Do Seasonal Climate Forecasts and Crop Insurance Matter for Smallholder Farmers in Zimbabwe? Using Contingent Valuation Method and Remote Sensing Applications

DO SEASONAL CLIMATE FORECASTS AND CROP INSURANCE MATTER FOR SMALLHOLDER FARMERS IN ZIMBABWE? USING CONTINGENT VALUATION METHOD AND REMOTE SENSING APPLICATIONS

DISSERTATION

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

By Ephias M. Makaudze, B.S., M.S., M.A.

******

The Ohio State University 2005

Dissertation Committee: Professor Brent Sohngen, Advisor Approved by Professor Timothy Haab Professor Mario Miranda ______Professor Carolyn Merry Advisor Agricultural, Environmental and Development Economics Graduate Program Abstract

As smallholder farmers in Zimbabwe face inevitable drought, the need to develop drought- mitigation strategies and risk transfer mechanisms becomes an important and challenging task for policy makers. Rather than treating drought as a natural disaster that warrants emergency declarations whenever it strikes, countries in Southern Africa could alter their policy to embrace drought as an integral part of their national policy framework. Although drought cannot be eliminated, its impact can be reduced through implementation of pro-active and pro-poor risk management policy programs. This study explored two potential policy programs. One program proposes wide-scale adoption of improved seasonal forecasts by smallholder farmers as a drought mitigation strategy, and the second program proposes implementation of area-yield drought- indexed insurance as a risk-transfer and risk-protection mechanism for the smallholder farmers.

To investigate whether adoption of seasonal forecasts and drought insurance is possible in Zimbabwe this dissertation explored three hypotheses: First, do seasonal forecasts really matter to smallholder farmers in Zimbabwe? Second, given the prevalence of food-aid in Zimbabwe, does drought insurance really matter for smallholder farmers? Third, given drought is a catastrophic risk, will a drought-index insurance scheme intended for smallholder farmers be viable and/or feasible?

The first two questions were empirically investigated via surveys based on the contingency valuation method (CVM). More than 1,000 smallholder farmers were surveyed throughout Zimbabwe’s agro-ecological regions II-V where willingness-to-pay (WTP) for the proposed programs was elicited. With respect to the first hypothesis, results showed that for the improved seasonal forecasts program, estimated WTP (Z$) based on a single-bound model ranged from Z$2,427 to Z$4,676. For a double-bound model, WTP ranged from Z$2,532 to Z$4,225. A distinct differential pattern in WTP was observed across districts and natural regions, where

ii households in wet districts revealed WTP that was consistently lower than those in drier districts. In fact WTP for households in natural region II was 36% and 30% lower than in regions IV and V, respectively. A similar pattern was observed for households in natural region III whose WTP was 17% and 9.3% lower than in regions IV and V, respectively. Because the perceived drought risk is more ominous in drier regions (IV and V) than in wet regions (II and III), households in the former are willing to sacrifice more for the provision of improved seasonal forecasts.

With respect to the second hypothesis, results showed that in the presence of food-aid, WTP or rather potential demand for drought insurance decreases by more than 35% for households in regions IV and V, while for regions II and III it decreases by 10.6%. The results imply that disincentive to purchase insurance in the presence of food-aid is greatest in drier regions IV and V and least in wet regions II and III. Across all regions/districts the demand for insurance is likely to decrease by more than 20% in the presence of food-aid. Thus, food-aid will discourage farmers from seeking more efficient drought risk protection mechanisms such as formal drought insurance.

With respect to the third hypothesis, results indicate that VCI showed appreciably high correlation with crop yields sufficient to consistently track yield losses and these results were fairly comparable with rainfall index. In addition, reasonable premium rates were recovered that are actuarially sound and inexpensive enough to attract participation of the rural poor. In as far as hedging against extreme drought events is concerned, a VCI-based contract could be sufficient. Basis risk becomes an issue, if the index is used to protect drought events of moderate intensity.

iii Dedicated to my parents and my late sister EnRose ”….for God has other plans……”

iv ACKNOWLEDGEMENTS

I wish to thank my major advisor Professor Brent Sohngen, who was my pillar of support throughout my graduate school at OSU. I would also like to thank Professor Tim Haab who cultivated my interest in discrete modeling and contingent valuation method. Professor Mario Miranda provided invaluable support and guidance especially in agricultural insurance and risk modeling. Further, I want to thank Professor Carolyn Merry for motivating my interests in remote sensing and GIS techniques.

I want to extend special thanks to Drs. Akin Adesina, Peter Hazell and Reneth Mano whose comments on earlier drafts gave more depth to this research study. I also owe gratitude to the following who helped me in various capacities: Dr Lovemore Rugube (Chairman, Department of Agricultural Economics,UZ) who provided office space and logistic support; Dr Kennedy Masamvu and Mr Blessing Siwela (SADC Remote Sensing Unit) who assisted in satellite data collection; Dr Leonard Unganai (Department of Meteorology) who provided meteorological data and Ms Charity Mutonhodza (Agricultural Research and Extension Services) who provided crop yield data. Special thanks go to my team of enumerators. I also would like to thank Dr Patrick Jeffers for proof reading some earlier chapters.

I would like to thank my wife Tsitsi for standing with me through thick and thin. To my three kids, Minanayashe, Kuzivakwashe and Kunashe I say your hugs were sources of encouragement. I also want to thank family members and relatives in Zimbabwe for their prayers and love and special thanks go to Rev Elias Makaudze and wife and Mr B.and W. Mrs Chinemhute.

Finally I want to thank the Rockefeller Foundation for financial support throughout my graduate study. Once again, God permitting, I dedicate to use my acquired expertise for the benefit of the rural poor in Zimbabwe and Africa.

v VITA

November, 11, 1964………………Born – Buhera, Zimbabwe 1987…………………………………………B.S. Agricultural Economics (Hons), University of Zimbabwe 1991-1993…..……………………………M.S. Agricultural Economics, Texas A&M University 2001…………………………………………M.A. Economics, The Ohio State University 1999—present………………………Rockefeller Fellow

PUBLICATIONS Research Publication J. Phillips, E. Makaudze and L. Unganai. 2001 ‘Current and Potential Use of Climate Forecasts for resource-poor Farmers in Zimbabwe’, American Society of Agronomy, Impact of El-Nino and Climate Variability on Agriculture, ASA special publication. No.3, Vol. 63

E. Makaudze, D. Bessler and S. Fuller. 1998 ‘A Time series analysis on Zimbabwe’s Maize supply to the Grain Marketing Board’, Development Southern Africa, Vol. 15, No. 3, (Spring)

E. Makaudze. 1997 ‘The challenge facing users of Methyl Bromide World-wide is to develop superior substitutes: What are alternative substitutes?’ Journal of Zimbabwe Tobacco, Vol. 6, No. 4 (April)

FIELDS OF STUDY

Major Field: Agricultural, Environmental, and Development Economics Minor Fields: Agricultural Insurance and Drought Risk Management

vi TABLE OF CONTENTS Page Abstract………………………………………………………………………………………………………………………………………………………………………………..ii Dedication………………………………………………………………………….…………………………………………………………………………….iv Acknowledgements…………………………………………………………….…………………………………………………………………………..v VITA………………………………………………………………………………………………………………………………………………………………….vi List of Tables…………………………………………………………………………..……………………………………………………………………….x List of Figures…………………………………………………………………………………………………………………………………………………xii

CHAPTERS: 1. Nature of Problem…………………………………………….……………………………………………………………………………….1 1.1 Background………………………………………………………………………………………………….……………………….1 1.2 Causes of Drought and its Impact on Zimbabwe’s Economy …………….....…..……………3 1.3 Early Warning Institutions and Drought Monitoring………………..……………………………..5 1.4 Review of Agricultural Policies in Zimbabwe…………………..…………………………………………7 1.4.1 Pre-Reform Agricultural Policies……………………………………………………………………………………7 1.4.2 Post-Reform Agricultural Policies…………………………………………………………………………………11 1.5 Lessons Learned………………………………………..……………………………………………………….……………..12 1.6 Defining the Problem Statement……………………………………………………………………………………14 1.7 Study Objectives………………………………………………………..……………………………………………………..16 2. Literature Review……………………………………………………………….…………………………………………………………..18 2.1 Remote Sensing and Drought Monitoring…………………………………………………………………..18 2.2 Agronomic Condition Assessment……………………………………………………………………………….21 2.3 Smallholder Drought Mitigation Strategies…………………………….…………………………………25 2.3.1 Inadequacy of Traditional Drought Mitigation Strategies……………………….. …………..26 2.4 Experience with Multiple Peril Crop Insurance……………………………………………………….27 2.5 Index-based Area-yield Drought Insurance……………………………………………………………….30

vii 3. Research Methods………………………………………………………………………………………………………….……………….32 3.1 Contingent Valuation Method………………………………………………………………………………………32 3.2 Adoption of Improved Seasonal Forecasts…………………………………………………..…………….35 3.3 Characterizing Demand for Drought Insurance………………………………………………………..36 3.3.1 Impact of Mitigation cost, r on Demand for Insurance…………………….….………………….36 3.3.2 Impact of Food-aid on Demand for Insurance……………………………………………………………39 3.4 Insurance Supply Function in the presence of Catastrophic risk, Food-aid and reinsurance …………………………………..…………………………………………………………40 3.4.1 Deriving Demand Function using MV utility Approach …..……………..……………………41 3.4.2 Deriving Short-run Supply function using MV utility Approach..……………………….43 3.4.3 Characterizing the Existence of Long-run Equilibrium .………………………………………..46 3.5 Designing an Index-based Indemnity Function…………………………………..….………………..48 3.5.1 What variable to use as Index?…...... ………………………50 3.5.2 How correlated is the Index with Crop Yield losses?...... ………………51 3.5.3 Designing and Pricing an Index-based Contract………..……..……………………………………..53 3.6 Data Requirements …………………………………………………………………………………………………………55 4. Exploratory Data Analysis…………………………………………………………………………………………….………………56 4.1 Data Collection Methods……………….………………………………………………………………………………56 4.1.1 Sampling Method….…………………………………………………………………………………………………………56 4.1.2 Focus Group Meetings and Pre-testing the Questionnaire…………………………………….57 4.1.3 Eliciting Farmers’ WTP for Seasonal forecasts…..……………………………………………………..57 4.1.4 Eliciting Farmers’ WTP for Drought Insurance………………………………………………………..58 4.2 Descriptive Analysis…………………………………………………………………………………………………….….58 4.2.1 Demographic Characteristics of sampled households……………………………………………..58 4.2.2 Basic Assets Ownership…………………………………………………………………………………………………59 4.2.3 Maize and Cotton Yield………………………………………………………………………………………………….60 4.2.4 Sources of Farm Income ………………………………………………………………………………………………...61 4.2.5 Local Maize, Private trader and GMB price ……………………………………………………………...63 4.2.6 Food-aid distribution during 2003/04 season…………………………………………………………...63 4.3 Seasonal Forecasts……………………………………………………………………………………………………………64 4.3.1 Confidence with the Forecasts………………………………………………………………………………….….65 4.3.2 Decision-making factors.…………………………………………………………………………………………….…66 4.4 Impact of Drought on Smallholder Farmers..………………………………………………………….…67 4.4.1 Drought coping strategies..…………………………………………………………………………………………...68

viii 4.4.2 Drought Vulnerability Factors….……………………………………………………………………………..…..69 4.5 Measuring Correlation between Crop Yield vs. VCI and Rainfall…….………………….70 4.5.1 Crop Yield vs. Drought Indices……………………………………………………………………………………..71 4.6 Crop Yield Loss Distribution…………………………………………………………………………………………78 4.6.1 Nonparametric density estimation…....…………………………………………………………………….….79 5. Discussion of Main Results………………………………………………………………………………………….…………….…89 5.1 Smallholder Farmers’ Preference for Seasonal Forecasts……………….……….…………….…89 5.1.1 Estimating a Random Utility Model with Log-Linear Income………………………….…..89 5.1.2 Estimating WTP for Seasonal Forecasts…………………………………………………….……………...93 5.2 Assessing Smallholders’ Preference for Drought Insurance……….…………….…………….94 5.2.1 Estimated Parameter Coefficients using Discrete Choice Models.…………..………..…95 5.3 Impact of Seasonal Forecasts on Household WP for Drought Insurance……………98 5.4 WTP for Drought Insurance……………………………………………………………………………….…………99 5.5 Rating Index Insurance across Districts………………….……………………………………….……….101 5.5.1 Smallholder Cotton and Maize Crop Yield Trend Analysis…………………….…….……..102 5.5.2 Yield Estimation………………………………………….…………………………………………………………….……108 5.5.3 Rating Index-based Insurance………………………………………………………………………………….…109 6. Summary and Conclusion…………………………………………………………………………………………………………...126 Appendices……………………………………………………………………………………………………………………………………..132 A Drought-related Impacts .…………………………………………………………………………………………….132 B Mathematical Derivations…………………………………………………………………………………………….133 C Existence of Long-run equilibrium………………………………………………………………………….…134 D Questionnaire………………………………………………………………………………………………………………....135 E Map of Agro-ecological Zones in Zimbabwe..………………………………………………………...148 Bibliography…………….…………………………………………………………………………………………………………………….149

