01 0.12 -0.22 0.04 0.04 0.06 PS 0.15 -0.03 0.10 0.02 0.21 -0.32 0.15 -0.09 -0.05 0.57 -0.31 HG 0.12 -0.31 -0.14 0.3 0.07 -0.1 0.36 -0.09 0.4 -0.19 0.51 MILK -0.12 0.11 0.20 -0.10 -0.34 0.03 0.65 -0.26 0.70** 0.19 0.13 ECM -0.05 0.11 0.18 -0.08 -0.32 0.00 0.64 -0.30 0.67* 0.22 0.05 ECMbw 0.01 0.25 0.21 -0.24 -0.37 0.07 0.54 -0.34 0.57 0.37 -0.26 mFA 0.35 -0.30 -0.15 0.32 0.48 -0.35 -0.26 0.10 -0.40 0.04 -0.19 mPR 0.50 -0.52 -0.31 0.43 0.49 -0.25 0.01 -0.26 -0.03 -0.05 0.02 mDM 0.26 -0.22 -0.22 0.25 0.44 -0.24 -0.43 0.04 -0.44 -0.02 -0.33 mRE -0.23 0.11 -0.12 -0.14 -0.29 0.43 -0.13 0.03 0.08 -0.50 0.25 tAI 0.00 -0.12 -0.18 -0.02 0.09 0.08 -0.18 0.12 -0.43 0.05 -0.13 A Region was also not included in any models due to VIF > 10, but it is a categorical variable; thus, its correlations with other variables were unable to be tested. Significant levels were adjusted by the Bonferroni method. B Abbreviations, see Table 7.1

172

Model building and model selection

After the initial independent variables were identified for each model, the Bayesian model averaging (BMA) method was applied to select the final multi-multivariate linear regression models using R package ‘BMA’ (Raftery et al., 2019). BMA method was chosen because it accounts for the uncertainty of variables included in the linear regression model by averaging over the best models identified, based on posterior model probability (Raftery et al., 2019). The final models included only the explanatory variables, which are suggestively (P  0.1) or significantly (P  0.05) associated with the outcome variables. The final model was summarized and further evaluated by looking at standardized residuals and leverage to ensure model assumptions were met (Richards et al., 2014). Only the regression coefficients and corresponding P-values remaining in the final models are reported.

7.3 Results 7.3.1 Variables associated with milk production

The detailed models for each of the milk production outcome variables (Table 7.2) are presented in Table 7.4. The R2 was moderate for the mPR model (55%), high for mDM model (72%), and very high for other models (> 84% for the mFA, MILK, ECM and ECMbw models).

For housing management variables, increased per cent of shed sides open was associated with increases in the corrected milk yields (ECM and ECMbw), and milk components mFA and mDM but not mPR. Increased altitude was associated with increased ECM but decreased mPR. Increased floor area per cow was associated with increased MILK but decreased mFA and mPR. Increased ridge roof height was associated with increases in ECM and mFA. Increased eave height was associated with decreased mDM. Increased mat area per cow was associated with decreased MILK.

For nutritional variables, increases in dietary DMIbw, DM and fat were associated with increases in MILK, ECM and ECMbw. However, increased dietary DM was associated with decreases in mFA and mDM. Increased dietary Na was associated with decreased MILK (but not associated with ECM or ECMbw) and with increases in mFA and mDM. Increased dietary NFC was associated with decreases in mFA, ECM, and ECMbw.

For animal variables, higher BSW was associated with increases in MILK, ECM, and ECMbw. However, higher ZEB was associated with decreases in ECMbw and mPR. JER was not found to be associated with any milk yield or component variables. Increased BW was only associated with increased ECM. Increased DIM was associated with decreased MILK but with increased mFA.

173

Increased number of lactations was associated with increased ECMbw. Increased BCS was associated with decreased MILK but with increased mPR.

Increased PS was associated with increases in mFA, mPR, ECM, and ECMbw.

174

Table 7.4 Multivariate regression models identifying the variables affecting cow milk yield (MILK, kg/cow/d), milk fat (mFA, %), milk protein (mPR, %), milk dry matter (mDM, %) energy corrected milk (ECM, kg/cow/d), and energy corrected milk adjusted for body weight (ECMbw, kg/100kg BW/d) MILK mFA mPR mDM ECM ECMbw Variables A Coef (SE) B P C Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P Intercept -31.69 (5.09) < 0.001 4.37 (0.74) < 0.001 3.3 (0.55) < 0.001 12.42 (0.92) < 0.001 -35.96 (4.77) < 0.001 -4.74 (0.79) < 0.001 Altitude ------3E-4 (1E-4) 0.002 -- -- 0.002 (0.001) 0.006 -- -- FloorCow 0.29 (0.06) < 0.001 -0.03 (0.01) 0.002 -- -- -0.07 (0.02) 0.002 ------MatCow -0.41 (0.2) 0.051 ------RidgeHei -- -- 0.18 (0.05) 0.002 ------0.67 (0.34) 0.060 -- -- EaveHei ------0.31 (0.14) 0.031 ------SideOpen -- -- 0.009 (0.002) < 0.001 -- -- 0.013 (0.004) 0.009 0.040 (0.01) 0.007 0.008 (0.002) 0.004 DMIbw 5.08 (0.85) < 0.001 -0.37 (0.16) 0.029 ------4.73 (0.82) < 0.001 1.11 (0.15) < 0.001 DM 0.30 (0.05) < 0.001 -0.03 (0.01) 0.006 -- -- -0.10 (0.02) < 0.001 0.26 (0.06) < 0.001 0.05 (0.01) < 0.001 Fat 1.58 (0.39) < 0.001 ------1.05 (0.4) 0.016 0.32 (0.08) < 0.001 NFC -- -- -0.04 (0.01) 0.004 -0.03 (0.01) 0.014 -- -- -0.15 (0.06) 0.024 -0.03 (0.01) 0.029 Na -6.53 (2.73) 0.026 1.68 (0.5) 0.003 -- -- 2.76 (1.06) 0.016 ------ZEB ------0.07 (0.02) < 0.001 ------0.05 (0.02) 0.02 BRW 0.84 (0.13) < 0.001 ------0.85 (0.13) < 0.001 0.18 (0.03) < 0.001 DIM -0.011 (0.004) 0.007 0.002 (0.001) 0.018 ------Lactations ------0.27 (0.08) 0.002 BW 0.04 (0.01) < 0.001 ------0.03 (0.005) < 0.001 -- -- BCS -2.63 (1.09) 0.024 -- -- 0.40 (0.17) 0.027 ------PS -- -- 0.61 (0.1) < 0.001 -- -- 1.11 (0.21) < 0.001 1.34 (0.64) 0.049 0.28 (0.12) 0.024 R2, % 94 84 55 72 92 88 A See Table 7.1 for the abbreviations of all variables. R2, Coefficient of determination of the model. B Coef, Coefficient; SE, Standard error. C The independent variables included in each model but had no significant effect (P > 0.10) included JER and the variables with ‘--’ sign in P column.

175

7.3.2 Variables associated with cow conformation

The models identifying the variables associated with body conformation showed moderate R2 ranging from 51% to 58% (Table 7.5). The variables associated with HG were also those associated with BW, which was to be expected as BW was estimated from HG. Increases in mat area per cow, eave height, and the number of lactations were associated with increases in both HG and BW. In contrast, increases in ZEB and JER were associated with decreases in both HG and BW. Increases in eave height, dietary fat, BRW, and PS were associated with increased BCS. In contrast, increases in dietary P and JER were associated with decreased BCS.

Table 7.5 Multivariate regression models identifying the variables affecting cow heart girth (HG, cm), body weight (BW, kg), and body condition score (BCS) HG BW BCS Variables A Coef (SE) B P C Coef (SE) P Coef (SE) P Intercept 172.34 (5.82) < 0.001 396.11 (43.71) < 0.001 2.19 (0.4) < 0.001 MatCow 1.68 (0.8) 0.046 12.66 (6.02) 0.045 -- -- EaveHei 3.11 (1.41) 0.036 23.77 (10.56) 0.033 0.14 (0.05) 0.012 Fat ------0.10 (0.06) 0.089 P ------1.31 (0.7) 0.073 ZEB -0.71 (0.31) 0.033 -5.64 (2.36) 0.025 -- -- BRW ------0.04 (0.02) 0.057 JER -0.37 (0.15) 0.019 -2.63 (1.12) 0.026 -0.02 (0.01) 0.005 Lactations 3.5 (1.67) 0.046 26.85 (12.53) 0.042 -- -- PS ------0.22 (0.07) 0.004 R2, % 58 58 51 A See Table 7.1 for the abbreviations of all variables. B Coef, coefficient; SE, Standard error. C The independent variables included in each model but had no significant effect included Alti, FloorCow, RidgeHei, SideOpen, DMIbw, DM, NFC, Na, DIM, Lignin, and the variables with ‘--’ sign in P column.

7.3.3 Variables associated with the level of heat stress, reproduction, and udder health of the cows

The R2 was low for the mRE model (44%) but high for the PS model (75%) and tAI model (76%) (Table 7.6). A decreased tAI was associated with increases in altitude, DMIbw, dietary fat, and BW, and decreases in mat area per cow, NFC, and DIM. A decreased PS was associated with increases in

176 altitude, eave height, and floor area per cow, but with decreases in NFC, lignin, P, and BCS. An increase in mRE was associated with increases in altitude and DMIbw.

Table 7.6 Multivariate regression analysis identifying the variables affecting panting score (PS), times of artificial inseminations per conception (tAT, times), and milk electrical resistance (mRE, units)

PS tAI mRE Variables A Coef (SE) B P C Coef (SE) P Coef (SE) P Intercept -3.35 (1.18) 0.009 12.89 (2.88) < 0.001 304.48 (44.2) < 0.001 Altitude -0.0008 (0.0001) < 0.001 -0.001 (0.0003) 0.034 0.03 (0.01) < 0.001 FloorCow -0.04 (0.02) 0.050 ------MatCow -- -- 0.49 (0.13) < 0.001 -- -- EaveHei -0.43 (0.10) < 0.001 ------DMIbw -- -- -1.27 (0.55) 0.030 26.6 (13.7) 0.062 Fat -- -- -0.88 (0.23) < 0.001 -- -- NFC 0.03 (0.02) 0.063 0.09 (0.03) 0.013 -- -- Lignin 0.35 (0.09) < 0.001 ------P 4.53 (1.09) < 0.001 ------DIM -- -- 0.008 (0.002) 0.003 -- -- BW -- -- -0.015 (0.003) < 0.001 -- -- BCS 0.6 (0.25) 0.023 ------R2, % 75 76 44 A See Table 7.1 for the abbreviations of all variables. B Coef, coefficient; SE, Standard error. C The independent variables included in each model but had no significant effect included RidgeHei, SideOpen, DM, Na, ZEB, BRW, JER, Lactations, PS, and the variables with ‘--’ sign in P column. PS was not in the model for PS. Lignin and P were not in the models for tAI and mRE.

7.4 Discussion

This study aimed to build explanatory models to prioritise potential housing, nutrition, and animal variables affecting the productivity and welfare of dairy cows in SDFs. Except for the mRE model, with a relatively low R2 of 44%, the R2 of all other models ranged from moderate (51%) for BCS to very high (94%) for MILK, indicating a wide range of interventions worthy of future research. The most inclusive models were for ECM and ECMbw, and since these indicators have already been corrected for variation in mFA and mPR (ECM) and BW (ECMbw), the variables affecting them deserve the most immediate attention. The variables most associated with ECM and ECMbw were

177 altitude, ridge roof height, per cent of shed sides open, DMIbw, dietary DM concentration, dietary fat concentration, genetic proportions of BSW and ZEB, BW, and PS.

7.4.1 Housing variables

Housing management variables, of all the variable groupings, were identified as the worthiest of future intervention research. Each 100 m increase in altitude was associated with an increase of 0.2 kg in daily ECM per cow, 0.08 unit decrease in mean PS, 0.1 times decrease in mean tAI and 3 units increase in mean mRE. These results are consistent with a case study of SDFs in Indonesia which reported that the daily milk yield of the lowland farms (8.3 kg/cow/d) was much lower than that of the highland farms (13.5 kg/cow/d), although this was only a simple comparison between regions without accounting for the effects of confounding variables associated with each region (Moran and Doyle, 2015a).

Besides altitude, increased eave or ridge height of the cowshed roof was associated with increased ECM, mFA, mDM, HG, BW, and BCS, and decreased PS, all of which are desirable. Similarly, increased floor area per cow was associated with increased MILK and decreased PS and increased per cent of shed sides open was associated with an increased mFA and mDM, and decreased PS. These associations are logical as altitude, roof height, floor area per cow, or per cent of shed sides open are all major variables affecting cow comfort by improving airflow, which should enhance the opportunity for evaporative cooling (Renaudeau et al., 2012; Moran, 2012; Fournel et al., 2017). The current study was conducted under conditions likely to induce heat stress - in autumn when the THI ranged from 71.9 to 85.6 units, which is considered hot to very hot for the cows (Zimbleman et al., 2009). These results consolidate those in Chapter 6, which indicated that increased altitude, floor area per cow, roof heights, and per cent of shed sides open increased ventilation and decreased THI inside the cowshed. Many studies have found that decreased THI was associated with decreased heat stress, which then increased feed intake, milk production, and cow fertility (Preez et al., 1990; Ravagnolo et al., 2000; Bouraoui et al., 2002; Könyves et al., 2017). Thus, our results suggest that building farms in high altitude regions, increasing floor area per cow, and increasing eave and ridge roof heights could be effective strategies to increase milk production of Vietnamese SDF cows.

Mats over concrete flooring are supplied for cows in Vietnam SDFs to make them more comfortable when resting. However, in the current study, it has been found that increasing mat area per cow could be more problematic than beneficial as it was associated with a decrease of 0.41 kg/cow/d in MILK and an increase of 0.5 times in tAI. With respect to this issue, Ashbaugh (2010) reported that many Vietnamese SDF farmers did not provide bedding on top of the concrete floors for cows because of hygiene issues, which then caused mastitis. This might indicate that the current type of mats (mainly 178 polyethylene foam mats) used in SDFs might not be suitable. However, the negative association between mat area and MILK might indicate that mat area is an indicator of other variables that were not included in the model.

7.4.2 Nutritional variables

Increased DMIbw and dietary DM concentration were associated with increased MILK, ECM, ECMbw, mRE and with decreased tAI, which are all desirable. The cows were in heat-stressed conditions, and the feed intake of heat-stressed cows is commonly depressed as an adaptive response of cows to reduce metabolic heat production (Gaughan and Mader, 2009; Renaudeau et al., 2012). Consequently, a decrease in feed intake was expected, and the decrease in feed intake explains the decrease in milk production associated with cows in heat-stressed conditions (Knapp and Grummer, 1991; West, 2003; West et al., 2003). Although the mechanisms that mediate the effect of heat stress on the reduced milk yield can be multifactorial, at least half of the reduction in milk yield can be explained by the decrease in feed intake caused by heat stress (Tao et al., 2020). The feed intake can be improved by using more concentrates or supplying cows with higher-quality grasses or roughage (less fibre) instead of the current roughage such as Napier grass or rice straw that farmers commonly used for their cows (Chu et al., 2005; Lam et al., 2010; Phong and Thu, 2016). Reducing dietary fibre results in decreasing bulk density of the diet, thus encouraging intake (West, 2003; Renaudeau et al., 2012). Cummins (1992) reported that during heat stress, dry matter intake and milk yield of cows given diets containing 14% to 17% ADF (DM basis) were higher than those of cows offered diets containing 21% ADF.

Increased dietary fat concentrations were associated with increased MILK, ECM, ECMbw, and BCS and associated with decreased PS, which are all desirable. These results are in line with those of other researchers who reported that increasing dietary fat was an effective strategy to eliminate the effects of heat stress on the dairy cow during summer. Lactating cows commonly experience negative energy balance during heat stress (West, 2003; Shwartz et al., 2009). Thus, a strategy to feed cows during heat stress is to increase dietary nutrient density, especially energy density by supplementing their diets with fat or increasing concentrate to compensate for the decline in feed intake.

In contrast to fat, increased dietary NFC, supplied mainly by the concentrate component of the diet, was negatively associated with mFA, mPR, ECM, ECMbw and positively associated with PS and tAI. These results were hard to explain because NFC is also a main source of NEL and so is contrary to our hypothesis. However, consistent with our results are those of Drackley et al. (2003), who conducted a study to compare productivity and heat stress level of HOL cows offered either a control diet (fat: 2.63% DM, NFC: 40.76% DM), high-fat diet (fat: 6.04% DM, NFC: 38.10% DM), or high 179

NFC diet (fat: 2.70% DM, NFC: 46.26% DM) during summer in the midwest of the United States. This study showed that mFA, fat corrected milk yield (3.5% fat), and the efficiency of milk production of the cows offered the high-fat diet were higher than those of cows offered the high NFC diet (Drackley et al., 2003). Respiration rate and rectal temperature were also lower in cows provided with high-fat diets than in cows provided with high NFC diets (Drackley et al., 2003).

Higher dietary lignin was associated with higher PS (more heat-stressed cows), which was to be expected. Diets rich in lignin are also likely to be rich in ADF (Adesogan et al., 2019). In the current study, dietary lignin was also and positively correlated with dietary ADF (r = 0.68, P < 0.05) and negatively associated with dietary NEL (r = -0.56, P < 0.05). Generally, the heat increment from the metabolic utilization of dietary fibre is considered higher than that from the metabolic utilization of starch or fat because the fermentation of fibre generates more heat and is less efficient (Renaudeau et al., 2012). Thus, under heat stress conditions, high fibre diets make cows more heat-stressed. It has also been suggested that feeding cows a diet containing high-quality NDF (low lignin concentration and high digestibility) instead of a diet with low-quality NDF, can reduce the negative effects of heat stress on their productivity, body temperature, and feeding behaviour (Arieli et al., 2004; Soriani et al., 2013). Interventions that specifically reduce the lignin but not necessarily the NDF concentration in SDF cows’ diets should be targeted for future research.

Although dietary Na concentration was not associated with either ECM or ECMbw, each per cent (DM basis) increase in dietary Na concentration was associated with a decrease of up to 6 kg of MILK per cow per day. This is inconsistent with the results of Schneider et al. (1986), who reported that increased dietary Na was associated with increased feed intake and milk yield of cows in hot weather. In Vietnamese SDFs, there could be two possible explanations for the negative associations between dietary Na and milk yields. Firstly, cows in Vietnam may not be supplied with enough drinking water during hot weather. Water is crucial for milk excretion, and the need for water increases when dietary Na is increased, especially during hot weather (West, 2003; Meyer et al., 2004; Appuhamy et al., 2016). Two studies on SDFs in Vietnam showed that 51% of SDFs provided less than 30 litres of water for a cow per day, while only around 29 to 35% provided fresh water ad libitum for the cows (Suzuki et al., 2006; Lam et al., 2010). Our results in Chapter 6 also showed quite similar results. A second explanation for the negative associations between dietary Na and milk yields is that during the current study, it was observed that, in the hope of increasing feed intake, Vietnamese SDFs farmers commonly supplemented Na in the form of NaCl rather than NaHCO3, whereas a study by

Coppock et al. (1982) showed that NaHCO3 is likely to be more effective than NaCl in increasing cows’ feed intake and milk yield during heat stress. Similarly, Gaughan & Mader (2009) found that

180 during hot conditions, adding NaCl to the diet can decrease the dry matter intake of cattle. Therefore, the dietary Na finding suggests the importance of research that tests the effect of ad libitum versus restricted provision of drinking water on cows in Vietnamese SDFs.

Each per cent increase in dietary P concentration was associated with an increase of up to 4.3 units in PS. To our knowledge, evidence of an association between dietary P and heat stress has not previously been reported. This association might reflect the effects of dietary CP on PS. Crude protein was not included in the PS model due to its variance inflation factor  10, but the correlation coefficient between P and CP was high. The CP concentration of the diet (16.5% DM) was also in excess relative to energy concentration (1.4 MCal/kg DM) when compared with the concentrations suggested by NRC (2001). When cows are offered diets with excess protein, the excess protein leads to an increase in metabolic heat generated during the excretion of excess nitrogen as urea (Huber et al., 1994; Dunshea et al., 2013), which then might cause an increase in PS.

7.4.3 Animal variables

A higher genetic proportion of BSW in the herds was associated with increases in MILK, ECM, ECMbw, and BCS, which are all desirable. These results were consistent with the results of other researchers. A study by El-Tarabany et al. (2017) reported that pure BSW and F1HOL_BSW cows are more adaptable to the subtropical conditions in Egypt than pure HOL cows, as shown by a slower rate of reduction in milk yield when THI changed from a low to a high level. That study also showed that F1HOL_BSW had a lower incidence of clinical mastitis, metritis and feet problems than pure HOL; and that pure BSW cows had a lower incidence of retained placenta and metritis than pure HOL (El-Tarabany et al., 2017).

A higher genetic proportion of JER was not associated with milk productivity traits but was associated with decreases in HG or BW, and BCS. This was expected because mature BW of JER (408 to 454 kg) is often smaller than mature BW of HOL (590 - 680 kg) and BSW (509 – 537 kg) (Capper and Cady, 2012; Piccand et al., 2013).

A higher genetic proportion of ZEB was associated with decreases in mPR, ECMbw, HG, and BW. The negative association between the genetic proportion of ZEB with ECMbw was consistent with the findings of a study by Branton et al. (1961), who reported that in a hot and humid climate, all crossbreeds of HOL with Red Sindhi (a ZEB breed) and crossbreeds of BSW with Red Sindhi yielded less milk and fat compared with their pure HOL or pure BSW mates during both winter and summer. Similarly, the negative associations between the genetic proportion of ZEB with HG and BW are understandable because the mature BW of Vietnamese ZEB cattle were very small (249 - 281 kg)

181

(Duy et al., 2013). A study by Branton et al. (1961) also showed that the mean BW of F1HOL_RSI (1/2 HOL + 1/2 Red Sindhi) and B1HOL_RSI (3/4 HOL + 1/4 Red Sindhi) crossbred cows were smaller than the BW of pure HOL at various ages from birth to 90 days postpartum.

Thus, the associations between the genetic proportion of breeds and the productivity and welfare indicators in the current study suggest that increasing the genetic proportion of BSW rather than increasing the genetic proportion of ZEB could be a more appropriate breeding strategy for Vietnamese SDFs. Unfortunately, HOL were excluded from the models due to VIF > 10; thus, it was impossible to infer the effects of HOL genetic proportions directly.

The strong positive association between BCS and PS is consistent with a study by Cincović et al. (2011), who reported that high BCS cows (BCS > 4, 5-point scale) had less ability to acclimatise to heat stress than normal and thin cows, indicated by lower milk yield and quality, and higher rectal temperature and respiration rate compared to other groups. Brown-Brandl et al. (2006) also reported that finished feedlot Angus and Charolais cattle with BCS  8 (9-point scale) had an average 6.8% higher respiration rate than the lean animals with BCS < 8.

7.4.4 Some limitations

A principal aim of the current study was to identify potential causes of the low productivity and welfare of Vietnamese SDF cows. However, it should be noted that similar to many epidemiological, public health, or social studies, it is by nature an observational study conducted in an inherently noisy environment rather than a randomized experimental study where the confounders are controlled (Glass et al., 2013). Thus, the associations found in our multivariate regression models may not be causal. An observed association can be due to the effects of one or more of the following: chance (random error), bias (systematic error), confounding, reverse causality or true causality (Lucas and Mcmichael, 2005; Barratt et al., 2009). To minimize these negativities, we applied two strategies. Firstly, we collected as many biologically and theoretically sound independent variables as we could to be included in each model (Glass et al., 2013). Secondly, to judge whether an observed statistical association represents a cause-effect relationship between explanatory and response variables (Ward, 2009; Geneletti et al., 2011), we applied as many as possible of Hill’s (1965) criteria that are commonly applied to epidemiological studies. These criteria include: the strength of association, consistency, specificity, temporality, biologic gradient, plausibility, analogy, coherence, and experiment (Hill, 1965). According to this method, briefly, explanations from the literature were searched and considered for each observed association to infer the likelihood of that association being a true cause-effect relationship. Then, intervention strategies were suggested based only on the most

182 likely cause-and-effect relationships between explanatory variable and outcome variables. In the current study, we did not consider the associations between dietary NFC with milk production, between dietary P with PS, and between tAI and DIM as causal associations. Thus, further studies are needed to find clear explanations for these associations.

It also should be noted that changes in variables that could have casual effects need to be large enough to be of practical significance. For example, regarding the effect of altitude, each 100 m increase in altitude was associated with a 0.2 kg increase in ECM. So, when considering regions in which to build cowsheds, one region must be some hundred meters higher than the others to have significant benefit. Similarly, each metre increase in cowshed ridge height was associated with a 0.67 kg increase in ECM, so an increase of 2 to 3 m in ridge height could significantly improve ECM. Further, some effects appeared to be significant but had high standard errors, for example, the effects of eave height on mDM, mat area per cow on BW, or floor area per cow on PS. Thus, the impact of changing those variables in the targeted production and welfare parameters remains uncertain. Therefore, the interventions' cost-effective and practical perspectives need to be considered, but these perspectives were not included in the current study.

In addition, the current study built the multivariate models only for some selected production and welfare indicators. Many other important production and welfare indicators such as reproduction parameters, mobility and cow cleanness were not studied. Also, due to the limited number of observations, only 32 SDFs, this study did not consider interactions between the predictor variables.

7.5 Conclusion

For Vietnamese SDFs, the following research priorities were identified. Improving shed design to cool the cows ranked as the foremost opportunity to simultaneously improve milk production and cow welfare. Shed improvement research should target reductions in ambient temperature and increased airflow. If possible, build farms in highland regions, but even in highland regions, responses to increases in eave and ridge roof heights, percentage of shed side open, and floor area per cow require testing. Next are dietary interventions. Research needs to target increased dietary dry matter intake and the effects of increased dry matter and fat concentration in the diet on milk production. Research is also needed to test whether a decrease in dietary lignin and phosphorus concentrations can reduce panting score without reducing cows’ milk yield and milk fat concentration. Finally, genetic research should test whether increasing the genetic proportion of BSW but not ZEB in individual cows can improve milk production.

183

Chapter 8 Application of infrared thermal technology to assess the level of heat stress and milk yield reduction of cows in tropical smallholder dairy farms

Abstract

Panting score (PS) is a common research tool used to assess the physiological state of cows exposed to heat stress, but it is subjective. Infrared temperature (IRT), measured by either infrared thermometers or cameras, may be a more objective and reliable alternative. There are very few studies evaluating the associations between PS, IRT and milk production. The applicability of IRT compared to PS as a means of assessing heat stress and milk yield reduction in dairy cows in tropical smallholder dairy farms (SDFs) was investigated. In autumn 2017, SDFs located across four typical dairy regions of Vietnam were each visited once to collect farm (n = 32) and individual cow data (n = 345). For each SDF, heat load index (HLI) inside the cowsheds, an indicator of environmental heat load calculated from ambient temperature, humidity, and wind speed, were measured. For each cow, PS (0 indicates a cow breathing normally, not panting; 4.5 indicates an extremely heat-stressed cow with excessive panting, tongue fully extended, and excessive drooling), IRTs of the cow body, single-day energy corrected milk yield (ECM), body weight, and body condition score were measured. Cow genotype, age, lactation number, and days in milk were recorded. The IRTs of the cows’ inner vulval lip (IVuT) were measured with an infrared thermometer; and the IRTs of the cows’ vulval surface (OVuT), inner tail base surface (ITBT), ocular area, muzzle, armpit area, paralumbar fossa area, fore udder, rear udder, forehoof and hind hoof were also measured with an infrared camera. Multivariate mixed-effects models were used to assess the associations between HLI with PS and IRTs, and associations between PS and IRTs with ECM while accounting for the effects of other cow variables. All IRTs correlated positively with PS (Pearson correlation, r = 0.23 to 0.50, P < 0.001). Each unit increase in HLI was associated with increases of 0.07 units in PS and 0.09 to 0.23°C in IRT (P < 0.05). Each °C increase in IVuT, OVuT, and ITBT was associated with decreases of 0.75, 0.87, and 0.70 kg/cow/d in ECM, respectively (P < 0.05) while PS and other IRTs were not significantly associated with ECM. Thus, all IRTs showed potential to assess the heat stress level of cows; and IVuT, OVuT and ITBT, but not PS and other IRTs, showed potential to predict ECM reduction in cows during heat stress. First cross (F1) Holstein Brown Swiss and F1 Holstein Jersey showed lower PS and yielded higher ECM than the third backcross (B3) Holstein Zebu (7/8 Holstein + 1/8 Zebu) and pure Holstein (P < 0.05). Thus, F1 Holstein Brown Swiss and F1 Holstein Jersey could be more suitable for tropical SDFs than B3 Holstein Zebu and pure Holstein.

184

8.1 Introduction

In dairy cattle, heat stress is associated with a reduction in feed intake, milk production, and reproduction, which in turn cause economic losses and raise welfare concerns (Kadzere et al., 2002; Hansen, 2007; Polsky and von Keyserlingk, 2017). Climate change is predicted to make heat stress a more severe issue, especially in the tropics (Morton, 2007; Henry et al., 2012; Sejian et al., 2015). Heat stress can be assessed using environment-based indicators such as temperature, humidity, a temperature-humidity index (THI), a heat load index (HLI) (Herbut et al., 2018), or animal-based indicators such as panting score (PS), body temperature, respiration rate, heart rate, dry matter intake, or behaviour (Gaughan et al., 2002; Wang et al., 2018; Galán et al., 2018; Rashamol et al., 2019). Although environment based indicators can help anticipate the impact of heat stress on cows, they do not reflect the physiological changes experienced by the cow as animal-based indicators do (Liu et al., 2019).

Among animal-based indicators, PS, rectal temperature, and respiration rate are most commonly used to assess heat stress in cattle (Galán et al., 2018; Liu et al., 2019). While rectal temperature and respiration rate are general indicators of animal health, PS was specifically developed to assess the heat stress levels of the cow. A PS of 0 indicates normal panting (normal breath, no drooling), and a PS of 4.5 indicates excessive panting (fast breath from the flank, tongue fully extended, excessive drooling, neck extended, and head held down). When environmental THI < 68 equivalent or HLI < 70, cattle are usually not heat-stressed and have rectal temperatures of 38.5 ± 0.5ºC; respiration rates of 20 ± 10 breaths/min, and PS of zero (Jackson and Cockcroft, 2002; Gaughan et al., 2008; Zimbleman et al., 2009). When THI and HLI are above the heat stress thresholds, PS, rectal temperature, and the respiration rate of the cattle increase (Regan and Richardson, 1938; Gaughan et al., 2008). Measuring PS is simple and non-invasive, without the need to restrain animals or buy equipment. Thus, PS has been widely studied and applied to dairy and beef cattle in both experimental and commercial conditions (Mader et al., 2006; Gaughan et al., 2009; Alfonzo et al., 2016; Unruh et al., 2017). The panting score appears to be suitable for SDFs where restraining facilities are limited. However, scoring PS is subjective, and its accuracy depends on the experience of the observers. The measurements of rectal temperature and respiration rate, on the other hand, are objective. However, these measurements are quite complicated, invasive, time-consuming, and animals usually need to be restrained, and these disadvantages limit their application under commercial conditions. As a result, rectal temperature and respiration rate are often used to assess heat stress during heat stress experiments but are rarely used on commercial farms.

185

Infrared thermometers and infrared thermal cameras, which both apply infrared thermal technology, can be alternative devices to the traditional rectal thermometer to measure the temperature of cattle. Both are non-invasive devices that can measure the infrared temperature (IRT) of cow surface with little or no direct contact with the animals (Hoffmann et al., 2013; Tattersall, 2016). Compared with PS, IRT is much more objective. When comparing infrared thermometer and infrared thermal camera, the infrared thermometer is cheaper and easier to use, but users need to stand close to, or maintain certain contact with animals. On the other hand, infrared thermal cameras are expensive, require expertise to use, and require time to process the captured images (Unruh et al., 2017). The advantages of infrared thermal cameras are their capacity to record moving animals, measure temperatures in many locations simultaneously, and the potential to capture the images automatically (Hoffmann et al., 2013). A few studies have shown the potential application of infrared thermal cameras for recognising heat stress in beef heifers (Unruh et al., 2017), Nellore cattle (Martello et al., 2016), purebred lactating Holstein cows, Girolando cows (crossbreeds of Holstein X Gir) (Daltro et al., 2017) and Jersey heifers (Salles et al., 2016). These studies have reported positive correlations between the infrared temperature of the animal body surface and PS, physiological parameters of the cows, and the temperature-humidity index (THI) of the environment. However, these studies were research projects where confounding variables were managed. To our knowledge, IRT technology has not been studied within a production system in SDFs where breeds, management, and the microclimate are diverse. Also, studies assessing the associations of PS and IRT with the milk production of the cows remain very scarce.

Therefore, this study aimed to investigate the possibility of using PS and infrared temperature as tools for assessing the heat stress status and productivity response of the cows in SDFs under hot conditions. To do this, the relationships between changes in the shed microclimate as measured by heat load index (HLI), PS (visual assessment) and body temperatures (obtained via infrared thermal devices), and cow milk production measured by energy corrected milk yield were investigated. It was hypothesized that within an SDF: (1) increased HLI will be associated with increased PS and infrared temperatures of the cows; and (2) increased PS and infrared temperatures, in turn, will be associated with decreased milk production.

186

8.2 Materials and methods 8.2.1 Farm visit

This study was conducted on 32 SDFs located in four typical dairy regions of Vietnam, which were a southern lowland (SL), a southern highland (SH), a northern lowland (NL), and a northern highland region (NH) (see Chapter 3). These four regions were selected to represent Vietnamese smallholder dairy regions because of their high population of dairy cows and the diverse weather conditions between the provinces. During a period between 24 August and 7 October 2017, each SDF was visited for an afternoon and again the next morning. At each visit, weather data and individual lactating cow data (n = 344) were obtained.

Infrared temperatures of the cows were obtained within this study, while cowshed microclimate data and other individual cow data, including panting score, milk yield, age, days in milk, lactation, body weight, and body condition score, were obtained from previous chapters. All the measurements of the cows in a given SDF and the microclimate data within the cowshed of that SDF were gathered on the same date.

8.2.2 Microclimate data

The microclimate data previously presented in Chapter 6 are used here. Briefly, the microclimate parameters including wind speed (WS, m/s), ambient temperature or dry bulb temperature (AT, °C), relative humidity (RH, %), black globe temperature (GT, °C), and dew point temperature (Tdp, °C) inside the cowshed of each SDF were measured at 0600 h, 0800 h, 1000 h, 1100 h, 1400 h, 1600 h, and 1800 h using a Kestrel 5400 Heat Stress Tracker (NIELSEN-KELLERMAN, USA).

Temperature-humidity index (THI, units) was calculated using the equation of Yousef (1985):

THI = AT + (0.36  Tdp) + 41.2, where: AT = dry bulb temperature (oC); Tdp = dew point temperature (oC).

Heat load index (HLI, units) was calculated using the following equations (Gaughan et al., 2008):

(2.4 – WS) HLIBGT  25 = 8.62 + 0.38  RH + 1.55  GT – 0.5  WS + e

HLIBGT < 25 = 10.66 + 2.8  RH + 1.3  GT – WS, where: RH = relative humidity (%); GT globe temperature (oC); WS = wind speed (m/s); e = base of natural logarithm.

THI and HLI measurements were then averaged to obtain average daytime THI and HLI at each SDF.

187

8.2.3 Individua cow data

Infrared temperature measurements

To facilitate easy identification and the recording of the data of individual cows, cow identification (ID) was written with a large black marker pen on A5 size paper, and this was attached with non- toxic, water-soluble glue to the left side of the cows (Figure 8.1 a).

a) Write and attach ID b) Score cow panting c) Inner vulva d) Inner tail base (ITBT), for each cow (PS) temperature (IVuT) outer vulva (OVuT)

e) Muzzle (MuzT) f) Ocular area (EyeT) g) Paralumbar fossa (ParT) h) Fore hoof (FHoT)

i) Hind hoof (HHoT) j) Fore udder (FUdT) k) Rear udder (RUdT) l) Armpit (ArmT)

Figure 8.1 Cow identification (a), scoring panting score (b), and measurements of infrared temperature (c to l) When analysing thermal images with FLIR Tool, within each shape, the solid red triangle shows the hottest point (maximum temperature), and the solid blue triangle shows the coolest point (minimum temperature).

The average temperature of the inner vulval lip (IVuT, °C) of each cow was the average of morning and afternoon temperatures of the inner vulval lip, which were measured between 0600 h and 0700 h and between 1500 h and 1600 h, respectively, using a Combi Infrared Thermometer (Combi Corporation, Japan). The inner vulva was chosen to measure temperature because the tail usually obscures this area, so it is usually clean, and the temperature in this area is less affected by

188 environmental variables such as direct sunlight or wind. When obtaining the temperature, the probe of the thermometer was placed between the vulval lips, pointed to the vulval wall at a position of less than 1 cm from outside the vulval lips, and the measurements were triplicated (Figure 8.1 c).

An infrared camera (FLIR E50 IR camera - 240X180 pixel resolution, FLIR Systems Inc.) was used to capture infrared thermal images of the cows at ten targeted body areas (Figure 8.1 d to l). The cow images were taken from a distance of 1.0 to 1.5 m (Montanholi et al., 2015) and within 20 minutes after afternoon milking (usually between 1530 h to 1730 h) when cows were usually standing and eating. The reflected apparent temperature was set at 20°C, and the emissivity was set at 0.98, which is the emissivity of human and mammal skins (Kastberger and Stachl, 2003; Hoffmann et al., 2013; Sathiyabarathi et al., 2016). The relative humidity and environmental temperature were obtained using a Kestrel 5400 Heat Stress Tracker (NIELSEN-KELLERMAN, USA) when the first thermal image was captured. The captured images were then used to determine the IRTs of the outer vulval surface (OVuT, °C), rear udder (RUdT, °C), inner tail base surface (ITBT, °C), ocular area (EyeT, °C), muzzle (MuzT, °C); armpit (ArmT, °C), paralumbar fossa area (ParT, °C), fore udder (FUdT, °C), forehoof (FHoT, °C), and hind hoof (HHoT, °C) (Figure 8.1 d to l). When analysing the images, a circle was drawn around the eyes, whereas a rectangle was drawn around other regions. Although maximum, minimum, and average IRTs were obtained from each shape, only the maximum IRT data were used in the subsequent statistical analysis.

Panting score

The panting score (PS) data of the cows were obtained from Chapter 3. Briefly, the day panting score (PS) was the average of morning and afternoon PS, which were scored between 0500 h to 0600 h and between 1400 h and 1500 h, respectively. A PS scale from 0 to 4.5 was used (0 indicates a cow breathing normally, not panting; 4.5 indicates excessive panting with fast breath from the flank, tongue fully extended, excessive drooling, neck extended, and head held down) (Gaughan et al. 2009) (Figure 8.1 b).

Milk yield and other general cow data

Milk yield, age, days in milk, lactation, body weight, and body condition score of each cow were obtained from Chapter 3. Briefly, the single-day milk yield of each cow was obtained by adding morning and afternoon milk yields obtained directly after each milking. The morning and afternoon milk output of each cow was also sampled to analyse milk fat and protein concentrations at the Nutrition Laboratory, Vietnam National University of Agriculture. Single-day milk yield, fat, and

189 protein concentrations were used to calculate single-day energy corrected milk yield (ECM, 3138 KJ per kg ECM) using the equation of Tyrrell and Reid (1965).

The heart girth of each cow was measured using a tape measure and was used to estimate cow body weight (BW) using the equation of Goopy et al. (2017). The body condition scores (BCS) of the cows were obtained by the methods described in Chapter 3. Body condition score (BCS, 5-point scale, 1 = lean, 5 = obese) was determined independently by two team members and averaged, using the guideline of Edmonson et al. (1989). The mean (± SD) obtained for BW and BCS were 479 ± 74 kg and 2.83 ± 0.50, respectively.

Age, number of lactations, and days in milk of each cow were obtained by asking the farmer and checking their record books where possible. The mean (± SD) obtained for age (years), number of lactations, and days in milk of those cows were 4.5 ± 1.7 years, 2.3 ± 1.4 lactations, and 189 ± 121 days, respectively.

The breed of each cow was obtained from Chapter 4, where cow breed was identified using the genomic data. Cows included in this study belonged to one of the nine following crossbreeds: F1HOL_ZEB (3 cows), B1HOL_ZEB (4 cows), B2HOL_ZEB (43 cows), B3HOL_ZEB (77 cows), pure HOL (166 cows), F1HOL_JER (20 cows), B1HOL_JER (7 cows), F1HOL_BSW (4 cows), and B1HOL_BSW (9 cows); where HOL, ZEB, JER, and BSW stands for Holstein, Zebu, Jersey, and Brown Swiss respectively, and F1, B1, B2, B3, and pure HOL represent the proportion of Holstein in a cow was > 1/4 to <3/4,  3/4 to < 7/8,  7/8 to < 15/16,  15/16 to < 31/32, and  31/32, respectively.