ix LIST OF TABLES Table Page 1.1 Land Classification in Zimbabwe………………….……………………………..…………..………….………………………2 1.2 Area under Irrigation: Large-scale vs. Smallholder Sector……..………..……………………….……………..2 1.3 Nominal Producer Prices in Z$ for Controlled Commodities……...………………………………….……..8 1.4 Maize and Cotton Production by Smallholder vs. Large-Scale Sector (1980-2000)……………………………………………………………………..…………………………….…………………….10 2.1 Characteristics of Selected Multispectral Remote Sensing System…………………..…..…………...20 2.2 Bands and Indices Commonly used for Vegetation Monitoring……………………..………..………….23 4.1 Demographic Statistics of the Sampled Households………………………………………………………………59 4.2 Household Basic asset ownership for Season 2003/04….……………………………………………………....60 4.3 Average Cotton and Maize Yield during best and worst Season…………………………………………....61 4.4 Crops and Livestock as Source of Income……………………………………………………………………………..….62 4.5 Average Farm Income during 2002/03 season………………………………………………………………………….62 4.6 Comparison of Local Maize, Private trader vs. GMB Prices……………………………….…………………63 4.7 Food-aid Distribution during season of 2003/04……………………………………………………………………..64 4.8 Households who received Seasonal forecasts during 2003/04………………………...…………………..65 4.9 Assessing Farmers’ Level of Confidence with the Forecasts…………………………………………………66 4.10 Decision-making factors ……………………………………………………………………………………………………………...67 4.11 Impact of Drought on Smallholder Farmers………………………………………………………………………….….68 4.12 Farmers’ Drought Coping Strategies…………………………………………………………………………………….……69 4.13 Drought vulnerability factors… …………………………………………………………………………………………………...70 4.14 Measuring correlation between crop yields vs. identified indices for the period 1980-00…………………………………………………………………………………………………………………………………72 5.1 Variable definition and description for (a) seasonal forecasts, (b) drought insurance (c) and SUR Models………………………………………………………………………………………………………………………90 5.2 Parameter estimates of Logit vs Bivariate Models …………………………….…………………………………...93 5.3 Estimated WTP for Improved Seasonal Forecasts……………………………………………………………….…94 5.4 Comparison of Single-bound vs. Double-bound Models……………………...……………………………...97 5.5 Impact of Mitigation costs on WTP for drought Insurance…………………………………………………98

x 5.6 Estimated WTP for Drought Insurance across Districts………………………………………………….……99 5.7 Aggregated WTP for Drought Insurance across NR II-V……………………….……………………………100 5.8 Change in WTP in the presence of food-aid relative to without food-aid case...…………….101 5.9 Trends in Smallholder Crop Yields and Rainfall (1980-99)…………………….……………………………103 5.10 SUR on Cotton and Maize yield with VCI and Quadratic Trend……………………………………...106 5.11 SUR on Cotton and Maize yield with Rainfall and Quadratic Trend…………………………….…107 5.12 Illustrating detrended yield for district # 1 (1980-00)……………………………………………………….…109 5.13 Descriptive statistics for VCI and Rainfall Index by districts………………………………………….….111 5.14 Rating area-yield insurance using VCI……………………………………………………………………………………..113 5.15 Rating area-yield insurance using Rainfall index….………………………………………………………………..114 5.16 Drought Classification……………………………………………………………………………………..…………………………..116

xi LIST OF FIGURES Figure Page 1.1 SST_anomalies vs. rainfall_anomalies (1980-2002)………..…………….…………………………..……………….4 1.2 GDP Growth rate and Rainfall anomaly for the period1970-1998)………….……………………………..5 2.1 Response of Wheat across NOAA-AVHRR Reflective Channels……………………………………….24 3.1 Characterizing the existence of Long-run Equilibrium under Catastrophic risk..…………..49 3.2 Evolution of correlation Coefficient for Channels 5/7 vs. Different Growth Stages of Rice…….….…………………………………………………………………………………………………………………………………………..52 4.1 Illustrating VCI Profile during good, average and bad season for districts 1 & 7.………….…74 4.2 Illustrating VCI Profile during good, average and bad season for districts 3 & 6..……………75 4.3 Illustrating VCI Profile during good, average and bad season for districts 2 & 5..…………..76 4.4 Illustrating VCI Profile during good, average and bad season for districts 4 & 9..…………..77 4.5 Estimated Rainfall Kernel density Distribution for NR II………………………………….…………..……..81 4.6 Estimated Rainfall Kernel density Distribution for NR III……………………………………………………82 4.7 Estimated Rainfall Kernel density Distribution for NR IV…………………………………………………...83 4.8 Estimated Rainfall Kernel density Distribution for NR V……………………………………………………..84 4.9 VCI Kernel density distribution for NR II………….………………………………………………………….………...85 4.10 VCI Kernel density distribution for NR III……………….…………….……………………………………………..…86 4.11 VCI Kernel density distribution for NR IV…………………………….….…………………………………………..…87 4.12 VCI Kernel density distribution for NR V…………………………….….………………………………………..…….88 5.1 National Maize and Cotton Production Trend by Smallholders (1969-99)……………………..104 5.2 Index-based Indemnnity Schedule for Standard Contract…………………………………….…..………..110 5.3 Portfolio loss vs. Optimal Indemnity for district 1 using VCI and Rainfall Index..………….117 5.4 Portfolio loss vs. Optimal Indemnity for district 2 using VCI and Rainfall Index..………..118 5.5 Portfolio loss vs. Optimal Indemnity for district 3 using VCI and Rainfall Index..………..119 5.6 Portfolio loss vs. Optimal Indemnity for district 4 using VCI and Rainfall Index..……..…120 5.7 Portfolio loss vs. Optimal Indemnity for district 5 using VCI and Rainfall Index..……..…121

xii 5.8 Portfolio loss vs. Optimal Indemnity for district 6 using VCI and Rainfall Index..……..…122 5.9 Portfolio loss vs. Optimal Indemnity for district 7 using VCI and Rainfall Index..……..…123 5.10 Portfolio loss vs. Optimal Indemnity for district 8 using VCI and Rainfall Index...……….124 5.11 Portfolio loss vs. Optimal Indemnity for district 9 using VCI and Rainfall Index...…….…12

xiii CHAPTER 1

NATURE OF THE PROBLEM 1.1. Background Agriculture constitutes a predominant sector of Zimbabwe’s economy and contributes about 14% to gross domestic product (GDP). Zimbabwe’s agriculture is a dualistic system consisting of a large-scale commercial-oriented farming sector versus a rural subsistent-oriented smallholder farming sector. The large-scale farmers are high input users who heavily rely on modern inputs and technology, such as fertilizers, herbicides, insecticides and machinery equipments. These farmers enjoy access to credits and insurance markets. The smallholder farmers, on the other hand, practice traditional farming that largely relies on simple implements, such as hoe and pick, animal draft power, own labor, and often have limited access to credit and insurance markets.

Smallholder farmers comprise over 70% of Zimbabwe’s rural population and more than 80% of these farmers are located in agro-ecological zones1 III, IV and V (Table 1.1 and Map 1). These farmers derive their livelihoods largely through subsistence farming. Compounded by lack of irrigation, most smallholders practice dry land farming. Less than 1% of the total cultivated area within the smallholder sector is under irrigation (Table 1.2) and hence the livelihoods of smallholder farmers are heavily affected by climatic variability. In sum, the interplay factors, such as climatic variability, lack of irrigation facilities, limited access to credit and insurance markets, and general macroeconomic instability combine to make smallholder agriculture a risky venture.

1 Zimbabwe agricultural land is sub-divided into five agro-ecological zones numbered I to V whose potential declines as zone number increases.

1 NR Average Annual Area Area Rural Ideal Type of Precipitation (mm) (km) (%) population(%) farming

I >1000 7000 2 1 Specialized and diversified farming

II 750-1000 58,600 15 8 Intensive grain production and ideal for rain fed agriculture

III 750-800 72,900 18 17 Semi-intensive farming and suitable for drought resistant crops

IV 450-750 147,800 38 45 Semi-extensive farming and drought resistant, short season crops

V <450 104,400 27 29 Extensive farming and highly drought resistant crops

Table 1.1: Land Classification in Zimbabwe Source: Ministry of Agriculture Annual report, 2000

Crop LSS (%) SSS (%) Maize (white) 8 0.5 Maize (yellow) 16 0 Cotton 41 0 Groundnuts 86 0 Soybeans 31 0 Wheat 100 0 Tobacco 15 0.1 Average 36 0.3

Table 1.2: Area under Irrigation: Large Scale vs. Smallholder Sector Source: Masters, 1991

2 Drought in particular is one of the major climatic risks that affect millions of smallholder farmers in developing countries. In Southern Africa for example, smallholders have experienced the ravages of drought in all forms; from mild to severe, from severe to catastrophic and from short- to long-term. Whenever a severe drought occurs smallholder farmers are left severely food insecure and often resort to government and non-government organizations (NGOs) for free food-handout and assistance. Drought in no uncertain terms is one factor that has perpetuated food insecurity and poverty in most parts of the developing countries.

1.2 Causes of Drought and its Impact on Zimbabwe’s Economy To institute pro-food security and risk management policies one has to understand the factors causing drought in Zimbabwe. Until recently, drought occurrence in Southern Africa has become closely linked to El Niño2 events that occur in the eastern tropical Pacific (Mason and Jury, 1997, Eastman 1996). When sea surface temperatures (SST) anomalies are high, this causes warming of the ocean resulting in a huge mass of warm water in the central and eastern tropical Pacific. This in turn affects atmospheric circulation, disturbs the normal pattern of air pressure, tropical rainfall and movement of trade winds leading to changes in weather patterns around the globe (Ropelewski and Halpert 1987, NOAA 1997). Southern Africa is that part of the world where El Niño events have a significant influence on rainfall patterns that are strongest during the peak summer rainfall months of December-March (Mason etal, 1997) and often culminates in severe droughts (WMO 1995, Kogan 1998). For instance, major drought events that occurred in Zimbabwe during the periods of 1983/84, 1986/87, 1991/92, 1997/98 are closely associated with El Niño events. The atmospheric component used to describe the El Niño event is called the “Southern Oscillation”. The coupling phenomenon involving both the atmosphere and ocean components is often referred to as El Niño Southern Oscillation (ENSO).

Figure 1.1 illustrates the relationship between SST versus rainfall anomalies in Zimbabwe. Positive SST anomalies underlie El Niño episodes and are inversely associated with rainfall anomalies and vice-versa during La- Niña episodes3. This relationship is somewhat observed during drought years 1983/84, 1986/87, 1991/92, 1997/98 in Figure.1.1. A good illustration is the 1991/92 drought with a record rainfall anomaly greater than –1.5 (Figure 1.1). While the greatest El Niño event was recorded during the 1997/98 season with an anomaly greater than 2.5, it did not transform into the worst drought like the one that occurred in 1991/92. This observation

2 The term El Niño was coined to describe the annual warm ocean current which comes from the tropical Pacific ocean close to the equator and runs along the coast of Peru and Ecuador about Christmas time 3 Term used to describe the opposite effects of El Niño

3 emphasizes that the relationship between El Niño and rainfall is more complex than what is observed prima facie.

3 2.5 2 1.5 1 0.5 0 -0.5 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 00 01 02 -1 -1.5 -2

SST_anomaly Rainfall_anomaly

Figure 1.1: SST_anomalies vs. rainfall_anomalies (1980-2002)

Following the ‘scientific discovery’ of El Niño and its influence on rainfall patterns in Zimbabwe, many institutions have become actively involved in monitoring El Niño episodes with the sole objective of providing strong evidence of an impending drought. Because the economy of Zimbabwe is largely dependent on agriculture, drought has a tendency of retarding economic growth. Figure.1.2 illustrates the ‘sensitivity’ of the Zimbabwe’s economy to rainfall anomalies. It is noticeable especially for the period after 1980 that an increase or decrease in economic growth rate is closely associated with rainfall patterns and somewhat occurs with lagged effects. For example, a good rainfall season in 1981/82 was subsequently followed with record economic growth rate of about 11.4% in 1982/83 while a bad rainfall season in 1991/92 subsequently resulted in a growth rate shrinking to a record level of 10.3% in 1992/93.

4 15

70 72 74 76 78 80 82 84 86 88 90 92 94 96 98 -5

-10

-15

GDP_grow th rate Rainfall_anomaly

Figure 1.2: GDP growth rate and rainfall anomaly for the period 1970-1998

Heal and Lin (1996) estimated that an El Niño-induced drought results in a 9.4% decline in agricultural production in Zimbabwe. Cane, Eshel and Buckland (1994) also illustrated a close relationship between crop production in Zimbabwe and El Niño cycles. They illustrated that El Niño episodes are closely related to rainfall and maize yield in Zimbabwe with correlation coefficients of 0.64 and 0.71 respectively. Lecomte and Thiao (1995) found that maize yields in Zambia and South Africa dropped significantly during the 1982-83 and 1991-92 El Niño events. Overall, a drought in southern Africa can cause a 5-10% drop in GNP and can result in substantial social disruption (Watson, 1996).

1.3 Early Warning Institutions and Drought Monitoring Various institutions have been put in place by the government in collaboration with other Southern African Development Community (SADC) countries, so as to deal with the issue of drought at both the national and regional level. A few such institutions include the National Early Warning unit (NEWU), Regional Early Warning unit (REWU), Regional Early Warning Systems (REWS), Drought Monitoring Centre (DMC) under the National Meteorology Department, SADC Regional Remote Sensing Project (RRSP), the Famine Early Warning System (FEWS) project of USAID and various ministerial committees. These institutions are briefly discussed below.

5 NEWU housed within the Department of AGRITEX under the Ministry of Agriculture was established in 1987 (Stack, 1998) and its core functions include providing advance information on anticipated production, shortages and requirements of basic food commodities. It also provides continuous monitoring of the food security situation, price monitoring, crop forecasting, input supplies and food stocks. In addition, NEWU’s other task is to identify food deficit areas, as well as assessing household food requirements. In essence, NEWU’s key function is to increase awareness of policy and decision-makers of food security situations on the ground.

REWS and REWU are other early-warning institutions, similar to NEWU. However, these institutions operate under the auspices of SADC. Like NEWU these institutions main role is to provide individual governments within SADC with advance information pertaining to availability of food crop, crop failures and shortfalls, and food stocks and projections.

FEWS is another early warning institution introduced by USAID in 1992, following the disastrous drought that occurred during the 1991/92 season. FEWS’s main task is to assist the Department of Social Welfare (under Ministry of Labor, Manpower Planning and Social Welfare) in identifying vulnerable households that qualify for public assistance. It also develops vulnerability assessment methods useful for targeting of food relief. FEWS works in collaboration with REWU and NEWU in developing national estimates for food relief and allocations.