8.2.4 Statistical analyses

All statistics were performed using the base and additional packages of R software (R Core Team, 2018). Cows were the experimental units in all analyses. While all individual cow data were measured at the animal level, THI and HLI were measured at the farm level. Thus, when analysing the data using cows as the experimental units, the cows within an SDF were assigned the same HLI and THI values as measured in that SDF. Descriptive statistics

Descriptive statistics for quantitative variables were calculated for each region using the R package ‘psych’ (Revelle, 2019).

Statistical comparisons and correlations

190

A Normality test showed that all quantitative variables in this study were normally distributed; thus, One-way ANOVA tests followed by Tukey–Kramer tests (P < 0.05) using R package ‘agricolae’ (Mendiburu, 2019) were applied to compare the means of variables between regions. Pearson correlation coefficients and corresponding significance levels between variables were calculated using the R package ‘psycho’ (Makowski, 2018).

Multivariate regression analysis

Twenty-four multivariate linear mixed effect models were fitted using a restricted maximum likelihood approach to evaluate the effects of the independent variables on the dependent variables using the R package ‘lmerTest’ (Kuznetsova et al., 2017). Farm identification was fitted as a random effect in all models to account for possible clustering. Four farms with less than five lactating cows per farm (a total of 12 cows) were removed to ensure the quality of the model. The matrix notation describing the model was:

y = X + Wc + e, where: y was the vector of dependent variables,  was the vector of fixed effects, c was the vector of 2 the random farm effect (28 farms) [c ~ N(0, Iσ c)], e was the vector of residual random farm effects 2 [e ~ N(0, Iσ e)], and X and W were the incidence matrices of the fixed effects and random environmental effects, respectively. Random farm effects and residual effects were assumed to be 2 2 independently distributed. The parameters of the model σ c and σ e were estimated by the restricted maximum likelihood (REML) method.

The first 12 models were used to evaluate the effect of environmental heat load measured by HLI on PS and 11 IRT measurements. In these models, HLI, region, breed, age, lactation, DIM, BW, BCS, and ECM of the cows were fitted as fixed effects () while each of the PS and 11 IRT measurements was an outcome variable (y). Although both THI and HLI were indicators of environmental heat load, the calculation of THI included only AT and RH, whereas the calculation of HLI included AT, RH, and WS (Yousef, 1985; Gaughan et al., 2008). Thus, we have chosen HLI rather than THI to be included as a fixed effect in building models.

The other 12 models were used to evaluate the effect of PS and 11 IRT measurements on ECM. In these models, ECM was fitted as an outcome variable (y) while region (4 regions), cow breed, age, lactation, DIM, BW, BCS and each of the PS and IRT measurements were fitted as fixed effects ().

When fitting models, Satterthwaite’s approximation method was used to calculate degrees of freedom and P-values for both F-statistics and t-statistics. To avoid multicollinearity, the independent variables

191 with the variance inflation factor (VIF) > 5 were removed first (Kock and Lynn, 2012). Based on this approach, the variable ‘age’ was removed from all models. Then, the manual backward elimination technique was applied to remove one by one the variables with F-statistics P-value > 0.1. To choose the final models, models were compared using likelihood ratio tests for the significance of dropped independent variables. Distributions of standard residuals of the final models were plotted to confirm that normality and homoscedasticity assumptions were met (Barberg et al., 2007). The results of the finally chosen models were presented. Marginal R2 of the final models, indicating the percentage of the total model variance explained by the fixed effects, was calculated as the ratio of fixed-effect variance to the total variance, which is the sum of fixed effect, random effect, and residual variances (Bartoń, 2019). Conditional R2 of the final models, indicating the percentage of the model variance explained by both fixed and random effects, was calculated as the ratio of the sum of the fixed and random effect variances to the total variance (Bartoń, 2019).

After choosing the final models, whenever the effects of the categorical variables, either breed or region, were significant, the least-square means of different levels of the categorical variables were compared.

8.3 Results 8.3.1 Correlations between main variables

The THI, HLI, ECM, PS and IRT measurements of cows in four regions are presented in Table 8.1. The overall means (± SE) for THI and HLI were 79.4 ± 1.9 and 86.4 ± 3.3, respectively. The means of THI and HLI in SL and NL were significantly higher than those in SH and NH (P < 0.05). The overall mean (± SE) of ECM was 15.7 kg/cow/d. The overall mean (± SE) of PS was 1.3 ± 0.2. Except for the means of EyeT, RUdT, ITBT, and OVuT, which were similar across the four regions, the means of PS and all other IRT measurements were highest in SL, followed by NL and NH, and lowest in SH (P < 0.03).

When comparing the different body regions, ITBT were highest (mean = 38.8oC), followed by FUdT (38.5oC), RUdT (38.2oC), and OVuT (38.0oC). In contrast, FHoT was lowest (35.5oC), followed by HHoT (35.2oC) and MuZT (36.4oC).

Table 8.2 presents the Pearson correlation coefficients (r) between variables included in this study. ECM correlated significantly and negatively with THI, HLI, PS, OVuT, EyeT, and FUdT (r = -0.12 to -0.41, P < 0.05); but correlated very weakly and insignificantly with other infrared temperature measurements (r = 0.00 to -0.12, P > 0.05). All IRT measurements correlated significantly and positively with either THI (r = 0.14 to 0.48, P < 0.05), HLI (r = 0.14 to 0.44, P < 0.05), or PS (r =

192

0.23 to 0.50, P < 0.001). All IRT measurements were significantly and positively correlated with each other (P < 0.001) with r ranging from 0.56 in the FUdT/HHoT pair to 0.78 in the ITBT/OVuT pair. PS was significantly associated with THI (r = 0.54, P < 0.001) and HLI (r = 0.56, P < 0.001). Among the infrared temperature measurements, MuzT and ParT showed the highest correlation coefficients with HLI, THI, and PS, whereas IVuT showed the lowest correlation coefficients with HLI, THI, and PS.

Table 8.1 Comparing means of the temperature-humidity index, heat load index, panting score, and infrared thermal temperatures across four main dairy regions A

Region, Mean Parameters Abbreviation P Mean  SEM SL SH NL NH THI, units THI 82.5a 75.5b 82.9a 76.7b < 0.001 79.4  1.9 HLI, units HLI 92.4a 80.0b 91.9a 81.2b < 0.001 86.4  3.3 Energy corrected milk ECM 13.1b 15.1b 15.6b 19.2a < 0.001 15.7  1.3 Panting score PS 1.8a 0.8c 1.4b 1.3b < 0.001 1.3  0.2 IRT measurement, oC Inner vulva IVuT 37.8a 37.0b 36.9b 37.4ab 0.015 37.3  0.2 Muzzle MuzT 37.0a 35.4b 36.8ab 36.4ab 0.025 36.4  0.3 Eye EyeT 38.2 37.4 37.9 37.8 0.095 37.8  0.2 Fore hoof FHoT 35.6a 32.3b 34.7a 35.4a 0.002 34.5  0.8 Armpit ArmT 37.6a 36.1b 36.8ab 37.4ab 0.025 37.0  0.3 Paralumbar fossa ParT 37.7a 35.7b 36.7ab 36.7ab 0.007 36.7  0.4 Fore udder FUdT 38.9a 38.0b 38.3ab 38.6ab 0.030 38.5  0.2 Rear udder RUdT 38.6 37.8 38.0 38.5 0.125 38.2  0.2 Hind hoof HHoT 36.2a 33.3b 35.4ab 35.9a 0.005 35.2  0.7 Inner tail base surface ITBT 39.2 38.4 38.7 39.0 0.178 38.8  0.2 Outer vulva OVuT 38.5 37.6 37.7 38.1 0.091 38.0  0.2 A P-values are given for One-way ANOVA tests (superscript letters are given for Tukey–Kramer test, P < 0.05). Means with the same letter are not significantly different from each other.

193

Table 8.2 Matrix of Pearson correlations between temperature-humidity index, heat load index, panting score, and infrared thermal temperatures A THI HLI ECM PS IVuT OVuT RUdT ITBT EyeT MuzT ArmT ParT FUdT FHoT

HLI 0.98*** 1 ------ECM -0.35*** -0.37*** 1 ------PS 0.54*** 0.56*** -0.14* 1 ------IVuT 0.14* 0.17** -0.08 0.23*** 1 ------OVuT 0.27*** 0.25*** -0.19** 0.37*** 0.49*** 1 ------RUdT 0.17** 0.14* -0.05 0.36*** 0.31*** 0.73*** 1 ------

ITBT 0.22*** 0.18** -0.12 0.38*** 0.43*** 0.78*** 0.76*** 1 ------EyeT 0.36*** 0.33*** -0.16** 0.35*** 0.38*** 0.61*** 0.63*** 0.68*** 1 - - - - - MuzT 0.48*** 0.44*** -0.10 0.45*** 0.34*** 0.66*** 0.65*** 0.63*** 0.70*** 1 - - - - ArmT 0.22*** 0.20** -0.04 0.40*** 0.33*** 0.67*** 0.70*** 0.70*** 0.58*** 0.69*** 1 - - - ParT 0.47*** 0.44*** -0.07 0.50*** 0.39*** 0.69*** 0.66*** 0.69*** 0.64*** 0.78*** 0.77*** 1 - - FUdT 0.24*** 0.22*** -0.13* 0.35*** 0.34*** 0.65*** 0.76*** 0.72*** 0.67*** 0.66*** 0.66*** 0.65*** 1 - FHoT 0.26*** 0.23*** -0.02 0.37*** 0.31*** 0.63*** 0.66*** 0.59*** 0.59*** 0.64*** 0.69*** 0.70*** 0.60*** 1 HHoT 0.22*** 0.20*** 0.02 0.38*** 0.32*** 0.60*** 0.62*** 0.57*** 0.59*** 0.57*** 0.64*** 0.68*** 0.56*** 0.75***

A Abbreviations as in Table 8.1. *, Significant at P  0.05; **, Significant at P  0.01; ***, Significant at P  0.001.

194

8.3.2 Associations of heat load index with inner infrared vulva temperature, outer infrared vulva temperature, and panting score

Results of multivariate linear mixed effect models showed that PS and all IRT measurements except HHoT had significant and positive associations with HLI (P  0.008) (Table 8.3 and Appendix 4).

Table 8.3 Multivariate linear mixed effect models identifying the variables significantly or suggestively associated with panting score (PS); and infrared temperatures of cow inner vulva (IVuT, °C), outer vulva (OVuT, °C), and inner tail base (ITBT, °C) A

Variable B PS IVuT OVuT ITBT Fixed effect Coef (SE) C P Coef (SE) P Coef (SE) P Coef (SE) P Intercept -4.71 (1.44) 0.004 28.93 (2.82) <.001 28.38 (3.49) <.001 30.10 (3.43) <.001 Heat load index 0.07 (0.02) <.001 0.10 (0.030) 0.003 0.12 (0.04) 0.005 0.11 (0.04) 0.007 (HLI) Region: South Low Reference -- D Reference -- Reference -- Reference -- Region: South High -0.31 (0.25) 0.225 0.52 (0.50) 0.306 0.29 (0.61) 0.634 0.42 (0.60) 0.494 Region: North Low -0.41 (0.13) 0.005 -0.71 (0.26) 0.012 -0.63 (0.32) 0.062 -0.46 (0.32) 0.160 Region: North High 0.17 (0.22) 0.435 0.91 (0.43) 0.046 1.09 (0.53) 0.052 1.07 (0.52) 0.051 -2.8E-4 -8.0E-4 -1.4E-3 -1.1E-3 Days in milk 0.070 0.013 0.001 0.001 (1.5E-4) (3.2E-4) (3.9E-4) (3.3E-4)

Lactation number 0.05 (0.01) <.001 ------0.05 (0.03) 0.080 ECM (kg/cow/d) -- -- -0.02 (0.01) 0.016 -0.03 (0.01) 0.007 -- -- Body condition score ------0.22 (0.08) 0.006 Breed: F1HOL_BSW Reference -- Reference -- Reference -- Reference -- Breed: ------0.28 (0.19) 0.146 ------B1HOL_BSW Breed: F1HOL_JER 0.30 (0.18) 0.091 ------Breed: B1HOL_JER 0.39 (0.20) 0.048 ------Breed: F1HOL_ZEB 0.27 (0.25) 0.292 ------Breed: B1HOL_ZEB 0.31 (0.18) 0.092 ------Breed: B2HOL_ZEB 0.47 (0.17) 0.006 ------Breed: B3HOL_ZEB 0.50 (0.17) 0.003 ------Breed: Pure HOL 0.53 (0.16) 0.001 ------Random effect Variance (SD) Variance (SD) Variance (SD) Variance (SD) Farm 0.05 (0.23) 0.21 (0.46) 0.32 (0.57) 0.32 (0.56) Residual 0.10 (0.31) 0.34 (0.59) 0.42 (0.65) 0.33 (0.58) Explanatory power, % Conditional R2 66.43 54.58 58.29 61.34 Marginal R2 47.29 26.79 26.01 23.78 A Variable cow age was excluded from all models due to VIF > 5; The independent variables excluded from each final model due to no significant effect (P > 0.1) included body weight and the variables with ‘--’ sign in the P column. B Abbreviations of variables: ECM, energy corrected milk; HOL, Holstein; ZEB, Zebu; BSW, Brown Swiss; JER, Jersey; F1, B1, B2, B3, and Pure indicate that the genetic proportion of Holstein was > 1/4 to <3/4,  3/4 to < 7/8,  7/8 to < 15/16,  15/16 to < 31/32, and  31/32, respectively. C Coef (SE), Coefficient (Standard error). D “--” indicates no estimate.

195

Each unit increase in HLI was associated with an increase of 0.07 unit in PS and increases from 0.09°C to 0.23°C in IRT measurements except HHoT when the other predictors were held fixed (Table 8.3 and Appendix 4). HHoT had a suggestive association with HLI (P = 0.070) (Appendix 4). The breed of the cows only showed significant associations with PS (P < 0.001) but not with any infrared temperatures (P > 0.1) (Table 8.3).

The changes in confidence intervals of PS, IVuT, OVuT, and ITBT over the range of HLI after adjusting for the effects of all other factors in the models are presented in Figure 8.2. When HLI changed from the lowest value (74.03) to the highest value (97.27), the changing patterns of IRT measurements were similar to the changing pattern of PS (Figure 8.2).

a) PS b) IVuT

c) OVuT d) ITBT

Figure 8.2 Effect plots predicting the effects (with confidence intervals) of heat load index (HLI) on the panting score (PS), infrared temperatures of inner vulva (IVuT), outer vulva (OVuT), and inner tail base (ITBT) The bars on the horizontal axis indicate the actual HLI records in the dataset.

It is necessary to consider the thresholds to interpret the meaning of IRT measurements when assessing the heat stress status of the cows. Gaughan et al. (2008) suggested that based on PS, cows can be classified into: normal (PS < 0.4), slightly heat-stressed (PS = 0.4 to less than 0.8), moderately

196 heat-stressed (PS = 0.8 to less than 1.2), and highly heat-stressed (PS  1.2). Using the PS thresholds suggested by Gaughan et al. (2008) and the prediction model for PS in Table 8.3, the thresholds for HLI can be calculated. The calculated thresholds for HLI then can be used to estimate thresholds for IRTs using other prediction models for IRTs in Table 8.3 and Appendix 4. For example, when using the prediction model for PS (first column of Table 8.3) and holding other predictors of the model fixed, a PS value of 0.4 (the first threshold of Gaughan et al., 2008) gave an HLI threshold = (0.4 + 4.71) / 0.07 = 73 units. Using the prediction model for IVuT (second column of Table 8.3) and holding other predictors of the model fixed, the HLI threshold of 73 gave an IVuT threshold = 28.93 + 0.10*73 = 36.2°C. Other thresholds for HLI and IRTs were calculated using the same methods, and Table 8.4 presents the thresholds calculated for some selected measurements. When using a measurement, the heat stress level of a cow can be categorized as: normal when that measurement < threshold 1, slightly heat-stressed when threshold 1  measurement < threshold 2, moderately heat-stressed when threshold 2  measurement < threshold 3, and highly heat-stressed when measurement  threshold 3. For example, using the thresholds for IVuT, SDF cows can be considered normal when IVuT < 36.2°C, slightly heat-stressed when 36.2°C  IVuT < 36.8°C, moderately heat-stressed when threshold 36.8°C  IVuT < 37.4°C, and highly heat-stressed when IVuT  37.4°C.

Table 8.4 Thresholds to assess environmental head load and heat stress levels of cows A Measurement Threshold 1 Threshold 2 Threshold 3 Reference Gaughan et al. PS_Reference 0.4 0.8 1.2 (2008) Estimations in the current study HLI, units 73.0 78.7 84.4 IVuT, °C 36.2 36.8 37.4 OVuT, °C 37.1 37.8 38.5 ITBT, °C 38.1 38.8 39.4 EyeT, °C 36.9 37.4 37.9 A Abbreviations as in Table 8.1. Assessment of heat stress based on measurements: normal, measurement < threshold 1; slightly heat-stressed, threshold 1  measurement < threshold 2, moderately heat-stressed, threshold 2  measurement < threshold 3, highly heat-stressed, measurement  threshold 3.

The least-square means of PS, IVuT, OVuT and ITBT across regions after adjusting for the effects of all other factors in the models are presented in Figure 8.3. The corresponding results for other IRT measurements are not presented because they showed the same patterns as the presented IRT measurements. Geographical regions showed significant associations with all PS, IVuT, OVuT, and ITBT (Figure 8.3). Consistently, least-square means of PS, IVuT, OVuT, and ITBT were highest in NH and lowest in NL (P < 0.05). The least-square means of PS, IVuT, OVuT, and ITBT in SL were

197 all similar to those in SH, although the pattern of PS was different from the patterns of IRT measurements.

a) PS b) IVuT

c) OVuT d) ITBT

Figure 8.3 Plots comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of panting score (PS) (a), infrared temperatures of inner vulva (IVuT) (b), outer vulva (OVuT) (c), and inner tail base (ITBT) (d) of the cows across regions Least-square means carrying different letters were significantly different from each other (P < 0.05).

The least-square means of PS for the crossbreeds are shown in Figure 8.4. In all crossbreed pairs (HOL_BSW, HOL_JER, and HOL_ZEB), an overall trend was that F1HOL crossbreeds tended to have the lowest PS and the greater the percentage of HOL, the greater the PS. Comparing crossbreeds, the least squared mean of PS of F1HOL_BSW (0.81), B1HOL_BSW (1.08), F1HOL_JER (1.11), and B1HOL_ZEB (1.12) were lower than those of B3HOL_ZEB (1.31) and pure HOL (1.33) (P < 0.05).

198

Figure 8.4 Plot comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of panting score between different dairy cow crossbreeds Least-square means carrying different letters were significantly different from each other (P < 0.05).

8.3.3 Associations of panting score, inner vulva infrared temperature, outer vulva infrared temperature, and inner tail base infrared temperature with energy corrected milk

Results of multivariate linear mixed effect models showed that only IVuT, OVuT, and ITBT had significant associations (P  0.035) with ECM while PS and all other IRT measurements did not (Table 8.5). The effect patterns of IVuT, OVuT, and ITBT on ECM were very similar to each other. Each °C increase in IVuT, OVuT, and ITBT was significantly associated with decreases of 0.75, 0.87, and 0.70 kg of ECM/cow/d, respectively. The regression coefficients of the effects of PS and other IRT measurements on ECM were positive for PS (0.32, P = 0.593), close to zero for HHoT (0.00, P = 0.978) and ParT (0.07, P = 0.783), and negative for the rest of the infrared temperatures (ranging from -0.45 for EyeT to -0.13 for FHoT, P > 0.113), however these associations were not significant. Thus, they were not presented.

199

Table 8.5. Multivariate linear mixed effect models identifying the associations of panting score (PS); and infrared temperatures for cow inner vulva (IVuT, °C), outer vulva (OVuT, °C), and inner tail base (ITBT, °C) with energy corrected milk yield (ECM, kg/cow/d) A

Effect of PS Effect of IVuT Effect of OVuT Effect of ITBT Variable B on ECM on ECM on ECM on ECM Fixed effect Coef (SE) C P Coef (SE) P Coef (SE) P Coef (SE) P Intercept 15.59 (2.52) <.001 43.98 (12.02) <.001 49.78 (11.71) <.001 43.04 (13.31) 0.002 PS and other IRT -- -- D ------IVuT -- -0.75 (0.31) 0.017 -- -- OVuT -- -- -0.87 (0.29) 0.003 -- ITBT ------0.70 (0.33) 0.035 -0.017 -0.018 -0.017 -0.017 Days in milk <.001 <.001 <.001 <.001 (0.0019) (0.0019) (0.002) (0.0021) 0.013 0.013 0.011 0.012 Body weight (kg) <.001 <.001 0.003 0.002 (0.0035) (0.0035) (0.0038) (0.0039) Region: South Low Reference Reference Reference Reference Region: South High 1.59 (1.17) 0.185 1 (1.19) 0.406 0.65 (1.18) 0.584 0.93 (1.2) 0.444 Region: North Low 2.67 (0.98) 0.012 2.04 (1.02) 0.054 2.14 (1) 0.041 2.43 (1.03) 0.025 Region: North High 5.33 (1) <.001 5.09 (1) <.001 5.05 (0.99) <.001 5.1 (1.02) <.001 Breed: F1HOL_BSW Reference Reference Reference Reference Breed: -4.02 (2.29) 0.080 -3.94 (2.27) 0.084 -4.25 (2.26) 0.061 -4.32 (2.34) 0.066 B1HOL_BSW Breed: F1HOL_JER -2.96 (2.12) 0.165 -2.63 (2.12) 0.214 -3.25 (2.11) 0.125 -3.21 (2.13) 0.134 Breed: B1HOL_JER -4.43 (2.37) 0.063 -4.25 (2.36) 0.073 -4.76 (2.30) 0.04 -4.58 (2.33) 0.051 Breed: F1HOL_ZEB -4.32 (3.01) 0.153 -4.02 (3.00) 0.180 -4.54 (2.94) 0.123 -4.39 (2.98) 0.142 Breed: B1HOL_ZEB -5.91 (2.19) 0.007 -5.8 (2.18) 0.008 -6.31 (2.22) 0.005 -6.19 (2.23) 0.006 Breed: B2HOL_ZEB -3.82 (2.04) 0.062 -3.69 (2.02) 0.069 -4.17 (2.02) 0.04 -4.29 (2.06) 0.038 Breed: B3HOL_ZEB -5.33 (1.98) 0.007 -5.07 (1.97) 0.010 -5.63 (1.94) 0.004 -5.62 (1.97) 0.005 Breed: Pure HOL -5.32 (1.94) 0.006 -5.14 (1.93) 0.008 -5.21 (1.9) 0.007 -4.96 (1.92) 0.011 Variance Variance Variance Variance Random effect (SD) (SD) (SD) (SD) Farm 1.97 (1.41) 2.00 (1.41) 1.65 (1.28) 1.74 (1.32) Residual 13.96 (3.74) 13.38 (3.71) 13.10 (3.62) 13.43 (3.67) Explanatory power, % Conditional R2 48.28 49.09 51.00 50.15 Marginal R2 40.98 41.78 44.80 43.68 A Variable cow age was excluded from all models due to VIF>5; The independent variables excluded from each final model due to no significant effect (P > 0.1) included body condition score, lactation number, and the variables with ‘--’ sign in P column of each model. B, C, D Abbreviations as in Table 8.3.

The effects of days in milk, body weight, region, and breed on ECM were consistent across the models (Table 8.5). Comparing models, the model, including the variable OVuT had the highest explanatory power (conditional R2 = 51%); thus, this model was used to draw the effect plots that compare the

200 least-square of ECM across the regions and between crossbreeds (Figure 8.5). After accounting for effects of OVuT, days in milk, BW, and breed; cows in NH yielded highest ECM (19.7 kg/cow/d), followed by cows in NL (16.8 kg/cow/d) and SH (15.3 kg/cow/d), and cows in SL yielded lowest ECM (14.7 kg/cow/d) (P < 0.05) (Figure 8.5 a).

When comparing breeds (Figure 8.5 b), the overall trend was that the least square mean ECM of F1 crossbreeds tended to be higher than that of B1 crossbreeds. HOL_BSW tended to yield the highest ECM, followed by HOL_JER, while HOL_ZEB and pure HOL tended to yield the least ECM. After accounting for effects of OVuT, days in milk, BW and region, the least-square mean ECM of F1HOL_BSW (20.9 kg/cow/d) and F1HOL_JER (17.6 kg/cow/d) were highest and significantly higher than those of B1HOL_ZEB (14.6 kg/cow/d) and B3HOL_ZEB (15.2 kg/cow/d) (P < 0.05). The least-square mean ECM of F1HOL_BSW was also higher than that of pure HOL (15.7 kg/cow/d).

a) ECM and regions b) ECM and crossbreeds

Figure 8.5 Effect plots comparing least-square means (black dots) with confidence interval (red arrows) and range (blue shaded areas) of energy corrected milk yields (ECM) across regions (a) and between crossbreeds (b) Least-square means carrying different letters were significantly different from each other (P < 0.05).

8.4 Discussion 8.4.1 Pearson correlations between temperature-humidity index, panting score, infrared temperature measurements, and energy corrected milk yield

In the current study, significant negative associations between PS and all IRT measurements with HLI and THI were found; and there were significant negative correlations of ECM with THI, HLI, PS, OVuT, EyeT and FUdT (P < 0.05). These negative correlations were expected because the mean HLI (86.4 units) and mean THI (79.4 units) between 0600 h and 1800 h in the current study were

201 higher than the heat stress thresholds of THI = 68 units (Zimbleman et al., 2009) and HLI = 70 units (Gaughan et al., 2008). When heat stress thresholds are exceeded, the body temperature and respiration rate of cattle increase, feed intake decreases, and milk yields decrease (Regan and Richardson, 1938; Könyves et al., 2017; Habeeb et al., 2018; Liu et al., 2019). Moreover, in the current study, all IRT measurements were significantly and positively correlated with PS, which is a heat stress indicator commonly applied in monitoring heat stress in dairy and beef cattle (Mader et al., 2006; Gaughan et al., 2009; Alfonzo et al., 2016; Unruh et al., 2017).

The positive correlations between PS and IRT measurements with HLI and THI in the current study were also consistent with previous studies (Bouraoui et al., 2002; Gaughan et al., 2008; Daltro et al., 2017; Unruh et al., 2017). For example, a study in Brazil using pure HOL and crossbreeds of HOL and Gir cows showed that EyeT, FUdT and RUdT in the afternoon correlated significantly with PS (0.37, 0.54, and 0.41, respectively, P < 0.01) (Daltro et al., 2017). The same authors also reported positive correlations between EyeT, FUdT, and RUdT with THI, but these correlations were not significant (P > 0.01) (Daltro et al., 2017). In a study by Bouraoui et al. (2002), THI was positively correlated with respiration rate (r = 0.89), heart rate (r = 0.88), rectal temperature (r = 0.85), and cortisol (r = 0.31). Gaughan et al. (2008) reported positive correlations between HLI and PS (r = 0.93, P < 0.001) and between THI and PS (r = 0.61, P < 0.001). The correlation between HLI and PS (r = 0.56) and between THI and PS (r = 0.54) in the current study were lower than those reported by Gaughan et al. (2008). These differences could be because Gaughan et al. (2008) examined the variation of PS within cows over a wide range of HLI and THI, while the current study examined it over a more restricted range. The negative correlations between ECM with THI and HLI in the current study are in the same direction as the results of previous studies, which reported that each unit increased in THI was associated with a decline in milk yield of 0.08 to 0.41kg/cow/d (Preez et al., 1990; Ravagnolo et al., 2000; Bouraoui et al., 2002; Könyves et al., 2017).

8.4.2 Applicability of panting score and infrared technology in monitoring heat stress

Statistical models which describe the associations between HLI with PS and IRT measurements show that each unit increase in HLI is associated with an increase of 0.07 unit in PS and increases ranging from 0.09°C to 0.23°C in IRT measurements. The positive regression coefficient of PS on HLI found in the current study (0.07) is consistent with and slightly higher than the result of Veissier et al. (2018), who reported that within the HLI range of 50.8 to 78.8, each unit increase in HLI was associated with a 0.05 unit increase in PS in grazing dairy cows. The significant positive associations between all IRT measurements and HLI and the significant correlations between IRT measurements and PS, as discussed earlier, suggest that IRT measurements at 11 different positions on the cow body found by 202 using either an infrared thermometer or an infrared thermal camera were comparable with the PS when assessing the heat stress level of the cows in SDFs.

Table 8.4 was constructed to serve as a guideline for interpreting the heat stress level of cows based on HLI and IRT measurements. In Table 8.4, the thresholds for HLI (73.0, 78.7 and 84.4) calculated from the thresholds for PS suggested by Gaughan et al. (2008), animal-based indicators of heat stress, were quite close to the HLI thresholds of 70, 77, and 86 suggested by the same authors. This showed the consistency of HLI as an environment-based indicator of heat stress for cattle in different production conditions. However, in Table 8.4, it should be noted that the thresholds for IRTs were calculated from the thresholds for HLI, which is an environment-based indicator of heat stress, not an animal-based indicator of heat stress. Thus, further studies using animal-based indicators of heat stress such as rector temperature, respiration rate, or biological markers to validate the thresholds for IRTs identified in the current study are recommended.

Using PS is the cheapest method for assessing the level of heat stress in SDFs provided that farmers are appropriately trained. The next best option is IVuT measured by infrared thermometer mainly because the infrared thermometer is much cheaper and easier to use than a thermal camera. If an infrared thermal camera is available, measurements of IVuT, OVuT, and ITBT are most meaningful as they were associated with both THI and ECM.

Results of models describing associations between PS and IRT measurements with ECM indicated that some IRT measurements, including IVuT, OVuT, and ITBT, had significant associations with ECM, while PS did not. The reason for PS not having the significant associations with ECM might be because PS is an assessment form based on respiratory rate, deepness of panting and degree of drooling (Gaughan et al., 2008), which are all among the adaptive responses of cattle in dealing with high environmental heat load and maintaining their body temperature within a small range. If those adaptive responses were still efficient enough for the cows to maintain their core body temperature, their milk production might not decrease. In other words, the cows that panted more were not necessarily the cows that already had decreased milk production. On the other hand, IRT measured the heat load retained on the cow body surface after all of her adaptive responses were activated. Increased IRT may indicate that the cows fail to maintain their body temperature; thus, milk production decreases.

Data from the current study suggests that IVuT, OVuT and ITBT are good indicators to assess the heat stress status of the dairy cows in SDFs, and they may also be potential traits to use in the selection of heat-tolerant dairy cattle. This is because IVuT, OVuT, and ITBT were positively associated with HLI while negatively associated with ECM. The desirable cows for selection could be the cows that 203 can maintain low IRT in a hot environment. However, further randomized experiments are necessary to validate the associations between these IRT measurements and ECM. Experiments are also required to estimate levels of genetic variation in these IRT measurements and their genetic relationships with other important production and welfare traits such as those traits related to feed intake, reproductive success, the function of the immune system, and behaviour of the cattle (Kadzere et al., 2002; West, 2003; Hansen, 2007; Polsky and von Keyserlingk, 2017). 8.4.3 Suitable dairy crossbreeds for tropical smallholder dairy farms

The overall trend in Figure 8.4 suggests that the higher PS was associated with a higher percentage of HOL in crossbreeds, which is consistent with the results of other authors (Dalcin et al., 2016; Alfonzo et al., 2016; Daltro et al., 2017). For example, in the same climatic conditions (THI between 74.2 to 84.6 units), heart rate (46.6 beats/min), rectal temperature (39.0oC), respiration rate (52.9 breaths/min) and PS (0.38) of F1HOL_GIR were within the normal ranges of 40 to 60 beats/min for heart rate, 38.3 to 39.3°C for rectal temperature, 23 to 40 breaths/min for respiration rate, and 0 to 0.4 for PS, whereas, these physiological parameters of B1HOL_GIR and pure HOL were all higher than the upper threshold of the normal ranges and significantly higher than the corresponding measurements of F1HOL_GIR (Alfonzo et al., 2016). Similarly, a study by Daltro et al. (2017) showed that at an ambient temperature range of 20.7°C to 37.9°C and relative humidity nearly 95%, the rectal temperature (40.84°C), respiratory frequency (111.36 breaths/min), and cardiac frequency (99.22 beats/min) of pure HOL were higher than those of F1HOL_GIR and B1HOL_GIR. The lower PS of F1HOL_BSW, B1HOL_BSW, F1HOL_JER, and B1HOL_ZEB compared to that of B3HOL_ZEB and pure HOL indicated that in terms of heat stress measured by PS, the F1 and B1 crossbreeds of HOL, especially with BSW, would be more heat-tolerant of the tropical conditions of Vietnam than B3 or pure HOL.

For milk production, the trend that the F1 crossbreeds of HOL with other breeds tended to yield higher ECM than F1 crossbreeds of HOL and pure HOL might reflect the effect of heterosis, which was also observed in other studies (Tadesse and Dessie, 2003; Madalena et al., 2012; Coffey et al., 2016). In addition, the tendency to yield higher ECM of the HOL_ BSW crossbreeds and, to a lesser extent, the HOL_ JER crossbreeds compared to HOL_ZEB crossbreeds and pure HOL as observed in the current study, are in agreement with other studies which showed that in terms of milk production BSW, to a lesser extent JER, and their crosses with HOL were more tolerant of tropical conditions than pure HOL (Johnson and Vanjonack, 1976; West et al., 2003; El-Tarabany et al., 2017). Johnson and Vanjonack (1976) reported that at an ambient temperature of 34°C and relative humidity of 80%, the milk yield of HOL, BSW, and JER cows was 41%, 56%, and 76%, respectively, of their normal

204 milk yields at an ambient temperature of 24°C and relative humidity of 38%. Similarly, a study in a subtropical region of Egypt by El-Tarabany et al. (2017) reported that when THI changed from low to high levels, the reduction in daily milk yield and peak milk yield of pure BSW and F1HOL_BSW was much smaller than pure HOL. The higher ECM of F1HOL_BSW compared to B1HOL_ZEB, B3HOL_ZEB, and pure HOL; and the higher ECM of F1HOL_JER compared to B1HOL_ZEB and B3HOL_ZEB, suggested that in terms of milk production, F1HOL_BSW and F1HOL_JER would be more suitable for Vietnamese SDFs than B3HOL_ZEB and pure HOL. However, it should be noted that in the current study, the numbers of F1HOL_BSW (4 cows) and F1HOL_JER (20 cows) were small; thus the results should be interpreted with care and further control experiments are recommended to validate the results of the current study.

8.5 Conclusion

In the context of SDFs, all infrared temperatures showed the potential to predict the heat stress level of cows. However, three infrared temperatures, including IVuT, OVuT, and ITBT, showed more advances than PS in the ability to predict ECM reduction of cows during heat stress. Comparing crossbreeds, F1HOL_BSW and F1HOL_JER could be more suitable for Vietnamese SDFs than pure HOL and B3HOL_ZEB.

205

Chapter 9 Genomic selection and genome-wide association studies for productivity and heat tolerance traits in smallholder dairy cows

Abstract

Genomic selection (GS) and genome-wide association studies (GWAS) have not been investigated in Vietnamese dairy cattle, even for basic milk production traits, probably due to the scarcity of individual phenotype recording in smallholder dairy farms (SDFs). This study aimed to estimate heritability (h2) and test the applicability of GS and GWAS for milk production, body conformation, and novel heat tolerance traits using single test day phenotypic data. Thirty-two SDFs located in either the north (a lowland vs a highland) or the south (a lowland vs a highland) of Vietnam were each visited for an afternoon and the next morning to collect phenotype data of all lactating cows (n = 345). Tail hair from each cow was sampled for subsequent genotyping with a 50K SNP chip at that same visit. Milk production traits (single-test day) were milk yield (MILK, kg/cow/d), energy corrected milk yield adjusted for body weight (ECMbw, kg/100 kgBW/d), fat (mFA, %), protein (mPR, %) and dry matter (mDM, %). Conformation traits were body weight (BW, kg) and body condition score (BCS, 1 = thin to 5 = obese). Heat tolerance traits were panting score (PS, 0 = normal to 4.5 = extremely heat-stressed) and infrared temperatures (IRTs, °C) at 11 areas on the external body surface of the cow (inner vulval lip, outer vulval surface, inner tail base surface, ocular area, muzzle, armpit area, paralumbar fossa area, fore udder, rear udder, forehoof, and hind hoof). Univariate linear mixed models and a 10-fold cross-validation approach were applied for GS. Univariate single SNP mixed linear models were applied for the GWAS. Estimated h2 were high (> 0.40) for mFA and BW; moderate (0.21 to 0.40) for mPR, mDM, ECMbw, BCS, PS, and IRTs at rear udder, outer vulval surface, and forehoof; low (0.11 to 0.20) for other IRTs; and very low (0.06) for MILK. Accuracy of genomic estimated breeding values (GEBVs) was low (< 0.13) for IRTs at hind hoof and fore udder; and moderate to high (0.22 to 0.68) for all other traits. The most significant regions on chromosomes (BTA) associated with milk production traits were 0.47-1.18 Mb on BTA14; for BCS, 15.55 Mb on BTA27; for PS, 40.05-40.08 Mb on BTA11; and for IRTs, 48.95-48.96 Mb on BTA8. Moderate to high h2 and moderate accuracies of GEBVs for ECMbw, mFA, mPR, BCS, BW, PS, and IRTs at rear udder, outer vulval surface, and forehoof suggested that GS using single test day phenotypic data could be applied for these traits. However, a greater sample size is suggested to decrease the bias of GEBVs by GS and increase the power of detecting significant quantitative trait loci (QTLs) by GWAS.

206

9.1 Introduction

Genomic selection (GS) and genome-wide association studies (GWAS) are two approaches revolutionizing the breeding of dairy cattle and knowledge of the genetic architecture of the important traits (Meuwissen et al., 2001, 2016; Sharma et al., 2015; Weller et al., 2017). Both approaches are based on the assumption that linkage disequilibrium (LD) exists between known markers (commonly single nucleotide polymorphisms, or SNPs, distributing across the whole genome) and unknown mutations controlling the traits (Hayes and Goddard, 2010). Although GS and GWAS focus on making the most of the information from the SNP data, they have different purposes. GS uses SNP and phenotype data of animals in a training population to build a prediction model that integrates the effects of all genotyped SNPs on the trait (Meuwissen et al., 2001). The GS model is then used to estimate the genomic estimated breeding values (GEBVs) of selection candidates for that trait (Meuwissen et al., 2001). By contrast, GWAS tests the effect of each SNP in turn to determine those most likely to be associated with a trait. These SNPs can be used to identify the mutations, quantitative trait loci (QTLs), or biological pathways associated with that trait.

Both GS and GWAS are now applied widely in developed countries for traditional economically important traits in dairy cattle. These include milk yield and milk fat and protein concentrations; durability traits such as body condition score, body size, body weight, body conformation, udder traits, feet and leg traits, milking temperament traits, and longevity; and health and reproduction traits such as udder health, calving ease, female fertility, and workability (Schrooten et al., 2000; Cole et al., 2011; Streit et al., 2013; Joerg et al., 2014; Nayeri et al., 2016; Abo-Ismail et al., 2017; Weller et al., 2017; Wiggans et al., 2017; Zhang et al., 2017; Jiang et al., 2019; Yin and König, 2019). However, the application of GS and GWAS in low-income developing countries remains limited and is further complicated by dairy populations in those countries often comprising crossbreeds of European breeds with local Zebu cattle (Ducrocq et al., 2018; Mrode et al., 2019). The limited application of GS and GWAS is most likely due to the scarcity of individual pedigree and phenotype records of the cows and the limitations of local quantitative genetic expertise and research (Ducrocq et al., 2018; Mrode et al., 2019).

Besides traditional traits, to cope with global warming, geneticists are increasingly interested in discovering SNPs for traits such as heat tolerance (Bernabucci et al., 2014; Weller et al., 2017). This in turn depends on the improved definition of heat tolerance traits. A few studies have indicated the possibility and the effectiveness of GS for heat-tolerant cows by defining heat tolerance as the extent of reduction in milk yield in response to heat stress (Garner et al., 2016; Nguyen et al., 2016a). “Slick hair coat” has also been defined as a heat tolerance trait as a result of GWAS (Huson et al., 2014). 207

Panting score (PS); body temperatures such as rectal, vaginal, or tympanic determined by the use of traditional thermometers; or the temperature of external body surface determined by infrared technologies have also been proposed. Cows with the genetic potential to maintain a low panting score and regulate body temperature more effectively during heat stress could be expected to show a lower reduction in productivity and welfare parameters due to heat stress (Poikalainen et al., 2012; Dikmen et al., 2012; Kim et al., 2014; Howard et al., 2014). Some studies showed that tectal, vaginal, and tympanic temperatures are heritable (Seath, 1947; Dikmen et al., 2012; Howard et al., 2014; Otto et al., 2019), but to date, no study has estimated the heritability of PS and body infrared temperatures or genetic structure of these traits. Compared with rectal or vaginal temperatures, the PS and infrared temperature measurements require minimal contact with the animals and therefore reduce the potential to impede cow welfare and productivity and save labour. If PS and infrared temperature prove to be heritable, they could become valuable traits for the selection of heat-tolerant dairy cattle.

Vietnam is a typical example of a tropical developing country where the application of GS and GWAS in dairy cattle has yet to occur, despite the rapid growth of the dairy industry there. Most of the dairy cows there (97%) are in smallholder dairy farms (SDFs, averaging 20 cows) where the farmers often do not record individual cow pedigree and phenotype records of the cows (Cai and Long, 2002; Vang et al., 2003; Department of Livestock Production, 2009; Lam et al., 2010; Trach, 2017a). The most common genotypes are Holstein and crossbreeds of Holstein with tropically adapted breeds (Red Sindhi and Sahiwal), local breeds (Yellow and Lai Sind), and other European dairy breeds (Jersey and Brown Swiss) (Hayley, 2010; Lam et al., 2010). A national breeding program has not yet been officially implemented, even for basic traits such as milk yield, milk fat, or protein concentration; and Vietnam has depended heavily on the importation of dairy bulls, semen straws, and heifers (Cai and Long, 2002; National Institute of Animal Sciences, 2017). To implement dairy cattle selection in Vietnam, GS should be a solution as it does not require pedigree records (Meuwissen et al., 2001; Weller et al., 2017). However, GS still requires the phenotypic records of individual cows, which is a major impediment to GS in Vietnam. SDF farmers generally do not have the milking infrastructure to facilitate this. Most cows are milked by portable milking machines with no milk meters or weigh scales. Farmers also have no access to any organised private or government providing herd recording expertise – almost all milk recording is based on the milk processing companies that buy milk from the farmers to record the saleable yield per milking per entire herd per farm. Therefore, the collection of individual cow data would be expensive in terms of money, time and labour, and there is a lack of expertise to facilitate it. Consequently, if GS is to be effective in Vietnam, it is important to keep the number of milk recording test days to a minimum.

208

Therefore, the following study aimed to use single-test day phenotypic data of the traits to: (1) estimate the genomic heritability of traditional production and body conformation traits and the novel heat tolerance traits of panting score and infrared body temperatures, (2) assess the accuracy and bias of GS on those traits, and (3) conduct GWAS studies to determine the genetic architecture of those traits and determine if quantitative trait loci (QTLs) that have previously been shown to segregate in Holstein cattle (e.g. DGAT1, Grisart et al., 2002) are also segregating in Vietnamese dairy cows.

9.2 Materials and methods 9.2.1 Animal and phenotype data

Data used in this chapter were obtained from Chapters 3, 4, 6 and 8. Briefly, 32 SDFs located in four typical smallholder dairy regions of Vietnam, including a southern lowland region, a southern highland region, a northern lowland region, and a northern highland region, were visited between 24 August and 7 October 2017 to collect individual phenotypic data of all lactating cows per SDF (n = 345). These four dairy regions represent Vietnamese SDF regions, as indicated by the high population density of dairy cows there and the diverse climate between regions.

Each SDF was visited twice to correspond with milking in an afternoon and the following morning. At the visits, general information about individual cows, including age, number of lactations and days in milk, was obtained by interviewing the farmer and verifying by checking their record books where possible. The mean ± standard deviation (SD) obtained for age (years), number of lactations, and days in milk of those cows were 4.5 ± 1.7 years, 2.3 ± 1.4 lactations, and 190 ± 121 days, respectively.

Single test-day measurements of milk production traits and body conformation traits were obtained, as described in Chapter 3. Briefly, for milk production traits, a single day milk yield (MILK, kg/cow/d) for each cow was obtained by weighing and summing the afternoon and the following morning milk yields. Milk samples for each cow were also collected at each milking, analysed at the Food Chemistry Lab (Vietnam National University of Agriculture, Hanoi, Vietnam) and averaged for percentages of milk dry matter (mDM, %, which is the percentage of all milk constituents excluding water), milk fat (mFA, %), and milk protein (mPR, %) content. These data were used to calculate the yield of milk dry matter (yDM, kg/cow/d), fat (yFA, kg/cow/d), protein (yPR, kg/cow/d) and energy- corrected milk (ECM, kg/cow/d) using the equation of Tyrrell and Reid (1965), and ECM adjusted for body weight (ECMbw, kg/100 kg BW/d).

For body conformation traits, the heart girth (HG, cm) of the cows was measured using a tape measure (Asia Technology Service Company, Vietnam). Cow body weight (BW, kg) was estimated from HG using an optimized algorithm of Goopy et al. (2018) (BWc, kg): BW0.3595 = 0.02451 + 0.04894  HG.

209

The body condition score (BCS) of each cow was determined as an average of independent BCS estimations made by the same two trained observers using a 5-point body condition score chart described by Edmonson et al. (1989) with one for very thin condition, five for very fat condition, and using an increment of 0.25.

Single test-day measurements of panting score (PS) were obtained as described in Chapter 3. Briefly, the PS of each cow was assessed and expressed as an average of morning PS taken between 0500 h to 0600 h and afternoon PS taken between 1400 h to 1500 h. One person assessed the PS of all cows based on a scale from 0.0 to 4.5 as previously used by Gaughan et al. (2009). A PS of 0.0 means that a cow is breathing normally and not panting, whereas a PS of 4.5 means the cow is excessively panting with fast breath from the flank, tongue fully extended, excessive drooling, neck extended, and head held down (Gaughan et al., 2009).

Single test-day measurements of infrared temperatures on external body surfaces of the cows were obtained as described in Chapter 8. The inner vulva temperature (IVuT, °C) of each cow was obtained by measuring and averaging morning inner vulval lip temperature taken within the period 0600 h to 0700 h and afternoon vulval temperature taken within the period 1500 h to 1600 h using a Combi Infrared Thermometer (Combi Corporation, Japan). When measuring IVuT, the probe of the thermometer was placed between the vulval lips, pointed to the vulval wall at a position of less than 1 cm from outside the vulval lips, and the measurements were triplicated. A FLIR E50 IR camera (FLIR Systems Inc., USA) was used to capture infrared thermal images of the cows at different body surface areas. From the captured images, FLIR Tool software (FLIR Systems Inc., USA) was used to obtain maximum temperatures of the outer vulval surface (OVuT, °C), rear udder (RUdT, °C), inner tail base surface (ITBT, °C), ocular area (EyeT, °C), muzzle (MuzT, °C); armpit (ArmT, °C), paralumbar fossa area (ParT, °C), fore udder (FUdT, °C), forehoof (FHoT, °C), and hind hoof (HHoT, °C). The images of each cow were taken from a distance between 1.0 and 1.5 metres (Montanholi et al., 2015) within 20 minutes after afternoon milking (usually between 1530 h to 1730 h) when the cows were standing and eating. The emissivity of the camera was set at 0.98 (Hoffmann et al., 2013; Sathiyabarathi et al., 2016), the reflected apparent temperature was set at 20°C (Sathiyabarathi et al., 2016), and the relative humidity and environmental temperature were set as the real-time reading of the weather station at the starting time of the first measurement.

210

9.2.2 Genotype data

Samples and genotyping

Genomic data were obtained as described in Chapter 4. Briefly, tail hairs of all lactating cows (n = 345) from the 32 surveyed SDFs were sampled from the tail switch. The samples were preserved in dry and cool room conditions in Vietnam before being exported to the Neogen Australasia laboratory (The University of Queensland, Gatton campus, Australia).

DNA extraction and genotyping were conducted at the laboratory of Neogen Australasia Pty Limited (Neogen Australasia, 2018). The DNA was extracted and purified from follicles of the tail hair samples using Sbeadex Livestock Kits (LGC Genomics GmbH, Germany) following the kit producer procedures adapted by the Neogen Australasia Lab (SOP No. 325 - The University of Queensland). These 345 DNA samples were then genotyped with GeneSeek Genomic Profiler Bovine 50K SNP chips (Neogen Corporation, Lincoln, NE). The SNP chips were prepared using the producer’s assay protocol and scanned using the iScan System (Illumina Inc., San Diego, USA) to generate a genomic dataset.

Quality control

R Software (R Core Team, 2018) was used for quality control of genomic data. Quality control (Gondro et al., 2014) removed 4,364 SNPs with call rates lower than 95%, 1,893 SNPs with minor allele frequency lower than 5%, and 650 SNPs with heterozygosity deviating ± 3 SD from the mean heterozygosity of all SNPs. One cow’s sample with a call rate less than 95% and two cow samples with heterozygosity deviating ± 3 SD from the samples’ heterozygosity mean were removed. In addition, 12 cows from four SDFs with less than five lactating cows per SDF and three cows with missing phenotypic data were excluded. The final data set for analysis included 329 cows from 28 SDFs, genotyped for 36776 SNPs.

9.2.3 Genetic parameter estimates and genomic selection

Genomic heritability estimation

A univariate model was applied to estimate variance components of traits using the R package ‘Sommer’ (Covarrubias-Pazaran, 2019). The model in the matrix notation was:

y = X + Z + Wc + e (1) where: y was the vector of the traits observed;  was the vector of fixed effects, which included age, lactations, days in milk, days in milk squared;  was the vector of additive genetic effects [ ~ N(0, 2 G a)], wherein G was the genomic relationship matrix constructed from genomic information using 211 the first method of VanRaden (2008); c was the vector of the random environmental farm effect (28 2 2 farms) [c ~ N(0, Icσ c)]; e was the vector of residuals [e ~ N(0, Iσ e)]; and X, Z, and W were the incidence matrices of the fixed effects, random additive genetic effects, and random environmental effects, respectively. The additive genetic effect, random environmental farm effect and residuals were assumed to be independently distributed. Variance components were estimated using the restricted maximum likelihood (REML) approach.

For each trait, the data of all cows in the final complete dataset were used to fit Model 1 to estimate the heritability (h2) and calculate the phenotype adjusted for fixed effects (푦*). Heritability was calculated as the ratio of the additive genetic variance to the sum of additive and residual variance: 2 2 2 2 2 h = σ a / (σ a + σ e) (Falconer and Mackay, 1996). A h less than 0.20 was considered low; 0.21 to 0.40, moderate; and above 0.40, high (Bailey, 2014). The phenotype adjusted for fixed effects were calculated as: 푦* = y – Xβ.

Genomic selection

The accuracy and bias of the GEBVs were assessed using 10-fold cross-validation. After quality control, the entire data set of 329 samples was randomly partitioned into ten subsets of equal size. Nine subsets were used as a training set to determine the GEBVs of the retained validation set (10%) using the previously described model. This process was repeated ten times so that each subset was used only once as the validation set. The accuracy (Acc) of GEBVs for each trait was determined by Acc = Cor(퐺퐸퐵푉, 푦*)/√ℎ2, where Cor was the correlation between GEBVs and the phenotype adjusted for fixed effects (푦*) of each validation set, and √ℎ2 was the square root of genomic h2 of the trait (Garcia et al., 2018).

The regression coefficient (Coef) of 푦* on predicted GEBVs was used to measure inflation of GEBVs (Garcia et al., 2018). The bias of GEBVs was then quantified as 1 – Coef (Hsu et al., 2017). A Coef of 1 is theoretically expected for unbiased GEBVs, whereas a Coef > 1 indicates deflation or downward bias of GEBVs, and a Coef < 1 indicates inflation or upward bias (Garcia et al., 2018).

For each trait, the Cor between GEBVs and y*, Acc of GEBVs, and Coef of y* on GEBVs were estimated separately for each validation set and then summarized across ten validation sets to obtain the mean and SE of those estimated values.

9.2.4 Genome-wide association studies

A Univariate single SNP mixed linear model was applied for all GWAS for all traits, using R package ‘Sommer’ (Covarrubias-Pazaran, 2019). The model in the matrix notation was:

212

y1 = mg + X + Z + Wc + e (2)

where: y is the vector of the traits observed, m is a vector of SNP genotypes, and g is a scalar of the associated additive effect of the SNP. Other parameters were as described previously in model 1 for genomic selection.

In the GWAS, the F-test was used to test the significant associations between SNPs and traits. The null hypothesis, H0, was g = 0. To avoid many false-negative results caused by stringent Bonferroni correction, an SNP was considered significant when its P nominal was  5E-05 and considered suggestively significant when its P nominal was 5.0E-05 to 5.0E-04 (Bolormaa et al., 2011; Streit et al., 2013; Do et al., 2018). Suggestive and significant SNPs located within a distance of 1 Mb were grouped into an SNP cluster. All significant SNPs and SNP clusters were presented for each trait and used for candidate gene search. Quantile-quantile (Q–Q) plots were used to visualize the relationship between the expected and observed distributions of -log10(P) of SNP effects to check if potential confounders such as population structure had been adjusted (Schmid and Bennewitz, 2017).

Candidate genes analyses

Bovine ARS-UCD1.2 assembly (https://www.ncbi.nlm.nih.gov/genome/gdv/?org=bos-taurus) was used for both SNP mapping and candidate gene search (Rosen et al., 2018). To assign SNPs to candidate genes, the method described by Cordero-Solorzano et al. (2019) and Do et al. (2018) was used. Firstly, the cow gene dataset (ARS-UCD1.2) was downloaded from Ensembl genes 99 databases (http://www.ensembl.org/biomart/martview/d7d745c7e5cc571815f3e6357b3d7c83) (Zerbino et al., 2018). Then, the downloaded dataset of cow genes was used to search for the candidate genes, which are the nearby genes within a flanking distance of 0.5 Mb up- or downstream from significant SNPs or clusters (Welderufael et al., 2018; Do et al., 2018). The Cattle QTL Database (Cattle QTLdb, https://www.animalgenome.org/cgi-bin/QTLdb/BT/index) was used to search for published QTLs associated with candidate genes and traits of interest (Hu et al., 2019).

9.3 Results 9.3.1 Genomic selection models

Milk production traits

Mean h2 were very low for MILK (0.06); low for ECM (0.17), yPR (0.15), and yDM (0.16); moderate for ECMbw, mPR, mDM, mRE, and yFA (0.25 to 0.33); and high for mFA (0.48) (Table 9.1).

213

The accuracies estimated from the validation process of GEBVs ranged from low for mRE (0.13) to moderate to high for other milk production traits (0.26 to 0.68). However, the GEBVs for all milk production traits were biased. The GEBV for mRE was biased upward (Coef = 0.77), while the GEBVs of all other milk production traits were biased downward (Coef > 1) (Table 9.1). The GEBVs were least biased (Coef = 0.77 to 1.75, most close to 1) for the traits with moderate to high h2 (ECMbw, mFA, mPR, mRE, and yFA); moderately biased (Coef = 2.83 to 3.52) for mDM and the traits with low h2 (ECM, yPR and yDM; and very biased (Coef = 22.09) for MILK.

Table 9.1 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for milk production traits A

B 2 2 2 2 Traits σ a σ c σ e h ± SE Cor ± SE Acc ± SE Coef ± SE MILK 1.09 9.65 17.93 0.06 ± 0.13 0.16 ± 0.04 0.68 ± 0.18 22.09 ± 15.97 ECM 2.39 7.53 12.11 0.17 ± 0.14 0.15 ± 0.03 0.38 ± 0.08 3.29 ± 1.01 ECMbw 0.22 0.15 0.44 0.33 ± 0.15 0.27 ± 0.06 0.48 ± 0.10 1.75 ± 0.49 mFA 0.22 0.04 0.24 0.48 ± 0.14 0.35 ± 0.06 0.50 ± 0.09 1.39 ± 0.26 mPR 0.04 0.02 0.11 0.27 ± 0.14 0.20 ± 0.03 0.38 ± 0.06 1.30 ± 0.21 mDM 0.23 0.25 0.68 0.25 ± 0.14 0.30 ± 0.05 0.61 ± 0.10 3.52 ± 1.02 mRE 189.7 658.1 462.4 0.29 ± 0.15 0.07 ± 0.04 0.13 ± 0.06 0.77 ± 0.31 yFA 0.007 0.009 0.018 0.28 ± 0.15 0.14 ± 0.04 0.26 ± 0.07 1.48 ± 0.61 yPR 0.003 0.012 0.016 0.15 ± 0.14 0.14 ± 0.05 0.35 ± 0.12 3.02 ± 1.42 yDM 0.041 0.111 0.21 0.16 ± 0.14 0.14 ± 0.04 0.35 ± 0.10 2.83 ± 1.07

A 2 2 2 2 σ a, additive genetic variance; σ c, random farm variance; σ e, residual variance; h , heritability; Cor, correlation between GEBVs and adjusted phenotypes; Acc, accuracy of GEBVs; Coef, regression coefficient of phenotype adjusted for fixed effects on GEBVs; SE, standard error. B Abbreviations of traits: MILK, milk yield (kg/cow/d); ECM, energy corrected milk (kg/cow/d); ECMbw, energy corrected milk yield for body weight (BW) (kg/100kg BW/cow/d); mFA, fat concentration (%); mPR (%), protein concentration (%); mDM, milk dry matter concentration (%); mRE, milk electrical resistance; yFA, milk fat yield (kg/cow/d); yPR, milk protein yield (kg/cow/d); yDM, milk dry matter yield (kg/cow/d).

Body conformation traits

Heritability estimates were moderate for BCS (0.35) and high for HG (0.63) and BWc (0.58) (Table 9.2). The GEBVs for BCS, HG, and BW were moderately accurate (Acc = 0.34, 0.36, and 0.34, respectively) with minimal bias (Coef = 1.07, 0.94, and 0.92, respectively).

214

Table 9.2 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for cow body conformation traits A

B 2 2 2 2 Traits σ a σ c σ e h ± SE Cor ± SE Acc ± SE Coef ± SE BCS 0.07 0.03 0.13 0.35 ± 0.14 0.2 ± 0.03 0.34 ± 0.06 1.07 ± 0.21 HG 38.4 28.9 22.9 0.63 ± 0.14 0.28 ± 0.05 0.36 ± 0.07 0.94 ± 0.2 BW 1764.6 1511.0 1269.9 0.58 ± 0.14 0.26 ± 0.05 0.34 ± 0.06 0.92 ± 0.18

A 2 2 2 2 σ a, additive genetic variance; σ c, random farm variance; σ e, residual variance; h , heritability; Cor, correlation between GEBVs and adjusted phenotypes; Acc, accuracy of GEBVs; Coef, regression coefficient of phenotype adjusted for fixed effects on GEBVs; SE, standard error. B Abbreviations of traits: BCS, body condition score; HG, heart girth (cm); BW: body weight (kg).

Heat tolerance traits Heritability estimates were moderate (0.26 to 0.36) for PS, OVuT, RUdT, and FHoT; and low (0.11 to 0.20) for other heat tolerance traits (Table 9.3).

The accuracy estimated from the validation process of GEBVs was very low for HHoT (0.06), low for FUdT (0.12) and moderate for other heat tolerance traits (0.22 to 0.35). The GEBVs for all heat tolerance traits were biased downward (Coef > 1). The GEBVs were least biased for RUdT (Coef = 1.08, most close to zero); moderately biased for other moderately heritable traits including FHoT (Coef = 1.22), PS (Coef = 1.38) and OVuT (Coef = 1.72); and moderately to highly biased (Coef = 1.82 to 7.05) for other heat tolerance traits with low h2.

Table 9.3 Heritability, accuracy of GEBVs, and regression coefficient of phenotype adjusted for fixed effects on GEBVs for heat tolerance traits A

B 2 2 2 2 Traits σ a σ c σ e h ± SE Cor ± SE Acc ± SE Coef ± SE PS 0.03 0.22 0.07 0.29 ± 0.15 0.14 ± 0.06 0.26 ± 0.12 1.38 ± 0.59 IVuT 0.05 0.38 0.32 0.14 ± 0.14 0.12 ± 0.04 0.32 ± 0.10 3.31 ± 1.1 OVuT 0.11 0.59 0.31 0.26 ± 0.17 0.14 ± 0.05 0.27 ± 0.09 1.72 ± 0.68 ITBT 0.06 0.48 0.28 0.17 ± 0.17 0.09 ± 0.07 0.23 ± 0.17 3.03 ± 2.47 EyeT 0.05 0.31 0.25 0.17 ± 0.16 0.14 ± 0.06 0.35 ± 0.15 3.39 ± 1.8 MuzT 0.07 0.96 0.55 0.11 ± 0.16 0.10 ± 0.07 0.29 ± 0.21 5.53 ± 3.93 ArmT 0.18 1.02 0.73 0.20 ± 0.17 0.10 ± 0.07 0.23 ± 0.17 2.67 ± 1.93 ParT 0.11 1.30 0.5 0.19 ± 0.17 0.11 ± 0.08 0.25 ± 0.19 7.51 ± 7.06 FUdT 0.09 0.41 0.42 0.17 ± 0.16 0.05 ± 0.08 0.12 ± 0.19 1.82 ± 1.22 RUdT 0.20 0.50 0.35 0.36 ± 0.17 0.14 ± 0.04 0.23 ± 0.07 1.08 ± 0.34 FHoT 0.77 2.99 1.97 0.28 ± 0.16 0.12 ± 0.07 0.22 ± 0.13 1.22 ± 0.6

215

HHoT 0.32 1.82 1.49 0.18 ± 0.16 0.03 ± 0.11 0.06 ± 0.25 2.61 ± 1.79

A 2 2 2 2 σ a, additive genetic variance; σ c, random farm variance; σ e, residual variance; h , heritability; Cor, correlation between GEBVs and adjusted phenotypes; Acc, accuracy of GEBVs; Coef, regression coefficient of phenotype adjusted for fixed effects on GEBVs; SE, standard error. B Abbreviations of traits: PS, panting score; IVuT, inner vulval lip temperature (oC); OVuT, outer vulval temperature (oC); RUdT, rear udder temperature (oC); ITBT, inner tail base surface temperature (oC); EyeT, ocular area temperature (oC); MuzT, muzzle temperature (oC); ArmT, armpit temperature (oC); ParT, paralumbar fossa area temperature (oC); FUdT, fore udder temperature (oC), FHoT, forehoof temperature (oC); HHoT, hind hoof temperature (oC).

9.3.2 Genome-wide association studies

Milk production traits

Fourteen SNPs were significantly (P  5E-05) and 110 SNPs were suggestively (5E-05 < P  5E-04) associated with milk production traits. Manhattan plots (Figure 9.1 a, c, e, g, i, k, m) present the positions on Bos taurus autosomes (BTA1 to BTA29) and the -log10(P) of the significant SNPs (dots on or above green line) and suggestive SNPs (dots between grey and green lines) for every milk production trait. See Appendix 5 for the full names, positions, and the -log10(P) of these SNPs.

For each trait, some SNPs, either significantly or suggestively associated with the traits, were located less than 1 Mb from each other; thus, they made SNP clusters, which are also the peaks in the Manhattan plots (Figure 9.1). For examples, there is a clear peak of an SNP cluster in BTA14 for mFA. Q-Q plots (Figure 9.1 b, d, f, h, j, l, n) plotted observed against expected -log10(P) for milk production traits. No Q-Q plots for milk production traits showed any systematic deviation from the diagonal (Y = X), indicating that population stratification was not a problem in these GWAS analyses. The slight deviations in the upper right tails from the diagonal of the Q-Q plots line indicates the presence of some SNPs significantly associated with the traits.

216

Figure 9.1 Manhattan and Q-Q plots of association results for milk production traits

Y-axis was defined as -log10(P); X-axis represents the genomic positions and chromosomes; Grey line indicates suggestive threshold (P = 5E-04); Green line indicates significant threshold (P = 5E-05); Purple line indicates Bonferroni-corrected P threshold at a significant level of 0.05 (P = 1E-5.87). See Table 9.1 for abbreviations of the milk production traits.

Table 9.4 presents the position of all significant SNPs and SNP clusters associated with milk production traits, the total number of candidate genes located within a flanking distance of 0.5 Mb

217 up- or downstream from those significant SNPs or SNP clusters, and the gene symbols of the closest genes to the SNPs. For milk yield traits, some significant SNPs were associated with more than one milk yield trait. SNP BovineHD1400017689 at position 61.57 Mb on BTA14 was significantly associated with both MILK (P = 1E-4.56) and suggestively associated with ECM (P = 1E-3.64) and ECMbw (P = 1E-3.38). There were 15 genes located within flanking distance of 0.5 Mb from this SNP, which could be potential candidate genes for MILK, ECM, and ECMbw. Among those 15 genes, genes ATP6V1C1 and AZIN1 were located closest to this SNP (other genes were not reported). Similarly, SNP BovineHD0100039253 (closest genes TMEM108 and 7SK) at position 136.41 Mb on BTA1 was significantly associated with MILK and suggestively associated with ECM and ECMbw. Other three regions associated with both ECM and ECMbw were 10.11 Mb on BTA8 (SNP Hapmap58990-rs29020862), 45.83-45.87 Mb on BTA22 (SNPs BovineHD2200013344 and BovineHD2200013356), and 15.42 Mb on BTA24 (SNP BTA-24502-no-rs). Two SNP clusters, one at 77.34-77.35 Mb on BTA12 and the other at 47.55-47.72 Mb on BTA17 were suggestively associated with ECM. SNP MS-rs383507306 (within gene MSRB3) at 48.33 Mb on BTA5 was significantly associated with ECMbw.

Table 9.4 Single nucleotide polymorphisms (SNPs) and related genes most associated with milk production traits

A B Position Single SNP Total F Max of Trait BTA C D E Closest genes G in Mb or SNP cluster Genes - log10(P) MILK, 14 61.57 BovineHD1400017689 15 ATP6V1C1, (--), AZIN1 4.56, ECM, 3.64, ECMbw 3.38 MILK, 1 136.41 BovineHD0100039253 16 TMEM108, (--), 7SK 4.33, ECM, 3.64, ECMbw 3.47 ECMbw, 8 10.11 Hapmap58990-rs29020862 18 EXTL3, (--), FZD3 4.85, ECM 3.61 ECM 24 15.42 BTA-24502-no-rs 0 , (--), 4.47, ECMbw 3.36 ECMbw, 22 45.83-45.87 BovineHD2200013344, 7 WNT5A, (--), CACNA2D3 3.85, ECM BovineHD2200013356 3.58 ECM 17 47.55-47.72 BovineHD1700013679, 8 U6, (TMEM132D), GLT1D1 4.24 BovineHD1700013751 ECM 12 77.34-77.35 BovineHD1200023274, 13 TMTC4, (--), NALCN 3.67 BovineHD1200023302 ECMbw 5 48.33 MS-rs383507306 20 bta-mir-763, (MSRB3), 4.61 LEMD3 mFA, 14 0.47-1.18 ARS-BFGL-NGS-26520, 84 PPP1R16A, (FOXH1, 6.08, mDM ARS-BFGL-NGS-57820, KIFC2, CYHR1, TONSL, 3.49 Chr14_1653693, VPS28, SLC39A4, CPSF1, Chr14_1699016, ADCK5, SLC52A2,

218

Chr14_1757935, FBXL6, TMEM249, Chr14_2022745 SCRT1, DGAT1, HSF1, BOP1, SCX, MROH1, bta- mir-1839, HGH1, WDR97, MAF1, SHARPIN, CYC1, GPAA1, EXOSC4, OPLAH, SPATC1, GRINA, PARP10, PLEC, bta-mir- 2309, EPPK1, NRBP2, PUF60, SCRIB, IQANK1, FAM83H, MAPK15, CCDC166, ZNF623, bta- mir-193a-2, TSTA3, PYCR3, TIGD5, EEF1D, NAPRT, MROH6, GSDMD, ZC3H3), MAFA mDM, 2 68.01-68.06 BovineHD0200019767, 1 DPP10, (--), 4.49, mPR, BovineHD0200019786 3.90, mFA 3.81 mFA 14 2.14 Hapmap36620- 27 TSNARE1, (--), PTP4A3 4.44 SCAFFOLD50018_7571 mPR 8 9.38 BovineHD0800002969 14 MIR124-1, (--), KIF13B 4.46 mPR 26 22.68-23.04 ARS-BFGL-NGS-101647, 53 KCNIP2, (ARMH3, HPS6, 3.54 ARS-BFGL-NGS-11271 LDB1, PPRC1, SNORD22, NOLC1, ELOVL3, PITX3, GBF1, NFKB2, PSD), FBXL15 mDM 9 22.62 Hapmap52337-rs29022325 10 TPBG, (--), U2 4.36 mRE 25 31.59-31.61 ARS-BFGL-NGS-8024, 4 U6, (--), 5_8S_rRNA 4.74 BovineHD2500008754 mRE 2 48.86-48.99 ARS-BFGL-NGS-113042, 5 ACVR2A, (--), U6 3.33 BovineHD0200014156 A See Table 9.1 for abbreviations of milk production traits. B BTA, Bos taurus autosome. C Position of SNP cluster or single SNP on the bovine ARS-UCD1.2 assembly in megabase pairs (Mb). D Reference SNP identification number (rsID) of each SNP is available in Appendix 5. E Total genes, total number of related genes located within downstream and upstream 500 Mb flanking region of each SNP or SNP cluster. F Closest gene: genes (symbol) located inside SNP cluster or harbouring significant SNPs are inside ( ) and highlighted in bold; “--" indicates no gene was found; two genes nearest to significant SNPs or SNP clusters, but not located inside SNP cluster or harbouring significant SNPs, are not highlighted. G For a single SNP, -log10(P) for that SNP was presented; For SNP clusters, maximum -log10(P) of SNPs within that cluster was presented. For SNP clusters associated with two or more traits, maximum values of - log10(P) were presented respectively for each trait listed in the first column. SNPs or SNP clusters with - log10(P)  4.3 (or P  5E-05) were significant and highlighted in bold.

For milk concentration traits, an SNP cluster (including 5 SNPs ARS-BFGL-NGS-26520, ARS- BFGL-NGS-57820, Chr14_1653693, Chr14_1699016, Chr14_1757935, and Chr14_2022745) at the

219 region 0.47-1.18 Mb on BTA14 was significantly associated with mFA and suggestively associated with mDM (Table 9.4). This SNP cluster is also shown by a peak on BTA14 in the Manhattan plot (Figure 9.1 g). The SNP ARS-BFGL-NGS-57820 (P = 1E-6.08) also passed the Bonferroni-corrected P threshold (P = 1E-5.87 at significant level of 0.05). The region 0.47-1.18 Mb on BTA14 suggested up to 84 candidate genes for mFA with 49 genes (bold, inside parentheses in Table 9.4) located within the region. An SNP cluster (including BovineHD0200019767 and BovineHD0200019786, closest gene DPP10) at position 68.01 to 68.06 Mb on BTA2 was significantly associated with mDM and suggestively associated with both mPR and mFA. In addition, SNP Hapmap36620- SCAFFOLD50018_7571 (closest genes TSNARE1 and PTP4A3) at position 2.14 Mb on BTA14 was significantly associated with mFA. SNP BovineHD0800002969 (closest genes MIR124-1 and KIF13B) at position 9.38 Mb on BTA8 was significantly associated with mPR. SNP Hapmap52337- rs29022325 (closest genes TPBG and U2) at position 22.62 Mb on BTA9 was also significantly associated with mDM.

For mRE, an SNP cluster (including SNPs ARS-BFGL-NGS-8024 and BovineHD2500008754; closest genes U6 and 5_8S_rRNA) at 31.59-31.61 Mb on BTA25 was significantly associated with mRE. Another SNP cluster (including SNPs ARS-BFGL-NGS-113042 and BovineHD0200014156, closest genes U6 and 5_8S_rRNA) at 48.86-48.99 Mb on BTA2 was suggestively associated with this trait.

Some other SNPs were also commonly and suggestively associated with two or more milk production traits (Appendix 5). For example, BovineHD1200009951 at position 33.55 Mb on BTA12 was suggestively associated with ECM (P = 1E-3.78) and ECMbw (P = 1E-4.18). SNPs BTB-00387060 at position 28.87 Mb on BTA9 and ARS-USDA-AGIL-chr21-49531537-000460 at position 49.07 on BTA21 were suggestively associated with mFA and mDM. SNP BovineHD0800025442 at position 84.37 Mb on BTA8 was suggestively associated with mFA and mRE.

Cow body conformation traits

One SNP was significantly, and 33 SNPs suggestively associated with cow body traits (Manhattan plots in Figure 9.2 a, c, e, and Appendix 5). The position of all significant SNPs and SNP clusters (either significant or suggestive) associated with cow body conformation traits and the potential candidate genes for those traits are summarized in Table 9.5. Q-Q plots (Figure 9.2 b, d, f) for body conformation traits indicate that population stratification was not a problem in these GWAS analyses.

220

Figure 9.2 Manhattan and Q-Q plots of association results for cow conformation traits

Y-axis was defined as -log10(P); X-axis represents the genomic positions and chromosomes; Grey line indicates suggestive threshold (P = 5E-04); Green line indicates significant threshold (P = 5E-05). See Table 9.2 for abbreviations of body conformation traits.

For BCS, BovineHD2700004244 (within gene SNX25) at position 15.55 Mb on BTA27 and BovineHD0100018075 (closest genes bta-mir-2285de and IGSF11) at position 63.33 Mb on BTA1 were the SNPs significantly associated with this trait (Table 9.5). Besides, an SNP cluster (including SNPs BovineHD2700004476 and BovineHD2700004480, closest genes F11 and MTNR1A) at position 16.32-16.33 Mb also on BTA27 was significantly associated with BCS and BW.

Since BW was calculated from HG, the patterns of SNP effects for these traits were, unsurprisingly, similar (Figure 9.2 c and e). An SNP cluster (including SNPs BovineHD0200020103 and BovineHD0200020109, closest genes U6 and CCDC93) at position 69.40 – 69.42 Mb on BTA2 was suggestively associated with both HG and BW (Table 9.5). Seven SNPs located on BTA1, BTA5, BTA16, and BTA28 were suggestively associated with both HG and BW (Figure 9.2 c, e, and Appendix 5). No significant or suggestive SNPs associated with HG and BW were associated with BCS and vice versa.

221

Table 9.5 Single nucleotide polymorphisms (SNPs) and related genes most associated with cow conformation traits

A B Position Single SNP Total F Max of Trait BTA C D E Closest genes G in Mb or SNP cluster Genes - log10(P) BCS 27 15.55 BovineHD2700004244 25 CFAP97, (SNX25), 4.81 LRP2BP BCS 1 63.33 BovineHD0100018075 9 bta-mir-2285de, (--), 4.30 IGSF11 BCS 27 16.32-16.33 BovineHD2700004476, 18 F11, (--), MTNR1A 3.99 BovineHD2700004480 HG, 2 69.40-69.42 BovineHD0200020103, 3 (--), U6, CCDC93 3.71, BW BovineHD0200020109 3.83 A See Table 9.2 for abbreviations of the body conformation traits. B, C, D, E, F, G See Table 9.4 for other abbreviations.

Heat tolerance traits

Nine SNPs were significantly, and 195 SNPs suggestively associated with heat tolerance traits (Manhattan plots in Figure 9.3, Figure 9.4, and Appendix 5). The position of all significant SNPs and the SNP clusters (either significant or suggestive) associated with heat tolerance traits and the potential candidate genes for those traits are summarized in Table 9.6. All Q-Q plots in Figure 9.3 and Figure 9.4 for heat tolerance traits indicate that population stratification was not a problem in these GWAS analyses.

Significant regions were found for each heat tolerance trait (Table 9.6). Two SNP clusters at position 40.05-40.08 Mb on BTA11 (including SNPs ARS-BFGL-NGS-112026 and BovineHD1100011765, closest genes CCDC85A and VRK2) and position 14.34-14.45 Mb on BTA20 (including SNPs BovineHD2000004465 and BovineHD2000004488 within gene U4) were associated with PS. A cluster at position 48.95-48.96 Mb on BTA8 (including SNPs BovineHD0800014723 and BovineHD0800014726 within gene TMC1) was associated with IVuT. SNP BovineHD0500026395 at position 92.5 Mb on BTA5 (closest genes 5S_rRNA and LMO3) was associated with OVuT. SNPs Hapmap29069-BTA-45382 (closest genes EPOP and MLLT6) at 39.3 Mb on BTA19 and ARS- BFGL-NGS-17312 (closest genes MEOX1 and SOST) at 43.59 Mb also on BTA19 were associated with MuzT. SNP BovineHD1400023207 (closest genes U6 and SNX16) at 79.78 Mb on BTA14 was associated with ParT. SNP BovineHD2000016438 (within gene TRIO) at 58.79 Mb on BTA20 was associated with FHoT.

222

Figure 9.3 Manhattan and Q-Q plots of association results for the first group of cow heat tolerance traits

Y-axis was defined as -log10(P); X-axis represents the genomic positions and chromosomes; Grey line indicates suggestive threshold (P = 5E-04); Green line indicates significant threshold (P = 5E-05). See Table 9.3 for abbreviations of heat tolerance traits.

223

Figure 9.4 Manhattan and Q-Q plots of association results for the second group of cow heat tolerance traits

Y-axis was defined as -log10(P); X-axis represents the genomic positions and chromosomes; Grey line indicates suggestive threshold (P = 5E-04); Green line indicates significant threshold (P = 5E-05). See Table 9.3 for abbreviations of heat tolerance traits.

224

Two SNP clusters were suggestively associated with more than one heat tolerance trait (Table 9.6). An SNP cluster (including SNPs ARS-BFGL-NGS-91939, BovineHD1400013035 and BovineHD1400013068 within gene ZNF704) at position 44-44.08 Mb on BTA14 was suggestively associated with both OVuT and RUdT. Another SNP cluster (including SNPs ARS-BFGL-NGS- 57209, BovineHD1900009335, BovineHD1900009346, and chr19_31794868 within gene MYOCD) at position 31.13-31.16 Mb on BTA19 was suggestively associated with both OVuT and FHoT. Some single SNPs were also suggestively associated with two heat tolerance traits (Appendix 5). These SNPs included BovineHD2200007708 (at 26.57 Mb on BTA22) suggestively associated with both ITBT and EyeT, Hapmap51348-BTA-90742 (at 74.80 Mb on BTA9) suggestively associated with both MuzT and ParT, and chr7_93243389 (at 90.84 Mb on BTA7) suggestively associated with both ParT and FUdT.

Additional SNP clusters that were suggestively associated with heat tolerance traits (Table 9.6) included: an SNP cluster (including 5 SNPs) at position 45.34-45.38 Mb on BTA5 associated with IVuT, and an SNP cluster (including 5 SNPs) at position 70.84-70.92 Mb on BTA16 associated with MuzT. For other heat tolerance traits, including FUdT, RUdT, HHoT, only single SNPs were suggestively associated with them (Appendix 5).

Table 9.6 Single nucleotide polymorphisms (SNPs) and related genes significantly associated with heat tolerance traits

A B Position Single SNP Total F Max of Trait BTA C D E Closest genes G in Mb or SNP cluster Genes - log10(P) PS 11 40.05-40.08 ARS-BFGL-NGS-112026, 3 CCDC85A, (--), VRK2 4.95 BovineHD1100011765 PS 20 14.34-14.45 BovineHD2000004465, 18 ADAMTS6, (U4), 4.56 BovineHD2000004488 CWC27 IVuT 8 48.95-48.96 BovineHD0800014723, 8 ZFAND5, (TMC1), 4.75 BovineHD0800014726 ALDH1A1 IVuT 5 45.34-45.38 BovineHD0500013119, 18 RAP1B, (--), MDM1 3.71 BovineHD0500013127, BovineHD0500013129 OVuT 5 92.5 BovineHD0500026395 5 5S_rRNA, (--), LMO3 4.31 OVuT, 14 44-44.08 ARS-BFGL-NGS-91939, 12 ZBTB10, (ZNF704), 3.95, RUdT BovineHD1400013035, PAG1 3.51 BovineHD1400013068 OVuT, 19 31.13-31.16 ARS-BFGL-NGS-57209, 10 bta-mir-744, (MYOCD), 3.51, FHoT BovineHD1900009335, ARHGAP44 3.59 BovineHD1900009346, chr19_31794868 ITBT 8 3.31-3.31 BovineHD0800001033 1 , (--), U6 4.30

225

ITBT 5 104.41- BTA-74836-no-rs, 53 VWF, (ANO2), NTF3 3.68 104.43 Hapmap31228-BTA-74842 EyeT 5 7.96-8.09 BovineHD0500002260, 2 NAV3, (--), SYT1 4.14 BovineHD0500002307 EyeT 1 148.17- BovineHD0100043438, 14 bta-mir-2285a, (--), 4.08 148.28 BovineHD0100043469 SETD4 EyeT 12 20.24-20.59 20 MIR15A, (DLEU7, bta- 3.64 BovineHD1200006096, mir-10177, BovineHD1200006202 RNASEH2B, FAM124A), SERPINE3 MuzT 19 39.30 Hapmap29069-BTA-45382 62 EPOP, (--), MLLT6 4.60 MuzT 19 43.59 ARS-BFGL-NGS-17312 91 MEOX1, (--), SOST 4.56 MuzT 16 70.84-70.92 ARS-BFGL-NGS-109968, 19 NSL1, (BATF3, 4.29 BovineHD1600020671, FAM71A), ATF3 BovineHD1600020682, BTB-00659960, chr16_72753272 MuzT 1 61.82-61.82 ARS-BFGL-NGS-91410, 2 7SK, (--), bta-mir-2285de 3.63 BovineHD0100017693 ArmT 2 129.35- ARS-BFGL-NGS-2341, 40 E2F2, (ASAP3, U6, 3.84 129.51 BovineHD0200037810, TCEA3, ZNF436, Hapmap39569-BTA-49765 HNRNPR), HTR1D ArmT 17 70.43-70.44 ARS-BFGL-NGS-62627, 48 U4, (DEPDC5), 3.37 chr17_72557299 YWHAH ParT 14 79.78 BovineHD1400023207 2 U6, (--), SNX16 4.33 FHoT 20 58.79 BovineHD2000016438 7 OTULINL, (TRIO), U6 4.37 FHoT 5 44.91-45.08 23 CPSF6, (CPM, MDM2, 4 BovineHD0500013030, SLC35E3, NUP107), BTA-106792-no-rs RAP1B FHoT 6 11.04-11.08 BovineHD0600003010, 4 UGT8, (--), ARSJ 3.52 BovineHD0600003024 A See Table 9.3 for abbreviations of the body conformation traits. B, C, D, E, F, G See Table 9.4 for other abbreviations.