Another early-warning institution is the Regional Remote Sensing Project (RRSP), initiated in 1988 under SADC. RRSP aims to strengthen national as well as regional capabilities in the area of remote sensing in an effort to support early warning activities.

Finally, another important early warning institution is DMC, established in 1991. DMC operates under the Department of Meteorological Services and its main objective is to provide regular and timely information on drought. DMC places emphasis on providing long-term ENSO-based seasonal forecasts, targeted mainly for public policy-makers.

A drawback in the context of current public policy is that early warning means providing compelling evidence that a drought is impending. Thus, early warning activities are concentrated towards preparing for drought emergency and logistics. The system places little emphasis on accessing climate-related information to end-users, especially smallholder farmers. The

6 availability of climate information, especially to smallholder farmers, could be vital for efficient drought mitigation.

1.4 Review of Agricultural Policies in Zimbabwe As Zimbabwe faces the inevitable recurrence of drought, the need to develop drought-mitigation strategies and/or risk transfer mechanisms becomes an important and challenging task. Although Zimbabwe ‘experimented’ with various policy measures (discussed below), most of these policies apparently have failed to address the core problem causing food insecurity and poverty amongst smallholders. The issue at the center of smallholder’s misery and hardship is drought. Why are smallholder farmers so vulnerable to drought shocks? Once the underlying causes of vulnerability are understood, it becomes prudent to prescribe policy actions appropriate for reducing the exposure of smallholders to drought risk. The section below discusses the main agricultural policies implemented in Zimbabwe since independence in 1980. The discussion is divided into two parts- pre-reform and post-reform policies.

1.4.1 Pre-Reform Agricultural Policies4 (i) Incentive producer price Since independence Zimbabwe’s cornerstone policy has been to provide incentive producer prices to the farming community. The noble objective of producer price policy was to increase farm incomes and hence raise the standards of living especially for the smallholder farmers. For example, starting in the 1985/86 season the government included millets5 (pearl and finger) as controlled commodities and concomitantly raised the nominal producer prices to almost twice that of white maize (Table.1.3). This policy was deliberately aimed at not only encouraging millets production by the smallholders, but also improving their incomes.

The Ministry of Agriculture largely administered producer prices, in collaboration with a number of quasi-government marketing institutions, such as the Grain Marketing Board (GMB), Cotton Marketing Board (CMB), Dairy Marketing Board (DMB) and Cold Storage Commission (CSC). These institutions enjoyed a government-mandated monopoly.

4 Pre-reform era refers to the period of “command market” where marketing of the main agricultural commodities was controlled by the government 5 Millets (pearl and finger) are predominantly grown by smallholder farmers

7 Year Maize Seed Soybean Sunflower Groundnuts Sorghum Pearl Finger cotton millets Millets 1980 120 375 160 159 330 105 Nc Nc 1981 120 400 170 182 330 115 Nc Nc 1982 120 515 200 220 390 115 Nc Nc 1983 140 515 260 255 420 120 Nc Nc 1984 180 570 287 285 450 140 Nc Nc 1985 180 670 320 320 500 180 250 300 1986 180 750 340 340 750 180 250 300 1987 180 800 385 390 900 180 250 300 1988 195 850 420 430 1000 195 250 300 1989 215 925 435 455 1000 215 250 300 1990 225 117 485 505 1250 225 260 310 1990 270 135 560 580 1500 270 350 350 1991 550 262 950 995 1800 550 550 520 1992 900 320 1323 1472 2400 520 520 520 1993 1050 555 1600 1472 2925 520 520 520

Table 1.3: Nominal producer prices in Z$ for controlled commodities Source: Ministry of Agriculture, nc = not controlled until 1985

The price controlling function involved annual reviewing and fixing of floor prices. The producer price was maintained as a two-tier system involving pre- and post-planting prices. Post-planting prices were only price signals subject to revision if necessary while pre-prices were the likely prices to be offered during the pending season.

In return farmers were expected to make compulsory delivery of all controlled crops to the appropriate marketing boards. After crop deliveries, the marketing boards would sell the crops at government-stipulated prices to various end-users, such as commercial millers, stock-feeders, brewers and etc. However, this controlled marketing chain created numerous problems. First, producer prices were maintained at artificially low levels that threatened the viability of farm enterprises, especially commercial farmers. Second, ex-marketing board prices were also maintained at artificially low levels that resulted in most marketing boards selling at a loss. Third, marketing boards could not exercise flexibility in their marketing operations, even in the face of adverse economic conditions. Fourth, some of the controlled crops, such as millets, lacked ready markets and this forced the GMB in particular to foot huge storage costs. Finally, due to

8 sub-optimal decisions, most marketing boards accumulated huge annual deficits and had to survive through government subsidies.

By early 1990s, under the auspices of the International Monetary Fund (IMF) and the World Bank the government introduced the Economic Structural Adjustment Program (ESAP). Within the agricultural sector, the reforms were intended to achieve three objectives: to eliminate food subsidy payments to marketing boards; encourage a free competitive market system; and privatize or commercialize the marketing boards and grant them more, if not full, autonomy over their commercial activities.

The general belief was that liberalizing the markets would provide incentives for and hence entice wide-scale private trader participation. This could facilitate grain trading between surplus and deficit regions with the likelihood that farmers would enjoy higher returns. Unfortunately, for a number of reasons things did not materialize as anticipated. For instance, there was no concomitant infrastructural development and/or improvement of the road network, especially in remote rural areas, to facilitate grain movement. There was also a lack of commitment by the government itself to fully relinquish control on some sectors such as export and import control of maize. Market restrictions were selectively applied to large versus small traders. Hence, without wide-scale private trader participation and the limited role of former marketing boards under the new market scenario, smallholder farmers were faced with severely limited access to formal market outlets.

(ii) Collection-point Marketing System A second policy experimented with was the setting up of GMB administered collection-point marketing system throughout the rural areas. This marketing system resulted in significant reduction in transport cost that led to high market participation rate by smallholders. The latter in turn resulted in phenomenal increases in crop sales from the smallholder sector to marketing boards especially GMB and CMB. For example maize production increased by more than 140% for the period 1980 to 1985 with maize sales to the GMB rising from 360,000 tons to about 820,000 by 1985, an increase more than two-fold (Table.1.4). Smallholder cotton production increased almost ten-fold and cotton deliveries to CMB increased tremendously as well. This period is often dubbed Zimbabwe’s miracle-production boom.

9 However, despite the seemingly positive impact to the smallholder economy, the collection-point marketing system was discontinued by around 1989. Because the program was a key contributor to the huge trade deficit incurred by GMB that by 1988 had accumulated to almost Z$1.0 billion (US $200million, at 1990 prices), the program was unsustainable and hence terminated.

Year Smallholder Sector Large Scale Sector Credit to smallholders Mze Mze Mze Cot Cot Mze Mze Mze Cot prod Cot # Value Prod Yield sales Prod yield Prod yield sales (tons) yield ‘000 (Z$m) (tons) T/ha to (tons) (t/ha) (tons) (t/ha) GMB (t/ha) GMB (ton) 1980 600 0.67 - 12 0.8 910 3.28 - 146 1.90 - - 1981 1000 1.00 363 45 0.76 1833 5.05 1650 125 1.86 18 4.2 1982 595 0.54 369 27 0.53 1213 3.84 1021 108 1.68 30 10.1 1983 285 0.27 152 32 0.50 624 2.20 464 114 1.89 39 13.2 1984 670 0.59 390 70 0.70 678 3.02 465 151 2.06 50 23.4 1985 1558 1.53 819 110 0.85 1153 4.85 1008 164 2.00 66 32.0 1986 1348 1.26 682 98 0.85 1064 4.43 912 153 1.93 78 38.9 1987 627 0.59 156 82 0.60 466 3.17 247 197 1.82 77 60.0 1988 1609 1.40 756 137 0.85 643 4.29 440 202 1.68 70 49.4 1989 1188 1.15 654 123 0.80 743 4.42 510 147 1990 1262 1.30 423 103 0.67 731 4.09 357 102 1.59 57 41.4 1991 1019 1.10 372 138 0.70 566 3.24 233 123 0.77 44 33.4 1992 115 0.16 0.72 35.7 0.20 245 1.61 12 40 1.69 30 26.4 1993 1133 1.01 718 134.5 0.68 878 4.44 620 79 1.76 - - 1994 1313 1.12 536 110.8 0.61 1012 2.33 546 70 1.08 - - 1995 399 0.33 31 56.1 0.31 440 4.50 32 36 1.83 - - 1996 1387 1.27 576 157.6 0.72 922 4.70 296 73 1.76 - - 1997 1453 0.98 178 197.8 0.74 738 4.14 76 80 1.92 - - 1998 727 0.69 75 182.6 0.76 690 3.66 155 90 0.47 - - 1999 845 0.67 - 188.4 0.69 674 4.39 - 76 1.74 - - 2000 1240 1.03 - 263.4 0.81 908 4.40 - 110

Table 1.4: Maize and Cotton Production by Smallholder vs. Large-Scale Sector (1980-2000) Source: Ministry of Agriculture Statistical Bulletin, 2000.

(iii) Increased Access to Credit The third policy implemented soon after independence was a deliberate attempt to increase access to credit facilities for the smallholder farm sector, a once neglected sector under colonial rule. Through the government-controlled Agricultural Financial Corporation (AFC), credit facilities were opened for the smallholder farmers. By the early 1980s there was a sudden surge in demand for agricultural loans by smallholders (second last column of Table 1.4). By 1986/87 the number of smallholders who had successfully applied for AFC loans peaked to 78,000

10 representing a four-fold increase compared to 1980. Unfortunately this success story did not last long, for by 1992 the number has dwindled to 30,000. The high loan default rate forced the AFC to curtail its loan services to smallholder farmers. The farmers on the other hand attribute their failure to service their loans due to drought. Currently less than 1% of the smallholder farmers are receiving loans from AFC (Chimbwanda 2002, personal communication). With the majority of smallholders lacking financial capital, these farmers cannot make meaningful investments beyond subsistence level.

1.4.2 Post-Reform Agricultural Policies6 (i) Free Crop Pack Program The post-reform policies are mainly intended to address either directly or indirectly the problem of drought. One such policy is the free crop pack program. This policy was initiated in 1992/93 as a post-drought recovery measure. The program was entirely meant for smallholder farmers who are considered more vulnerable to drought risk than their large-scale counterparts. The program involved distributing free inputs, mainly maize seeds and fertilizer packs. Although the benefits of this program have not been fully established, it is questionable whether the program had a significant impact on smallholder crop production because of inequitable distribution, inadequacy and in some cases ill-suitedness of seed crops in some regions.

(ii) Strategic Grain Reserve Program The second post-reform policy is strategic grain reserve. Under this policy, a physical stock of about 640,000 tonnage of maize is maintained by GMB on behalf of the government in any given year (Ministry of Agriculture 1995/96 policy statement). This policy arose following the call by the IMF and World Bank to fully commercialize the functions of GMB and the need to separate social from commercial functions of the board. The strategic grain reserve stock operates as a buffer intended to moderate grain prices, especially during periods of grain scarcity common during drought years. Under this program, storage costs are borne by the government.

A drawback with this policy is that, stock of 640,000 tons is not sufficient to meet a huge demand for grain that normally arises following a drought year. Normally national consumption averages 1.2 million per year and hence reserve stock of 640,000 is far insufficient. Further, the reserve stock may be too small to allow significant price moderating effects.

11 (iii) Food-Aid Program7 Perhaps the most active policy that has been in operation since 1980 is drought food relief or food-aid program. This is a free food transfer program operated mostly on emergency basis with the objective of averting a looming starvation or famine crisis. The merit of the program is that it provides direct delivery of food to households in dire need of it. However because much of the food-aid is given away indiscriminately, inequitably or simply in unascertained ways it is questionable whether the most vulnerable groups do get enough benefits to survive through the crisis (Sen, 1986). Further, the program is riddled with numerous problems. Food-aid costs are often exorbitant especially following an extreme drought event. The program is often fraught with moral hazards due to lack of targeting and often subject to political abuse. The program is inherently vulnerable to severe logistic failures especially if it is a wide-scale operation. More important, the program nurtures a culture of dependency that may discourage recipients from seeking efficient self-reliant risk management tools.

The main occupation of the current food-aid program is solely to provide immediate relief to the starving population as fast as possible. This program tends to overshadow the need to develop sustainable long-term policies aimed at providing a long-term solution to the problem of food insecurity. In Sen’s words (1986) food-aid “conjures up the picture of a battle already half lost and focuses the attention on emergency operations narrowly aimed at containing large-scale mortality”. Effective pro-food security policies call for much more, than simply rushing food to the victims when they start dying of starvation. It involves a network of decisions relating to diverse policy areas, such as the improving incomes, the stabilization of food prices, facilitating and promoting the development of credit markets and risk-transfer institutions, and the general rehabilitation of the rural economy.

1.5 Lessons Learned From the brief policy review above it is clear that Zimbabwe has experimented with a number of agricultural policies, aimed directly or indirectly at ensuring food security at both the national and household level. Valuable lessons can be drawn from this policy experience:

Lesson .1: Given the right mix of policies smallholders are capable of producing beyond subsistence level

6 refer to policies implemented after ESAP (1990/91) 7 is not strictly a post-reform policy

12 The smallholder’s ‘miracle production boom’ realized in the mid-1980s serves as a cue that given a supportive and conducive agricultural policy environment, smallholders have the capability to produce beyond subsistence level.

Lesson .2: Inflation has eroded producer price and farmers’ incomes Currently, the greatest threat to smallholders’ farm income is inflation. With an inflation rate of over 350% (during 2002/03) per year, real producer prices have been severely eroded to levels lower than 1980. Low producer prices translate into low farm incomes. Further with the majority facing limited access to formal market outlets, many will continue to subsist below the poverty datum line.

Lesson.3: Food-aid is not a solution to problem of food insecurity Food-aid remains the most active policy program being pursued by the government to deal with household food insecurity. While food-aid partly addresses the problem of food insecurity at least within the short-run, it is not in itself a solution. A solution to the problem of food insecurity does not lie in the mere provision of free food handout to affected households.