9.4 Discussion

The sample size in the current study was too small to draw meaningful conclusions in most cases. However, despite the small sample size, some remarkably high h2 estimates were achieved for body conformation traits, and some SNPs and genes for productivity and heat tolerance were indicated.

9.4.1 Genomic selection

To our knowledge, the current study is the first to assess the applicability of GS for milk production, body conformation, and heat tolerance traits in Vietnamese SDF cows. A single test day’s data may be sufficient on which to base a GS program for finding traits in each of these categories. It may also 226 be the first international study to estimate the h2 of heat tolerance traits (PS and infrared temperature of external body surface areas) in dairy cows.

Milk production traits

The h2 estimates for MILK (0.06) were much lower than those reported for MILK by other studies conducted in either tropical or cooler countries. In tropical countries, h2 for MILK has previously been reported as in the range 0.12 to 0.26 in Thai multi-breed dairy cows (Jattawa et al., 2015), Ethiopian Holsteins (Ayalew et al., 2017), Brazilian Holsteins (Petrini et al., 2016), and Brazilian Jerseys (Sabedot et al., 2018); and as from 0.15 to 0.46 in Korean Holsteins (Kim et al., 2009). In cooler countries, h2 of approximately 0.50 have been reported (United Kingdom Holsteins; Eaglen et al., 2013). However, Pegolo et al. (2018) found a very low h2 (0.09) for MILK in Italian Brown Swiss.

More consistent were the h2 comparisons for milk solids between the current and other studies. Heritability estimates for mFA (0.48), mPR (0.27), yFA (0.28), and yPR (0.15) in the current study were higher than the heritabilities ranging from 0.22 to 0.24 for mFA, and 0.14 to 0.25 for yFA in Thai crossbred dairy cattle (Koonawootrittriron et al., 2009; Jattawa et al., 2015; Wongpom et al., 2017). In countries such as Brazil, Korea, Holland, the United Kingdom, and North American, heritability estimates for mFA ranged from 0.15 to 0.50; mPR, from 0.07 to 0.53; yFA, from 0.24 to 0.52; and yPR, from 0.14 to 0.34 (Stoop et al., 2008; Vanraden et al., 2009; Kim et al., 2009; Luan et al., 2009; Toghiani, 2012; Eaglen et al., 2013; Cho et al., 2013; Campos et al., 2015; Petrini et al., 2016).

Few studies have reported h2 estimates for mRE, mDM, yDM, and ECMbw. The mRE is a reciprocal of milk electrical conductivity, which is commonly applied in commercial dairies as an indicator trait for mastitis (Van der Merwe et al., 2005; Hogeveen et al., 2010), and it is suggested as a potential selection trait for the improvement of cow udder health (Norberg et al., 2004a). The h2 estimate for mRE (0.29) in the current study was higher than that reported in Italian Brown (0.23, Mauro et al., 2005) but similar to that in Holstein cattle (0.26 to 0.36, Norberg et al., 2004b). Although seldom reported in the literature, mDM is an important trait for SDFs in Vietnam as the milk-collecting companies use it to define milk price for farmers. Similarly, ECMbw could be an important trait for selecting dairy cows for SDF systems as it indirectly reflects the efficiency of milk production per unit of cow space allowance or cow maintenance (metabolic energy) requirement. The moderate h2 for mDM (0.25) and ECMbw (0.33) suggests them as potential traits for selection programs.

The accuracy of GEBVs for production traits in the current study is consistent with that reported elsewhere. GEBVs accuracy for MILK (0.68) and for mFA (0.35) were higher than the range of 0.32

227 to 0.40 for MILK and 0.24 to 0.34 for yFA previously reported in Thai crossbred dairy cattle (Jattawa et al., 2015; Wongpom et al., 2019). However, the accuracies of GEBVs for all milk production traits in current studies were lower than those previously reported in China, North America, and France (Boichard et al., 2012). In Chinese Holsteins, accuracy ranged from 0.61 to 0.76 for MILK, 0.67 to 0.71 for mFA, 0.49 to 0.59 for mPR, 0.49 to 0.64 for yFA, and 0.63 to 0.75 for yPR (Ding et al., 2013). In North America, accuracy in Holstein bulls ranged from 0.56 to 0.58 for MILK, 0.69 to 0.78 for mFA, 0.62 to 0.69 for mPR, and 0.65 to 0.68 for yFA, and 0.57 to 0.58 for yPR (Vanraden et al., 2009). The lower accuracy and higher bias of GEBVs in the current study compared to other studies are understandable as the accuracy and bias of GEBV depend on the size of the training dataset. For example, it can be estimated that for a trait with an h2 of 0.3, the reference dataset would need around 3,000 individuals to achieve an accuracy of 0.6 and around 10,000 individuals to achieve an accuracy of 0.8 (Hayes et al., 2009a), whereas there were only 329 cows used in the current study. MILK in the current study had a moderate correlation between GEBV and adjusted phenotype of 0.16, but the accuracy of GEBV for this trait was very high (0.68). This very high accuracy was likely inflated by its very low h2 (0.06). The GEBVs for MILK were also very biased (Coef = 22.09). Consequently, to apply GS for MILK in future studies, multiple test day phenotypic data and a much larger sample size of cows will be required. However, from among the other milk production traits examined, the moderate h2 (0.27 to 0.48) and the moderate accuracy (0.26 to 0.50) of GEBVs for mFA, mPR, ECMbw, and yFA suggest that GS using single test day data is possible.

Cow body conformation traits

The moderate h2 (0.35) and the moderate accuracy (0.34) of GEBV for BCS in this study suggests that selection for this trait in Vietnamese SDF cows is possible. The h2 compares favourably with the range reported across many other studies: 0.13 to 0.18 for Canadian Holsteins, Jersey, and Ayrshire cows (Doormaal and Miglior, 2012); 0.27 to 0.28 for United Kingdom Holsteins (Pryce et al., 2001); and 0.29 to 0.58 for Irish Holsteins (Berry et al., 2003a). BCS has been included in the national breeding goals for dairy cattle in Canada (Doormaal and Miglior, 2012) and New Zealand (DairyNZ, 2019). It reflects the cow’s energy balance and welfare status (Roche et al., 2009; Ferguson and Matthews, 2011). High BCS is associated with improved female fertility, disease resistance and longevity (Doormaal and Miglior, 2012; Bastin and Nicolas Gengler, 2013). In general, the breeding goal for BCS is to shift the entire curve of BCS upward in a lactation, especially during mid-lactation (Doormaal and Miglior, 2012; Bastin and Nicolas Gengler, 2013).

The h2 estimate for HG in the current study (0.63) was considerably higher than the range of 0.33 to 0.40 previously reported in Chinese Holstein cows (Zhang et al., 2017), 0.22 in Canadian Holsteins

228

(Abo-Ismail et al., 2017) and 0.20 in United Kingdom Holstein cows (Eaglen et al., 2013). Similarly, the h2 estimate for BW (0.58) in the current study was higher than previously reported in Chinese Holstein heifers (0.20 to 0.47, Yin and König, 2018) and Brazilian Holsteins (0.26, Campos et al., 2015), and at the higher end of the range for Irish crossbred and pure Holsteins (0.39 to 0.60, Berry et al., 2002, 2003). The high h2 and the moderate accuracy (0.34) of GEBV for BW suggest that GS for this trait in Vietnamese SDF cows is possible. In Australia, BW has been included in dairy cow selection programs (Byrne et al., 2016).

Heat tolerance traits

To the best of our knowledge, the current study is the first to estimate the h2 of PS and infrared body temperature in dairy cows. Heritabilities for all the infrared temperatures, except MuzT, ranged from 0.14 to 0.36. These were very similar to what had previously been reported for rectal temperature measured with a traditional thermometer (0.13 to 0.31, Seath, 1947; Dikmen et al., 2012; Otto et al., 2019). The h2 for FUdT (0.35), PS (0.29), FHoT (0.28) and OVuT (0.26) were within the range of 0.21 to 0.68 previously reported for tympanic and vaginal temperatures (Howard et al., 2014).

Whilst there appear to be no previous reports on the h2 of infrared body temperatures for dairy cows, there are some on other animal species. Loyau et al. (2016) reported relatively low infrared thermal temperature h2 measured at the wing, comb, and shank external surfaces of hens ranging from 0.09 to 0.22 at ambient temperature between 18-22°C, and from 0.00 to 0.20 at a higher ambient temperature range of between 28 and 30°C.

Among heat tolerance traits, the moderate h2 and the moderate accuracy (0.23 to 0.27) of GEBVs for RUdT, FHoT, PS, and OVuT suggest that these traits are most meaningful and so should be included in further GS studies.

9.4.2 Genome-wide association studies

The GWAS results in the current study validate the associations of some SNPs and genes with milk production and body conformation that have been well-documented in previous studies on non- Vietnamese cattle in larger herds and cooler countries. Novel SNPs and genomic regions significant for those traditional traits and new heat tolerance traits were also found in the current study.

Milk production traits

The significant region (0.47-1.18 Mb) on BTA14 for mFA and mDM in the current study harbours many genes highly associated with milk production traits, such as DGAT1, GRINA, CYHR1, FOXH1, TONSL, PPP1R16A, ARHGAP39, MAF1, OPLAH, MROH1, ZNF7, ZNF34,

229

MAFA, KIFC2 (Grisart et al., 2004; Jiang et al., 2010; Streit et al., 2013; Minozzi et al., 2013; Nayeri et al., 2016; Do et al., 2018; Atashi et al., 2019). The region 68.01-68.06 Mb on BTA2 associated with mFA, mPR, and mDM in the current study was close to SNP rs110811091 (position 68.20 Mb) associated with MILK in Holsteins (Meredith et al., 2012). The gene DPP10 closest to this region was previously reported as a potential candidate gene for MILK, yPR, and mFA in a large-scale GWAS in US Holsteins (Jiang et al., 2019). The region 2.14 Mb on BTA14 associated with mFA in the current study was located at less than 500 Mb from the genes MAFA and SLURP1, which were reported as candidate genes for 305‐day milk yield and the lactation curve parameters in primiparous and multiparous Holstein cows (Atashi et al., 2019).

To the best of our knowledge, no GWAS have been performed for mRE before the current study. The most significant region associated with mRE was the region 31.59-31.61 Mb on BTA25. This region is close to SNP rs109903786 (31.4 Mb) which has previously been associated with milking speed in Holsteins (Marete et al., 2018). The majority of other GWAS concerning cow udder health focused on somatic cell score, a strait associated with mRE. These GWAS showed that many regions located across many chromosomes, including BTA1, BTA3, BTA4, BTA5, BTA6, BTA9, BTA10, BTA13, BTA15, BTA17, BTA18, BTA20, BTA21, BTA22, BTA23, BTA24, BTA25, and BTA26, were associated with somatic cell score (Wijga et al., 2012; Meredith et al., 2012, 2013; Wang et al., 2015; Welderufael et al., 2018; Kurz et al., 2019). However, all significant and suggestive SNP clusters associated with mRE in the current study are located at unique positions not previously associated with somatic cell score.

Cow body conformation traits

Region 69.40-69.42 Mb on BTA2 associated with HG and BW in the current study is close to QTL:157020 at position 69.55 Mb for yearling BW in Hanwoo cattle (Li et al., 2017) and QTL:106627 at position 66.4-70.0 for weaning BW in Blonde d'aquitaine cattle (Michenet et al., 2016). The Cattle QTL Database (Hu et al., 2019) reports 32 QTLs influencing BCS on BTA3, BTA4, BTA5, BTA6, BTA9, BTA12, BTA13, BTA14, BTA17, BTA19, BTA26, and BTA29. However, we did not find any significant and suggestive regions aligning with the QTL regions reported on the Cattle QTL Database. Oliveira et al. (2019) reported QTL:178032 and QTL:178534 near position 15.55 on BTA27, which were associated with BCS in the current study, but in their study, those QTLs were related to milk yield in Jersey cattle, not BCS.

230

Heat tolerance traits

To the best of our knowledge, there has not previously been any GWAS on PS and infrared temperature of dairy cattle. In the current study, the most significant regions for PS were found to be 40.05-40.08 Mb on BTA11; IVuT, 48.95-48.96 Mb, on BTA8; OVuT, 92.5 Mb on BTA5; ITBT, 3.31-3.31 Mb on BTA8; MuzT, 39.30 and 43.59 Mb on BTA19; ParT, 79.78 Mb on BTA14; and FHoT, 58.79 Mb on BTA20. In a study to identify the SNPs associated with thermoregulation in lactating Holsteins, Dikmen et al. (2015) identified some regions on BTA4 (64.3 Mb), BTA6 (45.1 Mb) and BTA24 (28.9 Mb) associated with rectal temperature; regions of BTA6 (45.1 Mb) and BTA24 (28.9 Mb) associated with respiration rate; and some regions on BTA5 (89.5 Mb), BTA26 (20.3 Mb) and BTA29 (48.3 Mb) associated with sweating rate. However, none of the significant regions associated with PS and body infrared temperature traits in our study overlapped with the regions related to body thermoregulation reported in the study of Dikmen et al. (2015). 9.4.3 Some limitations

A major limitation of this study was that the sample size used was very smaller, at least one to two orders of magnitude lower than that usual for GS and GWAS. Despite that, some remarkable high h2 estimates were achieved, and some SNPs and genes for productivity and heat tolerance were discovered and validated. The small sample size is the likely reason for the high SEs of the heritability estimates (Bolormaa et al., 2011), high SEs of the biases of GEBVs, relatively low accuracies of GEBVs for some traits (Hayes et al., 2009a), and low power of detecting significant SNPs in the GWAS (Spencer et al., 2009; Visscher et al., 2017).

A further limitation is that only single test-day measurements for each trait and THI were obtained, and those single test-day measurements were derived from the cows at a wide range of lactation number and days in milk. Some studies indicate that heritability estimates for milk production traits change widely during (Kim et al., 2009; Krattenmacher et al., 2019) and between lactations (Widyas et al., 2019). For example, h2 estimates for ECM were 0.58 in mid-lactation compared to 0.68 in early lactation in the study of Krattenmacher et al. (2019), and pedigree-based h2 estimates for imported Holsteins in Indonesia were 0.32, 0.42 and 0.63 for first, second and third lactations, respectively (Widyas et al., 2019). The heritability estimate for milk electrical conductivity (a reciprocal trait of mRE) also changed during a lactation and ranged from 0.36 at the beginning and end of the lactation to 0.26 in mid-lactation (Norberg et al., 2004b). Similarly, the heritability estimates of BCS changed widely during a lactation (Pryce et al., 2001; Berry et al., 2003a; Bastin et al., 2010), and so did the heritability estimates of BW (Berry et al., 2003a). In addition, when assessing the accuracies of GEBVs, to avoid inflated accuracy resulting from close family relationships between training and test 231 animals, partitioning animals into training and validation sets should be based on the family so that highly related animals are in the same validation set (Pszczola et al., 2012). However, due to the lack of pedigree data, cows in the current study were randomly partitioned into training and validation sets, and this could be a source of bias for the GEBVs in the current study.

9.5 Conclusion

This study confirmed the potential to apply GS to improve milk production, body conformation, and heat tolerance traits in Vietnamese smallholder dairy herds using single-test day measurements of the traits. Moderate to high h2 and moderate accuracy of GEBVs for some milk production traits (mFA, mPR, ECMbw, and yFA), all body conformation traits (BCS, HG, BW), and some heat tolerance traits (RUdT, FHoT, PS, and OVuT) suggest that GS using single-test day phenotypic measurements could be applied for these traits. For other traits, especially MILK, more test days are required.

The GWAS results validated the importance of SNPs and genes for milk production and body conformation previously reported in non-Vietnamese and large herd studies, indicating consistency of importance across herd size and climate for these. Novel SNPs and genes for heat tolerance on BTA5, BTA8, BTA11, BTA14, BTA19, and BTA20 are also proposed.

A larger sample size is suggested for further GS and GWAS in Vietnamese SDF cows to confirm these results.

232

Chapter 10 General discussion, implications of research, and conclusion

Improving the productivity and welfare of cows in smallholder dairy farms (SDFs) is crucial to the development of modern dairy systems in Vietnam. However, Vietnamese SDF farmers rarely monitor aspects of cow diets, comfort, pedigree, reproduction, or individual milk yield and composition; and so seven studies (Chapters 3 to 9 in this thesis) are described to suggest possible interventions to improve the productivity and welfare of Vietnamese SDF cows and to broadly evaluate opportunities for future research across these interrelated areas. Chapter 3 assessed the production and welfare status of SDF cows across regions and showed that among productivity and welfare variables assessed, excessive heat stress was of most concern, followed by the high number of inseminations per conception and low body condition score. Chapters 4, 5, and 6 focused on examining the limitations imposed by cow genotype, nutrition, and housing conditions. Chapter 7 focused on developing multivariate regression models that examined simultaneously the associations of the breed, nutrition, and housing condition variables identified earlier with the key productivity and welfare indicators identified in Chapter 3. This enabled a ranking of the relative importance of these limitations to indicate heat stress as the most prominent. Since heat stress was highlighted in Chapters 3, 6 and 7, Chapter 8 focused on testing the applicability of infrared temperatures (IRTs) measured at different external body surface areas of SDF cows to rapidly and non-invasively assess the level of heat stress relative to milk yield reduction. The final study (Chapter 9) focused on testing the applicability of genetic technologies for selecting cows already in Vietnamese SDF herds, within the context of the difficulties of gathering the required phenotypic information, to indicate that one test- day worth of data may be sufficient for a meaningful start. Genome-wide association studies (GWAS) were also reported in Chapter 9 to identify the genomic regions associated with milk production and heat tolerance traits in SDF cows.

10.1 Status of production and welfare

Before suggesting any strategies for improvement, it is crucial to measure current levels of productivity and welfare in SDF cows. Chapter 3 showed that the most concerning productivity issues were low energy corrected milk yield (ECM, 15.7 kg/cow/d) compared to milk yield of Holstein herds in commercial dairy farms also in Vietnam, such as farms of TH True milk company (20 kg/cow/d) (Duteurtre et al., 2015), or in temperate countries such as Korea and England (24.66 to 28.6 kg/cow/d; Eaglen et al., 2013; Shin et al., 2017); the relatively low milk fat (mFA, 3.66%) and protein concentration (mPR, 3.27%) compared to those of Holstein cows in temperate countries such as Germany and Holland (3.95 to 4.36% for mFA; 3.34 to 3.51% for mPR; Stoop et al., 2008; Gieseke et al., 2018); and the high number of inseminations per conception in the lowland regions (tAI, 1.9 to 233

3.2 times per conception) compared to that in Holstein herds in developed countries such as Australia and Denmark (1.4 to 1.9 times per conception) (Talukder et al., 2015; Lehmann et al., 2017). The most apparent welfare concerns were high heat stress level (96% of cows were moderately to highly heat-stressed in the afternoon) and low body condition score (BCS, 2.8) compared to the expected BCS range for Holstein cows (2.88 to 3.17; Barberg et al., 2007). Further studies were conducted to determine the most likely reasons for these limitations.

10.2 Breeds, nutrition, and housing

Suboptimal genotype, nutrition, and housing were hypothesised as the reasons for the low milk productivity, low welfare conditions, and the high levels of heat stress observed in Vietnamese SDF cows.

Level of Inbreeding

Inbreeding can cause a reduction in ECM and BW, and is a common concern in Vietnamese SDFs as cows are often bred without pedigree records (Cai and Long, 2002; Vang et al., 2003). However, the Chapter 4 results indicate that inbreeding is not a major problem. The SDF cows had a similar level of genetic diversity as the reference HOL populations from the USA, New Zealand, and France.

Genotypes

Using six reference genomic datasets for breeds, the overall genetic proportion of Holstein (HOL), Jersey (JER), Brown Swiss (BSW), and Zebu (ZEB) in the 344 SDF cows tested across the regions were 85.0%, 6.0%, 5.3%, and 4.5%, respectively (Chapter 4). Approximately 48% of them were pure HOL, 22% were B3HOL_ZEB, 12% were B2HOL_ZEB, 6% were F1HOL_JER, 3% were B1HOL_BSW, 2% were B1HOL_JER, 1% were B1HOL_ZEB, 1% were F1HOL_BSW, and 1% were F1HOL_ZEB. The proportion of high-performance dairy breeds present in the SDF cows was greater than anticipated; 70% were pure (HOL) or very close to pure (B3HOL_ZEB), and up to 95.5% contained a combination of HOL, BSW, and JER genetics. These results were far different from those of other studies based on farmers’ reports which estimated that only 5 to 15% of Vietnamese dairy herds were pure HOL (Tuyen, 2009; National Institue of Animal Science, 2016). The abundance of high-performance dairy breed genes presents a major problem as HOL, BRW, and JER genetics originate from temperate countries, and so in the tropical conditions of Vietnam, they are expected to suffer from heat stress (Polsky and von Keyserlingk, 2017). That expected high level of heat stress was observed in Chapter 3 and represented a likely explanation for the low ECM levels recorded. Chapter 3 also showed that the ECM (19.2 kg/cow/d) of cows in the northern highland (NH, a cool

234 region) was much higher than the ECM (15.6 kg/cow/d) of cows in the northern lowland (NL, a hotter region).

Given that the genetic base of the SDF cows was determined as having a higher proportion of dairy rather than ZEB infused dairy breeds than expected, the hypothesis was re-focused on sub-optimal nutrition as the potential cause of low performance.

Feeding regimes and diets

The diets for lactating cows in all regions were found to be excessive in protein, acid or neutral detergent fibre (ADF, NDF), lignin, and key minerals; but insufficient in dietary net energy concentration for lactation (NEL, Chapter 5). This imbalance could explain the low ECM, BW, and BCS of the cows, as indicated in Chapter 3. Nutritional modelling with PCDairy indicated the crude protein of diets was sufficient for the production of up to 26.8 kg of fat corrected milk per cow per day (FCM, kg/cow/d), whereas dietary NEL was sufficient for the production of a maximum 18.6 kg FCM/cow/d. The 18.6 kg potential was also higher than the 15.7 kg of ECM/cow/d observed. This discrepancy could be a consequence of the additional energy requirement induced by heat stress (Chapter 3) together with the energy cost of dealing with the extra fibre in the diet (West, 2003). Therefore, concerning dietary composition, the lack of NEL and the excess of fibre in the diets were defined as the most likely nutritional limitations to milk production. When cows are not supplied with enough nutrients, especially for milk production, they do not have nutrients for growth and body reserve, and so the lower BW and BCS of SDF cows observed was to be expected (NRC, 2001).

Chapter 5 also indicated that most (29 out of 32) SDFs did not weigh diet ingredients, and none knowingly took differences in the dry matter between feed types into account when offering diets. This could have been the reason for the excessive use of roughage in the lactating cow diets, which in turn reduced dietary NEL and increased dietary fibre concentrations. Napier grass, corn silage, and dry rice straw were the most common roughage types used in the diets, while analysis showed that all the main roughage was excessive in fibre concentrations (ADF, NDF, and lignin).

An especially concerning finding that could limit feed intake and therefore explain the lower milk yields relative to dietary NEL was the limited supply of water for the cows on some farms, particularly those in the southern lowland region (SL). Cows require water for milk production, and their requirement increases when they are heat-stressed (West, 2003). However, the Chapter 5 study showed that only 15 out of all 32 SDFs supplied water ad libitum to the cows.

While analysis of breed compositions did not give clear explanations for the low performance of Vietnamese SDF cows, the analysis of lactating cow diets pointed out some nutritional explanations

235 for that. The next consideration was to focus on understanding the risk of heat stress and cowshed microclimate in relation to cowshed design.

Cowshed design and microclimate within

The high panting score (PS) of the cows reported in Chapter 3 indicated a high level of heat stress across regions, even on the highland farms. Consequently, a detailed study of the environmental drivers of heat stress was implemented (Chapter 6). Ambient temperature (AT, °C), relative humidity (RH, %), wind speed (WS, m/s) within the cowsheds were determined and used to calculate complex environment-based heat stress indicators; temperature-humidity index (THI, calculated from AT and RH), and heat load index (HLI, calculated from AT, RH, and WS). As expected for a tropical climate, the average AT (27.7°C) and RH (81.2%) across regions during daytime were high. In addition, average WS across regions (0.40 m/s) was also extremely low compared to suggested values of between 1 to 3 m/s for optimal cow comfort (Curt and Mcfarland, 2017; The Dairyland Initiative, 2020). High AT, high RH and low WS within cowsheds led to very high THI (79.4 units) and HLI (86.4 units) during the daytime across regions. Based on the guidelines for interpreting HLI (Gaughan et al., 2008) and THI (Zimbleman et al., 2009), those high THI and HLI values indicated that the microclimate within the cowsheds was very hot, and they are consistent with the PS results in Chapter 3.

Altitude, roof height, floor area per cow, and the per cent of shed sides open were the variables most associated with AT, WS, THI and HLI within the cowsheds. Specifically, each 100 m increase in altitude was associated with decreases of 0.4°C in AT, 1.3 unit in HLI, and 0.8 unit in THI. Each metre increase in the eave height was associated with decreases of 0.78°C in AT, 0.14 m/s in WS, 3.31 units in HLI, and 1.42 units in THI. Each m2 increase in floor area per cow tended to be associated with a decrease of 0.12°C in AT, and each 10% increase in cowshed sides open tended to be associated with a decrease of 0.5 unit in HLI. The simultaneous use of roof soakers and fans was also associated with decreases of 1.3°C, 3.2 units, and 2.5 units, respectively in AT, THI, and HLI within cowsheds during the hottest time of the day (1000 h to 1600 h), compared with those at 1100 h. Whilst it was expected that building sheds in the highlands was a better option than building them in the lowlands, all regions could benefit from increasing roof height, floor area per cow, the per cent of shed sides open; and using roof soakers and fans to reduce the risk of heat stress.

10.3 Multivariate analysis of strategies to improve productivity and welfare

Dairy farms are multidimensional and complex systems where the variables, including cow genotypes, nutrition and housing condition, can simultaneously influence cow productivity and

236 welfare, with some variables having stronger effects than the others. Chapters 4, 5, 6 evaluated genotype, nutrition, and the housing conditions of SDF cows separately, whereas Chapter 7 considered these variables simultaneously to estimate the relative weightings of each for associations with milk production, heat stress, BW, and BCS of the cows. Heat stress was identified as the more important priority to address in SDF cows, compared to nutrition or genetics, and this is consistent with a range of multivariate type reviews on high merit cows in country systems at risk of heat stress (West, 2003; Das et al., 2016; Lees et al., 2019).

Nutritional strategies

In Chapter 7, it was indicated that increasing dry matter intake as a percentage of BW (DMIbw, % BW), dietary dry matter concentration (DM, % as fed) and dietary fat concentration (% DM), and simultaneously reducing dietary lignin concentration (% DM) could each increase milk production and reduce heat stress in SDF cows. Increased DMIbw was also associated with decreased tAI. Decreased dietary lignin was associated with decreased PS. In diets, fat is an energy source, and excessive lignin is commonly associated with excessive dietary fibre (NDF, ADF). Thus, the results in Chapter 7 are consistent with those in Chapter 4, which indicates that the diets are insufficient in energy and excessive in fibre.

During heat stress, cows' feed intake is often decreased, and cows’ energy requirement is increased (West, 2003). Thus, some reviews recommend the following strategies when feeding cows during hot weather: (1) ensuring sufficient drinking water, (2) increasing dietary fat and starch concentrations to increase dietary energy density to compensate for reduced dry matter intake, (3) feeding lower fibre diets to decrease the extra dietary heat increment from the fermentation of fibre compared to starch or fat, (4) avoiding excesses of dietary crude protein to minimize metabolic heat production when excreting excessive nitrogen as urea (West, 2003; Das et al., 2016; Lees et al., 2019). The nutritional limitations found in Chapter 5 and the strategies suggested in Chapter 7 align with each of these.

Chapter 7 results in conjunction with the dietary balance results of Chapter 5 indicate the opportunity to rebalance diets in all regions. For example, if considering a target of 25 kg FCM/cow/d per cow, current dietary CP concentrations (16.5% DM, Chapter 5) are more than adequate (PCDairy indicated a target of 15.7%). Similarly, Ca, P, and Na concentrations (0.78, 0.44, and 0.30, respectively) were higher than the targets of 0.58, 0.37, and 0.18% DM, respectively, required for 25kg of FCM. Some downward adjustment could be considered for these minerals to minimize the economic loss and environmental pollution caused by the excretion of excess, or they could be left as is if it is easier for the farmer to do that. However, consideration should be given to increasing the current NEL concentration (1.4 Mcal/kg DM), which was much lower than the PC Dairy indicated target of 1.59 237

Mcal/kg DM. To meet the NEL requirement but still ensure rumen health, current dietary fat concentrations (3.6% DM) could be increased to a maximum of 6.0% DM as suggested by NRC (2001) and PCDairy (Robinson and Ahmadi, 2015). In contrast, the current ADF concentration (26.2 to 28.2% DM) should be decreased to a range of 19 to 21% DM, and the current NDF concentration (43.9 to 47.4% DM) should be decreased to the range of 27 to 32% DM as suggested by Encyclopedia of Dairy Sciences (Lean, 2011a) and PCDairy (Robinson and Ahmadi, 2015). Current dietary lignin concentration (5.8 to 6.8% DM) should be decreased as much as possible. As discussed previously, the roughage that SDF farmers using such as Napier grass, corn silage, and dry rice straw, provided was all excessive in fibre. Thus, the amount of these components of the diet should be decreased. Also, further studies are required to reduce fibre concentrations in roughage, especially corn silage.

Heat stress abatement strategies

Improving the microclimate within the cowsheds was indicated as a key priority in Chapter 7. High altitude regions were associated with higher ECM, lower PS, and lower tAI of the cows compared to the low altitude regions. Similarly, the increased eave and ridge height of the cowshed roof were associated with increased ECM, mFA, mDM, BW, and BCS and associated with decreased PS. Increased floor area per cow was associated with increased MILK and decreased PS. Increased percentage of shed side open was associated with increased mFA and mDM and decreased PS. These results were expected as Chapter 6 indicated that increases in altitude, eave and ridge roof heights, floor area per cow, and percentage of shed side open were associated with decreases in AT, THI, and HLI and an increase in WS. When AT, THI, and HLI are decreased, cows are expected to be less heat-stressed, and as a result, the feed intake, milk production, and fertility of the cows are expected to improve (Preez et al., 1990; Ravagnolo et al., 2000; Bouraoui et al., 2002; Könyves et al., 2017). It will be difficult for SDF farmers to build cowsheds in high altitude regions as they would need to buy new farms. However, increasing eave and ridge roof height, increasing floor area per cow, increasing the percentage of shed side open, and installing fans and roof soakers are the more readily adoptable strategies that SDF farmers across regions can apply to improve microclimate within the cowsheds.

Chapters 3 and 6 indicated that cows in all regions, including the highlands, were suffering from heat stress; thus, cowsheds in all regions could be modified to advantage cow welfare and productivity. Current floor area per cow in SH, NL, and SL (5.2 to 7.5 m2/cow) could be increased to approximately 8 to 11 m2/cow, as recommended by global welfare regulations (NFACC, 2009; Moran and Doyle, 2015b; Red Tractor Assurance for Farms, 2017; RSPCA, 2018). Current ridge heights (3.6 m) and eave heights (2.8 m) across regions were considered low and should be increased. Although no studies

238 that have tested for relationships between optimum roof height and productivity or welfare parameters appear to be available, target optimal roof heights for SDFs of 5 m for eave height and 9 m for ridge height, to ensure sufficient ventilation and convenience for machinery, have been suggested (Moran and Chamberlain, 2017). Currently, only cowsheds in NL have fans, while cowsheds in other regions have almost no fans, and only 75% of the cowshed sides in the lowland regions are open. The limited use of fans and the opening of shed sides across regions could be reasons for very low WS across regions (0.02 m/s). Thus, SDFs across regions should use more fans and increase areas of shed sides open to improve current WS to suggested targets of between 1 to 3 m/s (Curt and Mcfarland, 2017; The Dairyland Initiative, 2020). In addition, for the lowlands, until the late afternoon (1800 h), the accumulated heat load units (AHLU, 83.7 units in NL and 93.1 units in SL) were much higher than the highest AHLU threshold (50 units) suggested by Gaughan et al. (2008). Therefore, for the SDFs in the lowlands, the heat abatement methods (e.g. fans and roof soakers) should not just be applied during the hot noontime but also should be applied in the late afternoon and night time to cool the cows down completely before a new day.

Besides modifying cowsheds to improve microclimate, other welfare concerns were also noted. Tie- up compared to loose housing occurred in all SDFs in NL and SL, and the lack of suitable flooring (e.g. rubber matting compared to bare concrete) was an additional issue in cowshed design that could easily be improved. Thus, building loose housing instead of tie-up housing sheds and the supply of suitable bedding materials for cows instead of leaving the cows lying on the bare concrete floor for extended periods are recommended.

10.4 Infrared technology for assessing heat stress

Chapters 3 and 6 indicated that of the issues measured in SDFs, heat stress was the most significant associated with productivity. Heat stress can be assessed using either animal-based indicators such as PS (as in Chapter 3), rectal temperature and respiration rate, or environment-based indicators such as THI and HLI (Chapter 6). However, environment-based indicators such as THI and HLI do not directly reflect the physiological changes experienced by the cows and, therefore, cannot be used as traits for the selection of heat-tolerant cows. Cow PS, on the other hand, directly reflects the heat stress changes experienced by the cows and can be considered as a trait, but measurement of PS is often subjective. Therefore, the study in Chapter 8 was conducted to test whether infrared temperatures (IRTs) measured at different areas of the cow body could be used as heat stress indicators. If IRTs show associations with known variables driving heat stress at least as good as the panting score, they should be further assessed as traits for selecting heat-tolerant cattle in future experiments. The advantage of IRTs compared with traditional temperature measurements, such as 239 rectal temperature, is that the measurement of IRTs is non-invasive, requiring little or no direct contact with the animal (Hoffmann et al., 2013; Tattersall, 2016). The non-contact and non-invasive attributes of IRTs are advantageous for their use on Vietnamese SDFs because Vietnamese SDFs often do not have the facility to restrain cows to, for example, carefully take a rectal temperature with a traditional rectal thermometer.

Chapter 8 tested the association of the environment-based heat stress indicator (HLI) with IRT at the inner vulval lip measured by a relatively simple and cheap infrared thermometer, and the associations of HLI with IRTs at the outer vulval surface, inner tail base surface, ocular area, muzzle, armpit area, paralumbar fossa area, fore udder, rear udder, forehoof, and hind hoof measured by a more expensive thermal camera. Associations of those IRTs were also assessed with energy corrected milk yield (ECM). Although all of the abovementioned IRTs tested were found to be of use in assessing the heat stress level of the SDF cows, the best were found to be the IRTs at the inner vulval lip, outer vulval surface, and inner tail base surface. For SDF farmers, using IRT of inner vulval lip could be the easiest and cheapest approach because IRT of inner vulval lip is measured using infrared thermometers, which are cheap to buy and use, whereas IRTs of other areas are measured using thermal cameras, which are expensive, complicated and mainly suitable for detailed research studies.

Further discovering the genetic background of IRTs and PS, the results of GWAS (Chapter 9) indicated that the single nucleotide polymorphisms (SNPs) and genes associated with IRTs and PS were located at the Bos taurus chromosome (BTA) 5, BTA8, BTA11, BTA14, BTA19, and BTA20.

10.5 Breeding and selection strategies

Backcrossing current dairy herds with imported dairy sires of high genetic merit, mainly HOL and directly importing pure HOL heifers, are common genetic strategies for improving milk yield in Vietnam (Trach, 2004; Lam et al., 2010; Tue et al., 2010; Moran, 2015c; Phong and Thu, 2016) However, the results of Chapters 7 and 8 indicate an opportunity for change. As shown in Chapter 8, increasing the concentration of HOL genetics to B3HOL_HOL or purer did not show an improvement in ECM. By contrast, B3HOL_HOL and purer HOL cows yielded the lowest ECM and were most heat-stressed, as indicated by the highest PS compared to other crossbreeds. Since B3HOL_ZEB or pure HOL accounted for 70% of tested SDF cows (Chapter 4), this high per cent of B3HOL_ZEB and pure HOL might limit the milk yield of the Vietnamese SDF herds.

HOL_BSW and HOL_ JER crossbreeds appear to be more suitable for Vietnamese SDFs than ZEB_HOL and purer HOL in terms of milk production and heat tolerance. Chapter 8 indicated that the heat stress level, based on PS, of F1HOL_BSW, B1HOL_BSW, F1HOL_JER, and B1HOL_ZEB

240 was lower than that of B3HOL_ZEB and pure HOL; and ECM of F1HOL_BSW and F1HOL_JER were higher than that of B1HOL_ZEB and B3HOL_ZEB. In Chapter 7, the higher genetic proportion of BSW in herds was associated with increases in MILK, ECM, and BCS, which are all desirable. In contrast, a genetic proportion of ZEB in the herds was associated with decreases in mPR and BW. An increased genetic proportion of JER in the herds was associated with decreased BW and BCS of the cows. All of these results suggest that HOL_BSW crossbreeds and, to a lesser extend, HOL_ JER crossbreeds might be more suitable for Vietnamese SDFs than ZEB_HOL crossbreeds and purer HOL in terms of milk production, heat tolerance, BW and BCS. This is consistent with the findings of other researchers who suggested that pure BSW and F1HOL_BSW were more adaptable to the subtropical conditions than pure HOL (Johnson and Vanjonack, 1976; West et al., 2003; El-Tarabany et al., 2017). However, it should be noted that the current estimates of breed composition effects have low accuracy due to the very small numbers of HOL_BSW, HOL_ JER crossbreeds that were observed in this study. Thus, further studies including a greater number of HOL_BSW and HOL_ JER crossbreed samples are needed to confirm the effects of breed composition.

Genomic selection (GS) is a long-term sustainable strategy to improve the productivity and welfare of cows. Chapter 9 focussed on testing the applicability of GS on Vietnamese SDF cows to select for milk production traits (MILK, ECM, mFA, mPR, mDM), body conformation traits (BW, and BCS), ECM adjusted for BW (ECMbw, kg ECM/100 kgBW/d), and heat tolerance traits (PS and IRTs) using only single test day measurements of each trait. Estimated h2 was high (> 0.40) for mFA and BW; moderate (0.21 to 0.40) for mPR, mDM, ECMbw, BCS, PS, and IRTs at the rear udder, outer vulval surface, and forehoof; low (0.11 to 0.20) for other IRTs; and very low (0.06) for MILK. Accuracy of genomic estimated breeding values (GEBVs) was low (< 0.13) for IRTs at hind hoof and fore udder; and moderate to high (0.22 to 0.68) for all other traits. The moderate to high heritabilities (0.26 to 0.58) and moderate accuracy (0.23 to 0.50) of GEBVs for some milk production traits (ECMbw, mFA, mPR, and yFA), all body conformation traits (BCS and BW), and some heat tolerance traits (PS and IRTs at the rear udder, outer vulval surface, and forehoof) indicate that the selection of SDF cows for those traits could be implemented with minimal phenotypic data collection.

The advantage of the GS approach compared to the breeding approach in Vietnam is that GS does not require knowledge of the pedigree of the cows; pedigree is often unavailable in Vietnam. According to the General Statistics Office of Vietnam (2017a), the total number of dairy cows in Vietnam was 301,649 in 2017. Assuming that the cows in the Chapter 4 study were representative of the entire Vietnamese dairy herd, then the majority of the genetic proportion of Vietnamese dairy cows were dairy breeds, including HOL (85.0%), JER (6.0%), and BSW (5.3%); but Vietnamese

241 dairy cows also had a small genetic proportion of ZEB (4.5%). These cows represent a diverse genetic base for selecting cows that are both high producing and have some adaptation to Vietnamese conditions. What Vietnam needs is an internal selection programme rather than the most current strategies which promote genetic development based on the continued importation of pregnant heifers, or bulls, semen or embryos (Trang, 2019; Tue, 2020).

10.6 Limitations and future directions 10.6.1 Limitations

Firstly, across all the research chapters, the collected data were only for a single day, whereas, for example, weather conditions might vary widely between days. Similarly, the number of cows used in the genetic studies was extremely small. Usually, a dramatically larger sample size is employed for GS and GWAS studies. For example, for a trait with a heritability equal to 0.3, the reference dataset would need around 3000 individuals to get an accuracy of 0.6 and around 10000 individuals to get an accuracy of 0.8 (Hayes et al., 2009a). The small sample size often decreases the accuracy and increases the bias of GEBVs, and decreases the power of GWAS in detecting the significant SNPs discussed in Chapter 9 (Comeron et al., 2003; Klein, 2007; Hayes et al., 2009a; Spencer et al., 2009; Visscher et al., 2017). Currently, a project to implement GS on Vietnamese SDF herds, with an increased sample size, is under development because of the potential shown by the current study.

Secondly, the results determined in Chapters 7, 8, and the GWAS in Chapter 9 were associative; the important parameters that were defined might not be causal (Lucas and Mcmichael, 2005; Barratt et al., 2009; Glass et al., 2013). Therefore, randomized experimental studies are suggested, when possible, to confirm the significant associations found in our multivariate regression models.

Thirdly, in Chapter 5, the measurements of the roughage intake might be biased, as farmers usually offered roughage to lactating cows and dry cows together, and lactating cows might consume more than dry cows. The intake roughage was calculated by dividing the total amount of roughage supplied to all cows, including lactating cows and dry cows. In addition, on three SDFs, the roughage was given to cows during the night when there was no observer; thus, those amounts were not directly measured by the research team. Instead, in those instances, the amount was taken to be the estimate given by the farmer. Also, not all feed types that farmers used were analysed via the DairyOne laboratory for nutritive value; assumed values were used for some. These variables might affect the accuracy of the roughage intake measurements and nutrient concentration calculations.

Finally, the main aspect of this research project was to determine the severity and consequences of heat stress in SDF cows. To do that, the original plan was to conduct the project at the hottest and

242 most humid time of the year, the summer of 2017. However, obtaining approvals, especially from local government authorities in Vietnam, took a longer time than expected, so the data collection was delayed until autumn 2017. Therefore, the heat stress aspect of the study did not reflect the hottest periods of the year. The microclimate is expected to be even more extreme, and the cows are expected to be more heat-stressed in the summer. However, in winter and spring times of the year, when the microclimate is expected to be cooler, and the SDF cows might be less stressed.

10.6.2 Future directions

Future genomic research should aim to increase the sample size of the training dataset for future GS studies to increase the accuracy and decrease the bias of GEBVs for milk production, body conformation, and heat tolerance traits so that GS for those traits can be successfully implemented. This increased sample size would also benefit future genome-wide association studies to increase the ability to detect significant SNPs for heat tolerance traits.

Additional genetic research could be to conduct a randomized experimental study to test whether increasing the proportion of BSW but not ZEB, in individual cows, through differential use of artificial insemination will improve milk production in the offspring.

Future nutritional research could be to conduct randomized experimental studies to test whether a decrease in dietary fat and lignin can reduce cow panting score and IRT measures such as IRTs at inner vulval lip, outer vulval surface, inner tail base surface, or rear udder, without also reducing milk yield and fat concentration. Such research would also validate the associations between HLI with IRTs of inner vulval lip, outer vulval surface, inner tail base surface, and the associations between IRTs and ECM. Future nutritional research could also examine the reasons for the low fibre concentrations of corn silage, find alternative roughage to Napier grass and rice straw with low fibre concentrations, discover types of oil-rich meal that should be added to the diets, and examine the effects of limited water supply on SDF cows and suggest efficient water supply methods.

For cow housing, even in highland farms, randomized experimental designs could be used to confirm optimum eave and ridge roof heights for SDF cowsheds to ensure measurable improvements in airflow for optimal evaporative cooling of cows and reductions in relevant IRT of external body surfaces of the cows and improved productivity.

10.7 Conclusions

The most important production and welfare problems defined in this thesis were extreme heat stress and relatively low ECM yield. Within those were the problems of low BW and BCS, high number of inseminations per conception, inadequate net energy and excess fibre in diets, and a lack of drinking 243 water. Excessive cow discomfort was also due to inadequate floor allowance and the lack of opportunity to move freely about pens due to the widespread use of tie-up rather than free-stall restraint systems.

The majority of SDF cows were surprisingly close, genetically, to pure HOL. They also retained a similar level of genetic diversity as the reference HOL populations sampled from the USA, New Zealand, and France. The genetic proportion of HOL, JER, BSW, and ZEB, averages across all herds, were 85.0%, 6.0%, 5.3%, and 4.5%, respectively.

An increasing genetic proportion of BSW in the herds could be more effective for improving milk yield, BW, and BCS of SDF cows than increasing the genetic proportion of ZEB. F1HOL_BSW and F1HOL_JER were more productive and less heat-stressed than B3HOL_ZEB and pure HOL under SDF conditions. Moderate to high h2 and moderate accuracy of GEBVs for key milk production traits (ECMbw, mFA, mPR, yFA), body conformation traits (BCS and BW), and heat tolerance traits (RUdT, FHoT, PS, and OVuT) suggest that GS using single-test day phenotypic measurements could be acceptable for the improvement of cow welfare and productivity in SDFs.

The GWAS results validated the importance of previously proposed SNPs and genes for milk production and body conformation. Novel SNPs and genes were also proposed for heat tolerance (at BTA5, BTA8, BTA11, BTA14, BTA19, and BTA20).

The lactating cow diets in all regions were generally low in net energy per kg of DM. Dietary strategies to improve milk production and welfare of cows include increasing the fat concentration in concentrates, decreasing the use of high lignin/NDF/ADF forage, and supplying drinking water ad libitum in purpose-built troughs. Drinking water should not be offered in the same trough as concentrates.

Compared to nutrition and genetic manipulations, the amelioration of heat stress was arguably the most important priority identified in this thesis to improve dairy cow productivity and welfare in Vietnam. Increasing shed roof height, floor allowance, percentages of shed sides open, and using roof soakers and the fans should all be implemented as soon as possible, while further well-designed experiments are recommended to be undertaken simultaneously to make sure that the very best interventions are applied in the long term. The IRTs that showed the best potential, compared to PS, to indicate milk yield decline in response to heat stress were IVuT, OVuT and ITBT.

244

List of References Abdel-Azim G 2016. R package ‘gdmp’ version 0.1.0: Genomic data management. Retrieved on 29 December 2017, from https://cran.r-project.org/web/packages/gdmp/index.html. Abdela N 2016. Sub-acute ruminal acidosis (SARA) and its consequence in dairy cattle: a review of past and recent research at global prospective. Achievements in the Life Sciences 10, 187–196. Abo-Ismail MK, Brito LF, Miller SP, Sargolzaei M, Grossi DA, Moore SS, Plastow G, Stothard P, Nayeri S and Schenkel FS 2017. Genome-wide association studies and genomic prediction of breeding values for calving performance and body conformation traits in Holstein cattle. Genetics Selection Evolution 49, 29. Adesogan AT, Arriola KG, Jiang Y, Oyebade A, Paula EM, Pech-Cervantes AA, Romero JJ, Ferraretto LF and Vyas D 2019. Symposium review: Technologies for improving fiber utilization. Journal of Dairy Science 102, 5726–5755. Agenäs S, Heath MF, Nixon RM, Wilkinson JM and Phillips C 2006. Indicators of undernutrition in cattle. Animal Welfare 15, 149–160. Ahmad SF, Panigrahi M, Chhotaray S, Pal D, Parida S, Bhushan B, Gaur GK, Mishra BP and Singh RK 2020. Revelation of genomic breed composition in a crossbred cattle of India with the help of Bovine50K BeadChip. Genomics 112, 1531–1535. Ahmadi A, Robinson P H and Chilibroste P 2013. PC dairy enterprise: a computer package for formulation and evaluation of rations for dairy cattle. In 2013 Cenference: sustainable agriculture through ICT innovation, Turin, Italy, 24–27 June. Akanno EC, Schenkel FS, Sargolzaei M, Friendship RM and Robinson JAB 2014. Persistency of accuracy of genomic breeding values for different simulated pig breeding programs in developing countries. Journal of Animal Breeding and Genetics 131, 367–378. Alexander DH, Novembre J and Lange K 2009. Fast model-based estimation of ancestry in unrelated individuals. Genome Research 19, 1655–1664. Alexander DH, Shringarpure SS, Novembre J and Lange K 2015. Admixture 1.3 Software Manual. Retrieved on 21 April 2019, from http://software.genetics.ucla.edu/admixture/admixture- manual.pdf. Alfonzo EPM, Barbosa da Silva MVG, dos Santos Daltro D, Stumpf MT, Dalcin VC, Kolling G, Fischer V and McManus CM 2016. Relationship between physical attributes and heat stress in dairy cattle from different genetic groups. International Journal of Biometeorology 60, 245– 253.

245

Almaw G, Zerihun A and Asfaw Y 2008. Bovine mastitis and its association with selected risk factors in smallholder dairy farms in and around Bahir Dar, Ethiopia. Tropical Animal Health and Production 40, 427–432. Ambazamkandi P, Thyagarajan G, Sambasivan S, Davis J, Shanmugam S and Joseph BA 2015. Shelter design for different livestock from a climate change perspective. In Climate change impact on livestock: Adaptation and mitigation (eds. V. Sejian, J. Gaughan, L. Baumgard and C. Prasad), pp. 399–424. Springer India, New Delhi, India. Aminah A and Chen CP 1989. Future prospects for fodder and pasture production. In Feeding dairy cows in the tropics (eds. Andrew Speedy and René Sansoucy). Food and Agricultural Organization of the United Nations, Rome, Italy. An T 2019. Development of small and medium-scale dairy farming with sustainable efficiency. Vietnam Livestock Magazine. Retrieved on 6 July 2019, from http://nhachannuoi.vn/phat- trien-chan-nuoi-bo-sua-quy-mo-vua-va-nho-hieu-qua-ben-vung/. Andersson BE and Jóhasson J 1993. Temperature regulation and environmental physiology. In ‘Dukes’ physiology of domestic animals (eds. M.J. Swenson and W.O. Reece), pp. 886–895. Cornell University Press, Ithaca, New York, USA. Angrecka S and Herbut P 2016. Impact of barn orientation on insolation and temperature of stalls surface. Annals of Animal Science 16, 887–896. Angrecka S, Herbut P, Nawalany G and Sokołowski P 2017. The impact of localization and barn type on insolation of sidewall stalls during summer. Journal of Ecological Engineering 18, 60–66. Appuhamy JADRN, Judy J V., Kebreab E and Kononoff PJ 2016. Prediction of drinking water intake by dairy cows. Journal of Dairy Science 99, 7191–7205. Arieli A, Rubinstein A, Moallem U, Aharoni Y and Halachmi I 2004. The effect of fiber characteristics on thermoregulatory responses and feeding behavior of heat stressed cows. Journal of Thermal Biology 29, 749–751. Ashbaugh HRBA 2010. A Descriptive survey of dairy farmers in Vinh Thinh commune, Vietnam. The Ohio State University. Asia Technology Service Company. Dairy cow tape measure. Retrieved on 18 August 2017, from http://channuoibo.vn/4-Thiet-bi-Chan-nuoi/Thuoc-do-khoi-luong-gia-suc/Thuo-c-do-kho-i- luo-ng-bo-su-a.html. Atashi H, Salavati M, De Koster J, Ehrlich J, Crowe M, Opsomer G, Consortium the G and Hostens M 2019. Genome-wide association for milk production and lactation curve parameters in

246

Holstein dairy cows. Journal of Animal Breeding and Genetics 137, 292–304. Aulchenko YS 2015. GenABEL tutorial. Retrieved on 30 September 2018, from http://finzi.psych.upenn.edu/library/GenABEL/doc/GenABEL-tutorial.pdf. Australian Department of Agriculture. Animal Welfare in Australia. Retrieved on 29 February 2020, from https://www.agriculture.gov.au/animal/welfare/animal-welfare-in-australia. Avendaño-Reyes L, Fuquay JW, Moore RB, Liu Z, Clark BL and Vierhout C 2010. Relationship between accumulated heat stress during the dry period, body condition score, and reproduction parameters of Holstein cows in tropical conditions. Tropical Animal Health and Production 42, 265–273. Ayalew W, Aliy M and Negussie E 2017. Estimation of genetic parameters of the productive and reproductive traits in Ethiopian Holstein using multi-trait models. Asian-Australasian Journal of Animal Sciences 30, 1550–1556. Baêta FC, Meador N, Shanklin and Johnson H 1987. Equivalent temperature index at temperatures above the thermoneutral for lactating dairy cows. In 1987 Summer Meeting of the American Society of Agricultural Engineers, pp. 1–21. The American Society of Agricultural Engineers, Joseph, Michigan, USA. Bailey E 2014. Heritability and the equine clinician. Equine Veterinary Journal 46, 12–14. Ban J van den 2014. Compound feed and dairy sector growth: A feed industry perspective. In Dairy asia: Towards sustainability, p. 31. Food and Agriculture Organization of the United Nations, Bangkok, Thailand. Bangdiwala SI and Shankar V 2013. The agreement chart. BMC Medical Research Methodology 13, 97. Barberg AE, Endres MI, Salfer JA and Reneau JK 2007. Performance and welfare of dairy cows in an alternative housing system in Minnesota. Journal of Dairy Science 90, 1575–1583. Barendse W 2017. Climate adaptation of tropical cattle. Annual Review of Animal Biosciences 5, 133–150. Barratt H, Kirwan M and Shantikumar S 2009. Epidemiology: association and causation. Retrieved on 18 August 2020, from https://www.healthknowledge.org.uk/public-health- textbook/research-methods/1a-epidemiology/association-causation. Bartoń K 2019. R Package ‘MuMIn’ version 1.43.15: Multi-Model Inference. Retrieved on 19 December 2019, from https://cran.r-project.org/web/packages/MuMIn/MuMIn.pdf. Bastin C, Loker S, Gengler N, Sewalem A and Miglior F 2010. Genetic relationships between body

247

condition score and reproduction traits in Canadian Holstein and Ayrshire first-parity cows. Journal of Dairy Science 93, 2215–2228. Bastin C and Nicolas Gengler 2013. Genetics of body condition score as an indicator of dairy cattle fertility: a review. Biotechnology, Agronomy, Society and Environment 17, 64–75. BC SPCA 2018. Standards for the raising and handling of dairy cattle. The British Columbia Society for the Prevention of Cruelty to Animals, Vancouver, Canada. Beede DK and Collier RJ 1986. Potential nutritional strategies for intensively managed cattle during thermal stress. Journal of Animal Science 62, 543–554. Bennewitz J and Meuwissen TH 2005. A novel method for the estimation of the relative importance of breeds in order to conserve the total genetic variance. Genetics Selection Evolution 37, 315– 337. Berman A, Folman Y, Kaim M, Mamen M, Herz Z, Wolfenson D, Arieli A and Graber Y 1985. Upper critical temperatures and forced ventilation effects for high-yielding dairy cows in a subtropical climate. Journal of Dairy Science 68, 1488–1495. Bernabucci U, Biffani S, Buggiotti L, Vitali A, Lacetera N and Nardone A 2014. The effects of heat stress in Italian Holstein dairy cattle. Journal of Dairy Science 97, 471–486. Bernard JK 2015. Nutritional challenges of lactating dairy cattle in a tropical climate. In The 6th International Seminar on Tropical Animal Production Integrated Approach in Developing Sustainable Tropical Animal Production, pp. 12–17. Yogyakarta, Indonesia. Berry DP, Buckley F and Dillon P 2007. Body condition score and live-weight effects on milk production in Irish Holstein-Friesian dairy cows. Animal 1, 1351–1359. Berry DP, Buckley F, Dillon P, Evans RD, Rath M and Veerkamp RF 2002. Genetic parameters for level and change of body condition score and body weight in dairy cows. Journal of Dairy Science 85, 2030–2039. Berry DP, Buckley F, Dillon P, Evans RD, Rath M and Veerkamp RF 2003a. Genetic relationships among body condition score, body weight, milk yield, and fertility in dairy cows. Journal of Dairy Science 86, 2193–2204. Berry RJ, Kennedy AD, Scott SL, Kyle BL and Schaefer AL 2003b. Daily variation in the udder surface temperature of dairy cows measured by infrared thermography: Potential for mastitis detection. Canadian Journal of Animal Science 83, 687–693. Bewley JM, Robertson LM and Eckelkamp EA 2017. A 100-year review: Lactating dairy cattle housing management. Journal of Dairy Science 100, 10418–10431.

248

Bickert WG and Light RG 1982. Housing systems. Journal of Dairy Science 65, 502–508. Blackshaw JK and Blackshaw AW 1994. Heat stress in cattle and the effect of shade on production and behaviour: a review. Australian Journal of Experimental Agriculture 34, 285–295. Boerner V, Johnston DJ and Tier B 2014. Accuracies of genomically estimated breeding values from pure-breed and across-breed predictions in Australian beef cattle. Genetics Selection Evolution 46, 61. Boichard D, Ducrocq V, Croiseau P and Fritz S 2016. Genomic selection in domestic animals: principles, applications and perspectives. Comptes Rendus Biologies 339, 274–277. Boichard DA, Guillaume FA, Baur AC, Croiseau PA, Rossignol MND, Boscher MY, Druet D, Genestout L, Colleau JJ, Journaux L, Ducrocq V and Fritz S 2012. Genomic selection in French dairy cattle. Animal Production Science 52, 115–120. Bolormaa S, Hayes BJ, Savin K, Hawken R, Barendse W, Arthur PF, Herd RM and Goddard ME 2011. Genome-wide association studies for feedlot and growth traits in cattle1. Journal of Animal Science 89, 1684–1697. Bonsma JC 1949. Breeding cattle for increased adaptability to tropical and subtropical environments. The Journal of Agricultural Science 39, 204–221. Bortolussi G, McIvor JG, Hodgkinson JJ, Coffey SG and Holmes CR 2005. The northern Australian beef industry, a snapshot. 2. Breeding herd performance and management. Australian Journal of Experimental Agriculture 45, 1075–1091. Bouraoui R, Lahmar M, Majdoub A, Djemali M and Belyea R 2002. The relationship of temperature- humidity index with milk production of dairy cows in a Mediterranean climate. Animal Research 51, 479–491. Branton C, McDowell RE, Frye JB and Johns DM 1961. Growth and production characteristics of Holstein-Friesian, Brown Swiss, and Red Sindhi crossbred females in Louisiana and Maryland. Journal of Dairy Science 44, 1344–1355. Britt JS, Thomas RC, Speer NC and Hall MB 2003. Efficiency of converting nutrient dry matter to milk in Holstein herds. Journal of Dairy Science 86, 3796–3801. Brouček J, Novák P, Vokřálová J, Šoch M, Kišac P and Uhrinčať M 2009. Effect of high temperature on milk production of cows from free-stall housing with natural ventilation. Slovak Journal of Animal Science 42, 167–173. Brown-Brandl TM, Eigenberg RA and Nienaber JA 2006. Heat stress risk factors of feedlot heifers. Livestock Science 105, 57–68.

249

Buffington DE, Collazo-Arocho A, Canton GH, Pitt D, Thatcher WW and Collier RJ 1981. Black globe-humidity index (BGHI) as comfort equation for dairy cows. Transactions of the ASAE 24, 711–714. Bui TN, Tran HC and Lebailly P 2013. Dairy milk production in the north of Vietnam: a case study in Moc Chau. International Journal of Advance in Management, Technology and Engineering Science 2, 137–143. Buniello A, Macarthur JAL, Cerezo M, Harris LW, Hayhurst J, Malangone C, McMahon A, Morales J, Mountjoy E, Sollis E, Suveges D, Vrousgou O, Whetzel PL, Amode R, Guillen JA, Riat HS, Trevanion SJ, Hall P, Junkins H, Flicek P, Burdett T, Hindorff LA, Cunningham F and Parkinson H 2019. The NHGRI-EBI GWAS Catalog of published genome-wide association studies, targeted arrays and summary statistics 2019. Nucleic Acids Research 47, D1005– D1012. Byrne TJ, Santos BFS, Amer PR, Martin-Collado D, Pryce JE and Axford M 2016. New breeding objectives and selection indices for the Australian dairy industry. Journal of Dairy Science 99, 8146–8167. Cai D Van, Anh HQ and Tri N Van 2005. Effects of heat stress on physiology and reproduction of crossed Holstein and imported pure Holstein in Southern areas. Institute of Agriculture Science for Southen Vietnam, Ho Chi Minh city, Vietnam. Cai D Van and Long VN 2002. Investigation of the status of the use of dairy sires and semen straws in the southern Vietnam. Retrieved on 27 July 2019, from http://www.trungtamqlkdg.com.vn/TTG_Res/Uploads/Acrobat/nghiencuukhoahoc/hientrangs udungtinh.pdf. Campos RV, Cobuci JA, Kern EL, Costa CN and McManus CM 2015. Genetic parameters for linear type traits and milk, fat, and protein production in Holstein cows in Brazil. Asian-Australasian Journal of Animal Sciences 28, 476–484. Campos G de los and Rodriguez PP 2016. CRAN - Package BGLR. Retrieved on 29 December 2017, from https://cran.r-project.org/web/packages/BGLR/index.html. Capper JL and Cady RA 2012. A comparison of the environmental impact of Jersey compared with Holstein milk for cheese production1. Journal of Dairy Science 95, 165–176. Cardoso CS, Hötzel MJ, Weary DM, Robbins JA and von Keyserlingk MAG 2016. Imagining the ideal dairy farm. Journal of Dairy Science 99, 1663–1671. Carpio DP Del, Lozano R, Wolfe MD and Jannink J-L 2018. Genome-wide association studies and

250

heritability estimation in the functional genomics era. In Population Genomics (ed. O.P. Rajora), pp. 361–425. Springer, Cham, Switzerland. Casas E, Kehrli ME and Jr. 2016. A review of selected genes with known effects on performance and health of cattle. Frontiers in Veterinary Science 3, 113. Cesaro JD, Nguyen MH, Pham DK, Hoang VQ and Duteurtre G 2014. Development of dairy production in Vietnam: smallholder farming and large scale investments. In Conference ‘Milk, factor of development’ from 21 to 23 May 2014. Rennes, France. Chagas LM, Bass JJ, Blache D, Burke CR, Kay JK, Lindsay DR, Lucy MC, Martin GB, Meier S, Rhodes FM, Roche JR, Thatcher WW and Webb R 2007. Invited review: new perspectives on the roles of nutrition and metabolic priorities in the subfertility of high-producing dairy cows. Journal of Dairy Science 90, 4022–4032. Cheruiyot EK, Bett RC, Amimo JO, Zhang Y, Mrode R and Mujibi FDN 2018. Signatures of selection in admixed dairy cattle in Tanzania. Frontiers in Genetics 9, 607. Chesterton N, Laven R, Nortje R and Mason W 2017. Preventing and managing lameness. Retrieved on 24 December 2017, from https://www.dairynz.co.nz/media/5787598/preventing-and- managing-lameness.pdf. Chiang CWK, Gajdos ZKZ, Korn JM, Kuruvilla FG, Butler JL, Hackett R, Guiducci C, Nguyen TT, Wilks R, Forrester T, Haiman CA, Henderson KD, Le Marchand L, Henderson BE, Palmert MR, McKenzie CA, Lyon HN, Cooper RS, Zhu X and Hirschhorn JN 2010. Rapid assessment of genetic ancestry in populations of unknown origin by genome-wide genotyping of pooled samples. PLoS Genetics 6, e1000866. Cho C, Cho K, Choy Y, Choi J, Choi T, Park B and Lee S 2013. Estimation of genetic parameters for milk production traits in Holstein dairy cattle. Journal of Animal Science and Technology 55, 7–11. Chu TKL, Yokogawa H and Kawaguchi T 2005. An analysis of the factors affecting dairy cow production in the South east of Vietnam. Journal of the Faculty of Agriculture, Kyushu University 50, 271–294. Church JS, Hegadoren PR, Paetkau MJ, Miller CC, Regev-Shoshani G, Schaefer AL and Schwartzkopf-Genswein KS 2014. Influence of environmental factors on infrared eye temperature measurements in cattle. Research in Veterinary Science 96, 220–226. Cincović MR, Belić B, Toholj B, Potkonjak A, Stevančević M, Lako B and Radović I 2011. Metabolic acclimation to heat stress in farm housed Holstein cows with different body condition

251

scores. African Journal of Biotechnology 10, 10293–10303. Coffey EL, Horan B, Evans R and Berry DP 2016. Milk production and fertility performance of Holstein, Friesian, and Jersey purebred cows and their respective crosses in seasonal-calving commercial farms. Journal of Dairy Research 99, 5681–5689. Cohen J 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20, 37–46. Cole JB, Wiggans GR, Ma L, Sonstegard TS, Lawlor Jr TJ, Crooker BA, Van Tassell CP, Yang J, Wang S, Matukumalli LK and Da Y 2011. Genome-wide association analysis of thirty one production, health, reproduction and body conformation traits in contemporary US Holstein cows. BMC Genomics 12, 408. Collier RJ, Dahl GE and Vanbaale MJ 2006. Major advances associated with environmental effects on dairy cattle. Journal of Dairy Science 89, 1244–1253. Combi Coporation Japan. Combi infrared ear thermometer user manual. Retrieved on 13 November 2018, from http://www.combi.com.hk/uploadfile/filedownload_list/2_1498207226.pdf. Comeron JM, Kreitman M and De La Vega FM 2003. On the power to detect SNP/phenotype association in candidate quantitative trait loci genomic regions: a simulation study. Pacific Symposium on Biocomputing 8, 477–489. Congress of Socialist Republic of Vietnam 2015. Veterinary Act (in Vietnamese). Conomos MP, Gogarten SM, Brown L, Chen H, Lumley T, Rice K, Sofer T, Thornton T and Maintainer Y 2019. R package ‘GENESIS’ version 2.17.0: Genetic estimation and inference in structured samples (GENESIS): Statistical methods for analyzing genetic data from samples with population structure and/or relatedness. Retrieved on 1 December 2019, from https://git.bioconductor.org/packages/GENESIS. Coppock CE, Grant PA, Portzer SJ, Charles DA and Escobosa A 1982. Lactating dairy cow responses to dietary sodium, chloride, and bicarbonate during hot weather. Journal of Dairy Science 65, 566–576. Cordero-Solorzano J, Parmentier HK, Arts JAJ, van der Poel J, de Koning DJ and Bovenhuis H 2019. Genome-wide association study identifies loci influencing natural antibody titers in milk of Dutch Holstein-Friesian cattle. Journal of Dairy Science 102, 11092–11103. Covarrubias-Pazaran G 2016. Genome-assisted prediction of quantitative traits using the R package sommer. PLoS ONE 11, e0156744. Covarrubias-Pazaran G 2019. R Package ‘sommer’ version 3.9.3: Solving mixed model equations in

252

R. Retrieved on 7 April 2019, from https://cran.r- project.org/web/packages/sommer/sommer.pdf. Cummins KA 1992. Effect of dietary acid detergent fiber on responses to high environmental temperature. Journal of Dairy Science 75, 1465–1471. Cuong VC, Cuong PK and Cuong PH 2006a. Test new diets for HF cows imported from the US raised in Moc Chau - Son La: initial results (in Vietnamese). National Institue of Animal Science, Hanoi, Vietnam. Cuong VC, Luu TX, Dat NQ, Hue PT, Trach NX and Hoa NX 2006b. Selection of 3/4 and 7/8 HF cows to generate a dairy herd which can produce more than 4000 kg of milk per lactation. National Institue of Animal Sciences, Hanoi, Vietnam. Cuong TH and Nga BT 2014. Actors’ benefit analysis on value chain of fresh dairy milk in Vietnam. Retrieved on 15 May 2018, from https://m.tailieu.vn/doc/phan-tich-loi-ich-cua-cac-tac-nhan- tren-chuoi-gia-tri-sua-bo-tuoi-o-viet-nam-2004137.html. Curt B and Mcfarland D 2017. Key considerations in fan cooling fan attributes. Dairy Business and Holstein World, 24–25. D & F control systems Inc 2017. Broad spectrum microtabs II. Retrieved on 12 August 2017, from http://www.aicompanies.com/product/broad-spectrum-microtabs/. Daetwyler HD, Calus MPL, Pong-Wong R, de los Campos G and Hickey JM 2013. Genomic prediction in animals and plants: simulation of data, validation, reporting, and benchmarking. Genetics 193, 347–365. Dairy Australia Limited 2014. Australian animal welfare standards and guidelines for cattle: a guide for dairy farmers. Dairy Australia Limited, Southbank, Victoria, Australia. Dairy One Forage Laboratory. The dairy one forage lab. Retrieved on 25 October 2017, from http://dairyone.com/analytical-services/feed-and-forage/about-forage/. Dairy One Forage Laboratory 2015. Analytical procedures. Retrieved on 20 June 2018, from https://dairyone.com/download/forage-forage-lab-analytical-procedures/. Dairy Vietnam 2017. Dairy farming situation in Vietnam. Dairy Vietnam Co., Ltd. Retrieved on 25 February 2018, from http://www.dairyvietnam.com/vn/Quan-ly-chan-nuoi-bo-sua/Tinh-hinh- chan-nuoi-bo-sua-nuoc-ta.html. DairyNZ 2019. All about breeding worth. Retrieved on 9 April 2020, from https://www.dairynz.co.nz/animal/animal-evaluation/interpreting-the-info/all-about-bw/. Dalcin VC, Fischer V, Daltro D dos S, Alfonzo EPM, Stumpf MT, Kolling GJ, Silva MVGB da and

253

McManus C 2016. Physiological parameters for thermal stress in dairy cattle. Revista Brasileira de Zootecnia 45, 458–465. Daltro D dos S, Fischer V, Alfonzo EPM, Dalcin VC, Stumpf MT, Kolling GJ, da Silva MVGB and McManus C 2017. Infrared thermography as a method for evaluating the heat tolerance in dairy cows. Revista Brasileira de Zootecnia 46, 374–383. Dao Phuong 2013. Ha Nam expands dairy farming towards a sustainable direction (in. Nhan Dan Online. Retrieved on 17 January 2018, from http://www.nhandan.com.vn/kinhte/tin- tuc/item/21825402-ha-nam-mo-rong-chan-nuoi-bo-sua-theo-huong-ben-vung.html. Das R, Sailo L, Verma N, Bharti P, Saikia J, Imtiwati and Kumar R 2016. Impact of heat stress on health and performance of dairy animals: A review. Veterinary World 9, 260–268. Decker JE, McKay SD, Rolf MM, Kim JW, Molina Alcalá A, Sonstegard TS, Hanotte O, Götherström A, Seabury CM, Praharani L, Babar ME, Correia de Almeida Regitano L, Yildiz MA, Heaton MP, Liu WS, Lei CZ, Reecy JM, Saif-Ur-Rehman M, Schnabel RD and Taylor JF 2015. Data from: Worldwide patterns of ancestry, divergence, and admixture in domesticated cattle. Dryad Dataset. Dekkers JCM 2012. Application of genomics tools to animal breeding. Current Genomics 13, 207– 212. Dennis G, Sherman BT, Hosack DA, Yang J, Gao W, Lane HC and Lempicki RA 2003. DAVID: Database for annotation, visualization, and integrated discovery. Genome Biology 4, R60. Department of Livestock Production 2009. Atlas of Vietnam animal husbandry. Department of Livestock Production 2017. No Title. Retrieved on 2 November 2017, from http://cucchannuoi.gov.vn/. Devendra C 2001. Smallholder dairy production systems in developing countries: characteristics, potential and opportunities for improvement - review. Asian-Australasian Journal of Animal Sciences 14, 104–113. Dikmen S, Alava E, Pontes E, Fear JM, Dikmen BY, Olson TA and Hansen PJ 2008. Differences in thermoregulatory ability between slick-haired and wild-type lactating Holstein cows in response to acute heat stress. Journal of Dairy Science 91, 3395–3402. Dikmen S, Cole JB, Null DJ and Hansen PJ 2012. Heritability of rectal temperature and genetic correlations with production and reproduction traits in dairy cattle. Journal of Dairy Science 95, 3401–3405. Dikmen S, Cole JB, Null DJ and Hansen PJ 2013. Genome-wide association mapping for

254

identification of quantitative trait loci for rectal temperature during heat stress in Holstein cattle. PLoS ONE 8, 1–7. Dikmen S and Hansen PJ 2009. Is the temperature-humidity index the best indicator of heat stress in lactating dairy cows in a subtropical environment? Journal of Dairy Science 92, 109–116. Dikmen S, Khan F a, Huson HJ, Sonstegard TS, Moss JI, Dahl GE and Hansen PJ 2014. The SLICK hair locus derived from Senepol cattle confers thermotolerance to intensively managed lactating Holstein cows. Journal of Dairy Science 97, 5508–5520. Dikmen S, Wang X -z., Ortega MS, Cole JB, Null DJ and Hansen PJ 2015. Single nucleotide polymorphisms associated with thermoregulation in lactating dairy cows exposed to heat stress. Journal of Animal Breeding and Genetics 132, 409–419. Ding X, Zhang Z, Li X, Wang S, Wu X, Sun D, Yu Y, Liu J, Wang Y, Zhang Y, Zhang S, Zhang Y and Zhang Q 2013. Accuracy of genomic prediction for milk production traits in the Chinese Holstein population using a reference population consisting of cows. Journal of Dairy Science 96, 5315–5323. Do DN, Schenkel FS, Miglior F, Zhao X and Ibeagha-Awemu EM 2018. Genome wide association study identifies novel potential candidate genes for bovine milk cholesterol content. Scientific reports 8, 13239. Dodds KG, Auvray B, Newman SAN and McEwan JC 2014. Genomic breed prediction in New Zealand sheep. BMC Genetics 15, 92. Doormaal B Van and Miglior F 2012. Official genetic evaluations for body condition score. Canadian Dairy Network, Canada. Dorshorst B, Henegar C, Liao X, Almén MS, Rubin CJ, Ito S, Wakamatsu K, Stothard P, Van Doormaal B, Plastow G, Barsh GS and Andersson L 2015. Dominant red coat color in Holstein cattle is associated with a missense mutation in the coatomer protein complex, subunit alpha (COPA) gene. PLoS ONE 10, e0128969. Drackley JK, Cicela TM and LaCount DW 2003. Responses of primiparous and multiparous Holstein cows to additional energy from fat or concentrate during summer. Journal of Dairy Science 86, 1306–1314. Draminski 2017. Draminski mastitis detector. Draminski, Olsztyn, Poland. Ducrocq V, Laloe D, Swaminathan M, Rognon X, Tixier-Boichard M and Zerjal T 2018. Genomics for ruminants in developing countries: From principles to practice. Frontiers in Genetics 9, 251. Dũng CA 2014. Development of dairy farming in Vietnam - some technical difficulties and solutions.

255

In Development of the dairy industry and dairy farming in Vietnam, 24 September 2014, pp. 1–8. Ho Chi Minh city, Vietnam. Dung NNX, Manh LH and Nhi NTM 2007. The composition & nutritive value of feed plants planted in Cantho city. Can Tho University Journal of Science 7, 183–192. Dunshea FR, Leury BJ, Fahri F, Digiacomo K, Hung A, Chauhan S, Clarke IJ, Collier R, Little S, Baumgard L and Gaughan JB 2013. Amelioration of thermal stress impacts in dairy cows. Animal Production Science 53, 965–975. Duteurtre G, Cesaro Nguyen J-D, Huong M, Duy P, Nguyen K and Luan N 2015. Multiscale assessment of the livestock development pathways in Vietnam the TH milk company (Vietnam): Is such a large-scale investment sustainable? REVALTER Research team in livestock transition in Vietnam, Hanoi, Vietnam. Duy N Van, Tuong ND, Van NTB, Huyen NT, Hiep DT, Phuong NV and Ton VD 2013. Growth performance of crossbred [Brahman x (Vietnamese Yellow x Sindhi)] cattle raising in Red River Delta. Journal of Animal Husbandry Sciences and Technics 6, 61–68. Eaglen S, Coffey M, Woolliams J and Wall E 2013. Direct and maternal genetic relationships between calving ease, gestation length, milk production, fertility, type, and lifespan of Holstein-Friesian primiparous cows. Journal of Dairy Science 96, 4015–4025. Edmonson AJ, Lean IJ, Weaver LD, Farver T and Webster G 1989. A body condition scoring chart for Holstein dairy cows. Journal of Dairy Science 72, 68–78. Effa K, Hunde D, Shumiye M and Silasie RH 2013. Analysis of longevity traits and lifetime productivity of crossbred dairy cows in the tropical highlands of Ethiopia. Journal of Cell and Animal Biology 7, 138–143. El-Tarabany MS, Roushdy EM and El-Tarabany AA 2017. Production and health performance of Holstein, Brown Swiss and their crosses under subtropical environmental conditions. Animal Production Science 57, 1137–1143. Endelman JB 2011. Ridge regression and other kernels for genomic selection with R package rrBLUP. The Plant Genome Journal 4, 250–255. Endelman J 2017. R package ‘rrBLUP’ version 4.5: Ridge regression and other kernels for genomic selection. Retrieved on 29 December 2017, from https://cran.r- project.org/web/packages/rrBLUP/index.html. Erbe M, Hayes BJ, Matukumalli LK, Goswami S, Bowman PJ, Reich CM, Mason BA and Goddard ME 2012. Improving accuracy of genomic predictions within and between dairy cattle breeds

256

with imputed high-density single nucleotide polymorphism panels. Journal of Dairy Science 95, 4114–4129. Ertl J, Legarra A, Vitezica ZG, Varona L, Edel C, Emmerling R and Götz K-U 2014. Genomic analysis of dominance effects on milk production and conformation traits in Fleckvieh cattle. Genetics Selection Evolution 46, 40. Falconer DS and Mackay TFC 1996. Introduction to quantitative genetics. Longmans Green, Harlow, Essex, UK. Ferguson D and Matthews L 2011. Animal welfare objective measures research program - validation of body condition score as a practical indicator of welfare in cattle. Meat & Livestock Australia, North Sydney, New South Wales, Australia. Fernandes Júnior GA, Rosa GJM, Valente BD, Carvalheiro R, Baldi F, Garcia DA, Gordo DGM, Espigolan R, Takada L, Tonussi RL, de Andrade WBF, Magalhães AFB, Chardulo LAL, Tonhati H and de Albuquerque LG 2016. Genomic prediction of breeding values for carcass traits in Nellore cattle. Genetics Selection Evolution 48, 7. Ferrão LFV, Ortiz R and Garcia AAF 2017. Genomic selection: state of the art. In Genetic Improvement of Tropical Crops, pp. 19–54. Springer International Publishing, New York City, USA. Fiedler M, Berg W, Ammon C, Loebsin C, Sanftleben P, Samer M, von Bobrutzki K, Kiwan A and Saha CK 2013. Air velocity measurements using ultrasonic anemometers in the animal zone of a naturally ventilated dairy barn. Biosystems Engineering 116, 276–285. Flaba J, Georg H, Graves RE, Lensink J, Loynes J, Ofner-schrock E, Ryan T, Ventorp M and Zappavigna P 2014. The design of dairy cow and replacement heifer housing. CIGR International Commission of agriculture and biosystems engineering, Gainesville, Florida, USA. Fleming A, Abdalla EA, Maltecca C and Baes CF 2018. Invited review: Reproductive and genomic technologies to optimize breeding strategies for genetic progress in dairy cattle. Archiv Fuer Tierzucht 61, 43. FLIR Systems I 2015. Temperature guns versus thermal imaging technology | FLIR Systems. Retrieved on 27 August 2020, from https://www.flir.com.au/discover/rd-science/temperature- guns-versus-thermal-imaging-technology/. FLIR Systems Inc. User’s manual FLIR Ex series. FLIR Systems Inc, Estonia USA. Flori L, Fritz S, Jaffrézic F, Boussaha M, Gut I, Heath S, Foulley J-L and Gautier M 2009. The

257

genome response to artificial selection: A case study in dairy cattle. PLoS ONE 4, e6595. Food and Agriculture Organization 2011. Guide to good dairy farming practice. Food and Agriculture Organization and International Dairy Federation, Rome, Italy. Fournel S, Ouellet V and Charbonneau É 2017. Practices for alleviating heat stress of dairy cows in humid continental climates: a literature review. Animals 7, 37. Francis RM 2017. Pophelper : an R package and web app to analyse and visualize population structure. Molecular Ecology Resources 17, 27–32. Frichot E and François O 2015. LEA: An R package for landscape and ecological association studies. Methods in Ecology and Evolution 6, 925–929. Friendly M and Meyer D 2016. Discrete data analysis with R: visualization and modeling techniques for categorical and count data. CRC Press, Taylor & Francis Group, Boca Raton, Florida, USA. Frkonja A, Gredler B, Schnyder U, Curik I and Sölkner J 2012. Prediction of breed composition in an admixed cattle population. Animal Genetics 43, 696–703. Galán E, Llonch P, Villagrá A, Levit H, Pinto S and del Prado A 2018. A systematic review of non- productivity-related animal-based indicators of heat stress resilience in dairy cattle. PLoS ONE 13, e0206520. Galas DJ and McCormack SJ 2003. An historical perspective on genomic technologies. Current Issues in Molecular Biology 5, 123–128. Gao Y, Gautier M, Ding X, Zhang H, Wang Y, Wang X, Faruque MO, Li J, Ye S, Gou X, Han J, Lenstra JA and Zhang Y 2017. Species composition and environmental adaptation of indigenous Chinese cattle. Scientific Reports 7, 16196. Garcia-Peniche TB, Cassell BG, Pearson RE and Misztal I 2005. Comparisons of Holsteins with Brown Swiss and Jersey cows on the same farm for age at first calving and first calving interval. Journal of Dairy Science 88, 790–796. García-Ruiza A, Coleb JB, VanRaden PM, Wiggans GR, Ruiz-Lópeza FJ and Tassell CP V. 2016. Changes in genetic selection differentials and generation intervals in US Holstein dairy cattle as a result of genomic selection. PNAS - Proceedings of the National Academy of Sciences, 3995–4004. Garcia ALS, Bosworth B, Waldbieser G, Misztal I, Tsuruta S and Lourenco DAL 2018. Development of genomic predictions for harvest and carcass weight in channel catfish 06 Biological Sciences 0604 Genetics. Genetics Selection Evolution 50, 66. Garner JB, Douglas ML, Williams SRO, Wales WJ, Marett LC, Nguyen TTT, Reich CM and Hayes