Lesson .4: Lack of risk-sharing policy mechanisms One issue that is not well addressed from the policy perception is how to minimize the vulnerability of smallholders to drought shocks. Currently, the government tends to narrowly treat drought as a natural disaster rather as an integral part of a national policy framework. This is evidenced by the lack of risk-sharing policy mechanisms that seek to minimize the impact of drought, not only to smallholders but the farming community as whole. While drought risk can never be eliminated, its impact can be reduced through implementation of policies that seek to efficiently mitigate its impact.

What are rather missing in Zimbabwe are agricultural policies that promote the evolvement of formal risk-sharing institutions. Risk-sharing institutions could play a crucial role in reducing the vulnerability of smallholders to drought risk. As well argued by Zeller and Sharma (2000), the myth that the poor smallholders are unable to save or insure has led to implementation of non-supportive and non-facilitative policies that have neglected the savings and insurance services that are essentially relevant to the poor. One plausible avenue to reduce the vulnerability of smallholders to drought shocks lies in the way farmers can protect themselves against

13 drought. Instead of relying on ad-hoc food-aid, smallholder farmers could be encouraged to seek more formal risk-sharing tools, such as drought insurance.

1.6 Defining the Problem Statement This study investigates two policy questions: First, would a policy that encourages adoption of improved seasonal climate forecasts by smallholder farmers be a more efficient drought mitigation strategy? Second, would a policy that implements area-yield drought-index insurance based on remotely sensed vegetation indices be a feasible risk-transfer and risk-sharing mechanism for the smallholder farmers?

In view of increasingly better understanding of ENSO events, it is now possible to predict major El Niño episodes at lead times of about 3-6 months (Mason, 1996). Although El Niño-based seasonal forecasts are being disseminated in Zimbabwe (Phillips et al, 2001) prior to farming season, it is not certain how farmers indeed use these forecasts. While studies in developed countries revealed that farmers benefit substantially from using improved forecasts (Mjelde et al 2000, Solow et al 1998, Easterling et al 1987), the benefit for using these forecasts in developing countries is rudimentary.

Do seasonal forecasts really matter to the smallholder farmers? Perhaps or perhaps not; perhaps seasonal forecasts matter because farmers can use them for efficient mitigation purposes or perhaps forecasts do not matter because there are other more pressing problems than simply the non-availability of seasonal forecasts. To verify this hypothesis, surveys based on the contingent valuation method (CVM) are conducted to assess the farmers’ willingness-to-pay (WTP) for the provision of improved seasonal forecasts8. A hypothetical market situation is created where a farmer is to receive, but at a cost, improved seasonal forecasts scaled down by specific region. The advantage is that improved forecasts are likely to give better prediction than the current broad El Niño-based forecasts. Therefore, farmers’ preferences for seasonal forecasts can then be evaluated thereof using the CVM approach (see section 3.1).

Since drought is a catastrophic risk, will drought insurance intended for smallholder farmers be feasible and/or viable? Skees, Hazell and Miranda (1999) suggested that an insurance scheme intended for smallholders ought to meet the following minimum conditions: first, the scheme ought to be affordable and accessible to the majority of the households including the poor;

14 second, the scheme should compensate for the catastrophic income losses so as protect consumption; third, the scheme should be practical to implement; fourth, the scheme ought to be provided by the private sector with little or no government subsidies; and finally the scheme must avoid moral hazards and adverse selection.

This study advocates for an area-yield drought-index insurance based on the remotely sensed vegetation condition index (VCI). Such a scheme has the potential to meet the minimum conditions suggested by Skees et al: (i) since the insured receives indemnity payments on the basis of an index determined by remotely sensed vegetation indices, the problem of moral hazards is eliminated; (ii) although the scheme can be complemented with ground data, the farmer is not strictly required to demonstrate any loss and hence this is likely to result in significant reduction in administration costs, thus in turn enhancing the feasibility of the scheme; (iii) because drought is a catastrophic risk, it may be prudent for the government to facilitate the emergence and development of an insurance market by offering subsidized reinsurance, and (iv) because satellite data are cheap to obtain and often come with high temporal and spatial resolution this makes crop yield monitoring and loss assessment relatively easy such that implementing the scheme should not be difficult.

My prime concern is whether the insurance scheme will be feasible. Feasibility among other things depends on: first, the demand for insurance services by smallholders; the stronger the demand the higher the chances that the scheme can survive. Related to the demand for insurance is the ability and willingness to pay premium rate high enough to cover operational costs and yet low enough to attract a large clientele that includes the rural poor. Second, the greatest challenge lies in whether the insurance provider will be able to withstand catastrophic risk within both the short and long run. One way could be for the insurer to participate in reinsurance markets or for the government to offer subsidized reinsurance. Third, the provision of food-aid to smallholders may decrease both the demand for and supply of insurance services and hence precipitating the failure of insurance market. These factors are likely to have a bearing on the feasibility of insurance scheme suggested and some of these aspects are further explored in Chapter.3.

8 The location-specific forecasts are hereafter referred to as improved seasonal forecasts

15 1.7 Study Objectives This study intends to attain three objectives. The first objective is to assess the economic value of improved seasonal forecasts to smallholder farmers using the willingness-to-pay (WTP) approach. With improved understanding and possibility of predicting major El-Niño events at lead times of about 3-6 months, is a policy that advocates for wide-scale adoption of seasonal forecasts by smallholder farmers a more effective drought mitigation strategy? To attain this objective, surveys based on the contingency valuation method (CVM) were conducted throughout the country’s agro-ecological regions II-V (see Chapter 4).

Similar to the first, the second objective is to use the WTP approach to assess the potential demand for formal drought-index insurance by smallholder farmers. Currently there is no formal crop insurance for smallholders and rather farmers informally insure themselves through such activities as food aid, remittances, migration and the selling of assets, especially livestock.

The third objective involves assessing the feasibility of offering formal drought insurance to smallholder farmers on the basis of remotely sensed vegetation indices. If vegetation indices, such as the normalized difference vegetation index (NDVI) and vegetation condition index (VCI), are correlated with crop yield, are easy to observe and measure on a regular basis via satellites sensors, not subject to manipulation by individual farmers, can area-yield drought-index insurance based on these indices be feasible?

Following these objectives, the study intends to verify the following hypotheses: (i) seasonal forecasts are important to smallholders for efficient drought mitigation purposes, provided they are tailored by specific region or locality; and throughout the study these region-specific forecasts are referred to as improved seasonal forecasts, (ii) smallholder farmers posit strong demand for formal drought insurance, (iii) because drought is a catastrophic risk, it is the cost not the lack of demand that constrains provision of insurance services to smallholders, and (iv) in view of the latter hypothesis, if cost-effective remedies, such as area-yield insurance based on remotely sensed vegetation indices are used, supply of drought insurance to smallholder farmers may be feasible.

The key assumptions of the study are: (i) Smallholder agriculture is rain-fed. In other words smallholder farmers largely practice dry-land farming that does not involve irrigation. (ii) Smallholder farmers exhibit risk-averse behavior and hence they are willing to purchase drought

16 insurance. The latter behavior explains why farmers are willing to sacrifice small premium payments so as avoid potentially huge losses. (iii) Because the insuring firm faces catastrophic risk, its too exhibits risk-aversion and hence its willingness to seek reinsurance. (iv) A risk- sharing scheme is Pareto efficient in that by engaging into a risk-sharing arrangement both insured (farmer) and insurer (firm) will achieve improvement in their respective welfare. Finally, remotely sensed vegetation indices can be measured with minimal atmospheric, geometric and radiometric distortion.

17 CHAPTER 2

LITERATURE REVIEW

This chapter provides a literature review focusing on four issues pertinent to the study. The first part reviews the capability of remote sensing techniques in monitoring drought and assessing agronomic condition. The second part presents a review on smallholder drought mitigation strategies and the inadequacy of traditional self-insurance mechanisms. The third part reviews experience with publicly provided multiple peril crop insurance (MPCI) programs and their expensive failure. The last part reviews a new approach to insurance using index and area-based contracts.

2.1 Remote Sensing and Drought Monitoring Drought is a complex natural hazard that affects more people than any other hazard in Zimbabwe and likewise most of Southern Africa. It differs from other natural hazards, such as floods, earthquakes or tornadoes, in the sense that its effects often accumulate slowly over a considerable period of time and hence drought is often referred to as a ‘creeping phenomenon’ (Tannehill, 1947). The importance of drought lies in its impacts that can be direct or indirect, singular or cumulative, immediate or delayed (Asfaw, 1989).

Three characteristics that distinguish one drought from another are severity, duration and spatial coverage. Accurately monitoring drought entails quantifying its spatial extent, severity and duration and often this is no easy task. High spatial and temporal variation makes it extremely difficult to monitor and predict drought intensity. Traditionally, the approach has been to use meteorological variables to assess drought severity (e.g., use of standardized precipitation index, soil moisture index, Palmer drought severity index). However, meteorological networks are generally less dense for an accurate representation of a high spatial climatic variability.

18 Recent satellite technological advances permit evaluation of drought from a new perspective. The advantage with remote sensing technology is that it provides repeated measurements at a particular spatial scale and spectral bandwidths that allows dynamic environmental conditions, such as vegetation cover, to be monitored with considerable accuracy. Hence remote sensing has proven a powerful tool for evaluating the temporal and spatial aspects of drought (Johnson 1993, Peters et al, 2002). Interest in satellite observation and subsequent evaluation of drought stems mainly from the following aspects: (i) remote sensing offers a unique vantage point, (ii) provides a synoptic view, (iii) forms a permanent record or data archive, (iv) offers extra visual information and (v) is usually cost-effective (Lillesand etal. 2000). In sum remote sensing, as a monitoring system, is capable of providing three essential functions- control, warning and forecasting.

In recent years, many international organizations and national governments have shown growing interest in using satellite data for drought early warning and crop yield assessment (Johnson et al, Yin and Williams 1997). Remotely sensed data, especially from the Advanced Very High Resolution Radiometer (AVHRR) sensor aboard the National Oceanic and Atmospheric Administration (NOAA) series of polar-orbiting satellites of the USA, has been used for drought early warning and food security purposes (Johnson et al, Kogan, 1998). Large area monitoring of vegetation dynamics has been made possible by polar orbiting and environmental satellites, such as SPOT, LANDSAT and NOAA (see Table 2.1). Today, these space borne satellites are extensively used in land-resource mapping and inventorying.

Remote sensing satellites are equipped with sensors that detect electromagnetic radiation. The satellites can be divided into two groups, viz. earth resource satellites and environmental satellites. Examples of earth resource satellites include SPOT and LANDSAT. One advantage with earth resource satellites, such as LANDSAT, is that due to high spatial resolution LANDSAT data can be used for dynamic monitoring of crop conditions particularly during the growing season. However, the low temporal resolution often limits the usefulness of LANDSAT data for monitoring crop conditions over large areas. For instance, clouds cover can “contaminate” the data, rendering the data of little use. The next LANDSAT pass covering the same area is 16 days later. Experience has shown that it is highly unlikely that cloud-free data can be acquired once a month (Dijk, 1986). Because the key to crop condition assessment depends on knowing the state of the plants at critical stages in their life cycles, cloud cover contamination at these critical times severely limits the usefulness of LANDSAT data.

19 Other types of satellites are the environmental satellites, which provide high frequency global coverage of weather systems and have a relatively low spatial resolution of 1-5 km2. There are two types of environmental satellites, viz. geo-stationary and polar orbiting. The geo-stationary satellites are approximately 36,000 km above the earth’s surface and these make a complete orbit above the equator every 24 hours. Images produced by geo-stationary satellites always represent the same part of the earth’s surface. Examples of geo-stationary satellites include the European Meteosat, the U.S. GOES and Japanese GMS. The polar-orbiting satellites orbit at an altitude of about 800-1400 km (Dijk 1986) and pass the same general area at the same local time very 12 or 24 hours.

Band Bandwidth IFOV Temporal Altitude Swath width (µm) Resolution (km) (m) (orbit days) (km) Landsat Multispectral Scanner (MSS) on ERTS 1,2 and Landsat 3, 4 and 5 4 0.50-0.60 79x79 18 917 185 5 0.60-0.70 6 0.70-0.80 7 0.80-1.10 8 10.4-12.6 240x240 Landsat Thematic Mapper (TM) on Landsat-7 1 0.45-0.52 30x30 16 705 185 2 0.52-0.60 30x30 3 0.63-0.69 30x30 4 0.76-0.90 30x30 5 1.55-1.75 30x30 6 10.4-12.5 120x120 7 2.08-2.35 30x30 NOAA-Advanced Very High Resolution Radiometer (AVHRR-12) Local Area Coverage (LAC) data 1 0.58-0.68 1100x1100 Twice a day 845-861 2700 2 0.73-1.10 1100x1100 3 3.55-3.93 1100x1100 4 10.3-11.3 1100x1100 5 11.5-12.5 1100x1100 French SPOT High Resolution Visible Sensor Systems (HRV) 1, 2 and 3 Multispectral mode: 1 0.50-0.59 20x20 26 832 60 2 0.61-0.68 20x20 3 0.79-0.89 20x20 Panchromatic Mode: 1 0.51-0.73 10x10 26 832 60

Table 2.1: Characteristics of Selected Multispectral Remote Sensing Systems source: Jensen, 1996.

20 Examples of polar-orbiting satellites are TIROS-N and NOAA-15. What makes the polar- orbiting satellites, such as the NOAA-AVHRR, popular in environmental monitoring is their high temporal resolution (Table.2.1). These satellites provide twice-daily coverage of the planet’s surface, making them ideal for early warning systems, drought monitoring, crop assessment and yield estimation. Another advantage with NOAA satellites is that given their daily coverage, they are likely to provide more cloud-free images compared to LANDSAT with repetitive coverage that occurs every 16 days (Table 2.1). Further these data are accessible at many receiving stations around the world and can be made available in near-real time for the evaluation of current conditions. However, the disadvantages with NOAA data pertain to their low spatial resolution, geometric and radiometric distortions.