258

BJ 2016. Genomic selection improves heat tolerance in dairy cattle. Scientific Reports 6, 1–8. Gaughan JB, Holt SM and Pritchard RH 2009. Assessment of housing systems for feedlot cattle during summer. The Professional Animal Scientist 25, 633–639. Gaughan JB and Mader TL 2009. Effects of sodium chloride and fat supplementation on finishing steers exposed to hot and cold conditions. Journal of Animal Science 87, 612–621. Gaughan JB, Mader T, Holt S, Hahn G and Young BA 2002. Review of current assessment of cattle and microclimate during periods of high heat load. Animal Production Australia 24, 77–80. Gaughan JB, Mader TL, Holt SM and Lisle A 2008. A new heat load index for feedlot cattle. Journal of Animal Science 86, 226–234. Gautier M, Laloë D and Moazami-Goudarzi K 2010. Insights into the genetic history of French cattle from dense SNP data on 47 worldwide breeds. PLoS ONE 5, e13038. Gebreyesus G, Bovenhuis H, Lund MS, Poulsen NA, Sun D and Buitenhuis B 2019. Reliability of genomic prediction for milk fatty acid composition by using a multi-population reference and incorporating GWAS results. Genetics Selection Evolution 51, 16. Gelasakis AI, Valergakis GE, Arsenos G and Banos G 2012. Description and typology of intensive Chios dairy sheep farms in Greece. Journal of Dairy Science 95, 3070–3079. Geneletti S, Gallo V, Porta M, Khoury MJ and Vineis P 2011. Assessing causal relationships in genomics: From Bradford-Hill criteria to complex gene-environment interactions and directed acyclic graphs. Emerging Themes in Epidemiology 8, 5. General Department of Vietnam Customs 2017. Import report. Retrieved on 21 December 2017, from https://www.customs.gov.vn/Lists/ThongKeHaiQuanLichCongBo/Attachments/1018/2017- T12K1-1N(VN-SB).pdf. General Statistics Office of Vietnam 2016. Livestock survey of General Statistic Office. Retrieved on 10 July 2019, from http://nhachannuoi.vn/phat-trien-chan-nuoi-bo-sua-quy-mo-vua-va- nho-hieu-qua-ben-vung/. General Statistics Office of Vietnam 2017a. Statistics on number of domestic animals and animal products of the whole country by 01/10/2017. General Statistics Office of Vietnam. Retrieved on 15 March 2018, from http://nhachannuoi.vn/thong-ke-so-luong-va-san-luong-gia-suc-gia- cam-ca-nuoc-01102017/. General Statistics Office of Vietnam 2017b. Weather data. Retrieved on 16 December 2019, from http://www.gso.gov.vn/default_en.aspx?tabid=515&idmid=5&ItemID=19299. Gieseke D, Lambertz C and Gauly M 2018. Relationship between herd size and measures of animal

259

welfare on dairy cattle farms with freestall housing in Germany. Journal of Dairy Science 101, 7397–7411. Gioi P Van, Duc N Van and Them TT 2006. Studying effects of some fixed factors on milk yield of nuclear and class I crossed HF in Vietnam. National Institue of Animal Sciences, Hanoi, Vietnam. Glass TA, Goodman SN, Hernán MA and Samet JM 2013. Causal inference in public health. Annual Review of Public Health 34, 61–75. Gobena M, Elzo MA and Mateescu RG 2018. Population structure and genomic breed composition in an Angus–Brahman Crossbred cattle population. Frontiers in Genetics 9, 90. Gondro C, Porto-Neto LR and Lee SH 2014. SNPqc – an R pipeline for quality control of Illumina SNP genotyping array data. Animal Genetics 45, 758–761. Gondro C, Werf J van der and Hayes B 2013. Genome-wide association studies and genomic prediction. Humana Press, New York, USA. Gonzalez-Mejia A, Styles D, Wilson P and Gibbons J 2018. Metrics and methods for characterizing dairy farm intensification using farm survey data. PLoS ONE 13, e0195286. Google. Google Earth. Google, Inc. Retrieved on 31 December 2019, from https://www.google.com/earth/. Goopy JP, Pelster DE, Onyango A, Marshall K and Lukuyu M 2018. Simple and robust algorithms to estimate liveweight in African smallholder cattle. Animal Production Science 58, 1758– 1765. Goudet J and Jombart T 2015. R Package ‘hierfstat’ version 0.04-22: Estimation and Tests of Hierarchical F-Statistics. Retrieved on 1 December 2019, from https://cran.r- project.org/web/packages/hierfstat/hierfstat.pdf. Grisart B, Coppieters W, Farnir F, Karim L, Ford C, Berzi P, Cambisano N, Mni M, Reid S, Simon P, Spelman R, Georges M and Snell R 2002. Positional candidate cloning of a QTL in dairy cattle: Identification of a missense mutation in the bovine DGAT1 gene with major effect on milk yield and composition. Genome Research 12, 222–231. Grisart B, Farnir F, Karim L, Cambisano N, Kim JJ, Kvasz A, Mni M, Simon P, Frère JM, Coppieters W and Georges M 2004. Genetic and functional confirmation of the causality of the DGAT1 K232A quantitative trait nucleotide in affecting milk yield and composition. Proceedings of the National Academy of Sciences of the United States of America 101, 2398–2403. Gruber B, Unmack PJ, Berry OF and Georges A 2018. dartr: An r package to facilitate analysis of

260

SNP data generated from reduced representation genome sequencing. Molecular Ecology Resources 18, 691–699. Grueneberg A and Campos de los G 2019. BGData - A suite of R packages for genomic analysis with big data. G3: Genes, Genomes, Genetics 9, 1377–1383. Guo X, Christensen OF, Ostersen T, Wang Y, Lund MS and Su G 2016. Genomic prediction using models with dominance and imprinting effects for backfat thickness and average daily gain in Danish Duroc pigs. Genetics Selection Evolution 48, 67. Guo G, Lund MS, Zhang Y and Su G 2010. Comparison between genomic predictions using daughter yield deviation and conventional estimated breeding value as response variables. Journal of Animal Breeding and Genetics 127, 423–432. Habeeb A, Gad A and Atta M 2018. Temperature-humidity Indices as indicators to heat stress of climatic conditions with relation to production and reproduction of farm animals. International Journal of Biotechnology and Recent Advances 1, 35–50. Habier D, Fernando RL and Dekkers JCM 2007. The impact of genetic relationship information on genome-assisted breeding values. Genetics 177, 2389–97. Habier D, Fernando RL and Dekkers JCM 2009. Genomic selection using low-density marker panels. Genetics 182, 343–53. Habier D, Fernando RL and Garrick DJ 2013. Genomic BLUP decoded: a look into the black box of genomic prediction. Genetics 194, 597–607. Habier D, Fernando RL, Kizilkaya K and Garrick DJ 2011. Extension of the bayesian alphabet for genomic selection. BMC Bioinformatics 12, 186. Hahn GL 1995. Environmental management for improved livestock performance, health and well- being. Japanese Journal of Livestock Management 30, 113–127. Hahn GL, Parkhurst AM and Gaughan JB 1997. Cattle respiration rate as a function of ambient temperature. Mid Central Conference, ASAE Paper 97, 121. Hai HT 2008. Decision No.10/2008/QD-TTg of Prime Minister on approving animal production development strategy through 2020. Retrieved on 18 June 2019, from http://cucchannuoi.gov.vn/chien-luoc-quy-hoach/chien-luoc-phat-trien-chan-nuoi/chien-luoc- phat-trien-chan-nuoi-den-2020.html. Hai HT 2012. Decision No. 124/2012/QD-TTg of Prime Minister on approving the master plan for agricultural production development through 2020 with a vision toward 2030. Retrieved on 18 June 2019, from

261

http://www.chinhphu.vn/portal/page/portal/chinhphu/hethongvanban?mode=detail&documen t_id=154513. Hansen PJ 2007. Exploitation of genetic and physiological determinants of embryonic resistance to elevated temperature to improve embryonic survival in dairy cattle during heat stress. Theriogenology 68, S242–S249. Hatem MH, Sadek RR and Samer M 2004. Cooling, shed height, and shed orientation affecting dairy cows microclimate. Misr Journal of Agricultural Engineering 21, 714–726. Hayes D 2012. Vietnam takes off. Dairy Industries International 77, 26–27. Hayes BJJ, Bowman PJJ, Chamberlain AJJ and Goddard MEE 2009a. Invited review: Genomic selection in dairy cattle: Progress and challenges. Journal of Dairy Science 92, 433–443. Hayes BJ, Bowman PJ, Chamberlain AC, Verbyla K and Goddard ME 2009b. Accuracy of genomic breeding values in multi-breed dairy cattle populations. Genetics Selection Evolution 41, 51. Hayes B and Goddard M 2010. Genome-wide association and genomic selection in animal breeding. Genome 53, 876–883. Hayes BJ, Lewin HA and Goddard ME 2013. The future of livestock breeding: genomic selection for efficiency, reduced emissions intensity, and adaptation. Trends in Genetics 29, 206–214. He J, Guo Y, Xu J, Li H, Fuller A, Tait RJ, Wu X and Bauck S 2018a. Estimating genomic breed composition of individual animals in ten cattle breeds: Comparison of SNP panels and statistical methodology. Proceedings of the World Congress on Genetics Applied to Livestock Production Electronic, 684. He J, Guo Y, Xu J, Li H, Fuller A, Tait RG, Wu X-L and Bauck S 2018b. Comparing SNP panels and statistical methods for estimating genomic breed composition of individual animals in ten cattle breeds. BMC Genetics 19, 56. Heinrichs J, Isaler VA and Jones CM 2017. Body condition scoring as a tool for dairy herd management. Penn State Extension. Retrieved on 1 June 2017, from https://extension.psu.edu/body-condition-scoring-as-a-tool-for-dairy-herd-management. Henderson CR 1986. Recent developments in variance and covariance estimations. Journal of Animal Science 63, 208–216. Henry B, Charmley E, Eckard R, Gaughan JB and Hegarty R 2012. Livestock production in a changing climate: adaptation and mitigation research in Australia. Crop and Pasture Science 63, 191. Henshall JM 2013. Validation of genome-wide association studies (GWAS) results. In Genome-Wide

262

Association Studies and Genomic Prediction. Methods in Molecular Biology (Methods and Protocols) (eds. C. Gondro, J. van der Werf and B. Hayes), pp. 411–421. Humana Press, Totowa, New York, USA. Herawati T, Priyanto D and Purwantari ND 2016. Economic value of dairy cattle business in the lowlands outside Java. In Proceedings of International Seminar on Livestock Production and Veterinary Technology, pp. 222–231. Herbut P 2013. Temperature, humidity and air movement variations inside a free-stall barn during heavy frost. Annals of Animal Science 13, 587–596. Herbut P, Angrecka S and Walczak J 2018. Environmental parameters to assessing of heat stress in dairy cattle-a review. International journal of biometeorology 62, 2089–2097. Hernández-Castellano LE, Nally JE, Lindahl J, Wanapat M, Alhidary IA, Fangueiro D, Grace D, Ratto M, Bambou JC and de Almeida AM 2019. Dairy science and health in the tropics: challenges and opportunities for the next decades. Tropical Animal Health and Production 51, 1009–1017. Hiep T, Bang NN and Thang CM 2014. Effects of corn silage diet on performance and methane emission of lactating cows (in Vietnamese). Journal of Animal Science and Technology 51, 10– 21. Hiep T, Bang NN, Trach NX and Thang CM 2016. Effects of dietary supplementation of cotton seed oil on intake, performance and enteric methane emissions from lactating cows. Journal of Animal Husbandry Sciences and Technics (ISSN 1859-476X) 207, 48–56. Hill AB 1965. The environment and disease: Association or causation? Proceedings of the Royal Society of Medicine 58, 295–300. Ho Chi Minh City Extension Center 2015. Application manual TMR for dairy farming (in Vietnamese). Agriculture Publishing House, Hanoi, Vietnam. Hoffmann G, Schmidt M, Ammon C, Rose-Meierhöfer S, Burfeind O, Heuwieser W and Berg W 2013. Monitoring the body temperature of cows and calves using video recordings from an infrared thermography camera. Veterinary Research Communications 37, 91–99. Hogeveen H, Kamphuis C, Steeneveld W and Mollenhorst H 2010. Sensors and clinical mastitis-the quest for the perfect alert. Sensors 10, 7991–8009. Hoggart CJ, Shriver MD, Kittles RA, Clayton DG and McKeigue PM 2004. Design and analysis of admixture mapping studies. American Journal of Human Genetics 74, 965–978. Howard JT, Kachman SD, Snelling WM, Pollak EJ. JJ, Ciobanu DC, Kuehn LA and Spangler ML

263

2014. Beef cattle body temperature during climatic stress: a genome-wide association study. International Journal of Biometeorology 58, 1665–1672. Hsu WL, Garrick DJ and Fernando RL 2017. The accuracy and bias of single-step genomic prediction for populations under selection. G3: Genes, Genomes, Genetics 7, 2685–2694. Hu Z-L, Park CA and Reecy JM 2019. Building a livestock genetic and genomic information knowledgebase through integrative developments of Animal QTLdb and CorrDB. Nucleic Acids Research 47, D701–D710. Huang A and Liu D 2016. R package ‘EBglmnet’ version 4.1: Empirical Bayesian lasso and elastic net methods for generalized linear models. Retrieved on 29 December 2017, from https://cran.r-project.org/web/packages/EBglmnet/index.html. Huang DW, Sherman BT and Lempicki RA 2009. Bioinformatics enrichment tools: Paths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Research 37, 1–13. Huber JT, Higginbotham G, Gomez-Alarcon RA, Taylor RB, Chen KH, Chan SC and Wu Z 1994. Heat stress interactions with protein supplemental fat, and fungal cultures. Journal of Dairy Science 77, 2080–2090. Hulsegge I, Schoon M, Windig J, Neuteboom M, Hiemstra SJ and Schurink A 2019. Development of a genetic tool for determining breed purity of cattle. Livestock Science 223, 60–67. Humer E, Petri RM, Aschenbach JR, Bradford BJ, Penner GB, Tafaj M, Südekum KH and Zebeli Q 2018. Invited review: Practical feeding management recommendations to mitigate the risk of subacute ruminal acidosis in dairy cattle. Journal of Dairy Science 101, 872–888. Huson HJ, Kim E-S, Godfrey RW, Olson TA, Mcclure MC, Chase CC, Rizzi R, O’brien AMP, Van Tassell CP, Garcia JF, Sonstegard TS, Arendonk J Van and Steibel J 2014. Genome-wide association study and ancestral origins of the slick-hair coat in tropically adapted cattle. Frontiers in Genetics 5, 1–12. Husson F and Josse J 2010. Multivariate data analysis: special focus on clustering and multiway methods. In useR! 2010, the R user conference. Gaithersburg, Maryland, USA. Husson F, Josse AJ, Jérôme A and Agrocampus P 2010. Principal component methods-hierarchical clustering-partitional clustering: why would we need to choose for visualizing data? TechnicalReport – Agrocampus. Retrieved on 27 December 2019, from http://www.agrocampus-ouest.fr/math/. Husson F, Josse J, Le S and Mazet J 2019. R package ‘FactoMineR’ version 2.0: multivariate exploratory data analysis and data mining. Retrieved on 27 December 2019, from

264

http://factominer.free.fr. Illumina 2015. Iscan system guide. San Diego, California, USA. Jackson PGG and Cockcroft PD 2002. Appendix 1: Normal Physiological Values. In Clinical Examination of Farm Animals (eds. P.G.G. Jackson and P.D. Cockcroft), p. 301. Blackwell Science, Hoboken, New Jersey, USA. Jattawa D, Elzo MA, Koonawootrittriron S and Suwanasopee T 2015. Comparison of genetic evaluations for milk yield and fat yield using a polygenic model and three genomic-polygenic models with different sets of SNP genotypes in Thai multibreed dairy cattle. Livestock Science 181, 58–64. Jiang L, Liu J, Sun D, Ma P, Ding X, Yu Y and Zhang Q 2010. Genome wide association studies for milk production traits in Chinese Holstein population. PLoS ONE 5. Jiang J, Ma L, Prakapenka D, VanRaden PM, Cole JB and Da Y 2019. A large-scale genome-wide association study in U.S. Holstein cattle. Frontiers in Genetics 10, 412. Jiang J, Zhang Q, Ma L, Li J, Wang Z and Liu J-F 2015. Joint prediction of multiple quantitative traits using a Bayesian multivariate antedependence model. Heredity 115, 29–36. Joerg H, Meili C, Ruprecht O, Bangerter E, Burren A and Bigler A 2014. A genome-wide association study reveals a QTL influencing caudal supernumerary teats in Holstein cattle. Animal Genetics 45, 871–873. Johnson HD and Vanjonack WJ 1976. Effects of environmental and other Stressors on blood hormone patterns in lactating animals1. Journal of Dairy Science 59, 1603–1617. Jordan ER 2003. Effects of heat stress on reproduction. Journal of Dairy Science 86, E104–E114. Kadzere CT, Murphy MR, Silanikove N and Maltz E 2002. Heat stress in lactating dairy cows: A review. Livestock Production Science 77, 59–91. Kaiser HF 1961. A note on Guttman’s lower bound for the number of common factors. British Journal of Statistical Psychology 14, 1–2. Kang H, Zhou L, Liu J, Kang H, Zhou L and Liu J 2017. Statistical considerations for genomic selection. Frontiers of Agricultural Science and Engineering 4, 268. Karimuribo ED, Fitzpatrick JL, Bell CE, Swai ES, Kambarage DM, Ogden NH, Bryant MJ and French NP 2006. Clinical and subclinical mastitis in smallholder dairy farms in Tanzania: Risk, intervention and knowledge transfer. Preventive Veterinary Medicine 74, 84–98. Kassambara A 2010. Practical guide to principal component methods in R. WIREs Computational Statistics 2, 433–459. 265

Kassambara A and Mundt F 2019. R Package ‘factoextra’ version 1.0.6: Extract and Visualize the Results of Multivariate Data Analyses. Retrieved on 27 December 2019, from https://github.com/kassambara/factoextra/issues. Kastberger G and Stachl R 2003. Infrared imaging technology and biological applications. Behavior Research Methods, Instruments, and Computers 35, 429–439. Kebreab E, Ahmadi A and Corner-Dolloff C 2019. Ration formulation software enhances farmer productivity, decreases emission intensity and trains nutritionists in Vietnam. Agrilinks. Retrieved on 30 August 2020, from https://www.agrilinks.org/post/ration-formulation- software-enhances-farmer-productivity-decreases-emission-intensity-and. Keenan K, Mcginnity P, Cross TF, Crozier WW and Prodöhl PA 2013. DiveRsity: An R package for the estimation and exploration of population genetics parameters and their associated errors. Methods in Ecology and Evolution 4, 782–788. Khaing T, Loh T, Saleh G, Halim M and Samsudin AA 2015. Feed intake, growth performance and digestibility in goats fed whole corn plant silage and Napier grass. Malaysian Journal of Animal Science 18, 87–98. Khanh NQ 2014. Experience of Vinamilk in Viet Nam. In Dairy Asia: Towards sustainability, p. 35. Food and Agriculture Organization of the United Nations, Bangkok, Thailand. Khatkar MS, Moser G, Hayes BJ and Raadsma HW 2012. Strategies and utility of imputed SNP genotypes for genomic analysis in dairy cattle. BMC Genomics 13, 538. Khoi NV and Dung T Van 2014. The dairy industry in Vietnam: A value chain approach. International Journal of Managing Value and Supply Chains 5, 1–17. Kim NY, Kim SJ, Park JH, Oh MR, Jang SY, Kim DH, Sung SH, Jeon BT and Moon SH 2014. Seasonal changes in the body surface temperature of Hanwoo (Bos taurus coreanae) steers. Animal Production Science 54, 1476. Kim B-W, Lee D, Jeon J-T and Lee J-G 2009. Estimation of genetic parameters for milk production traits using a random regression test-day model in holstein cows in Korea. Asian-Australasian Journal of Animal Sciences 22, 923–930. Klein RJ 2007. Power analysis for genome-wide association studies. BMC Genomics 8, 58. Knapp DM and Grummer RR 1991. Response of lactating dairy cows to fat supplementation during heat stress. Journal of Dairy Science 74, 2573–2579. Kock N and Lynn GS 2012. Lateral collinearity and misleading results in variance-based SEM: An illustration and recommendations. Journal of the Association for Information Systems 13, 546–

266

580. Kolling GJ, Stivanin SCB, Gabbi AM, Machado FS, Ferreira AL, Campos MM, Tomich TR, Cunha CS, Dill SW, Pereira LGR and Fischer V 2018. Performance and methane emissions in dairy cows fed oregano and green tea extracts as feed additives. Journal of Dairy Science 101, 4221– 4234. Kolver ES 2003. Nutritional limitations to increased production on pasture-based systems. Proceedings of the Nutrition Society 62, 291–300. Konig IR 2011. Validation in genetic association studies. Briefings in Bioinformatics 12, 253–258. Könyves T, Zlatković N, Memiši N, Lukač D, Puvača N, Stojšin M, Halász A and Miščević B 2017. Relationship of temperature-humidity index with milk production and feed intake of Holstein- Frisian cows in different year seasons. The Thai Journal of Veterinary Medicine 47, 15–23. Koonawootrittriron S, Elzo MA and Thongprapi T 2009. Genetic trends in a Holstein × other breeds multibreed dairy population in Central Thailand. Livestock Science 122, 186–192. Krattenmacher N, Thaller G and Tetens J 2019. Analysis of the genetic architecture of energy balance and its major determinants dry matter intake and energy-corrected milk yield in primiparous Holstein cows. Journal of Dairy Science 102, 3241–3253. Krause KM and Oetzel GR 2006. Understanding and preventing subacute ruminal acidosis in dairy herds: A review. Animal Feed Science and Technology 126, 215–236. Kuehn LA, Keele JW, Bennett GL, McDaneld TG, Smith TPL, Snelling WM, Sonstegard TS and Thallman RM 2011. Predicting breed composition using breed frequencies of 50,000 markers from the US Meat Animal Research Center 2,000 bull project. Journal of Animal Science 89, 1742–1750. Kuivanen KS, Alvarez S, Michalscheck M, Adjei-Nsiah S, Descheemaeker K, Mellon-Bedi S and Groot JCJ 2016. Characterising the diversity of smallholder farming systems and their constraints and opportunities for innovation: A case study from the Northern Region, Ghana. NJAS - Wageningen Journal of Life Sciences 78, 153–166. Kuleshov M V, Jones MR, Rouillard AD, Fernandez NF, Duan Q, Wang Z, Koplev S, Jenkins SL, Jagodnik KM, Lachmann A, McDermott MG, Monteiro CD, Gundersen GW and Ma’ayan A 2016. Enrichr: a comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Research 44, W90–W97. Kurz JP, Yang Z, Weiss RB, Wilson DJ, Rood KA, Liu GE and Wang Z 2019. A genome-wide association study for mastitis resistance in phenotypically well-characterized Holstein dairy

267

cattle using a selective genotyping approach. Immunogenetics 71, 35–47. Kuznetsova A, Brockhoff PB and Christensen RHB 2017. lmerTest package: tests in linear mixed effects models. Journal of Statistical Software 82, 1–26. Lam V, Wredle E, Thao NT, Man N Van and Svennersten-sjaunja K 2010. Smallholder dairy production in Southern Vietnam: Production, management and milk quality problems. African Journal of Agricultural Research 5, 2668–2675. Lean IJ 2011a. Feeds, ration formulation| Systems describing nutritional requirements of dairy cows. In Encyclopedia of Dairy Sciences (eds. J.W. Uquay, P.F. Fox and P.L.H. McSweeney), pp. 418–428. Elsevier, San Diego, California, USA. Lean I 2011b. Ration formulation | Systems describing nutritional requirements of dairy cattle. In Encyclopedia of Dairy Sciences (eds. J.W. Uquay, P.F. Fox and P.L.H. McSweeney), pp. 418– 428. Elsevier, San Diego, California, USA. Lees AM 2015. Biological responses of feedlot cattle to heat load. The University of Queensland. Lees JC, Lees AM and Gaughan JB 2018. Developing a heat load index for lactating dairy cows. Animal Production Science 58, 387–1391. Lees AM, Sejian V, Wallage AL, Steel CC, Mader TL, Lees JC and Gaughan JB 2019. The impact of heat load on cattle. Animals 9, 332. Legarra A, Robert-Granié C, Manfredi E and Elsen J-M 2008. Performance of genomic selection in mice. Genetics 180, 611–618. Lehmann JO, Mogensen L and Kristensen T 2017. Early lactation production, health, and welfare characteristics of cows selected for extended lactation. Journal of Dairy Science 100, 1487– 1501. LGC Genomics. Sbeadex livestock kit. Retrieved on 22 December 2018, from https://www.lgcgroup.com/LGCGroup/media/PDFs/Products/Extraction/Kits inserts/sbeadex- livestock-kit.pdf. Li Y, Gao Y, Kim Y-S, Iqbal A and Kim J-J 2017. A whole genome association study to detect additive and dominant single nucleotide polymorphisms for growth and carcass traits in Korean native cattle, Hanwoo. Asian-Australasian journal of animal sciences 30, 8–19. Li X, Wang S, Huang J, Li L, Zhang Q and Ding X 2014. Improving the accuracy of genomic prediction in Chinese Holstein cattle by using one-step blending. Genetics Selection Evolution 46, 66. Liang D, Wood CL, McQuerry KJ, Ray DL, Clark JD and Bewley JM 2013. Influence of breed, milk

268

production, season, and ambient temperature on dairy cow reticulorumen temperature. Journal of Dairy Science 96, 5072–5081. Liem LT 2016. Decision No. 4697/QD-UBND approving the Project on improving the quality of dairy herds in Ho Chi Minh City for the period 2016-2020 (in Vietnamese). People’s Committee of Ho Chi Minh City. Little S, King R, Lean I, Rabiee A and Breinhild K 2009. A review of 11 applied dairy nutrition models used in Australia Summary report. Retrieved on 25 July 2019, from http://www.rumen8.com.au/resources/G2M-models-summary-report-2009---final.pdf. Liu J, Li L, Chen X, Lu Y and Wang D 2019. Effects of heat stress on body temperature, milk production, and reproduction in dairy cows: a novel idea for monitoring and evaluation of heat stress — A review. Asian-Australasian Journal of Animal Sciences 32, 1332–1339. Liu JJ, Liang AX, Campanile G, Plastow G, Zhang C, Wang Z, Salzano A, Gasparrini B, Cassandro M and Yang LG 2018. Genome-wide association studies to identify quantitative trait loci affecting milk production traits in water buffalo. Journal of Dairy Science 101, 433–444. Liu Y, Nyunoya T, Leng S, Belinsky SA, Tesfaigzi Y and Bruse S 2013. Softwares and methods for estimating genetic ancestry in human populations. Human Genomics 7, 1–7. Loan CTK, Yokogawa H and Kawaguchi T 2004. The econonics of dairy cow raising in the south east of Vietnam. Journal of the Faculty of Agriculture, Kyushu University 2, 497–512. LokeshBabu DS, Vasant PJ, Jeyakumar S, Manimaran A, Kumaresan A, Pushpadass HA, Sivaram M, Ramesha KP, Kataktalware MA, Siddaramanna and Sathiyabarathi M 2018. Monitoring foot surface temperature using infrared thermal imaging for assessment of hoof health status in cattle: A review. Journal of Thermal Biology 78, 10–21. de Los Campos G, Hickey JM, Pong-Wong R, Daetwyler HD and Calus MPL 2013. Whole-genome regression and prediction methods applied to plant and animal breeding. Genetics 193, 327– 345. de los Campos G, Sorensen D and Gianola D 2015. Genomic heritability: what is it? PLoS Genetics 11, e1005048. Loyau T, Zerjal T, Rodenburg TB, Fablet J, Tixier-Boichard M, Pinard-Van Der Laan MH and Mignon-Grasteau S 2016. Heritability of body surface temperature in hens estimated by infrared thermography at normal or hot temperatures and genetic correlations with egg and feather quality. Animal 10, 1594–1601. Luan TT, Woolliams JA, Lien S, Kent M, Svendsen M and Meuwissen THEE 2009. The accuracy of

269

genomic selection in Norwegian Red cattle assessed by cross-validation. Genetics 183, 1119– 1126. Lucas RM and Mcmichael AJ 2005. Public health classics association or causation: evaluating links between ‘environment and disease’. Bulletin of the World Health Organization 83, 792–795. Luong NH, Tuyen DK, Vang ND, Giao HK, Hai NV, Noi V Van, Thao L Van, Ha TS, Hieu VN, Manh NS, Son NH and Huyen NTD 2006. Studying some technical indicators of Australian dairy cows imported to Vietnam (2002 - 2004). Scientific Reports of Vietnam National Institue of Animal Science, 1–13. Lwin M, Mon SLY, Yamanaka H, Nagano Y, Mannen H, Faruque MO, Kawabe K, Okamoto S and Shimogiri T 2018. Genetic diversities and population structures of four popular Myanmar local cattle breeds. Animal Science Journal 89, 1648–1655. Ly LV, Giao HK, Sanh M Van, Su V Van and Sat LM 1999. Monograph of conservation of livestock gene sources in Vietnam. Agriculture Publishing House, Hanoi, Vietnam. Ma YL and Du WK 2010. Relationship between body temperature and disease of cattle. Shandong Journal of Animal Science and Veterinary Medicine 31, 91–92. Maaruf K and Paputungan U 2017. Comparison of protein and cell wall degradation of selected tropical and temperate roughages. Livestock Research for Rural Development. 29, #120. Macmillan K, Gao X and Oba M 2017. Increased feeding frequency increased milk fat yield and may reduce the severity of subacute ruminal acidosis in higher-risk cows. Journal of Dairy Science 100, 1045–1054. Madalena FE, Peixoto MGCD and Gibson J 2012. Dairy cattle genetics and its applications in Brazil. Livestock Research for Rural Development 24, #97. Mader TL, Davis MS and Brown-Brandl T 2006. Environmental factors influencing heat stress in feedlot cattle. Journal of animal science 84, 712–719. Mader TL, Johnson LJ and Gaughan JB 2010. A comprehensive index for assessing environmental stress in animals. Journal of Animal Science 88, 2153–2165. Mai TH, Vu CC, Nguyen HL and Nguyen XT 2011. Efficacy of using total mixed ration (TMR) for feeding dairy cattle. Journal of Science and Development 9, 584–591. Makowski D 2018. The psycho package: an efficient and publishing-oriented workflow for psychological science. Journal of Open Source Software 3, 470. Mangiafico S 2019. R package ‘rcompanion’ version 2.3.7: Functions to support extension education program evaluation. Retrieved on 28 October 2019, from http://rcompanion.org.

270

Manh H and Bich H 2017. The Vinamilk company surveyed and invested in a dairy farm in Ha Nam (in Vietnamese). Ha Nam Online. Retrieved on 17 January 2018, from http://www.baohanam.com.vn/kinh-te/cong-ty-vinamilk-khao-sat-dau-tu-trang-trai-chan- nuoi-bo-sua-tai-ha-nam-1223.html. Marees AT, de Kluiver H, Stringer S, Vorspan F, Curis E, Marie-Claire C and Derks EM 2018. A tutorial on conducting genome-wide association studies: Quality control and statistical analysis. International Journal of Methods in Psychiatric Research 27, e1608. Marete A, Sahana G, Fritz S, Lefebvre R, Barbat A, Lund MS, Guldbrandtsen B and Boichard D 2018. Genome-wide association study for milking speed in French Holstein cows. Journal of dairy science 101, 6205–6219. Martello LS, Silva S da L e, Gomes R da C, Corte RRP da S and Leme PR 2016. Infrared thermography as a tool to evaluate body surface temperature and its relationship with feed efficiency in Bos indicus cattle in tropical conditions. International Journal of Biometeorol 60, 173–181. Mateo J 2016. Hanoi training demonstrates feed management software for enhancing dairy production in Vietnam. Retrieved on 22 December 2017, from https://vietpigs.net/2016/11/07/dairy-software-training/. Matukumalli LK, Lawley CT, Schnabel RD, Taylor JF, Allan MF, Heaton MP, O’Connell J, Moore SS, Smith TPL, Sonstegard TS and Van Tassell CP 2009. Development and characterization of a high density SNP genotyping assay for cattle. PLoS ONE 4, e5350. Mauro P, Gallo L, Carnier P, Daniele M, Zotto R and Cassandro M 2005. Genetic aspects of milk electrical conductivity in Italian Brown cattle. Italian Journal of Animal Science (Suppl. 3) 4, 169–171. McDonald JH 2014. Handbook of biological statistics (3rd ed.). Sparky House Publishing, Baltimore, Maryland, USA. Meat and Livestock Australia 2006. Tips and tools: Heat load in feedlot cattle. Retrieved on 30 June 2017, from https://www.mla.com.au/globalassets/mla-corporate/generic/about-mla/heatload- in-feedlot-cattle.pdf. Melka MG and Schenkel FS 2012. Analysis of genetic diversity in Brown Swiss, Jersey and Holstein populations using genome-wide single nucleotide polymorphism markers. BMC Research Notes 5, 161. Mendiburu F de 2019. R package ‘agricolae’ version 1.3-1: Statistical procedures for agricultural

271

research. Retrieved on 4 May 2019, from https://cran.r- project.org/web/packages/agricolae/index.html. Meredith BK, Berry DP, Kearney F, Finlay EK, Fahey AG, Bradley DG and Lynn DJ 2013. A genome-wide association study for somatic cell score using the Illumina high-density bovine beadchip identifies several novel QTL potentially related to mastitis susceptibility. Frontiers in Genetics 4, 229. Meredith BK, Kearney FJ, Finlay EK, Bradley DG, Fahey AG, Berry DP and Lynn DJ 2012. Genome-wide associations for milk production and somatic cell score in Holstein-Friesian cattle in Ireland. BMC Genetics 13, 21. Van der Merwe JS, Smit FJ, Durand AM, Kruger LP and Michael LM 2005. A cost-benefit model comparing the California milk cell test and milk electrical resistance test. Onderstepoort Journal of Veterinary Research 72, 309–314. Meuwissen TH, Hayes BJ and Goddard ME 2001. Prediction of total genetic value using genome- wide dense marker maps. Genetics 157, 1819–1829. Meuwissen T, Hayes B and Goddard M 2016. Genomic selection: A paradigm shift in animal breeding. Animal Frontiers 6, 6–14. Meyer U, Everinghoff M, Gädeken D and Flachowsky G 2004. Investigations on the water intake of lactating dairy cows. Livestock Production Science 90, 117–121. Meyer D, Zeileis A and Hornik K 2006. The strucplot framework: Visualizing multi-way contingency tables with vcd. Journal of Statistical Software 17, 1–48. Meyer D, Zeileis A, Hornik K, Gerber F and Friendly M 2017. R Package ‘vcd’ version 1.4-4: Visualizing Categorical Data. Retrieved on 30 November 2019, from https://orcid.org/0000- 0003-0918-3766. Mi H, Muruganujan A, Casagrande JT and Thomas PD 2013. Large-scale gene function analysis with the PANTHER classification system. Nature Protocols 8, 1551–1566. Michenet A, Barbat M, Saintilan R, Venot E and Phocas F 2016. Detection of quantitative trait loci for maternal traits using high-density genotypes of Blonde d’Aquitaine beef cattle. BMC Genomics 17, 88. Microsoft 2016. Microsoft Official Academic Course. Wiley, Hoboken, New Jersey, USA. Miglior F, Fleming A, Malchiodi F, Brito LF, Martin P and Baes CF 2017. A 100-Year Review: Identification and genetic selection of economically important traits in dairy cattle. Journal of dairy science 100, 10251–10271.

272

Miller GY, Bartlett PC, Lance SE, Anderson J and Heider LE 1993. Costs of clinical mastitis and mastitis prevention in dairy herds. Journal of the American Veterinary Medical Association 202, 1230–1236. Ministry of Planning and Management 2017. Website of Ministry of Planning and Management. Retrieved on 1 January 2017, from http://www.mpi.gov.vn/Pages/tinhthanh.aspx. Minozzi G, Nicolazzi EL, Stella A, Biffani S, Negrini R, Lazzari B, Ajmone-Marsan P and Williams JL 2013. Genome wide analysis of fertility and production traits in Italian Holstein cattle. PLoS ONE 8, e80219. Moghaddar N, Swan AA and van der Werf JH 2014. Comparing genomic prediction accuracy from purebred, crossbred and combined purebred and crossbred reference populations in sheep. Genetics Selection Evolution 46, 58. Montanholi YR, Lim M, Macdonald A, Smith BA, Goldhawk C, Schwartzkopf-Genswein K and Miller SP 2015. Technological, environmental and biological factors: referent variance values for infrared imaging of the bovine. Journal of Animal Science and Biotechnology 6, 27. Montesinos-López OA, Montesinos-López A, Luna-Vázquez FJ, Toledo FH, Pérez-Rodríguez P, Lillemo M and Crossa J 2019. An R package for Bayesian analysis of multi-environment and multi-trait multi-environment data for genome-gased prediction. G3: Genes, Genomes, Genetics 9, 1355–1369. Moraes LE, Strathe AB, Fadel JG, Casper DP and Kebreab E 2014. Prediction of enteric methane emissions from cattle. Global Change Biology 20, 2140–2148. Moran JB 2009. Key performance indicators to diagnose poor farm performance and profitability of smallholder dairy farmers in Asia. Asian-Australasian Journal of Animal Sciences 22, 1709– 1717. Moran J 2010. Business management for tropical dairy farmers. CSIRO Publishing, Clayton, Victoria, Australia. Moran J 2012. Managing high grade dairy cows in the tropics. CSIRO Publishing, Clayton, Victoria, Australia. Moran JB 2013. Addressing the key constraints to increasing milk production from small holder dairy farms in tropical Asia. International Journal of Agriculture and Biosciences 2, 90–98. Moran J 2015a. Improving livestock welfare in developing countries: Bringing Australian export dairy heifers into the ‘welfare tent’. Agricultural Science 27, 56–62. Moran JB 2015b. Developing a farm audit to address the problems of stock welfare on small holder

273

dairy farms in Asia. International Journal of Livestock Production 6, 26–34. Moran J 2015c. Tropical dairy farming: Feeding management for small holder dairy farmers in the humid tropics. CSIRO Publishing, Clayton, Victoria, Australia. Moran JB and Brouwer JW 2013. Interrelationships between measures of cow and herd performance and farm profitability on Malaysian dairy farms. International Journal of Agriculture and Biosciences 2, 221–233. Moran J and Chamberlain P 2017. Blueprints for tropical dairy farming: increasing domestic milk production in developing countries. CSIRO Publishing, Clayton, Victoria, Australia. Moran J and Doyle R 2015a. Chapter 9. Stock management on Asian small holder dairy farms. In Cow Talk: understanding dairy cow behaviour to improve their welfare on Asian farms, pp. 159–180. CSIRO Publishing, Clayton, Victoria, Australia. Moran J and Doyle R 2015b. Cow talk. Understanding dairy cow behaviour to improve their welfare on Asian farms. CSIRO Publishing, Clayton, Victoria, Australia. Moran J and Morey P 2015. Strategies to increase the domestic production of raw milk in indonesia and other South East Asian Countries. In International Seminar on Tropical Animal Production (ISTAP), pp. 1–11. Morton JF 2007. The impact of climate change on smallholder and subsistence agriculture. Proceedings of the National Academy of Sciences of the United States of America 104, 19680– 19685. Mrode R, Ojango JMK, Okeyo AM and Mwacharo JM 2019. Genomic selection and use of molecular tools in breeding programs for indigenous and crossbred cattle in developing countries: Current status and future prospects. Frontiers in Genetics 9, 694. Mrode AR and Thompson R 2013. Linear models for the prediction of animal breeding values 3rd edition. CABI, Oxfordshire, UK. Muir WM 2007. Comparison of genomic and traditional BLUP-estimated breeding value accuracy and selection response under alternative trait and genomic parameters. Journal of Animal Breeding and Genetics 124, 342–355. Munoz SR and Bangdiwala SI 1997. Interpretation of Kappa and B statistics measures of agreement. Journal of Applied Statistics 24, 105–111. National Animal Welfare Advisory Committee 2018. Code of welfare: dairy cattle. Minister for Agriculture, New Zealand Government, Wellington, New Zealand. National Institue of Animal Husbandry 2000. Composition and nutritive value of animal feeds in

274

Vietnam. Agriculture Publishing House, Hanoi, Vietnam. National Institue of Animal Science 2016. Developing dairy cow breeds in Vietnam - current situation and solutions. National Institue of Animal Science, Hanoi, Vietnam. National Institute of Animal Sciences 2017. Development of dairy cattle breeds in Vietnam. National Institute of Animal Science. Retrieved on 25 February 2018, from http://vcn.vnn.vn/phat-trien- giong-bo-sua-o-viet-nam-hien-trang-va-giai-phap_n58215_g773.aspx. National Research Council 1971. A guide to environmental research on animals. National Academy of Sciences, Washington, DC, USA. Nayeri S, Sargolzaei M, Abo-ismail MK, May N, Miller SP, Schenkel F, Moore SS and Stothard P 2016. Genome-wide association for milk production and female fertility traits in Canadian dairy Holstein cattle. BMC Genetics 17, 1–11. Neal WM, Becker RB and Arnold PTD 1935. The digestible nutrients of Napier grass and crotalaria intermédia silages, natal grass hay, and the dried refuses of grape fruit and orange canneries. Journal of Agricultural Research 51, 173–176. Nei M 1987. Molecular evolutionary genetics. Columbia University Press, New York City, USA. Neogen Australasia 2018. Neogen Australasia Pty Limited. Retrieved on 7 December 2019, from https://genomics.neogen.com/au/. Neogen Australasia Pty Limited 2015. Hair Sample Instructions. Retrieved on 7 December 2019, from https://www.neogen.com/genomics-how-to. Neylon JM and Kung L 2003. Effects of cutting height and maturity on the nutritive value of corn silage for lactating cows1,2. Journal of Dairy Science 86, 2163–2169. NFACC 2009. Code of practice for the care and handling of dairy cattle. National Farm Animal Care Council (NFACC), Ottawa, Ontario, Canada. Nga BT 2017a. An analysis of the fresh milk chain in Ha Nam province of Vietnam. Environmental Science: Water Research & Technology 1, 33–39. Nga BT 2017b. Efficiency of milk production in small holders of Cu Chi district, Ho Chi Minh city of Vietnam. Scholars Journal of Economics, Business and Management 4, 527–531. Ngo TD, Pham KC, Somers R, Cai D Van and Trach NX 2009. Nutrition and feeding management in dairy cattle. Vietnam Belgium Dairy Project, Ha Noi. Ngo SG and Trinh HD 2006. Initial application of temperature and humidity index THI on evaluate the feasibility of livestock development in the Northern Delta. The 10th Scientific Report of Vietnam Institute of Meteorology, Hydrology and Environment. Retrieved on 10 December 275

2017, from http://www.dairyvietnam.com/vn/Quan-ly-chan-nuoi-bo-sua/Buoc-dau-su-dung- chi-so-nhiet-am-THI-de-danh-gia-kha-nang-phat-trien-chan-nuoi-dai-gia-suc-va-gia-cam- khu-vuc-dong-bang-bac-bo.html. Nguyen N 2014. An update of the milk market in Vietnam. Lactose-free dairy opportunity. Arcada, 64pp. Nguyen TTT, Bowman PJ, Haile-Mariam M, Nieuwhof GJ, Hayes BJ and Pryce JE 2017a. Short communication: Implementation of a breeding value for heat tolerance in Australian dairy cattle. Journal of Dairy Science 100, 7362–7367. Nguyen TTT, Bowman PJ., Haile-Mariam M, Pryce JE and Hayes BJ 2016a. Genomic selection for tolerance to heat stress in Australian dairy cattle. Journal of Dairy Science 99, 2849–2862. Nguyen MH, Cesaro JD, Pham DK, Hoang VQ and Duteurtre G 2016b. Which is a sustainable development of dairy sector in Vietnam: Dairy household farms or intensive and large scale farms. Journal of Animal Science and Technology 61, 12–21. Nguyen KC, Nguyen VK and Hanzen C 2017b. Study on reproductive performance of Holstein x Lai Sind crossbred dairy heifers and cows at smallholdings in Ho Chi Minh city, Vietnam. Tropical Animal Health and Production 49, 483–489. Nguyen Xuan Dong 2017. Decision 1357 / QD-UBND 2017 approving the agricultural development planning of Ha Nam 2025 2035 (in Vietnamese). People’s Committee of Ha Nam province. Retrieved on 17 January 2018, from https://thuvienphapluat.vn/van-ban/Linh-vuc-khac/Quyet- dinh-1357-QD-UBND-2017-phe-duyet-quy-hoach-phat-trien-nong-nghiep-Ha-Nam-2025- 2035-360389.aspx. Nishio M and Satoh M 2014. Including dominance effects in the genomic BLUP method for genomic evaluation. PLoS ONE 9, e85792. Noi V Van, Them TT, Luong NH, Duc N Van, Son NH, Ha TS, Tan ND and Ha LT 2006. Growth, reproduction, and milk production abilities of first generation 75% HF. National Institue of Animal Science, Hanoi, Vietnam. Norberg E, Hogeveen H, Korsgaard IR, Friggens NC, Sloth KHMN and Løvendahl P 2004a. Electrical conductivity of milk: ability to predict mastitis status. Journal of dairy science 87, 1099–1107. Norberg E, Rogers GW, Goodling RC, Cooper JB and Madsen P 2004b. Genetic parameters for test- day electrical conductivity of milk for first-lactation cows from random regression models. Journal of Dairy Science 87, 1917–1924.

276

Novembre J 2016. Pritchard, Stephens, and Donnelly on population structure. Genetics 204, 391– 393. NRC 2001. Nutrient requirements of dairy cattle. National Academies Press, Washington, DC, USA. NRC 2016. Nutrient requirements of beef cattle. National Academies Press, Washington, DC, USA. Oetting WS, Jacobson PA and Israni AK 2017. Validation is critical for genome-wide association study–based associations. American Journal of Transplantation 17, 318–319. Ogle D, Wheeler P and Dinno A 2019. R package ‘FSA’ version 0.8.26: simple fisheries stock assessment methods. Retrieved on 29 December 2019, from https://rdrr.io/cran/FSA/man/FSA.html. Ogorevc J, Kunej T, Razpet A and Dovc P 2009. Database of cattle candidate genes and genetic markers for milk production and mastitis. Animal Genetics 40, 832–51. Oliveira HRR, Cant JPP, Brito LFF, Feitosa FLBLB, Chud TCSCS, Fonseca PASAS, Jamrozik J, Silva FFF, Lourenco DALAL and Schenkel FSS 2019. Genome-wide association for milk production traits and somatic cell score in different lactation stages of Ayrshire, Holstein, and Jersey dairy cattle. Journal of Dairy Science 102, 8159–817. Olson TA, Lucena C, Chase CC and Hammond AC 2003. Evidence of a major gene influencing hair length and heat tolerance in dairy cattle. Journal of Animal Science 81, 80–90. Oltenacu PA and Broom DM 2010. The impact of genetic selection for increased milk yield on the welfare of dairy cows. Animal Welfare 19, 39–49. Östensson K, Vo L, Sjögren N and Wredle E 2013. Prevalence of subclinical mastitis and isolated udder pathogens in dairy cows in Southern Vietnam. Tropical Animal Health and Production 45, 979–986. Otto PI, Guimarães SEF, Verardo LL, Azevedo ALS, Vandenplas J, Sevillano CA, Marques DBD, Pires M de FA, de Freitas C, Verneque RS, Martins MF, Panetto JCC, Carvalho WA, Gobo DOR, da Silva MVGB and Machado MA 2019. Genome-wide association studies for heat stress response in Bos taurus × Bos indicus crossbred cattle. Journal of Dairy Science 102, 8148– 8158. Owen E, Smith T and Makkar H 2012. Successes and failures with animal nutrition practices and technologies in developing countries: A synthesis of an FAO e-conference. Animal Feed Science and Technology 174, 211–226. Paraffin AS, Zindove TJ and Chimonyo M 2018. Perceptions of factors affecting milk quality and safety among large- and small-scale dairy farmers in Zimbabwe. Journal of Food Quality 3, 1–

277

7. Pegolo S, MacH N, Ramayo-Caldas Y, Schiavon S, Bittante G and Cecchinato A 2018. Integration of GWAS, pathway and network analyses reveals novel mechanistic insights into the synthesis of milk proteins in dairy cows. Scientific Reports 8, 566. Pérez P and de los Campos G 2014. Genome-wide regression and prediction with the BGLR statistical package. Genetics 198, 483–495. Pérez P, De los Campos G, Perez P and De los Campos G 2014. BGLR : A statistical package for whole genome regression and prediction. Genetics 198, 483–495. Petrini J, Iung LHS, Rodriguez MAP, Salvian M, Pértille F, Rovadoscki GA, Cassoli LD, Coutinho LL, Machado PF, Wiggans GR and Mourão GB 2016. Genetic parameters for milk fatty acids, milk yield and quality traits of a Holstein cattle population reared under tropical conditions. Journal of Animal Breeding and Genetics 133, 384–395. Phong LV and Thu NV 2016. Investigation of growth, production and dairy techniques at Song Hau Farm, Long Hoa and Evergrowth Dairy Cow Cooperative in the Mekong Delta. Can Tho University Journal of Science 2, 48–55. Piccand V, Cutullic E, Meier S, Schori F, Kunz PL, Roche JR and Thomet P 2013. Production and reproduction of Fleckvieh, Brown Swiss, and 2 strains of Holstein-Friesian cows in a pasture- based, seasonal-calving dairy system. Journal of Dairy Science 96, 5352–5363. Pilachai R, Schonewille JTT, Thamrongyoswittayakul C, Aiumlamai S, Wachirapakorn C, Everts H and Hendriks WH 2013. Diet factors and subclinical laminitis score in lactating cows of smallholder dairy farms in Thailand. Livestock Science 155, 197–204. Pinto S, Hoffmann G, Ammon C, Amon B, Heuwieser W, Halachmi I, Banhazi T and Amon T 2019. Influence of barn climate, body postures and milk yield on the respiration rate of dairy cows. Annals of Animal Science 19, 469–481. Pinzón-Sánchez C and Ruegg PL 2011. Risk factors associated with short-term post-treatment outcomes of clinical mastitis. Journal of Dairy Science 94, 3397–3410. Poikalainen V, Praks J, Veermäe I and Kokin E 2012. Infrared temperature patterns of cow’s body as an indicator for health control at precision cattle farming. Agronomy Research Biosystem Engineering 1, 187–194. Polat B, Colak A, Cengiz M, Yanmaz LE, Oral H, Bastan A, Kaya S and Hayirli A 2010. Sensitivity and specificity of infrared thermography in detection of subclinical mastitis in dairy cows. Journal of Dairy Science 93, 3525–3532.

278

Polsky L and von Keyserlingk MAG 2017. Invited review: Effects of heat stress on dairy cattle welfare. Journal of Dairy Science 100, 8645–8657. Polsky LB, Madureira AML, Filho ELD, Soriano S, Sica AF, Vasconcelos JLM and Cerri RLA 2017. Association between ambient temperature and humidity, vaginal temperature, and automatic activity monitoring on induced estrus in lactating cows. Journal of Dairy Science 100, 8590– 8601. Poncheki JK, Canha MLS, Viechnieski SL and de Almeida R 2015. Analysis of daily body weight of dairy cows in early lactation and associations with productive and reproductive performance. Revista Brasileira de Zootecnia 44, 187–192. Popescu S, Borda C, Diugan EA, Niculae M, Stefan R and Sandru CD 2014. The effect of the housing system on the welfare quality of dairy cows. Italian Journal of Animal Science 13, 2940. Porras-Hurtado L, Ruiz Y, Santos C, Phillips C, Carracedo A and Lareu M V 2013. An overview of STRUCTURE: applications, parameter settings, and supporting software. Frontiers in Genetics 4, 98. Preez JH du, Hattingh PJ, Giesecke WH and Eisenberg BE 1990. Heat stress in dairy cattle and other livestock under southern African conditions. III. Monthly temperature-humidity index mean values and their significance in the performance of dairy cattle - PubMed. Onderstepoort Journal of Veterinary Research 57, 243–248. Preston TR 1989. Matching livestock systems with available resources. In Feeding dairy cows in the tropics (eds. A. Speedy and R. Sansoucy). Food and Agricultural Organization of the United Nations, Rome, Italy. Pryce JE, Coffey MP and Simm G 2001. The relationship between body condition score and reproductive performance. Journal of Dairy Science 84, 1508–1515. Pszczola M, Strabel T, Mulder HAA and Calus MPLPL 2012. Reliability of direct genomic values for animals with different relationships within and to the reference population. Journal of Dairy Science 95, 389–400. Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MAR, Bender D, Maller J, Sklar P, de Bakker PIW, Daly MJ and Sham PC 2007. PLINK: A tool set for whole-genome association and population-based linkage analyses. American Journal of Human Genetics 81, 559–575. R Core Team 2018. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. Raftery A, Hoeting J, Volinsky C, Painter I, Yeung KY and Sevcikova H 2019. R Package ‘BMA’:

279

Bayesian Model Averaging. Rahayu AP, Hartatik T, Purnomoadi A and Kurnianto E 2018. Estimation of 305 day milk yield from cumulative monthly and bimonthly test day records in Indonesian Holstein cattle. IOP Conference Series: Earth and Environmental Science 119, 012028. Ramírez-Rivera EJ, Rodríguez-Miranda J, Huerta-Mora IR, Cárdenas-Cágal A and Juárez-Barrientos JM 2019. Tropical milk production systems and milk quality: a review. Tropical Animal Health and Production 51, 1295–1305. Rashamol VP, Sejian V, Pragna P, Lees AM, Bagath M, Krishnan G and Gaughan JB 2019. Prediction models, assessment methodologies and biotechnological tools to quantify heat stress response in ruminant livestock. International Journal of Biometeorology 63, 1265–1281. Ravagnolo O, Misztal I and Hoogenboom G 2000. Genetic component of heat stress in dairy cattle, development of heat index function. Journal of Dairy Science 83, 2120–2125. Raven LA, Cocks BG, Kemper KE, Chamberlain AJ, Vander Jagt CJ, Goddard ME and Hayes BJ 2016. Targeted imputation of sequence variants and gene expression profiling identifies twelve candidate genes associated with lactation volume, composition and calving interval in dairy cattle. Mammalian Genome 27, 81–97. Red Tractor Assurance for Farms 2017. Dairy Standards. Assured Food Standards, London, UK. Reed E, Nunez S, Kulp D, Qian J, Reilly MP and Foulkes AS 2015. A guide to genome-wide association analysis and post-analytic interrogation. Statistics in Medicine 34, 3769–3792. Regan WM and Richardson GA 1938. Reactions of the dairy cow to changes in environmental temperature. Journal of Dairy Science 21, 73–79. Rekant SI, Lyons MA, Pacheco JM, Arzt J and Rodriguez LL 2016. Veterinary applications of infrared thermography. American Journal of Veterinary Research 77, 98–107. Renaudeau D, Collin A, Yahav S, De Basilio V, Gourdine JL and Collier RJ 2012. Adaptation to hot climate and strategies to alleviate heat stress in livestock production. Animal 6, 707–728. Revelle W 2019. R package ‘psych’ version 1.9.12: Procedures for Psychological, Psychometric, and Personality Research. Retrieved on 29 December 2019, from https://personality- project.org/r/psych. Richards S, Vanleeuwen J, Shepelo G, Karuoya Gitau G, Kamunde C, Uehlinger F and Wichtel J 2014. Associations of farm management practices with annual milk sales on smallholder dairy farms in Kenya. Veterinary World 8, 88–96. Rincent R, Laloe D, Nicolas S, Altmann T, Brunel D, Revilla P, Rodriguez VM, Moreno-Gonzalez

280

J, Melchinger A, Bauer E, Schoen C-C, Meyer N, Giauffret C, Bauland C, Jamin P, Laborde J, Monod H, Flament P, Charcosset A and Moreau L 2012. Maximizing the reliability of genomic selection by optimizing the calibration set of reference individuals: comparison of methods in two diverse groups of maize inbreds (Zea mays L.). Genetics 192, 715–728. Riveiro JA, Mantecón AR, Álvarez CJ and Lavín P 2013. A typological characterization of dairy Assaf breed sheep farms at NW of Spain based on structural factor. Agricultural Systems 120, 27–37. Robinson PH and Ahmadi A 2015. PCDAIRY: least cost and ration analysis programs for dairy cattle, version number 2016. Roche JR, Friggens NC, Kay JK, Fisher MW, Stafford KJ and Berry DP 2009. Invited review: Body condition score and its association with dairy cow productivity, health, and welfare. Journal of Dairy Science 92, 5769–5801. de Roos APW, Hayes BJ and Goddard ME 2009. Reliability of genomic predictions across multiple populations. Genetics 183, 1545–1553. Rosen BD, Bickhart DM, Schnabel RD, Koren S, Elsik CG, Zimin A, Dreischer C, Schultheiss S, Hall R, Schroeder SG, Van Tassell CP, Smith TPL and Medrano JF 2018. Modernizing the bovine reference genome assembly. In Proceedings of the World Congress on Genetics Applied to Livestock Production, p. 802. Auckland, New Zealand. RSPCA 2018. RSPCA welfare standards for dairy cattle. Royal Society for the Prevention of Cruelty to Animals, West Sussex, UK. Rusdy M 2016. Elephant grass as forage for ruminant animals. Livestock Research for Rural Development 28, #49. Rushen J, Passillé AM de, Keyserlingk MAG von and Weary DM 2008. The welfare of cattle. Springer, Dordrecht, The Netherlands. Sabedot MA, Romano G de S, Pedrosa VB and Pinto LFB 2018. Genetic parameters for type score traits and milk production in Brazilian Jersey herds. Revista Brasileira de Zootecnia 47, e20170093. Salles MSV, da Silva SC, Salles FA, Roma LC, El Faro L, Bustos Mac Lean PA, Lins de Oliveira CE and Martello LS 2016. Mapping the body surface temperature of cattle by infrared thermography. Journal of Thermal Biology 62, 63–69. Sankararaman S, Sridhar S, Kimmel G and Halperin E 2008. Estimating local ancestry in admixed populations. American Journal of Human Genetics 82, 290–303.

281

Santos VS, Martins Filho S, Resende MDV, Azevedo CF, Lopes PS, Guimarães SEF and Silva FF 2016. Genomic prediction for additive and dominance effects of censored traits in pigs. Genetics and Molecular Research 15, gmr15048764. Sathiyabarathi M, Jeyakumar S, Manimaran A, Pushpadass HA, Sivaram M, Ramesha KP, Das DN, Kataktalware MA, Jayaprakash G and Patbandha TK 2016. Investigation of body and udder skin surface temperature differentials as an early indicator of mastitis in Holstein Friesian crossbred cows using digital infrared thermography technique. Veterinary World 9, 1386–1391. Schingoethe DJ, Linke KN, Kalscheur KF, Hippen AR, Rennich DR and Yoon I 2004. Feed efficiency of mid-lactation dairy cows fed yeast culture during summer. Journal of Dairy Science 87, 4178–4181. Schmid M and Bennewitz J 2017. Invited review: Genome-wide association analysis for quantitative traits in livestock – a selective review of statistical models and experimental designs. Archives Animal Breeding 60, 335–346. Schneider PL, Beede DK and Wilcox CJ 1986. Responses of lactating cows to dietary sodium source and quantity and potassium quantity during heat stress. Journal of Dairy Science 69, 99–110. Schrooten C, Bovenhuis H, Coppieters W and Van Arendonk JAM 2000. Whole genome scan to detect quantitative trait loci for conformation and functional traits in dairy cattle. Journal of Dairy Science 83, 795–806. Seath DM 1947. Heritability of heat tolerance in dairy cattle. Journal of Dairy Science 30, 137–144. Sejian V, Gaughan J, Baumgard L and Prasad C 2015. Climate change impact on livestock: adaptation and mitigation. Springer, New Delhi, India. Sempéré G, Moazami-Goudarzi K, Eggen A, Laloë D, Gautier M and Flori L 2015. WIDDE: a Web- Interfaced next generation database for genetic diversity exploration, with a first application in cattle. BMC Genomics 16, 940. Seo K, Mohanty TR, Choi T and Hwang I 2007. Biology of epidermal and hair pigmentation in cattle: a mini-review. Veterinary Dermatology 18, 392–400. Sharma A, Lee JS, Dang CG, Sudrajad P, Kim HC, Yeon SH, Kang HS and Lee S-H 2015. Stories and challenges of genome wide association studies in livestock — A review. Asian- Australasian Journal of Animal Sciences 28, 1371–1379. Shin D, Lee C, Park K, Kim H and Cho K 2017. Genome-association analysis of Korean Holstein milk traits using genomic estimated breeding value. Asian-Australasian Journal of Animal Sciences 00, 1–11.

282

Shmueli G 2010. To explain or to predict? Statistic Science 25, 289–310. Shwartz G, Rhoads ML, Vanbaale MJ, Rhoads RP and Baumgard LH 2009. Effects of a supplemental yeast culture on heat-stressed lactating Holstein cows. Journal of Dairy Science 92, 935–942. Silanikove N 2000. Effects of heat stress on the welfare of extensively managed domestic ruminants. Livestock Production Science 67, 1–18. Da Silva RG, Maia ASC and de Macedo Costa LL 2014. Index of thermal stress for cows (ITSC) under high solar radiation in tropical environments. International Journal of Biometeorology 59, 551–559. Silva MVB, dos Santos DJA, Boison SA, Utsunomiya ATH, Carmo AS, Sonstegard TS, Cole JB and Van Tassell CP 2014. The development of genomics applied to dairy breeding. Livestock Science 166, 66–75. Solberg TR, Sonesson AK, Woolliams JA and Meuwissen THE 2008. Genomic selection using different marker types and densities. Journal of Animal Science 86, 2447–2454. Soriani N, Panella G and Calamari L 2013. Rumination time during the summer season and its relationships with metabolic conditions and milk production. Journal of Dairy Science 96, 5082–5094. Spencer CCA, Su Z, Donnelly P and Marchini J 2009. Designing genome-wide association studies: sample size, power, imputation, and the choice of genotyping chip. PLoS Genetics 5, e1000477. Stachowicz K, Sargolzaei M, Miglior F and Schenkel FS 2011. Rates of inbreeding and genetic diversity in Canadian Holstein and Jersey cattle. Journal of Dairy Science 94, 5160–5175. Stoop WM, Van Arendonk JAM, Heck JML, Van Valenberg HJF and Bovenhuis H 2008. Genetic parameters for major milk fatty acids and milk production traits of Dutch Holstein-Friesians. Journal of Dairy Science 91, 385–394. Van Straten M, Shpigel NY and Friger M 2008. Analysis of daily body weight of high-producing dairy cows in the first one hundred twenty days of lactation and associations with ovarian inactivity. Journal of Dairy Science 91, 3353–3362. Streit M, Wellmann R, Reinhardt F, Thaller G, Piepho HP and Bennewitz J 2013. Using genome- wide association analysis to characterize environmental sensitivity of milk traits in dairy cattle. G3: Genes, Genomes, Genetics 3 3, 1085–1093. Su V Van and Binh DV 2001. Cattle breeding in Vietnam. In Development startegies for genetic evaluation for beef production in developing countries (eds. J. Allen and A. Na-Chiangmai), pp. 98–101. Australian Centre for International Agricultural Research (ACIAR), Canberra,

283

Australian Capital Territory, Australia. Sun X, Qu L, Garrick DJ, Dekkers JCM and Fernando RL 2012. A fast EM algorithm for BayesA- like prediction of genomic breeding values. PLoS ONE 7, e49157. Sun C, VanRaden PM, Cole JB and O’Connell JR 2014. Improvement of prediction ability for genomic selection of dairy cattle by including dominance effects. PLoS ONE 9, e103934. Suzuki K, Kanameda M, Ogawa T, Nguyen TTD, Dang TTS, Luu QH and Pfeiffer DU 2006. Productivity and socio-economic profile of dairy cattle farmers amongst rural smallholder communities in northern Vietnam. Livestock Science 101, 242–250. Syrstad O 1989. Dairy cattle cross-breeding in the tropics: Performance of secondary cross-bred populations. Livestock Production Science 23, 97–106. Szkiba D, Kapun M, von Haeseler A and Gallach M 2014. SNP2GO: Functional analysis of genome- wide association studies. Genetics 197, 285–289. Tadesse M and Dessie T 2003. Milk production performance of Zebu, Holstein Friesian and their crosses in Ethiopia. Livestock Research for Rural Development 15, #26. Talukder S, Celi P, Kerrisk KL, Garcia SC and Dhand NK 2015. Factors affecting reproductive performance of dairy cows in a pasture-based, automatic milking system research farm: a retrospective, single-cohort study. Animal Production Science 55, 31. Talukder S, Kerrisk KL, Ingenhoff L, Thomson PC, Garcia SC and Celi P 2014. Infrared technology for estrus detection and as a predictor of time of ovulation in dairy cows in a pasture-based system. Theriogenology 81, 925–935. Tang H, Peng J, Wang P and Risch NJ 2005. Estimation of individual admixture: Analytical and study design considerations. Genetic Epidemiology 28, 289–301. Tao S, Orellana Rivas RM, Marins TN, Chen YC, Gao J and Bernard JK 2020. Impact of heat stress on lactational performance of dairy cows. Theriogenology 150, 437–444. Tattersall GJ 2016. Infrared thermography: A non-invasive window into thermal physiology. Comparative Biochemistry and Physiology -Part A 202, 78–98. TCVN 5504 2010. TCVN 5504:2010 (ISO 2446-2008) Determination of fat content. Ministry of Science and Technology, Hanoi, Vietnam. TCVN 8082 2013. TCVN 8082:2013 (ISO 6731-2010) Milk, cream and evaporated milk - Determination of total solids content (Reference method). Ministry of Science and Technology, Hanoi, Vietnam. TCVN 8099 2015. TCVN 8099-1:2015 (ISO 8968-1:2014) Milk and milk products - Determination 284

of nitrogen content - Part 1: Kjeldahl principle and crude protein calculation. Ministry of Science and Technology, Hanoi, Vietnam. Tedeschi LO, Cavalcanti LFL, Fonseca MA, Herrero M and Thornton PK 2014. The evolution and evaluation of dairy cattle models for predicting milk production: An agricultural model intercomparison and improvement project (AgMIP) for livestock. Animal Production Science 54, 2052–2067. The Dairyland Initiative 2020. Heat abatement and ventilation of adult cow facilities. Retrieved on 15 March 2020, from https://thedairylandinitiative.vetmed.wisc.edu/home/housing- module/adult-cow-housing/ventilation-and-heat-abatement/. Thom EC 1959. The discomfort index. Weatherwise 12, 57–61. Thomasen JR, Guldbrandtsen B, Su G, Brøndum RF and Lund MS 2012. Reliabilities of genomic estimated breeding values in Danish Jersey. The Animal Consortium 6, 789–796. Thompson JR, Attia J and Minelli C 2011. The meta-analysis of genome-wide association studies. Briefings in Bioinformatics 12, 259–269. Thong L Van 2017. Application of high technologies in dairy production conference. In Application of high technologies in dairy production conference, 2nd June 2017. Hanoi, Vietnam. Tiem P Van 2014. Evaluation and selection of Holstein Friesian bulls in Vietnam. National Institue of Animal Science. Todde G, Murgia L, Caria M and Pazzona A 2016. A multivariate statistical analysis approach to characterize mechanization, structural and energy profile in Italian dairy farms. Energy Reports 2, 129–134. Toghiani S 2012. Genetic relationships between production traits and reproductive performance in Holstein dairy cows. Archiv Tierzucht 55, 458–468. Trach NX 2004. Reproductive performance and milk productivity of several types of dairy crossbred cattle raised in Mocchau and Hanoi statistical. Vietnam Journal of Animal Science 1, 12–14. Trach NX 2017a. Situation of dairy production in Vietnam and the main concerns. In Improving dairy cattle health and production in Vietnam on 19 December 2017. Hanoi, Vietnam. Trach NX 2017b. Animal Welfare: definition and practice (in Vietnamese). In National Conference and Exhibition on Animal and Veterinary Sciences on 11 to 12 March 2017. Can Tho University, Can Tho, Vietnam. Trach NX, Thom MT and Ban L Van 2007. Textbook of cattle and buffalo production. Agriculture Publishing House, Hanoi, Vietnam.

285

Trang Y 2019. Vinamilk additionally imported 1,600 HF and A2 dairy cows from the US. Dan Tri. Retrieved on 17 May 2020, from https://dantri.com.vn/kinh-doanh/vinamilk-nhap-them-1600- bo-sua-hf-va-a-2-tu-my-20190314154952778.htm. Tue T 2020. Import 4,500 HF high-yielding dairy cows from the US. Tuoi Tre Online. Retrieved on 17 May 2020, from https://tuoitre.vn/nhap-4-500-con-bo-sua-cao-san-hf-tu-my- 20200202153204273.htm. Tue N Van, Binh DV and Sanh mai Van 2010. Milk production of three types of crossbreeds 1/2HF, 3/4HF and 7/8HF (Holstein Friesian x laisind) kept in small holders, Bac ninh province. Journal of Animal Science and Technology 26, 9–16. Tuyen DK 2009. Development situation of dairy farming in Vietnam in the period 2001-2009 and forecast for 2010-2020 (in Vietnamese). Department of Livestock Production, Hanoi, Vietnam. Tyrrell HF and Reid JT 1965. Prediction of the energy value of cow’s milk. Journal of Dairy Science 48, 1215–23. Unruh EM, Theurer ME, White BJ, Larson RL, Drouillard JS and Schrag N 2017. Evaluation of infrared thermography as a diagnostic tool to predict heat stress events in feedlot cattle. American Journal of Veterinary Research 78, 771–777. Usamentiaga R, Venegas P, Guerediaga J, Vega L, Molleda J and Bulnes F 2014. Infrared thermography for temperature measurement and non-destructive testing. Sensors 14, 12305– 12348. USDA 2016. Dairy 2014: Dairy cattle management practices in the United States, 2014. USDA, Colorado, USA. Vallimont JE, Dechow CD, Daubert JM, Dekleva MW, Blum JW, Barlieb CM, Liu W, Varga GA, Heinrichs AJ and Baumrucker CR 2010. Genetic parameters of feed intake, production, body weight, body condition score, and selected type traits of Holstein cows in commercial tie-stall barns. Journal of Dairy Science 93, 4892–4901. Vang ND, Thuy LT, Ly LV and Thu N Van 2003. The Vietnam national country report on animal genetic resources. National Institute of Animal Husbandry, Hanoi, Vietnam. VanRaden PM 2008. Efficient methods to compute genomic predictions. Journal of Dairy Science 91, 4414–4423. VanRaden PM and Sanders AH 2003. Economic merit of crossbred and purebred US dairy cattle. Journal of Dairy Science 86, 1036–1044. Vanraden PM, Van Tassell CP, Wiggans GR, Sonstegard TS, Schnabel RD, Taylor JF and Schenkel

286

FS 2009. Invited review: Reliability of genomic predictions for North American Holstein bulls. Journal of Dairy Science 92, 16–24. Veissier I, Andanson S, Dubroeucq H and Pomiès D 2008. The motivation of cows to walk as thwarted by tethering. Journal of Animal Science 86, 2723–2729. Veissier I, Van laer E, Palme R, Moons CPH, Ampe B, Sonck B, Andanson S and Tuyttens FAM 2018. Heat stress in cows at pasture and benefit of shade in a temperate climate region. International Journal of Biometeorology 62, 585–595. Verbyla KL, Hayes BJ, Bowman PJ and Goddard ME 2009. Accuracy of genomic selection using stochastic search variable selection in Australian Holstein Friesian dairy cattle. Genetics Research 91, 307. Vietnam Japan Digital Scale Company 2017. Digital scale model OCS M 100. Retrieved on 18 August 2017, from http://www.canvietnhat.com/can-treo-ocs-m-100kg.html. Vinamilk 2017. Vinamilk đầu tư nhiều trang trại bò sữa công nghệ cao. Retrieved on 14 January 2018, from https://www.vinamilk.com.vn/vi/vung-nguyen-lieu/chi-tiet-tin/383/1515/vinamilk- dau-tu-nhieu-trang-trai-bo-sua-cong-nghe-cao. Visscher PM, Wray NR, Zhang Q, Sklar P, McCarthy MI, Brown MA and Yang J 2017. 10 Years of GWAS discovery: biology, function, and translation. The American Journal of Human Genetics 101, 5–22. Visser C, Lashmar SF, Van Marle-Köster E, Poli MA and Allain D 2016. Genetic diversity and population structure in South African, French and Argentinian Angora Goats from genome- wide SNP data. PLoS ONE 11, 154353. Vo L 2011. Milk production on smallholder dairy cattle farms in southern Vietnam. Swedish University of Agricultural Sciences Uppsala, 52pp. De Vries A 2020. Symposium review: why revisit dairy cattle productive lifespan? Journal of Dairy Science 103, 3838–3845. de Vries M, Bokkers EAM, Dijkstra T, van Schaik G and de Boer IJM 2011. Invited review: Associations between variables of routine herd data and dairy cattle welfare indicators. Journal of Dairy Science 94, 3213–3228. Vu NH, Lambertz C and Gauly M 2016. Factors influencing milk yield, quality and revenue of dairy farms in southern Vietnam. Asian Journal of Animal Sciences 10, 290–299. Wang X, Bjerg BS, Choi CY, Zong C and Zhang G 2018. A review and quantitative assessment of cattle-related thermal indices. Journal of Thermal Biology 77, 24–37.

287

Wang MH, Cordell HJ and Steen K Van 2019. Statistical methods for genome-wide association studies. Seminars in Cancer Biology 55, 53–60. Wang X, Ma P, Liu J, Zhang Q, Zhang Y, Ding X, Jiang L, Wang Y, Zhang Y, Sun D, Zhang S, Su G and Yu Y 2015. Genome-wide association study in Chinese Holstein cows reveal two candidate genes for somatic cell score as an indicator for mastitis susceptibility. BMC Genetics 16, 111. Ward AC 2009. The role of causal criteria in causal inferences: Bradford Hill’s ‘aspects of association’. Epidemiologic Perspectives and Innovations 6, 2. Warnes GR, Bolker B, Lumley T and Johnson RC 2018. R package ‘gmodels’ version 2.18.1: Various R programming tools for model fitting. from https://cran.r-project.org/package=gmodels. Weigel KA 2001. Controlling Inbreeding in Modern Breeding Programs. Journal of Dairy Science 84, E177–E184. Weigel KA, VanRaden PM, Norman HD and Grosu H 2017. A 100-Year Review: Methods and impact of genetic selection in dairy cattle—From daughter–dam comparisons to deep learning algorithms. Journal of Dairy Science 100, 10234–10250. Weir BS and Cockerham CC 1984. Estimating F-Statistics for the Analysis of Population Structure. Evolution 38, 1358. Weir BS and Goudet J 2017. A unified characterization of population structure and relatedness. Genetics 206, 2085–2103. Welderufael BG, Løvendahl P, de Koning D-J, Janss LLG and Fikse WF 2018. Genome-wide association study for susceptibility to and recoverability from mastitis in Danish Holstein cows. Frontiers in Genetics 9, 1–12. Welfare Quality® Consortium 2009. Welfare Quality® assessment protocol for cattle. Welfare Quality Consortium, Lelystad, Netherlands. Weller JI, Ezra E and Ron M 2017. Invited review: A perspective on the future of genomic selection in dairy cattle. Journal of Dairy Science 100, 8633–8644. Weng Z, Zhang Z, Zhang Q, Fu W, He S and Ding X 2013. Comparison of different imputation methods from low- to high-density panels using Chinese Holstein cattle. Animal 7, 729–735. West JW 2003. Effects of heat-stress on production in dairy cattle. Journal of Dairy Science 86, 2131– 2144. West JW, Mullinix BG and Bernard JK 2003. Effects of hot, humid weather on milk temperature, dry matter intake, and milk yield of lactating dairy cows. Journal of Dairy Science 86, 232–242.

288

Wickham H, Chang W, Henry L, Pedersen TL, Takahashi K, Wilke C, Woo K and Yutani H 2019. R package ‘ggplot2’ version 3.2.1: Create elegant data visualisations using the grammar of graphics. Retrieved on 1 December 2019, from https://github.com/tidyverse/ggplot2. Wickham H, Kalicinski M, Valery K, Leitienne C, Colbert B and Hoerl D 2016. R package ‘readxl’ version 0.1.1: read excel files. Retrieved on 31 December 2019, from http://rapidxml.sourceforge.net. Widyas N, Mahfudhoh N and Prastowo S 2019. Genetic parameters of milk yield from three consecutive lactation periods of imported Holstein cows. AIP Conference Proceedings 2202, 20076. Wientjes YCJ, Veerkamp RF and Calus MPL 2013. The effect of linkage disequilibrium and family relationships on the reliability of genomic prediction. Genetics 193, 621–631. Wiggans GR, Cole JB, Hubbard SM and Sonstegard TS 2017. Genomic selection in dairy cattle: the USDA experience. Annual Review of Animal Biosciences 5, 309–327. Wijga S, Bastiaansen JWM, Wall E, Strandberg E, de Haas Y, Giblin L and Bovenhuis H 2012. Genomic associations with somatic cell score in first-lactation Holstein cows. Journal of Dairy Science 95, 899–908. Wolc A, Arango J, Settar P, Fulton JE, O’Sullivan NP, Preisinger R, Habier D, Fernando R, Garrick DJ and Dekkers JC 2011. Persistence of accuracy of genomic estimated breeding values over generations in layer chickens. Genetics Selection Evolution 43, 23. Wolfenson D and Roth Z 2019. Impact of heat stress on cow reproduction and fertility. Animal Frontiers 9, 32–38. Wongpom B, Koonawootrittriron S, Elzo MA and Suwanasopee T 2017. Milk yield, fat yield and fat percentage associations in a Thai multibreed dairy population. Agriculture and Natural Resources 51, 218–222. Wongpom B, Koonawootrittriron S, Elzo MA, Suwanasopee T and Jattawa D 2019. Accuracy of genomic-polygenic estimated breeding value for milk yield and fat yield in the Thai multibreed dairy population with five single nucleotide polymorphism sets. Asian-Australasian Journal of Animal Sciences 32, 1340–1348. World Organisation for Animal Health 2011. OIE - Terrestrial animal health code. World Organisation for Animal Health (OIE), Paris, French. Wright S 1965. The interpretation of population structure by F-statistics with special regard to systems of mating. Evolution 19, 395–420.

289

Wright S 1978. Evolution and the genetics of populations. Vol 4: Variability within and among natural populations. Journal of Human Evolution 9, 157. Yaemkong S, Koonawootrittriron S, Elzo M and Suwanasopee T 2010. Milk quantity, quality and revenue in dairy farms supported by a private organization in Central Thailand. Livestock Research for Rural Development 22, #33. Yeamkong S, Koonawootrittriron S, Elzo MA and Suwanasopee T 2010. Effect of experience, education, record keeping, labor and decision making on monthly milk yield and revenue of dairy farms supported by a private organization in central Thailand. Asian-Australasian Journal of Animal Sciences 23, 814–824. Yi N and Xu S 2008. Bayesian LASSO for quantitative trait loci mapping. Genetics 179, 1045–1055. Yin T and König S 2018. Genetic parameters for body weight from birth to calving and associations between weights with test-day, health, and female fertility traits. Journal of Dairy Science 101, 2158–2170. Yin T and König S 2019. Genome-wide associations and detection of potential candidate genes for direct genetic and maternal genetic effects influencing dairy cattle body weight at different ages. Genetics Selection Evolution 51, 4. Yousef MK 1985. Stress physiology in livestock. CRC Press, Boca Raton, Florida, USA. Zeggini E and Ioannidis JPA 2009. Meta-analysis in genome-wide association studies. Pharmacogenomics 10, 191–201. Zeileis A, Meyer D and Hornik K 2007. Residual-based shadings for visualizing (conditional) independence. Journal of Computational and Graphical Statistics 16, 507–525. Zeng P, Zhao Y, Qian C, Zhang L, Zhang R, Gou J, Liu J, Liu L and Chen F 2015. Statistical analysis for genome-wide association study. Journal of Biomedical Research 29, 285–97. Zerbino DR, Achuthan P, Akanni W, Amode MR, Barrell D, Bhai J, Billis K, Cummins C, Gall A, Girón CG, Gil L, Gordon L, Haggerty L, Haskell E, Hourlier T, Izuogu OG, Janacek SH, Juettemann T, To JK, Laird MR, Lavidas I, Liu Z, Loveland JE, Maurel T, McLaren W, Moore B, Mudge J, Murphy DN, Newman V, Nuhn M, Ogeh D, Ong CK, Parker A, Patricio M, Riat HS, Schuilenburg H, Sheppard D, Sparrow H, Taylor K, Thormann A, Vullo A, Walts B, Zadissa A, Frankish A, Hunt SE, Kostadima M, Langridge N, Martin FJ, Muffato M, Perry E, Ruffier M, Staines DM, Trevanion SJ, Aken BL, Cunningham F, Yates A and Flicek P 2018. Ensembl 2018. Nucleic Acids Research 46, D754–D761. Zhang X, Chu Q, Guo G, Dong G, Li X, Zhang Q, Zhang S, Zhang Z and Wang Y 2017. Genome-

290

wide association studies identified multiple genetic loci for body size at four growth stages in Chinese Holstein cattle. PLoS ONE 12, e0175971. Zhang W, Gao X, Zhang Y, Zhao Y, Zhang J, Jia Y, Zhu B, Xu L, Zhang L, Gao H, Li J and Chen Y 2018. Genome-wide assessment of genetic diversity and population structure insights into admixture and introgression in Chinese indigenous cattle. BMC Genetics 19, 114. Zhang B, Kirov S and Snoddy J 2005. WebGestalt: An integrated system for exploring gene sets in various biological contexts. Nucleic Acids Research 33, W741–W748. Zhong S, Dekkers JCM, Fernando RL and Jannink J-L 2009. Factors affecting accuracy from genomic selection in populations derived from multiple inbred lines: a barley case study. Genetics 182, 355–364. Zimbleman RB, Rhoads RP, Rhoads ML, Duff GC, Baumgard LH. and Collier RJ 2009. A re- valuation of the impact of temperature humidity index (THI) and black globe humidity index (BGHI) on milk production in high producing dairy cows. In Proceedings of the 24th Annual Southwest Nutrition and Management Conference, pp. 158–168.