2.2 Agronomic Condition Assessment There are several ways in which remote sensing can be of vital use in mapping and monitoring agronomic conditions, such as vegetation condition monitoring and mapping vegetation type, monitoring soil moisture and moisture availability, and monitoring environmental stress. Numerous vegetation indices are extensively used for agronomic condition monitoring, especially vegetation stress and crop yield assessment. A few such indices are summarized in Table 2.2b. Satellite-based indices have the advantage in monitoring spatial and temporal variation of drought-related vegetation stress at regional, continental and even global scales due to their large area and frequent coverage. The basis for using these indices is the observation that the “greenness vigor” of vegetation is a good indicator of environmental conditions prevailing at any given time. Since in drier or semi-arid environments water availability is usually the limiting factor for vegetation development, “greenness vigor” should be a good indicator of the occurrence and severity of water stress (Vogt et al, 2000).

The normalized difference vegetation index (NDVI) derived from AVHRR is one such index that has been extensively used for vegetation monitoring, crop yield assessment and drought detection (Benedetti and Rossini, 1993; Maas, 1988). This is because NDVI helps compensate for changing illumination, surface slope, aspect and other extraneous factors (Lillesand etal 2000). The NDVI measures vegetation vigor caused by chlorophyll activity, sometimes referred to as “greenness vigor”. The NDVI is simply the difference in radiation measured in two different wavelength bands (NIR and R) divided by the sum of the radiation in these two bands (Table 2.2b). In theory, NDVI measurements range between -1 and +1. However, in practice the measurements generally range between -0.1 and + 0.7 (Goward et al., 1985). Clouds, water, snow,

21 ice and non-vegetated surfaces have negative NDVI values. Bare soils and other background materials produce NDVI values ranging from -0.1 to +0.1. The NDVI values for vegetation range from 0.01 to 0.66, with low values indicating poor vegetation conditions and possibly unfavorable weather impacts, while high values describe the opposite.

The vegetation indices ability to detect drought conditions is based upon unique spectral signatures manifested by different vegetation types via specific wave-bands (Table 2.2a). In the case of AVHRR, channel_1 is sensitive to the visible wavelength range (0.58-0.68 µm). As illustrated in Figure 2.1, healthy vegetation typically is highly absorptive in this wavelength so that reflected radiation detected by channel_1 is low. On the other hand, channel_2 measures reflected radiation in the near-infrared wavelength (0.72-1.1 µm). Healthy vegetation is highly reflective in the NIR wavelength. Combining these two channel responses provides a vegetation index that measures vegetation vigor. Stressed vegetation has a higher reflectance than healthy vegetation in the visible (channel_1) and a lower reflectance in the near-infrared (channel_2) region of the electromagnetic spectrum as observed in Figure 2.1. Soils, in general, exhibit higher reflectance within visible bands but lower reflectance in the near-infrared (Figure 2.1).

Notwithstanding the widespread use of AVHRR NDVI, it has been observed that a number of factors can influence NDVI observations that are unrelated to vegetation condition (Lillesand etal 2000). Among these factors are variability in incident solar radiation, radiometric response characteristics of the sensor, atmospheric effects and off-nadir viewing effects. Because the AVHRR scans over ±55o compared to ±7.7o for Landsat and ±2.06o for SPOT, this results in a substantial change in the size of the ground resolution cell along an AVHRR scan line.

Kogan (1998) suggested another measure called vegetation condition index (VCI), as defined in Table 2.2b. The VCI is an indicator of the vigor of the vegetation cover as a function of NDVI minima and maximum encountered for a given land cover and over a longer period. It normalizes NDVI (or any other vegetation index) according to its variability over many years and results in a consistent index for different land cover types (Vogt et al, 2000). It is an attempt to separate the short-term weather signal from the long-term signal, as reflected by the vegetation and this makes it better indicator of water stress conditions than NDVI (Kogan and Sullivan, 1993). The VCI may be thought of as being closely related to the vegetation condition in a specific region and if data are recorded over long periods, where extremes in climate variability are sampled,

22 then VCI may indicate potential crop yields (McVicar and Jupp, 1998). VCI is dimensionless and ranges from 0 to 100 with 0 indicating the worst condition and 100 the best.

(a) Bands Satellite/sensor Vegetation bands used in vegetation monitoring Red (μm) Near Infrared (μm) Landsat-MSS 0.60-0.70 (channel.5) (0.80-1.10 (channel.7) (1,2,3,4,5) Landsat-TM (4,5,7) 0.63-0.69 (channel.3) 0.76-0.90 (channel.4) SPOT-HRV (1,2,3,4) 0.61-0.69 (channel.2) 0.61-0.69 (channel.3) IKONOS 0.45-0.52 (Channe.1) 0.76-0.90 (channel.4) NOAA-AVHRR 0.55-0.68 (channel.1) 0.73-1.1 (channel.2) (b) Indices Vegetation Index Equation NDVI (normalized difference vegetation index) NIR−R NIR+R SAVI (soil adjusted vegetation index (1+L)*(NIR−R) (NIR−R+L) VCI (vegetation condition index) (NDVI −NDVI ) i min * 100 (NDVImax −NDVImax ) Soil adjusted ratio vegetation index NIR (Green−a /b) TSAVI (Transformed soil adjusted vegetation index) a*(NIR−a*G−b) (G+a*NIR−a*b) LAI (Leaf area index) LAI * (NDVI − NDVI ) max i min (NDVI − NDVI ) max min TCI (Temperature condition index) (BTmax −BT ) i * 100 (BTmax −BTmax ) (where BT=brightness temperature) Tasseled cap transformation This transformation identifies 4 indices: SBI (soil brightness index) GVI (green vegetation index) YVI (yellow vegetation index) NSI (non-such index)

Where: NIR is reflectance radiated in near- infrared and R is reflectance radiated in visible red wave band; G is reflectance in the green band; a=0.969; b=0.085;L=0.50

Table 2.2:(a) Bands and (b) Indices commonly used for Vegetation Monitoring Source: Adapted from Thenkabail (1992) and Guevara (2001)

23 Channel-1 Channel-2

0.4 Healthy wheat

Stressed 0.3 wheat Reflectance ( µ m

) 0.2

Soil

0.1

0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

Wavelength (µm)

Figure 2.1: Response of Wheat Across NOAA-AVHRR Reflective Channels Source: Adapted from Dijk.

24 Kogan (1997) suggested another index called temperature condition index (TCI). This index is mathematically defined in an analogous fashion to VCI with brightness value (BT) replacing NDVI (see Table 2.2b). TCI is based on the fact that the brightness temperature of the vegetation canopy or the soil surface will rise with increasing water stress and hence is an appropriate measure of tracking drought conditions.

In sum vegetation indices are correlated with photosynthetic activity in non-wilted plant foliage and are generally good predictors of plant canopy biomass, vigor or stress (Tucker 1979, Labus et al 2002). If sequential observations based on vegetation indices are taken frequently over a season, seasonal profiles are developed that show the progression of crop canopy emergence, maturation and senescence, which reflect crop performance and are related to crop yields (Benedetti et al 1993, Doraiswamy et al 2003). Thus by examining seasonal growth profiles over many growing seasons, one can identify critical times in crop-growth cycles essential to establish the degree of correlation between NDVI, VCI and grain yield.

Therefore, this study proposes that, since vegetation indices such as NDVI and VCI are correlated with crop yield, are easy to observe and measure on a regular basis via satellites sensors and not subject to manipulation by individual farmers, they may form ideal tools to use for area-yield drought-index insurance.

2.3 Smallholder Drought Mitigation Strategies Drought is a common phenomenon in most semi-arid parts of Zimbabwe, especially in agro- ecological zones IV and V (see Map.1). Not only are the smallholder farmers accustomed to it, but generally it is central to their farming operations. Faced with a drought-prone environment smallholders have often devised own strategies to ensure survival. As discussed by Swinton (1988), these drought risk mitigation strategies can be divided into two categories: (i) strategies that seek to minimize the risk of failing to produce the minimal subsistence quantity, and (ii) strategies that limit losses following a production failure. The first category can be classified as “risk-minimization” strategies and the second as “loss management” strategies. In the same vein, Matlon (1991) classifies these strategies as ex-ante and compensatory ex-post drought mitigation strategies.

Risk minimization strategies commonly practiced in semi-arid environments include crop, cultivar and land-type diversification, intercropping, large land-area cultivation and under-

25 investment in modern inputs, such as improved seeds, fertilizers, herbicides, fungicides and insecticides. According to Matlon (1991), these measures are taken in advance to reduce the household’s vulnerability to anticipated production or yield risk. These strategies are provided in the details in literature (see Binswanger et al., 1979, Swinton, 1988 Matlon, 1991).

Loss management strategies are measures taken after the drought shock had occurred. In essence these measures aim to ensure the survival of the household during the turbulent times and hence are primarily concerned with consumption smoothing. Examples of such measures include livestock sales, food-aid, borrowing from merchants, selling domestic assets, migratory labor, and seeking off-farm employment. Depletion of productive assets such as distress livestock selling, jeopardizes the continued existence of the household (Walker and Jodha, 1986).

Adesina and Sanders (1991) discuss another set of risk management tools, referred to as ‘sequential and adaptive’ strategy. Under extreme weather uncertainties smallholder farmers often develop sequential and adaptive decision-making strategies that take advantage of the more accurate information that becomes available as the season unfolds. Examples of such measures include staggered planting, re-seeding, and changing plant population.

2.3.1 Inadequacy of traditional drought mitigation strategies An important question is how efficient are these traditional strategies in mitigating drought risk? Evidence concerning the efficiency of most traditional risk mitigation strategies by smallholders is scanty (Walker etal, 1986). Despite a seemingly wide range of risk management alternatives at the disposal of smallholders, in practice only a few have significant impact. The failure of most traditional drought mitigation strategies in semi-arid regions arises due to: covariate risks, aridity and ‘thin’ cultivars choice and cost (Matlon, 1991; Gautam et al., 1994).

The efficiency of loss-management strategies in protecting consumption and/or consumption smoothing depends largely on the extent to which the respective income measures (agricultural wage income, transfers, non-farm income, sales of assets, food aid) are covariant with local production outcomes. Because of the catastrophic effects, drought-induced losses become correlated across smallholders and may cover large areas. This covariance effect often called catastrophic or systemic risk severely limits the scope for inter-household risk-spreading. Guatam et al (1994) cites covariate (systemic) risks as main obstacles that weakens traditional risk management tools and concluded that households are collectively unable to insure against

26 covariate risks. The covariability problem can be overcome through risk-sharing arrangements that cut across regions that do not simultaneously experience drought. Few informal strategies such as seasonal migration can achieve this.

Matlon (1991) elaborates that with respect to risk-minimization strategies, measures such as crop cultivar and land-type diversification are severely limited especially when applied to most semi-arid zones. This is because the low rainfall and shallow droughty soils restrict farmers’ flexibility with respect to crop or cultivar choice. Matlon concludes that under such conditions traditional risk-minimization strategies are inefficient.

The efficiency of ‘sequential and adaptive’ strategies depends on the latitude of flexibility a farmer can exercise in re-allocating and/or re-shifting resources. This strategy works better during early stages of the season before major resource commitments are made, but diminishes as the farm season progresses. For instance at mid-season, there is nothing much the farmer can do to drastically influence the final outcome. And again this tends to dilute the efficiency of sequential strategies as risk management tools.

Another important limitation of traditional risk management strategies is their cost. As an illustration, assume a farmer adopts crop diversification as a risk management strategy. Diversification has a tendency of discouraging the farmer from growing highly risky, but potentially high-income crops. Since the farmer is primarily concerned with risk avoidance, this can be plausibly achieved by growing low risk crops with potentially low return. Thus, crop diversification pursued as a risk management strategy is likely to reduce the average income (Guatam etal, 1994).

From the discussion above, one can conclude that most of the informal traditional drought mitigation strategies at the disposal of smallholders are indeed not very efficient.

2.4 Experience with Multiple Peril Crop Insurance Agricultural production is inherently a risky venture. As a result, farmers face a variety of risks such as unfavorable climate conditions, volatility in input and output prices, livestock disease outbreak, locusts and pest outbreaks. The impact of these risks is more pronounced to most smallholder farmers in developing communities.

27 In most countries governments have realized the crucial role that risk-sharing institutions, such as crop insurance, can play in mitigating risks. Risk-sharing arrangements aim to attain three objectives. First, risk sharing transfers the risk to institutions which can bear the risk or which are less averse. Second, risk-sharing enables risk to be pooled across regions, crops or other sectors of the economy as a strategy to lessen the covariate risks. Third, risk sharing can stabilize farm incomes (Hazell et al., 1986).

As an institutional response to agricultural risk, public crop insurance covering multiple risk and/or specific risks has been in existence in most developed countries (USA, Canada, Europe) for over a century (Mishra, 1996). Public multi-peril crop insurance (MPCI) has been used as a major risk assistance tool by many developed countries and recently in Asian countries. In Southern Africa, formal insurance institutions of this nature are virtually non-existent. Insurance markets have failed to develop in most rural markets for three reasons (Binswanger, 1986): First, due to asymmetric information, the costs involved in measuring expected yield, probability distributions, yield shortfalls in any given year may be excessive. If these costs are charged as part of the premiums, then the premium charges may be excessive resulting in low participation rate. Coinsurance, deductibles and insuring peril-specific rather than multi-peril risks are some of the practical measures that can be employed to counter these problems. Second, due to systemic risk, crops of insured farmers within and across regions may fail simultaneously and hence insurers must have high cash reserves to meet the likely huge indemnity payments. The latter provides a major disincentive for the establishment of formal private insurance. Third, often administration costs involved in monitoring crop yield performance and loss assessment of individual farmers are too exorbitant. Hence Binswanger (1986) concludes that crop yield insurance is plagued with numerous problems, such that it is the costs and not the absence of farmers’ demand that constrains the provision of privately provided crop insurance to smallholders.

The benefits of crop insurance are well known. Crop insurance has the potential to stabilize farmers’ income and save farmers from disasters that could arise from crop failure. In addition, crop insurance could encourage risk-averse farmers to adopt improved technologies that can lead to increased production. Thus crop insurance could lead to more efficient use of resources. Furthermore, crop insurance can potentially reduce the risk for credit agencies and hence help farmers have better credit rating for purposes of crop loans. Thus insurance, like collateral,

28 increases the expected return of the loan and hence insurance can be viewed as a partial substitute for collateral (Binswanger, 1986).

The experience with publicly provided multiple-peril crop insurance (MPCI) has been an expensive failure (Hazell, 1992, Roberts and Dick, 1991, Gudger 1991). A number of reasons are attributed to the failure of MCPI schemes. First, many risks covered under multi-peril insurance programs are inherently uninsurable9 and in turn this encourages moral hazards and adverse selection. Second, the administrative costs for most MCPI schemes are prohibitively high and hence contribute to huge losses. Third, often governments undermine public insurers for political reasons and this negatively impacts the viability of MPCI schemes. Finally, without a well- diversified portfolio, MPCI schemes are more susceptible to the problem of co-variability and hence resulting in huge losses, particularly during a catastrophic year.

The failure of multi-peril insurance schemes prompted the revision of these schemes. The new approach emphasizes the need to provide area-based and peril-specific insurance schemes (see section 2.5 below). Hail insurance provides a good example of a long standing, subsidy-free and private-driven insurance scheme that been successfully implemented in many countries. In this study, a drought-index insurance is advocated, which like hail insurance, is peril-specific and can be designed to cover a few crops. The idea of drought insurance is not new. It was introduced by Chakravarti (India, 1920) in lieu of moral hazards when he made the observation that ‘no insurance authority could ever maintain a supervising agency which would be able to watch and enforce that every insured field receives the required amount of care and attention at the hands of its cultivator. Unless some method can be devised by which this great difficulty is eliminated, a system of crop insurance would indeed be impossible’, (in Mishra, page 309). Chakravarti (1920) observes that despite ‘a direct system of crop insurance’ being desirable in India in the context of smallholder farmers, it may not be feasible. Instead Chakravarti (1920) argues that given an important characteristic, viz. the dependence on Indian Agriculture on rainfall, ‘it is not only possible but also practicable’ to introduce an indirect system of crop insurance which he specifically refers to as drought insurance. He concludes that an effective agricultural insurance program that hedges the peasantry against serious pecuniary loss in respect of their agricultural activities will render the country less liable to the ravages of famine.

9 For a risk to be insurable it must satisfy the following: (i) the likelihood of the event must be readily quantifiable; (ii) damage it causes must be easy to attribute and evaluate; (iii) probability of occurrence should not too high to make insurance unaffordable; (iv) neither the occurrence of the event or the damage it causes should be influenced by the insured’s behavior; (v) risks ought be independent

29 2.5 Index-based Area-yield Drought Insurance In light of the inherent limitations of traditional MPCI insurance programs (as discussed above), some new innovations have recently emerged. The new innovations include area-yield insurance programs, various exchange-traded area yield contracts, catastrophe options and catastrophe bonds. In particular, catastrophe options and catastrophe bonds can allow insurers to securitize correlated risks and hence circumvent the limitations of traditional insurance markets (Vedenov and Miranda, 2001). However, the drawback is that in most developing countries financial markets are not fully developed to include innovations of this nature, such as catastrophe options and bonds.

On the other hand, the innovation pertaining to area-yield insurance has the potential to address to the crop insurance needs of the rural poor. Essentially, the principle of area-yield insurance is that contracts, unlike MPCI, are written against specific perils or events defined and recorded at a regional level (e.g. county, district or ward). Specific perils could include drought, floods, and hailstorm. Insurance is sold in standard units (e.g. $1, $10 or $100) with a standard contract for each unit purchased called a Standard Unit Contract (SUC). The premium rate for a SUC is the same for all farmers/buyers who buy the same contract in a given region, and all farmers/buyers receive the same indemnity per SUC if the insured event occurs. Buyers of insurance could be anyone interested in hedging against a specific agricultural-related peril and is free to buy as many units of the insurance as desired.

However area-yield insurance requires long and reliable series of area-yield data that are often unavailable in many countries. Alternatively indices could be used instead. Data series for indices such as rainfall, soil moisture, temperature and evapo-transpiration are easily available and disaggregated to the desired level, such as county or district.

Skees etal (1999) discussed a number of attractive features associated with area-yield insurance: First, because farmers/buyers in a region pay the same premium and receive the same indemnity per SUC, it reduces the problems of adverse selection. In addition, because there is virtually nothing an insured farmer can do to influence the outcome of an index, this eliminates the problem of moral hazard. Hence, a farmer who purchases insurance has the same economic incentives to produce a profitable crop as the uninsured farmer. Second, this insurance program could be inexpensive to administer, since there are no individual contracts to write, no on-farm

30 inspections and no individual field-loss assessments. Further, the insurance could be easy to market since SUCs could be used as lottery tickets and presentation of the certificate would be sufficient to claim a payment. Third, by defining SUCs in small denominations makes them more affordable to poor households. Furthermore purchasers need not be farmers, nor should they be restricted to live or work in a specific region. Fourth, such an insurance scheme could be easily administered by private firms and provide an entry point for private insurers to develop other types of insurance products necessary for rural people. Finally, group borrowing in micro-finance could also find area-based index insurance an increasingly viable proposition for hedging against drought risk.

However, index-based insurance has its own problems. For example, an individual can suffer losses and not get any indemnity payment because the major event triggering the payment has not occurred. It is also possible for an individual to be paid when he/she had suffered no loss. This type of risk is referred to as basis risk and index-based insurance would not be attractive, if the basis risk becomes too high.

31 CHAPTER 3

RESEARCH METHODS

This chapter discusses the research methods used to address the research questions discussed in Chapter 1. Central to my investigation are two issues: do seasonal forecasts and drought insurance really matter to smallholder farmers? Second, given drought is a catastrophic risk, can area-yield insurance based on indices, such as rainfall and VCI indices be feasible? To empirically evaluate these hypotheses, surveys based on the contingent valuation method (CVM) are used to ‘recover’ farmers’ preferences or willingness-to-pay (WTP) for improved seasonal forecasts and drought insurance. The first section discusses CVM and discrete dichotomous double-bounded choice models used to estimate WTP. The second section is designed to investigate the feasibility of the proposed insurance scheme by deriving demand for and supply of drought insurance using expected and mean-variance (MV) utility framework. Using the derived demand function, we use comparative static to characterize the demand for insurance in the presence of food-aid and mitigation costs. In addition, the section characterizes the insurance supply function and conditions that support the existence of long-run insurance equilibrium in the presence of catastrophic risk, food-aid and subsidized reinsurance. The final section presents a general methodology for designing and pricing an index-based drought insurance contract.

3.1 Contingent Valuation Method Do seasonal forecasts really matter to the smallholder farmers and does drought insurance matter for smallholder farmers? To empirically evaluate farmers’ preferences for seasonal forecasts and drought insurance we elicited farmers willingness-to-pay10 (WTP) for the two programs using surveys based on contingent valuation method (CVM). CVM is discussed in detail by Mitchell

10 WTP measures maximum amount of income a person will pay in exchange for an improvement in circumstances, or maximum amount a person is prepared to pay to avoid a decline in circumstances

32 and Carson (1989) and has recently become a common tool for evaluating non-market goods. Our prime motive is to build an empirical model based on double-bounded dichotomous choice questions as discussed by Haab and McConnell (2001) that shall be used to estimate WTP both for seasonal forecasts and drought insurance by smallholders.

A double-bounded dichotomous choice model is constructed as follows; first, respondents are presented with an initial bid price. Second, based on the initial response the respondent is presented with a new bid price which can be lower if the initial response was a ‘no’ or higher if the response was ‘yes’. The double-bounded choice model, unlike the single choice model, increases efficiency mainly because: (i) the answer sequences of ‘yes-no’ or ‘no-yes’ places clear bounds on WTP, (ii) the ‘no-no’ and ‘yes-yes’ pairs result in gains in efficiency, and (iii) by increasing the number of responses it means more observations and more degrees of freedom are used to fit a given function (Haab et al, Hanemann et al., 1991). Following Haab and McConnell11 (2001) the indirect utility function for respondent j can be specified as: = Vij Vi (m j, h j,εij) (1)

Where i=1 denotes the final state or condition that prevails when the contingent valuation (CV) program is implemented and i=0 denotes the current status quo. The determinants of the utility

th function Vij are: mj that represents j respondent’s discretionary income; hj represents the k- dimensional vector of household characteristics and choice attributes (e.g., age, education,

cultivar, planting date, farming experience, input costs, etc), and εij is a random unobserved component of the utility function unknown to the researcher. In our case i=0 could be production occurring without incorporating improved seasonal forecasts while i=1 indicates production with improved forecasts. Equally, we could treat i=0 as production without drought insurance and i=1 as production where the farmer purchased insurance. For respondent j who answers yes to the

CV discrete choice question, eliciting a payment of $tj his utility function incorporating WTP say for improved forecasts exceeds the utility function without forecasts − > V1(m j t j , h j,ε1j ) V0(m j, h j,ε0 j ) (2) which can also be expressed as: − > V1[m j WTP(t j ), h j,ε1j ] V0[m j, h j,ε0 j ] (3)

11 For details see Haab and McConnel’s book Chapter 5.

33 Now the general form of the double-bounded model can be formulated as follows: let t1 be the first bid price and t2 be second. Then the bounds on WTP can be constructed as: ≤ < t1 WTP t2 For yes - no responses

> ≥ t1 WTP t2 For no - yes responses (4) ≥ WTP t2 For yes - yes responses

< WTP t2 For no - no responses The general econometric model for the double-bounded CVM is: = μ + ε WTPij i ij (5)

th where WTPij represents j respondent’s willingness to pay; i=1,2 indicates the first and second μ μ answers; and 1 and 2 are the means of the first and second responses. If we treat these means as a function of individual covariates thus: μ = β = β +ε i hij ⇒WTPij hij ij (6)

This general model accommodates the jth respondent to first and second responses to the posed CV question. To construct the likelihood function we first need to derive the probability of observing each of the two-bid response sequence (yes-yes, yes-no, no-yes, no-no). The probability that respondent j answers yes-no combination is: = ≥ < Pr[yes, no] Pr[(WTP1j t1),(WTP2 j t2)] (7) = μ +ε ≥ μ +ε < Pr[( 1 1j t1),( 2 2 j t2)] Likewise the probabilities for the other combinations can be constructed in an analogous fashion. Hence the jth respondent’s contribution to the likelihood function is

μ = μ +ε ≥ μ +ε < YN μ +ε > μ +ε ≥ YY L j[ | t] Pr[( 1 1j t1),( 2 2 j t2)] * Pr[( 1 1j t1),( 2 2 j t2)] * μ +ε < μ +ε < NN μ +ε < μ +ε > NY Pr[( 1 1j t1),( 2 2 j t2)] *Pr[( 1 1j t1),( 2 2 j t2)]

(8) where YY=1 for yes-yes answer, 0 otherwise, NY=1 for no-yes answer, 0 otherwise. This formulation is called the bivariate discrete choice model. If the errors are assumed to be normally σ 2 σ 2 distributed with 0 mean and respective variances of 1 and 2 , then the WTP1j and WTP2j also

34 μ μ σ 2 σ 2 have a bivariate normal distribution with means 1 and 2 , variances 1 and 2 and correlation coefficient ρ 12. Given the dichotomous choice responses to each question, the normally distributed bivariate model is referred to as the bivariate probit model. The likelihood function for the bivariate probit model is derived as follows: The probability of yes-no response is: t − μ t − μ Pr[(μ + ε ≥ t ),(μ + ε < t )] = Φ [( 1 1),( 2 2 ),−ρ] 1 1j 1 2 2 j 2 ε ε σ σ (9) 1 2 1 2 where Φ ε ε (.) is the standardized bivariate normal cumulative distribution function with zero 1 2 means, unit variances and correlation coefficient ρ . Similarly, the probability of the no-no response is t − μ t − μ μ +ε < μ +ε < = Φ 1 1 2 2 ρ Pr[( 1 1j t1),( 2 2 j t2)] ε ε [( σ ),( σ ), ] (10) 1 2 1 2 The probability of the yes-yes response is − μ − μ t1 1 t2 2 Pr[(μ + ε ≥ t ),(μ + ε ≥ t )] = Φ ε ε [( ),( ), ρ] (11) 1 1 j 1 2 2 j 2 1 2 σ σ 1 2 The probability of the no-yes response is t − μ t − μ μ +ε < μ +ε ≥ = Φ 1 1 2 2 −ρ Pr[( 1 1j t1),( 2 2 j t2)] ε ε [( σ ),( σ ), ] (12) 1 2 1 2 Thus the jth contribution to the bivariate probit likelihood function becomes t − μ t − μ μ = Φ 1 1 2 2 ρ L j[ 1 | t] ε ε [d1j ( σ ), d2 j ( σ ), d1jd2 j ] (13) 1 2 1 2 = − = − where d1 j 2y1 j 1 and d 2 j 2y2 j 1 , for y1j=1 if the response to the first question is yes,

and 0 otherwise , y2j=1 if the response to the second question is yes and 0 otherwise. This bivariate probit model is a general parametric model relating to the two-response survey. Once the parameters are estimated WTP as defined in equation (5) can be easily computed.

3.2 Adoption of Improved Seasonal forecasts To elicit the farmer’s WTP for improved seasonal forecasts, we looked at a hypothetical situation where the forecasts are made available to farmers in a specific area or region but at a cost.

2 2 ρ = σ / σ + σ 12 12 1 2 σ = covariance between errors for the two WTP functions 12

35 Smallholders’ WTP for seasonal forecasts is assessed via surveys (see section 3.5) based on CVM. Currently El-Niño-based seasonal forecasts are being disseminated in Zimbabwe, albeit at a small scale. A drawback with these forecasts is that they are not tailored by specific locality13 and this may decrease the user’s utility for the forecasts. This study examines a policy that calls for a wide-scale distribution of seasonal forecasts tailored by specific locality and that takes into consideration not only El -Niño effects, but also other locally prevailing weather variables, such as soil moisture, temperature and elevation. But to call for such a policy initiative, we need to assess the farmers’ WTP for the provision of improved seasonal forecasts for mitigation purposes. The presumption is that farmers attach greater value to improved region-specific forecasts compared to the current broad-based forecasts. Hence region-specific forecasts are likely to result in efficient mitigation.