291

Appendices

Appendix 1: Human research ethics approval certificate

292

Appendix 2 a: Animal research ethics approval certificate

293

Appendix 2 b: Animal research ethics approval certificate

294

Appendix 3: Nutrient composition of feeds commonly used for dairy cows in regions (on dry matters basis, otherwise stated) A

Price DM NEL TDN CP ADF NDF Fat NFC Starch Lignin Ca K Mg Na P S Cu Fe Mn Zn Feed Region-Feed name % Mcal Ref USD/ton % % % % % % % % % % % % % % ppm ppm ppm ppm code B AF /kg C 551 NL-Napier grass silage 44 31.0 1.20 58.0 6.6 42.5 71.5 2.0 10.3 0.6 5.70 0.36 2.90 0.30 0.03 0.29 0.10 11 413 91 45 2 570 NL_SH-Corn powder 3462 88.0 2.00 88.7 9.3 3.2 9.5 4.3 75.3 70.0 0.09 0.03 0.37 0.14 0.03 0.30 0.12 3 35 6 27 1 581 SH-Rice grain with husk 200 89.0 1.70 81.4 9.0 9.0 11.0 1.8 72.2 66.0 0.90 0.07 0.53 0.14 0.07 0.36 0.05 3 57 20 71 1 582 SL-Rice hay 170 65.0 1.20 58.4 14.8 38.0 61.2 2.1 13.1 9.0 5.90 0.30 3.64 0.15 0.01 0.40 0.20 13 2 25 15 2 600 SL-Fresh rice straw 0 34.0 0.84 37.0 4.3 50.0 78.0 1.4 8.0 6.3 12.00 0.15 0.40 0.83 0.12 0.08 0.08 13 2 33 16 2 605 NH-Fresh corn leaves 58 27.4 1.44 65.0 11.4 31.8 63.6 2.0 21.8 3.5 3.70 0.63 0.17 0.35 0.01 0.15 0.22 11 104 256 69 2 607 SH-Dried distillers grain 330 89.0 2.21 82.0 29.5 13.6 34.2 11.1 11.0 9.3 4.30 0.16 1.03 0.33 0.24 0.79 0.40 6 123 21 62 1 608 SH-Sweet potato tuber 220 30.0 1.86 80.0 5.5 5.2 11.3 1.1 78.4 69.3 1.10 0.12 1.22 0.09 0.02 0.15 0.11 7 700 51 43 1 A Abbreviations: SL, south lowland; SH, south higland; NL, north lowland; NH, north highland; USD, United States Dollar; AF, as fed; DM, dry matter; NEL, net energy for lactation; TDN, total digestible nutrients; CP, crude protein; ADF, acid detergent fibre; NDF, neutral detergent fibre; FAT, fat; NFC, nonfibre carbohydrate; Ca, calcium; P, phosphorus; K, potassium; Mg, magnesium; Na, sodium; S, sulphur; Cu, copper; Fe, iron; Mn, manganese; and Zn, zinc. B Feed code, code of feed in feed library of PCDairy-Vietnamese version. C Ref, references: 1) PCDairy-Vietnamese version (Robinson and Ahmadi, 2015); 2, Feed nutritive value books (National Institue of Animal Husbandry, 2000)

295

Appendix 4: Multivariate linear mixed effect models identifying the variables significantly and suggestively associated with infrared temperatures of rear udder (RUdT, °C), ocular area (EyeT, °C), muzzle (MuzT, °C); armpit (ArmT, °C), paralumbar fossa area (ParT, °C), fore udder (FUdT, °C), fore hoof (FHoT, °C), and hind hoof (HHoT, °C)

Variable RUdT EyeT MuzT ArmT ParT FUdT FHoT HHoT Fixed effect Coef (SE)B P Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P Coef (SE) P 23.54 Intercept 28.62 (3.47) <.001 30.34 (2.92) <.001 21.24 (3.9) <.001 20.67 (4.46) <.001 21.21 (4.59) <.001 30.2 (3.21) <.001 14.28 (7.56) 0.071 0.002 (6.77) Heat load index 0.12 (0.04) 0.004 0.09 (0.03) 0.006 0.18 (0.04) <.001 0.19 (0.05) 0.001 0.18 (0.05) 0.001 0.11 (0.03) 0.004 0.23 (0.08) 0.008 0.14 (0.07) 0.070 -10.5E-4 -5.2E-4 -10.4E-4 -10.3E-4 -11.3E-4 -10.1E-4 -26.5E-4 -23.7E-4 Days in milk 0.011 0.086 0.016 0.064 0.012 0.01 0.002 0.002 (4.1E-4) (3.0E-4) (4.3E-4) (5.6E-4) (4.4E-4) (3.9E-4) (8.6E-4) (7.4E-4) Region: South Low Reference --C Reference -- Reference -- Reference -- Reference -- Reference -- Reference -- Reference -- Region: South High 0.67 (0.6) 0.278 0.49 (0.51) 0.346 0.56 (0.68) 0.419 0.71 (0.77) 0.365 -0.04 (0.8) 0.958 0.56 (0.56) 0.327 -0.21 (1.32) 0.876 -0.8 (1.18) 0.503 -0.85 Region: North Low -0.36 (0.32) 0.276 -0.17 (0.27) 0.537 -0.02 (0.36) 0.951 -0.7 (0.41) 0.103 -0.85 (0.42) 0.056 -0.43 (0.3) 0.158 -0.69 (0.7) 0.337 0.187 (0.63) Region: North High 1.41 (0.53) 0.013 0.76 (0.44) 0.099 1.55 (0.59) 0.015 1.86 (0.68) 0.012 1.04 (0.7) 0.148 1.04 (0.49) 0.043 2.64 (1.15) 0.03 1.11 (1.03) 0.293 Lactation number -- -- -0.05 (0.023) 0.032 ------0.07 (0.03) 0.029 ------0.47 Body condition score -0.36 (0.1) <.001 -0.25 (0.07) <.001 -0.24 (0.1) 0.023 -0.34 (0.13) 0.011 -0.19 (0.11) 0.077 -0.43 (0.09) <.001 -- -- 0.011 (0.18) Body weight 1.0E-3 1.3E-3 3.0E-3 ------0.063 0.064 ------0.008 (calculated, kg) (1.0E-3) (0.8E-3) (1.3E-3) Variance Variance Variance Variance Variance Variance Variance Variance Random effect (SD) (SD) (SD) (SD) (SD) (SD) (SD) (SD) Farm 0.30 (0.55) 0.23 (0.48) 0.39 (0.63) 0.49 (0.70) 0.57 (0.75) 0.26 (0.51) 1.43 (1.20) 1.19 (1.09) Residual 0.52 (0.72) 0.29 (0.54) 0.60 (0.77) 0.87 (0.93) 0.60 (0.77) 0.49 (0.70) 2.68 (1.64) 1.71 (1.31) Explanatory power Conditional R2 (%) 51.34 55.92 60.09 53.9 67.68 49.76 51.65 53.47 Marginal R2 (%) 23.04 21.16 33.88 27.85 36.94 23.37 25.85 21.12 A Variable cow age was excluded from all models due to VIF > 5; The independent variables that were excluded from each final model due to no significant effect (P > 0.1) were energy corrected milk yield (kg) and the variables with ‘--’ sign in the P column. B Coef (SE), Coefficient (Standard error). C “--” indicates no estimate.

296

Appendix 5: Significant and suggestive SNPs associated with the studied traits

A B C D E F G No. Traits SNP Name SNP rsID BTA Position MAF -Log10(P) 1 MILK BovineHD1400017689 (9*) 134385209 14 61568053 0.22 4.56 2 MILK BovineHD0100039253 (4*) 110595833 1 136408359 0.25 4.33 3 MILK BovineHD0100012042 (3) 109167899 1 42519517 0.20 4.09 4 MILK BovineHD0700020773 134589945 7 68635812 0.28 4.01 5 MILK ARS-BFGL-NGS-24259 #N/A 29 17195274 0.22 3.93 6 MILK UA-IFASA-6647 (16) 109436130 14 6444552 0.25 3.68 7 MILK ARS-BFGL-NGS-100407 42270651 14 19990424 0.14 3.60 8 MILK ARS-BFGL-NGS-25002 109304813 29 48553592 0.41 3.49 9 MILK BovineHD2300000214 (12) 135061625 23 1417661 0.26 3.43 10 MILK BTA-97930-no-rs 41667267 7 64989418 0.44 3.32 11 ECM BTA-24502-no-rs (13) 41624447 24 15423390 0.09 4.47 12 ECM UA-IFASA-6647 (16) 109436130 14 6444552 0.25 4.26 13 ECM BovineHD1700013679 109989240 17 47547581 0.23 4.24 14 ECM BovineHD1700013751 135007183 17 47718063 0.26 3.79 15 ECM BovineHD1200009951 (8) 109323452 12 33554962 0.52 3.78 16 ECM ARS-BFGL-NGS-1284 43317139 2 133722150 0.25 3.72 17 ECM BovineHD1200023302 109552726 12 77354740 0.22 3.67 18 ECM BovineHD0700013589 133246321 7 45382201 0.41 3.66 19 ECM BovineHD0100039253 (4*) 110595833 1 136408359 0.25 3.64 20 ECM BovineHD1400017689 (9*) 134385209 14 61568053 0.22 3.64 21 ECM Hapmap58990-rs29020862 (15) 29020862 8 10109654 0.48 3.61 22 ECM BovineHD2200013344 (10) 108939153 22 45829824 0.25 3.58 23 ECM BovineHD2200013356 (11) 110407779 22 45865436 0.24 3.53 24 ECM ARS-BFGL-NGS-59301 109918052 6 115277519 0.34 3.50 25 ECM BovineHD0100012042 (3) 109167899 1 42519517 0.20 3.46 26 ECM BovineHD2300000214 (12) 135061625 23 1417661 0.26 3.46 27 ECM BovineHD1200023274 133844007 12 77336398 0.36 3.37 28 ECM ARS-BFGL-NGS-4091 #N/A 4 109634517 0.31 3.35 29 ECM BovineHD0800002426 132818377 8 7770094 0.24 3.32 30 ECM BovineHD2500004714 110713117 25 16567443 0.22 3.31 31 ECMbw Hapmap58990-rs29020862 (15) 29020862 8 10109654 0.48 4.85 32 ECMbw MS-rs383507306 #N/A 5 48333093 0.18 4.61 33 ECMbw BovineHD1200009951 (8) 109323452 12 33554962 0.52 4.18 34 ECMbw BTB-00316291 43523142 7 62888521 0.14 4.12 35 ECMbw BTA-42763-no-rs 41579996 18 20971908 0.40 4.00 36 ECMbw BovineHD1000002221 110958956 10 7143693 0.48 3.92 37 ECMbw BovineHD2400015344 135879931 24 53478938 0.27 3.90 38 ECMbw BovineHD2200013356 (11) 110407779 22 45865436 0.24 3.85 39 ECMbw ARS-BFGL-BAC-29836 109306311 24 54111123 0.17 3.84 40 ECMbw BovineHD0900021283 136213973 9 75365973 0.16 3.75 41 ECMbw BovineHD1700013026 42915870 17 45659934 0.26 3.71 297

42 ECMbw BovineHD2200013344 (10) 108939153 22 45829824 0.25 3.70 43 ECMbw ARS-BFGL-NGS-24479 41916457 19 44483294 0.48 3.56 44 ECMbw BTA-45898-no-rs 41637937 19 50054802 0.20 3.49 45 ECMbw BovineHD0100039253 (4*) 110595833 1 136408359 0.25 3.47 46 ECMbw BovineHD1400017689 (9*) 134385209 14 61568053 0.22 3.38 47 ECMbw Hapmap52717-rs29026394 29026394 5 39918212 0.36 3.38 48 ECMbw BovineHD0100013822 134496895 1 48636979 0.21 3.36 49 ECMbw BTA-24502-no-rs (13) 41624447 24 15423390 0.09 3.36 50 ECMbw BovineHD1000000336 133847041 10 1194756 0.44 3.33 ARS-USMARC-Parent-EF150946- 51 ECMbw #N/A 26 13194710 0.33 3.31 rs29023666_dup 52 mFA Chr14_2022745 #N/A 14 831004 0.19 6.08 53 mFA ARS-BFGL-NGS-57820 (1) 109146371 14 465742 0.19 5.65 54 mFA Chr14_1699016 #N/A 14 513203 0.26 5.54 Hapmap36620- 55 mFA 29024688 14 2142001 0.22 4.44 SCAFFOLD50018_7571 56 mFA Chr14_1653693 #N/A 14 468124 0.26 3.99 ARS-USDA-AGIL-chr21-49531537- 57 mFA #N/A 21 49078965 0.12 3.98 000460 (2) 58 mFA BovineHD0200019767 (5*) 137638571 2 68013296 0.14 3.81 59 mFA BTB-00387060 (14) 43594837 9 28874995 0.10 3.78 60 mFA Hapmap53146-ss46526085 41255232 6 117411122 0.15 3.62 61 mFA BovineHD0800025442 (7) 41595751 8 84373759 0.24 3.56 62 mFA BovineHD1600023682 41831896 16 79081755 0.24 3.56 63 mFA BovineHD0100045976 43282982 1 155493553 0.25 3.54 64 mFA BovineHD2800003715 137313865 28 12839224 0.23 3.54 65 mFA BovineHD0800011006 137445455 8 36730211 0.36 3.53 66 mFA Chr14_1757935 #N/A 14 572120 0.24 3.52 67 mFA BovineHD2100015390 42896144 21 53322768 0.42 3.45 68 mFA BovineHD0800010236 109747713 8 34351802 0.29 3.39 69 mFA BovineHD1300017524 110175702 13 60610043 0.26 3.39 70 mFA BTA-38756-no-rs 41636353 16 1124008 0.12 3.39 71 mFA BovineHD0400018332 109977926 4 66392130 0.35 3.38 72 mFA BovineHD2400002457 133692290 24 8403392 0.29 3.38 73 mFA BovineHD0400021616 43405704 4 77353189 0.26 3.36 74 mFA ARS-BFGL-NGS-26520 109617015 14 1184910 0.17 3.31 75 mPR BovineHD0800002969 109557540 8 9383625 0.11 4.46 76 mPR BovineHD0500035492 110913642 5 86848745 0.46 4.15 77 mPR ARS-BFGL-NGS-112625 109238076 13 73857548 0.41 4.11 78 mPR BovineHD0500000629 42546720 5 2380672 0.24 3.95 79 mPR BovineHD0200019786 (6) 42653046 2 68056765 0.17 3.90 80 mPR BovineHD1000010106 110028343 10 31101987 0.18 3.87 81 mPR BovineHD0100028921 135107459 1 100524143 0.21 3.78 Hapmap36074- 82 mPR 29009628 24 17943061 0.19 3.68 SCAFFOLD137719_496

298

83 mPR Hapmap43869-BTA-49512 41636478 2 119494517 0.25 3.68 84 mPR ARS-BFGL-NGS-31650 110084531 4 119689852 0.21 3.65 85 mPR BovineHD0200019767 (5*) 137638571 2 68013296 0.14 3.65 86 mPR ARS-BFGL-NGS-11271 110513274 26 23039524 0.20 3.54 87 mPR BovineHD0600013337 42544242 6 46960003 0.51 3.53 88 mPR ARS-BFGL-NGS-101647 110560119 26 22679137 0.29 3.52 89 mPR BovineHD1300023914 136322381 13 81687983 0.40 3.47 90 mPR ARS-BFGL-NGS-111840 109799695 8 11631891 0.39 3.45 91 mPR BovineHD0700000001 134214229 7 153780 0.24 3.38 92 mPR BovineHD0100029974 43256476 1 104784023 0.47 3.34 93 mPR BovineHD1100009590 136744671 11 32077023 0.23 3.34 94 mPR BovineHD2300011087 133545932 23 38537811 0.27 3.33 95 mPR BovineHD1900016959 132861792 19 58769970 0.47 3.31 96 mDM BovineHD0200019767 (5*) 137638571 2 68013296 0.14 4.49 97 mDM Hapmap52337-rs29022325 29022325 9 22622058 0.22 4.36 98 mDM Hapmap55185-rs29018283 29018283 16 33179783 0.19 4.18 99 mDM BTB-00387060 (14) 43594837 9 28874995 0.10 3.97 100 mDM BovineHD0400022069 109393982 4 79105163 0.49 3.66 101 mDM BovineHD0200019786 (6) 42653046 2 68056765 0.17 3.60 102 mDM ARS-BFGL-NGS-57820 (1) 109146371 14 465742 0.19 3.49 103 mDM BovineHD0500011510 108961958 5 40025504 0.22 3.47 104 mDM BovineHD0500025555 137673171 5 89551264 0.22 3.44 ARS-USDA-AGIL-chr21-49531537- 105 mDM #N/A 21 49078965 0.12 3.32 000460 (2) 106 mRE BovineHD2500008754 42946243 25 31605747 0.34 4.74 107 mRE ARS-BFGL-NGS-8024 42409466 25 31592006 0.32 4.36 108 mRE BovineHD2100001815 133083191 21 8097675 0.24 4.03 109 mRE BovineHD1600000152 43096017 16 896630 0.41 3.95 110 mRE BovineHD2400017827 109539813 24 59349263 0.44 3.95 111 mRE ARS-BFGL-NGS-112522 110725254 13 78396102 0.25 3.64 112 mRE BovineHD0300011011 136922091 3 35345961 0.27 3.64 113 mRE BovineHD0500033097 109251990 5 113870788 0.24 3.58 114 mRE Hapmap43389-BTA-87911 41662716 6 75838737 0.17 3.58 115 mRE BovineHD0800025442 (7) 41595751 8 84373759 0.24 3.48 116 mRE BovineHD1400023327 133864802 14 80186536 0.50 3.47 117 mRE BovineHD0200039887 134610379 2 135502021 0.37 3.45 118 mRE BovineHD1100020808 109928521 11 72839192 0.51 3.45 119 mRE ARS-BFGL-NGS-107899 110049475 18 1177337 0.18 3.40 120 mRE BovineHD2900005278 42167840 29 17584375 0.47 3.40 121 mRE BovineHD0200039615 133604483 2 134765709 0.52 3.36 122 mRE ARS-BFGL-NGS-113042 109941542 2 48993143 0.22 3.33 123 mRE BovineHD4100009957 41682035 13 15840042 0.23 3.31 124 mRE BovineHD0200014156 132906739 2 48858590 0.28 3.30

299

125 BCS BovineHD2700004244 136885568 27 15547774 0.42 4.81 126 BCS BovineHD0100018075 133722624 1 63334223 0.10 4.3 127 BCS BovineHD2700004476 109948697 27 16325987 0.48 3.99 128 BCS BovineHD0300008932 110921406 3 27833026 0.26 3.68 129 BCS BovineHD1300022504 41711120 13 77003923 0.24 3.68 130 BCS BovineHD1700013848 135033310 17 47924478 0.22 3.55 131 BCS BovineHD1900017112 109961681 19 59345427 0.46 3.53 132 BCS BovineHD0300035459 43102842 3 120937313 0.09 3.50 133 BCS BovineHD0400017221 110062264 4 62461121 0.49 3.44 134 BCS BovineHD0100043669 136404382 1 148837863 0.38 3.43 135 BCS BovineHD1800000373 134219189 18 1615849 0.22 3.38 136 BCS BovineHD0100028122 109247309 1 158134933 0.24 3.36 137 BCS BovineHD0300015738 109317776 3 51915567 0.15 3.33 138 BCS BovineHD2700004480 42114792 27 16330605 0.29 3.31 139 HG BovineHD0500026126 (5) 137548207 5 91587346 0.25 3.85 140 HG chr5_49669089 (8) #N/A 5 49439093 0.15 3.84 141 HG Hapmap44646-BTA-38361 (9) 41635148 16 29015336 0.17 3.84 142 HG BovineHD0200020103 (2) 134576806 2 69396120 0.26 3.71 143 HG BovineHD0200020109 (3) 43332709 2 69421911 0.26 3.71 144 HG BTA-63705-no-rs (7) 41651430 28 20214922 0.26 3.69 145 HG BovineHD1600008781 (6) 42397333 16 30155855 0.26 3.60 146 HG BovineHD1600002332 135414806 16 7603624 0.13 3.56 147 HG ARS-BFGL-NGS-66662 (1) #N/A 1 148909835 0.26 3.41 148 HG BovineHD0500004265 (4) 110442615 5 14322876 0.48 3.35 149 HG BovineHD1200008104 132964195 12 27132591 0.20 3.31 150 HG BTB-00670329 41838044 17 6733751 0.37 3.31 151 BW BTA-63705-no-rs (7) 41651430 28 20214922 0.26 4.08 152 BW chr5_49669089 (8) #N/A 5 49439093 0.15 3.84 153 BW BovineHD0200020103 (2) 134576806 2 69396120 0.26 3.83 154 BW BovineHD0200020109 (3) 43332709 2 69421911 0.26 3.83 155 BW Hapmap44646-BTA-38361 (9) 41635148 16 29015336 0.17 3.82 156 BW BovineHD0500026126 (5) 137548207 5 91587346 0.25 3.63 157 BW BovineHD1600008781 (6) 42397333 16 30155855 0.26 3.63 158 BW ARS-BFGL-NGS-66662 (1) #N/A 1 148909835 0.26 3.34 159 BW BovineHD0500004265 (4) 110442615 5 14322876 0.48 3.32

160 PS ARS-BFGL-NGS-112026 110156130 11 40079795 0.22 4.95 161 PS BovineHD2000004465 41940721 20 14337270 0.16 4.56 162 PS chr18_43763877 #N/A 18 43594730 0.27 4.27 163 PS BovineHD0100030007 135569709 1 104905020 0.18 4.25 164 PS BovineHD0900029637 133397027 9 100251768 0.48 4.07 165 PS BovineHD2000004488 41934462 20 14454343 0.16 3.92 166 PS BovineHD1400018134 41736693 14 62804326 0.19 3.89

300

167 PS BovineHD0800004576 137056575 8 14654799 0.44 3.88 168 PS BovineHD0100000832 110738206 1 3260370 0.18 3.79 169 PS BovineHD0800029257 137257870 8 97341930 0.39 3.79 170 PS BovineHD0200029230 135475145 2 101324699 0.16 3.68 171 PS BovineHD0900023133 137756982 9 81974419 0.23 3.67 172 PS BovineHD0300025963 110032171 3 89690667 0.51 3.62 173 PS BovineHD1100011765 133951700 11 40050575 0.22 3.61 174 PS BTB-01619293 42730076 8 33688162 0.20 3.59 175 PS BovineHD0300012380 136806941 3 40495260 0.27 3.54 Hapmap36162- 176 PS 29021987 18 52610995 0.08 3.52 SCAFFOLD261827_5746 177 PS BovineHD0300012098 134124059 3 39419610 0.44 3.50 178 PS BovineHD0400011913 133235075 4 43202870 0.23 3.50 179 PS BovineHD0800005917 133350398 8 19001439 0.23 3.46 180 PS BovineHD1900003389 110333045 19 12425980 0.16 3.44 181 PS BovineHD4100011730 41632440 14 75954451 0.22 3.43 182 PS BTA-20306-no-rs 41579284 2 116659433 0.25 3.42 183 PS ARS-BFGL-NGS-18019 110379081 18 48272264 0.24 3.37 184 PS ARS-BFGL-NGS-18303 #N/A 4 76900239 0.08 3.33 185 PS BovineHD0600015326 134148513 6 54354205 0.17 3.33 186 PS BovineHD0600014164 42887379 6 49799216 0.24 3.31 187 IVuT BovineHD0800014723 137520698 8 48949453 0.14 4.75 188 IVuT BovineHD0800014726 135840375 8 48956781 0.16 4.25 189 IVuT BovineHD1900003810 135073357 19 14049254 0.44 4.03 190 IVuT BovineHD1900011049 134584197 19 37404259 0.26 4.01 191 IVuT ARS-BFGL-NGS-16008 109552006 26 13419517 0.24 3.95 192 IVuT ARS-BFGL-NGS-3757 109725437 3 100156027 0.25 3.88 193 IVuT BovineHD1900004073 109294634 19 14908169 0.25 3.83 194 IVuT BovineHD1900011435 43719933 19 39222569 0.42 3.83 195 IVuT BovineHD2100015365 110705710 21 53202184 0.31 3.72 196 IVuT BovineHD0500013129 133720216 5 45376389 0.23 3.71 197 IVuT BovineHD0500025003 133949844 5 87726564 0.27 3.71 198 IVuT Hapmap25018-BTA-125160 81115197 10 14581464 0.20 3.70 199 IVuT BovineHD1800007011 136388971 18 22790033 0.39 3.69 200 IVuT BovineHD0500013119 132757610 5 45336636 0.35 3.61 201 IVuT ARS-BFGL-NGS-33535 109294126 8 8377201 0.38 3.56 202 IVuT ARS-BFGL-NGS-32769 110865656 19 21777851 0.25 3.53 203 IVuT ARS-BFGL-NGS-114938 42068792 25 32600127 0.37 3.52 204 IVuT BovineHD0500013127 134789295 5 45373112 0.23 3.52 205 IVuT BovineHD2200004336 137198203 22 15976044 0.24 3.48 206 IVuT ARS-BFGL-NGS-101117 110198589 8 48938261 0.24 3.44 207 IVuT BovineHD1900009350 133832527 19 31166010 0.40 3.32 208 OVuT BovineHD0500026395 137212660 5 92498091 0.48 4.31 209 OVuT BovineHD0100024795 132673311 1 86605783 0.15 4.23

301

210 OVuT BovineHD1400013068 110344589 14 44084230 0.26 3.95 211 OVuT BovineHD1400013035 109754076 14 43999824 0.42 3.91 212 OVuT ARS-BFGL-NGS-91939 (2) 109900017 14 44079419 0.25 3.64 213 OVuT BovineHD2400003128 109166035 24 10741327 0.47 3.61 214 OVuT chr19_31794868 #N/A 19 31155156 0.20 3.51 215 OVuT BovineHD1500018880 134555676 15 65069563 0.24 3.46 216 OVuT BovineHD0900014381 110051018 9 51449826 0.27 3.44 217 OVuT ARS-BFGL-NGS-57209 (1) 110236385 19 31151670 0.29 3.40 218 OVuT BTB-01741857 42854990 28 2669298 0.15 3.40 219 OVuT BovineHD1900009346 (4) 136147412 19 31156520 0.30 3.39 220 OVuT BovineHD1900009335 (3) 136600914 19 31129839 0.32 3.37 221 OVuT Hapmap47921-BTA-34862 41630560 14 45867755 0.40 3.31 222 ITBT BovineHD0800001033 137124509 8 3311336 0.26 4.30 223 ITBT BTA-93171-no-rs 41664246 2 13083457 0.40 4.28 224 ITBT BovineHD1900017537 (5) 109293743 19 60517058 0.25 4.12 225 ITBT Hapmap40672-BTA-121035 41568084 15 72561262 0.13 4.01 226 ITBT BovineHD1000017897 136127633 10 61751597 0.18 3.97 227 ITBT BTA-74836-no-rs 41654499 5 104430699 0.35 3.68 228 ITBT BovineHD1000010019 110283228 10 30588472 0.20 3.65 229 ITBT BovineHD0200019896 135090181 2 68542547 0.24 3.61 230 ITBT BovineHD1800011923 109097687 18 40378221 0.14 3.56 231 ITBT BovineHD0600007256 137437801 6 24953980 0.29 3.53 232 ITBT ARS-BFGL-BAC-29180 109960821 2 12027596 0.25 3.42 233 ITBT BovineHD1400023151 134465744 14 79607778 0.31 3.42 234 ITBT BovineHD2200007708 (6) 109322273 22 26571218 0.24 3.42 235 ITBT BovineHD1900009523 109126530 19 31695319 0.25 3.41 236 ITBT BovineHD1400022953 135097756 14 78805344 0.11 3.40 237 ITBT Hapmap31228-BTA-74842 81110166 5 104410553 0.43 3.40 238 ITBT AX-22704931 #N/A 20 39815312 0.13 3.39 239 ITBT BovineHD0800003414 109889211 8 10693162 0.47 3.38 240 EyeT BovineHD0200037109 110136625 2 127206382 0.13 4.28 241 EyeT BovineHD2200007708 (6) 109322273 22 26571218 0.24 4.16 242 EyeT BovineHD0500002307 110157123 5 8087973 0.50 4.14 243 EyeT BovineHD0200037552 109197644 2 128627263 0.24 4.12 244 EyeT BTA-19847-no-rs 41628433 1 145340231 0.14 4.10 245 EyeT BovineHD0100043438 109106543 1 148172257 0.25 4.08 246 EyeT BovineHD1000001418 133131250 10 4706527 0.48 3.93 247 EyeT BovineHD1200006202 110397243 12 20594293 0.45 3.64 248 EyeT BovineHD0100043469 135648542 1 148282126 0.23 3.60 249 EyeT BovineHD1200006096 109843112 12 20235614 0.25 3.60 250 EyeT BovineHD1900004722 133653570 19 16591256 0.17 3.60 251 EyeT BovineHD0500002260 109300557 5 7958891 0.24 3.52 252 EyeT BovineHD1300019320 109342821 13 67324498 0.16 3.52

302

253 EyeT BovineHD0500005487 109860253 5 18927598 0.22 3.48 254 EyeT BovineHD1900017537 (5) 109293743 19 60517058 0.25 3.39 255 EyeT BovineHD0100028531 43256298 1 99215719 0.47 3.38 256 EyeT ARS-BFGL-NGS-91225 110274027 7 6568731 0.22 3.36 Hapmap34737- 257 EyeT 43706923 3 24575444 0.23 3.34 BES11_Contig221_1196 258 MuzT Hapmap29069-BTA-45382 41602218 19 39296143 0.37 4.60 259 MuzT ARS-BFGL-NGS-17312 110855079 19 43593946 0.25 4.56 260 MuzT BTB-00659960 41822522 16 70836094 0.24 4.29 261 MuzT BovineHD2800001260 109309454 28 3182010 0.28 4.02 262 MuzT BovineHD1600020671 137578402 16 70869841 0.48 3.96 263 MuzT BovineHD1300001379 136390366 13 5360261 0.13 3.92 264 MuzT BovineHD0900021133 109835240 9 74925057 0.20 3.83 265 MuzT BovineHD1800015898 132734213 18 53771600 0.25 3.79 266 MuzT chr16_72753272 #N/A 16 70858661 0.47 3.79 267 MuzT Hapmap51348-BTA-90742 (8) 41658343 9 74797172 0.22 3.77 268 MuzT BovineHD0400028880 109333592 4 101713234 0.51 3.75 269 MuzT BovineHD0900022063 133139030 9 78374229 0.28 3.73 270 MuzT ARS-BFGL-NGS-42019 42980676 28 45311398 0.25 3.67 271 MuzT ARS-BFGL-NGS-109968 110394333 16 70862193 0.47 3.65 272 MuzT ARS-BFGL-NGS-91410 #N/A 1 61817159 0.18 3.63 273 MuzT Hapmap59218-rs29014793 29014793 2 27402031 0.25 3.61 274 MuzT ARS-BFGL-NGS-114136 #N/A 19 35834667 0.17 3.59 275 MuzT BovineHD0400000470 137773212 4 2112821 0.21 3.57 276 MuzT BovineHD2500003591 110129589 25 12772459 0.21 3.57 277 MuzT BovineHD1600020682 41823566 16 70920043 0.48 3.52 278 MuzT ARS-BFGL-NGS-60450 110222344 28 43155287 0.15 3.51 279 MuzT ARS-BFGL-NGS-114383 110983360 19 47080183 0.39 3.50 280 MuzT BovineHD1100006106 109859972 11 20062091 0.22 3.49 281 MuzT BovineHD4100014243 135011262 19 46866311 0.17 3.47 282 MuzT BovineHD0500024232 137546422 5 85146444 0.36 3.46 283 MuzT BovineHD2800013247 42980656 28 45309977 0.43 3.40 284 MuzT BovineHD1100012988 110993102 11 44885968 0.29 3.38 285 MuzT BovineHD0100017693 133597687 1 61817959 0.13 3.32 286 ArmT ARS-BFGL-BAC-5688 109044741 1 147809393 0.34 4.03 287 ArmT ARS-BFGL-NGS-101258 110412185 5 64842324 0.21 3.94 288 ArmT BovineHD0400002990 110289147 4 9945724 0.22 3.86 289 ArmT BovineHD0200037810 134498813 2 129514375 0.23 3.84 290 ArmT BovineHD2500001294 110631609 25 5417066 0.10 3.75 291 ArmT ARS-BFGL-NGS-2341 #N/A 2 129347211 0.16 3.70 292 ArmT Hapmap39569-BTA-49765 41637024 2 129451897 0.23 3.68 293 ArmT ARS-BFGL-BAC-12583 43101493 13 57500003 0.34 3.54 294 ArmT BovineHD4100011169 110657645 14 11157034 0.32 3.50 295 ArmT BovineHD0300011908 137368765 3 38574729 0.17 3.43

303

296 ArmT BovineHD0300013050 109959926 3 42769882 0.25 3.43 297 ArmT ARS-BFGL-NGS-62627 109808146 17 70434930 0.43 3.37 298 ArmT chr17_72557299 #N/A 17 70437800 0.43 3.37 299 ArmT BovineHD1400014784 134835216 14 49974558 0.42 3.33 300 ArmT ARS-BFGL-NGS-12655 110941353 19 39273060 0.27 3.32 301 ArmT BovineHD0600019862 136613146 6 69694946 0.23 3.30 302 ParT BovineHD1400023207 133244421 14 79782432 0.46 4.33 303 ParT Hapmap49707-BTA-68355 43709344 3 75226273 0.44 4.02 304 ParT BovineHD0300022955 41257694 3 79300975 0.41 4.01 305 ParT BovineHD1000012551 110444662 10 40901338 0.42 4.01 306 ParT BovineHD1700021148 110118588 17 70435488 0.24 4.00 307 ParT Hapmap51348-BTA-90742 (8) 41658343 9 74797172 0.22 3.78 308 ParT ARS-BFGL-NGS-16109 41724795 14 6962216 0.49 3.77 309 ParT chr7_93243389 (7) #N/A 7 90844398 0.16 3.73 310 ParT INRA-638 110438768 3 61800267 0.25 3.73 311 ParT BTB-00902072 42066935 1 33805826 0.08 3.71 312 ParT BovineHD0300030975 110752065 3 107241504 0.24 3.64 313 ParT BovineHD2700001971 136547169 27 7372233 0.44 3.62 ARS-USDA-AGIL-chr18-58106372- 314 ParT #N/A 18 57780512 0.33 3.52 000357 315 ParT BovineHD2800002369 134928620 28 7904578 0.18 3.50 316 ParT BovineHD1600018169 136805873 16 63034806 0.32 3.45 317 ParT ARS-BFGL-NGS-3402 110888654 27 8397592 0.39 3.40 318 ParT BovineHD2700002312 135021425 27 8419936 0.39 3.40 319 ParT BovineHD0300018432 136942142 3 61226587 0.17 3.33 320 FUdT Chr12_15613746 #N/A 12 15564761 0.22 3.90 321 FUdT BTB-00357945 43565716 8 72515403 0.24 3.61 322 FUdT BovineHD1500022429 110279884 15 76060265 0.40 3.57 323 FUdT BovineHD0200021468 136378419 2 74526076 0.48 3.40 324 FUdT BovineHD2100018925 109788902 21 62702505 0.48 3.39 325 FUdT chr7_93243389 (7) #N/A 7 90844398 0.16 3.34 326 FUdT Hapmap42592-BTA-57478 41644360 24 16874416 0.37 3.32 327 RUdT Hapmap40731-BTA-41334 41570570 24 24396084 0.21 4.25 328 RUdT BovineHD2300014503 134574677 23 49838000 0.21 4.11 329 RUdT BovineHD0800030080 43577102 8 99700118 0.27 3.98 330 RUdT BovineHD2100010047 137170529 21 34638191 0.17 3.84 331 RUdT BTA-110429-no-rs 41574556 11 87691339 0.24 3.63 332 RUdT Hapmap52867-rs29023496 29023496 21 26581000 0.24 3.62 333 RUdT BTA-81053-no-rs 41654637 8 39659087 0.23 3.56 334 RUdT ARS-BFGL-NGS-91939 (2) 109900017 14 44079419 0.25 3.51 335 RUdT BTB-00775794 41937398 20 19061537 0.18 3.44 336 RUdT BTB-00750203 41911936 19 37634187 0.38 3.42 337 RUdT BovineHD1500013875 42492500 15 47748578 0.17 3.31 338 FHoT BovineHD2000016438 133688493 20 58788183 0.24 4.37

304

339 FHoT ARS-BFGL-NGS-16266 109472737 20 58795440 0.44 4.09 340 FHoT BovineHD0500013030 135678859 5 45080266 0.23 4.00 341 FHoT ARS-BFGL-NGS-109563 109585633 13 57184961 0.12 3.98 342 FHoT BTA-106792-no-rs 41567990 5 44905098 0.22 3.62 343 FHoT BovineHD1900009346 (4) 136147412 19 31156520 0.30 3.59 344 FHoT BovineHD0600003010 42664656 6 11038502 0.20 3.52 345 FHoT BovineHD1100012056 110207403 11 41309494 0.48 3.51 346 FHoT BovineHD2000003394 133898629 20 10747267 0.19 3.44 347 FHoT BovineHD1400023832 132681180 14 81726995 0.43 3.42 348 FHoT BovineHD2700003488 110803530 27 12983873 0.24 3.42 349 FHoT ARS-BFGL-NGS-57209 (1) 110236385 19 31151670 0.29 3.41 350 FHoT chr19_31794859 #N/A 19 31155147 0.20 3.38 351 FHoT BovineHD0600003024 42796952 6 11078997 0.36 3.33 352 FHoT BovineHD0700004607 110852922 7 15312197 0.24 3.33 353 FHoT chr5_49669089 #N/A 5 49439093 0.15 3.33 354 FHoT BovineHD1900009335 (3) 136600914 19 31129839 0.32 3.32 355 HHoT ARS-BFGL-NGS-91573 110460688 4 64951300 0.16 4.10 356 HHoT ARS-BFGL-NGS-29072 110043561 9 50477436 0.23 3.85 357 HHoT ARS-BFGL-NGS-114201 41904278 19 26942729 0.21 3.78 358 HHoT BovineHD0200030665 43717935 2 105777668 0.41 3.77 359 HHoT BovineHD1600014652 133362619 16 51796918 0.17 3.49 360 HHoT BovineHD1900017487 137055422 19 60383571 0.47 3.42 361 HHoT BovineHD1700018658 41853198 17 62517763 0.34 3.37 362 HHoT ARS-BFGL-NGS-42664 110836011 18 47905510 0.21 3.35 363 HHoT ARS-BFGL-NGS-83284 109856572 23 28851988 0.38 3.33 A Abbreviations of traits: MILK, milk yield; ECM, energy corrected milk; ECMbw, energy corrected milk adjusted for body weight (BW); mFA, fat concentration; mPR, protein concentration; mDM, milk dry matter concentration; mRE, milk electrical resistance; yFA, fat yield; yPR, protein yield; yDM, milk dry matter yield; BCS, body condition score; HG, heart girth; BW: body weight; PS, panting score; IVuT, inner vulval lip temperature; OVuT, outer vulval temperature; RUdT, rear udder temperature; ITBT, inner tail base surface temperature; EyeT, ocular area temperature; MuzT, muzzle temperature; ArmT, armpit temperature; ParT, paralumbar fossa area temperature; FUdT, fore udder temperature, FHoT, fore hoof temperature; HHoT, hind hoof temperature. B A SNP associated with more than one traits was marked with a unique number in side ( ). C rsID, reference SNP identification number. “#N/A” indicates no identification was found. D BTA, Bos taurus autosome. E Position of SNP cluster or single SNP on the bovine ARS-UCD1.2 assembly in mega base pairs (Mb). F MAF, minor allele frequency. G P-value, SNPs with P-value  5.0E-04 were reported, significant SNPs (P  5.0E-05) in bold and red.

305

306