3.3 Characterizing Demand for Drought Insurance Crop insurance markets for smallholder farmers in Zimbabwe do not exist. The absence of formal insurance markets that could offer efficient risk sharing alternatives for smallholders amounts to market failure This market failure persists despite smallholder farmers generally showing strong demands for insurance services (Binswanger, 1986). Therefore, to characterize demand for drought insurance under non-existence of formal markets we used non-market contingent valuation method discussed earlier. Using CVM farmers are presented with a hypothetical insurance market and are asked to respond to questions that can be used to recover measures of WTP (see appended questionnaire, A.3). By deriving the hypothetical demand for drought insurance, we gain more insights on how the demand is affected by provision of free food-aid and mitigation cost r. Assuming a rational, utility-maximizing smallholder farmer, our interest is look at the two scenarios: (i) demand for drought insurance when formal mitigation is available via the adoption of improved seasonal forecasts at cost r, and (ii) demand for drought insurance when smallholders correctly anticipate ex-post government disaster food relief.

3.3.1 Impact of Mitigation costs, r on Demand for Insurance Let’s assume a risk-averse farmer with initial wealth W, faces loss of L with probability, Φ . Assume a monopolist insurer charges a premium rate of π per dollar of coverage and the decision to buy insurance is made well before seasonal forecasts are made available. Further, assume the farmer decides to use location-specific seasonal forecasts for drought mitigation, but at cost r. In

13 Seasonal forecasts could be based on the district level

36 addition, assume forecasts are credible enough such that for farmers who use the forecasts, they will reduce losses with appreciable certainty. Let H(r) be the maximum possible loss, if farmer spends r ex-ante. If the farmer does not undertake mitigation then his loss will be L and hence H(0)=L. Specifically assume H ′(r) < 0 implies that as mitigation expenditures increase, the maximum potential loss will decrease and H ′′(r) ≥ 0 .

If we assume the insurer’s objective is to maximize profits, then his maximization problem can be set as: Max N.(π - ρ)α(π )L (14) π where (π - ρ) is the profit of providing one dollar of insurance to the farmer and α(π ) is the coverage level as a function of π demanded by farmers. Taking the first order derivative with respect to π gives the optimal premium rate:

α (π ) π * = ρ − i i α ′ π ( i ( )) α (π ) = ρ + i ⇒ [1 − ρ α ′ π ] (15) ( i ( )) = ρ + 1 ⇒ [1 η ] for η ≠ 0 where η is the elasticity of demand facing the monopolist insurer. The result shown in (15) characterizes price equilibrium charged by the monopolist. Assuming α is a normal14 good, premium charges by the monopolist will depend on η , such that the greater elasticity of demand faced by the monopolist the lower the premium chargesπ and vice-versa. Observe that under perfectly competitive markets, the insurer would charge the actuarially fair price of π = ρ .

To derive the farmer’s maximization problem in the presence of mitigation cost, first let’s derive the optimal condition when there is no mitigation cost. Suppose the farmer’s utility function, U(.) is a von Neumann-Morgenstern utility function, such that U′(.) > 0, U′′(.) < 0 . The farmer’s

demand for insurance is reflected by coverage levelα(π ) , that shows demand as a function of premium rate π. Without mitigation cost, individual farmer i will choose α ∈[0,1] so that:

∂α 14 Since the coverage levelα is assumed a normal good then < 0 ⇒ η < 0 . ∂π

37 Max E[U] = (1-Φ)U(W -απL) + ΦU(W -απL +αL − L) α Taking the first order derivative with respect to α gives the optimal condition:

π *(1− Φ) U ′(W −π *α*L − (1−α*)L) i = i i i (16) Φ −π * ′ −π *α * (1 i ) U (W i i L) The optimal condition derived in (16) shows that with no mitigation or food-aid, the farmer equates the ratio of marginal utility of substitution between the loss and No-loss state to the ratio of the premiums weighted by respective probabilities. Now, let’s consider the case where farmer is incurring mitigation cost r, taking π as given, his utility maximization problem subject to the constraints that α ∈[0,1] and r ≥ 0 becomes: Max E[U] = (1-Φ)U[W -απL - r) + ΦU(W -απL +αL − H (r) − r] α, r FOC:

π * (1− Φ) U ′[W − π * α* L − H (r* ) + α* L − r* ] α : m = m m m m m (17) Φ − π * ′ − π * α* − * (1 m ) U [W m mL rm] ∂E[U] r: = −(1- Φ)U′[W -απL - r] - Φ(H ′(r) +1)U′[W -απL + αL − H (r) − r] ∂r

∂E[U (r* ,α * )] Applying Kuhn-Tucker conditions, r * ≥ 0 and by assumption m m ≤ 0 hence m ∂ * rm ′ * * * (1− Φ) U [W − π α L − r ] H ′(r* ) +1= m m m (18) m Φ ′ − π * α* − * + α* − * U [W m mL H(rm ) mL rm] Combining equations (17) and (18) gives: 1 − 1 H ′(r* ) = − ⇒ r* = H ′ 1(− ) (19) m π * m π * m m Equation (17) gives the optimal condition for the demand for insurance in the presence of mitigation costs. Observe that the optimality condition looks similar as obtained in (16) except that marginal utility function under both loss and no-loss states is influenced by additional variables, H(r) and r.

For the result derived in (19), observe the negative inverse relationship between r and π , which implies as the premium charges increase, farmers are likely to spend more on mitigation, but for a

38 ∂r given seasonal information set Ω , > 0 farmers are likely to mitigate more with an increase ∂π Ω in premiums. This result is expected, since an increase in π makes insurance more expensive and hence providing incentives for farmers to increase investment in mitigation, ceteris paribus.

min Since H(0)=0 we can place a lower bound π such that for a given seasonal forecasts m Ω information set:

⎡ − ⎢ 1 H ′ 1 − = 0 ⇒ r* = 0 ⎢ min m ⎢ π ⎣ m Ω

This means there is a premium range for a given forecast information set Ω such that * [π < π min ] = 0 and the optimal response for the farmer is to set r =0. As argued by Kelly m Ω m etal (2003), expenditures on mitigation will be zero for events of low probability. For our case, farmers may choose not to invest in drought mitigation, if the forecasts predict a good year. Thus the optimal amount spent on mitigation is:

⎡ − ⎤ * = ′−1⎢ 1 ⎥ r max(0, H ⎢ ⎥) m π * ⎣⎢ m ⎦⎥

3.3.2 Impact of Food-aid on Demand for Insurance This section looks at the impact of food-aid, if farmers correctly anticipate it. Assume that after a loss occurs, government pays a fraction (1−δ ) of the uninsured losses. This means the government pays (1−δ )(H(r) −αL)) of uninsured losses. The maximization problem of the insurer remains the same as in equation (15). Because the farmer correctly anticipates food-aid his expected utility becomes

Max E[U] = (1- ρ)U[W -απL - r) + ρU(W -απL +αL − H(r) − r + (1−δ )(H (r) −αL)] α, r (18) FOC:

39 − π * − Φ ′ − π * ′ g (1 ) U (.) g Φ U (.) α : = L(1− δ ) ls ⇒ = L(1− δ ) ls Φ − π * U ′ − π * 1− Φ U ′ (.) (1 g ) nl 1 g nl ′ where: U ls indicates the ‘loss-state’ marginal utility function with the same arguments shown in ′ (18) and U nl indicates ‘no-loss’ state. By further simplification and re-arrangement (see appendix, A.2) ∂ π log( ) < ∂δ 0 (19)

′ ′ − (1− ρ) U (.) 1 ⎡(1− ρ) U (.) ⎤ nl -1= H ′(r)δ ⇒ − ⎢ nl +1 ⎥ = H ′(r) r : ρ U ′ (.) δ ⎢ ρ U ′ (.) ⎥ ls ⎣ ls ⎦

′ 1 − ⎡(1− ρ) U (.) ⎤ But by assumption H ′(r) < 0 ⇒ H ′ 1⎢ nl +1⎥ = r δ ⎢ ρ U ′ (.) ⎥ ⎣ ls ⎦ Hence ∂ r < ∂δ 0 (20) Equations (19) and (20) underlie important implications: the result in (19) implies that the presence of food-aid is likely to reduce premium rates charged by monopolist insurers, and results in (20) implies that a farmer is likely to spend less on drought mitigation, if he correctly anticipates government free food-aid. The latter result shows that food-aid may discourage farmers from pursuing more efficient self-reliant drought-risk mitigation strategies.

3.4 Insurance Supply Function in the Presence of Catastrophic Risk, Food-aid and Reinsurance While many studies have established that smallholder farmers posit strong demands for insurance services (Binswanger, 1986, Sakurai 1997, Zeller and Sharma, 1998), it is the cost of supplying insurance services that has proven difficult. Costs of supplying crop insurance has proven costly partly for three reasons: (i) incentive problems due to asymmetric information that gives rise to problems of moral hazards and adverse selection, (ii) prohibitively high administrative costs, and (iii) catastrophic risk that characterizes crop insurance.

40 Offering crop insurance has proven difficult due to systemic or catastrophic risk. Catastrophic events, such as drought or floods, result in systemic losses correlated across farmers and geographical regions. Such risks cannot be eliminated through risk pooling (Duncan and Myers, 2000) and as a result the insurer has to bear some cost. Exposure to catastrophic risk may be reduced, if the insurer: holds large reserves; diversifies across different types of catastrophic risks by different geographic areas; reinsures with larger well diversified reinsurance firms or uses market-based catastrophic futures to hedge catastrophic risk. More recently some market-based mechanisms for managing catastrophic risk have been suggested, for example, catastrophe bonds and futures (Miranda and Glabuar 1997; Duncan and Myers, 2000). While such market-based mechanisms hold promise, these are difficult to implement in developing countries where financial markets are not only thin, but not well developed.

In this section, the aim is to characterize the long-run equilibrium of insurance supply in the presence of catastrophic risk, food-aid, and subsidized reinsurance using the mean-variance15 (MV) utility approach, suggested by Duncan and Myers (2000). Notice that MV is preferred to the expected-utility framework (used in section 3.3) because the latter does not lead to neat derivation of the desired equilibrium conditions.

3.4.1 Deriving Demand function using MV Utility Approach Suppose the insurance market is characterized by large a number of N individual farmers, each with a potential income, M. Each farmer faces a stochastic loss L with known probability Φ and 0, with probability (1− Φ ). In the event of loss the end-of-season income for each farmer will be (M-L). All farmers face the same marginal probability distribution for L, but the loss for each pair of farmers may be correlated. Assume that the insurer is offering contracts to farmers to insure against their loss L. The contract can be described by (π ,α, n ), where α ∈[0,1] is coverage level,

specified by the farmer; π is insurance premium per unit of coverage level, and n describes the number of contracts held by the insurer. The farmer wants to choose a coverage level α that maximizes her preference or utility function defined on the end-of-season income. Thus the end- of-season net wealth (ω ) for a farmer who purchases insurance is:

15 The mean variance utility level has well known undesirable properties, but nonetheless it is used since it helps illustrate most of the fundamentals being investigated in this study.

41 ω = Μ − πα + α L − H (r) − r + (1 − δ )(H (r) −α L) (21) d d d { d { { { mitigation−cos t 123 1424 434 premium indemnity max−loss fraction paid uninsured loss payments with mitigation as food −aid where the subscript d refers to the demand for insurance. From equation (21), the farmer pays the πα premium d regardless of the state of nature and if drought occurs, he suffers a loss H(r) , but α assuming the farmer adopts seasonal forecasts he pays r as mitigation cost. The amount d L of the loss is reimbursed by insurance as indemnity. In addition, both the farmer and insurer correctly anticipates that in the event of a loss, the government provides free food-aid, which − δ α covers a fraction (1 ) of his uninsured loss (H(r)- d L ). To derive demand for insurance, the mean-variance (MV) utility approach can be used: U = E(M ) − 0.5λVar(M ) 123 1424 434 mean variance where U denotes the utility function, which is increasing in expected net income but decreasing in the variance of income and E denotes the expectations operator. The parameter 0.5 λ represents the degree of the farmer’s absolute risk aversion. From (21), the farmer’s MV utility function can be set-up as: = Μ −πα +α − − + −δ −α − λ Μ − 2 U E[ d d L H(r) r (1 )(H(r) d L)] 0.5 E[( E(M ) ] and this reduces to: = Μ −πα + Φα − Φ − + Φ −δ −α − λ − Φ 2 σ 2 − α ρ σ σ +α 2σ 2 U d d L H(r) r (1 )(H(r) d L) 0.5 (1 ) [ H 2 d H,L H L d L] (22) From (22) take the first derivative with respect to: ∂U = −π + ΦL − (1−δ )ΦL − 0.5λ(1− Φ)2δ 2[−2ρσ σ + 2α σ 2 ] α ∂α H L d L d : d −π + Φ − −δ Φ − λ − Φ 2δ 2 − ρσ σ + α σ 2 = ⇒ L (1 ) L 0.5 (1 ) [ 2 H L 2 d L ] 0 1 1 1 1 ⇒ α = {ρσ σ − [π − ΦL + (1−δ )ΦL]} (23) d σ 2 H L λ − Φ 2 δ 2 L (1 )

ρ Assume var[H(r)] and var(L) are the same and H,L =1 (perfect correlation), then equation (23) simplifies to: 1 1 1 ⇒ α =1− [π − ΦL + (1−δ )ΦL] (24) d λσ 2 − Φ 2 δ 2 L (1 )

42 Equation (24) represents the demand for drought insurance by smallholders in the presence of both seasonal forecasts and government food-aid. From (24) the following comparative statics are derived:

∂α 1 1 ∂α d = −.− [−ΦL] ⇒ d < 0 ∂δ {δ 25(a) + (.)12−3 ∂σ 2 {+ {+ L

∂α 1 1 ∂α d = −.− 2 [π > ΦδL] ⇒ d > 0 25(b) 123σ ∂σ 2 + L (.) 142+ 43 ∂σ 2 L {+ {+ L

∂α 1 ∂α d = −. [−Φ(1+δ )] ⇒ d > 0 ∂ { ∂ 25(c) L − (.)142− 43 L {+ Results in (25) have several important implications: Equation 25(a) shows that the demand for α δ insurance, d , decreases as food-aid, (1- ) increases. Hence, as inferred earlier, the presence of free food-aid is likely to reduce the demand for insurance. This result agrees with what we observed in (20). Using the MV utility approach, the mitigation cost, r does not directly enter the demand function (23) or (24), but rather indirectly through the mitigation loss variance term σ 2 σ 2 = σ 2 H (r) . Since it was assumed that H (r) L are the same, then 25(b) indicates that increase in

σ 2 loss variance, L will increase the demand for insurance, while 25(c) shows that an increase in expected loss L will increase the demand for insurance. To summarize our results, the demand for insurance is reduced in the presence of government food-aid, while mitigation cost r does not directly affect insurance under the MV the approach, but rather indirectly via var[H(r)] , which shows that as expected loss variance increases the demand for insurance increases as well.

3.4.2 Deriving Short-run Supply function using MV Utility Approach The next task is to derive the short-run insurance supply function in the presence of catastrophic risk, food-aid and subsidized reinsurance following Duncan and Myer’s model. Suppose a single monopoly firm that has access to reinsurance, supplies drought insurance. Assume crop losses are correlated across crop and geographic regions resulting in catastrophic risk. Due to catastrophic risk, it becomes rational to think of the insuring firm as being risk averse rather

43 than risk neutral. Suppose the end-of-season profit for selling insurance to N farmers and reinsuring a proportion φ of its policies is given by

n n n ω = N . α . (1 − φ) . (π − c) − α L + α φ L + α γ L s { s s ∑ i s ∑ i s ∑ i (26) farmers { 123 12−3 i=1 i=1 i=1 coverage reinsured profit level 142 43 142 43 142 43 level proportion total indemnity amount received subsidy received as reinsurance from the government

Equation (26) can be simplified as: n ω = α −φ π − −α −φ −γ s n s (1 )( c) s (1 ) ∑ Li (27) i=1 Where the subscript s refers to the supply of insurance, π is the premium charged per unit of coverage (as before), and c is the insurance costs per unit of coverage. The insurance firm remits a proportion, φ (forφ ∈[0,1]) of its premium to a reinsurer. In return, the reinsurer has the responsibility to pay the primary insurer (monopolist firm) some proportion (φ +γ ) of the indemnities. The parameter γ represents subsidy paid by the reinsurer to the primary insurer

and 0 ≤ γ < (1−φ) . If γ > 0 , then the reinsurer covers a higher proportion of indemnities than it receives in the form of premium and hence implying a subsidy. On the other hand, if γ = 0 , then there is no subsidy. Assume the government is the reinsurer and the values φ and γ are set exogenously by the government. The insurer offers a coverage level α s to farmers at premium level π (as before). As in the case of farmers, assume the risk-averse insurance firm has a linear MV utility function with risk parameter 0.5ψ 16. Therefore, using equation (27) the mean and variance of the firm’s profit is computed as ~ = α −φ π − Φ − +γΦ − ψ α 2σ 2 −φ −γ 2 + − ρ V n s[(1 )( L c) L] 0.5 n s L (1 ) .[1 (n 1) ] (28) where ρ is the correlation coefficient between losses of any two farmers and is defined as:

Cov(Li , L j ) ρ = (29) var(Li ) var(L j )

For purposes of exposition, assume ρ is the same for every pair of farmers and let’s restrict attention only to positive values of ρ so as focus on catastrophic risk. The correlation parameter ρ measures the degree of catastrophic risk, in the sense that as ρ increases, it increases loss between all of the individuals in the market, resulting in a higher portfolio risk for the primary insurer.

44 The firm offers a coverage level α s that maximizes equation (28) for a given premiumπ and n number of policies. The first order condition with respect to α s :

α [(1−φ)(π − ΦL − c) +γΦL]−ψnα σ 2 (1−φ −γ )2.[1+ (n −1)ρ] = 0 s : s L 1 1 1 1 ⇒ α s = [(1− Φ)(π − ΦL − c) +γΦL] (30) ψ σ 2 −φ −γ 2 [1+ (n −1)ρ] L (1 ) Equation (30) gives the short-run supply of drought insurance coverage in the presence of catastrophic risk. From the supply function we can obtain the following comparative statics: ∂α n −1 1 ∂α s = − [π > ΦL + c] ⇒ s < 0 31(a) ∂ρ + − ρ 2 (.)1424 434 ∂ρ [1 (n 1) ] { + 1424+ 434 +

∂α −.− 2 1 ∂α s = [π > ΦL + c] ⇒ s > 0 31(b) ∂φ −φ −γ 3 (.) 1424 434 ∂φ (1 ) { + 1424+ 434 +

∂α −.− 2 1 ∂α s = [(ΦL] ⇒ s > 0 31(c) ∂γ −φ −γ 3 (.) 123 ∂γ (1 ) { + 1424+ 434 +

∂α −.ρ 1 ∂α s = [ω > ΦL + c] ⇒ s < 0 31(d) ∂n + − ρ 2 (.) 1424 434 ∂n [1 (n 1) ] { + 1424− 434 +

Results in 31(a)-(d) imply the following: 31(a) shows that the short-run insurance supply function decreases with increase in catastrophic risk ρ , this result confirms the conventional wisdom that catastrophic risk reduces supply of insurance; 31(b) indicates that short-run insurance supply increases with increase in reinsured proportion, φ , this result may help emphasize that the reinsurance market are preconditions for supply of crop insurance in the catastrophic risk; 31(d) shows that an increase in government subsidy γ increases the supply of insurance, and finally 31(d) shows that the supply of insurance decreases with an increase in the number of the insured farmers, n. The latter result underpins an important fact that due to catastrophic risk, risk-pooling becomes ineffective and this contrasts the non-catastrophic risk insurance markets, such as automobile where risks are nearly stochastically independent and

16 Following Duncan et al (2000), assume that the insurance firm by virtue of it being larger and more diversified than farmers, it is less risk responsive than the farmers.

45 risk pooling is more effective as the number of insured n increases. Using the results derived in sections 3.4.1 and 3.4.2, we characterize long-run equilibrium in the presence of catastrophic risk, food-aid and subsidized reinsurance.

3.4.3 Characterizing the Existence of the Long-run Equilibrium Assume the long-run monopolistic competitive equilibrium is characterized by zero profits that result from to entry or exit of contesting firms. Specifically, assume the long-run reservation ~ utility level b, such that the long-run insurance supply function V equals the reservation utility ~ ~ level and hence the long-run equilibrium isV = b . By combining results from (24) and (30) a monopolistic competitive equilibrium of an insurance model under catastrophic risk, food-aid and subsidized reinsurance is characterized by a premium level π e , coverage level α e and number of policies ne that satisfy: 1 1 1 α e − ρ + [π − ΦL + (1−δ )ΦL] = 0 (32) λσ 2 − Φ 2 δ 2 L (1 ) 1 1 1 1 α e − [(1− Φ)(π e − ΦL − c) +γΦL] = 0 (33) ψ σ 2 −φ −γ 2 [1+ (ne −1)ρ] L (1 ) 1 1 1 1 ~ ne (α e )2 − neα e [(1− Φ)(π e − ΦL − c) +γΦL]−b = 0 (34) ψ σ 2 −φ −γ 2 [1+ (ne −1)ρ] L (1 )

Equation (32) is the demand for drought insurance by an individual farmer; equation (33) defines the short-run supply of drought insurance by the insuring firm and equation (34) defines the ~ long-run equilibrium condition, such that the utility of each firm (V ) equals the reservation ~ utility level b . By solving (32)-(34) simultaneously the long-run equilibrium is determined and

e characterized by premiumπ e , coverage level α and number of policies ne sold by the firm. Further, by solving for π e in (32) and substituting it in (33) gives:

(1− Φ)[(1− Φ)2λσ 2δ + LΦ(δ −1) − c]+γΦL α e = L (35) ψ −φ −γ 2 + e − ρ σ 2 [ (1 ) [1 (n 1) ] L

Equation (35) gives the equilibrium coverage level α e for a given ne that equates the demand and short-run supply functions. A graphic analysis helps to characterize the equilibrium under a

46 monopolistic firm. Assume φ =γ =0, meaning there is no reinsurance and subsidized reinsurance. As shown in Figure 3.1, demand for insurance (24) is graphed in the northeast quadrant, with premiumπ being linearly and negatively related to coverage levelα . From (35), as n → ∞ thenα → 0 and as n → 0 , α → k (constant). This equilibrium relationship between α and n is graphed in the southeast quadrant. The southwest quadrant shows the relationship between b ~ ~ and n which can be derived from the long-run relationship, V = b . From (32) substitute for π into (28) and similarly from (33) substitute for α into (28); this gives equilibrium as the ~ θ θ = σ 2 ρ λ ψ γ φ δ function denoted as V (n; ) where (c, L , , , , , , ) is a vector of parameters. This equilibrium is graphed as function of n in the southwest quadrant in Figure 3.1. From (36), ~ ~ observe that if n=0, then V (0;θ) = 0 and that V (n;θ) is a monotone function in n. Assuming no ~ reinsurance (φ =γ =0), as n → ∞ then V (n;θ) asymptotically converges to a limiting value that ~ equates to b as shown in (37).

2 2 2 ~ 0.5n{(1− Φ)[(1− Φ) λσ δ + ΦL(δ −1) − c]} ~ V (n;θ) = L = b (36) σ 2ψ + − ρ L [1 (n 1) ]

2 2 2 ~ {(1− Φ)[(1− Φ) λσ δ + ΦL(δ −1)]} ~ lim V = L = b (37) n→∞ σ 2ψρ 2 L

Using results in (24)-(37) we analyze the long-run monopolistic competitive equilibrium in crop insurance without reinsurance or subsidy (the stand-alone position). From Figure 3.1, assuming the insurer is a monopoly firm, be is the long-run equilibrium reservation level, ne the number of contracts and π e equilibrium coverage level. Observe that equilibrium ( (π 0,b0, n0) denotes competitive equilibrium. The monopoly firm offers a lower coverage α e (< α 0) level, charges a higher premium π e (>π 0 ), has higher reservation preference level be (> b0 ) and offers more contracts ne (> n0 ) . In a market characterized by large insurance firms wielding greater monopoly power, it is most likely that the insurer will demand a higher reservation preference ~ level b m , such that b m > b and under a situation, insurance markets fail to exist. Another ~ important result is that as long as b m > b regardless of the number of contracts n, the firm may

47 sell, and equilibrium will not exist. As observed by Duncan and Myer (2000), even if the number of contracting farmers goes to infinity, this can not induce firm(s) to participate in such a market because the expected profits are not large enough to compensate for an increase in risk. From ~ Figure 3.1 equilibrium breaks down if and only if the reservation preference level exceeds b and irregardless of n, equilibrium will not exist.

To analyze the impact of food-aid δ on long-run equilibrium, from (37) observe the comparative ∂ ~ θ V (n, ) > static ∂δ 0 (see appendix). This result implies that increase in food-aid does not support ~ equilibrium and by raising the reservation preference level beyond b leads to a breakdown of the equilibrium (Figure.3.1). To analyze the effect of subsidized reinsurance (γ >0) from (37), ~ ∂V (n,θ) < 0 observe the comparative static ∂γ (see appendix). This implies that subsidized ~ reinsurance reduces reservation preference level b0 < b m < b and hence supports long-run equilibrium. Finally, to characterize long-run equilibrium under catastrophic risk ρ , from (37) ~ ∂V (n,θ) < 0 observe ∂ρ meaning an increase in catastrophic risk reduces the reservation preference ~ level and by lowering it below b or b m thus supporting long-run equilibrium. Hence the lowering of catastrophic risk will help to sustain the long-run equilibrium.

3.5 Designing an Index-based Indemnity Function Area-yield insurance based on remotely sensed vegetation indices such as VCI has the potential to offer more effective yield-loss coverage than the traditional crop insurance based on individual yield-loss performance. The latter has not only proven costly but also bedeviled by the problems of moral hazards and adverse selection. Because smallholder farmers are widely scattered monitoring individual yield performance and loss assessment could be cumbersome and costly. Rather, with recent advances in remote sensing technology which come with high spatial and temporal resolution, crop yield and loss performance can be adequately monitored and predicted with reasonable accuracy at low cost.

48 π

π e π 0

~ b be b0 D m

α e α 0 ~ α V

Figure.3.1 Characterizing the existence of long-run equilibrium under catastrophic risk.

(π e,be, ne ) denotes monopolistic long-run equilibrium, (ω0,b0, n0) denotes competitive long-run equilibrium Source: Adapted from Duncan and Myers, (2000)

49 This section designs an index-based indemnity function. But, when designing an index-based insurance contract, we need to decide on three important questions: (i) what variable(s) to use as an index(ces) for measuring the variable of interest (i.e., area-yield losses); (ii) how correlated is the index with our variable of interest and basis risk, and (iii) how to structure the indemnity schedule?

3.5.1 What variable to use as an Index? In our case, an ideal index is one that is highly correlated with area-based crop yield and hence able to accurately track yield losses. In addition, an ideal index is one that is easy to observe and measure on a regular basis and not subject to manipulation by an individual farmer or insurer. Indices, such as satellite-derivatives (NDVI, VCI, TCI, etc), and weather-related derivatives (temperature, rainfall, soil moisture, etc) tend to meet the ideal index criterion defined above.

As discussed in detail in Chapter 2, NDVI measures vegetation vigor caused by chlorophyll activity and as a result it is highly correlated with photosynthetic activity in wilting and non- wilted plant foliage. Thus, NDVI is a good predictor of plant canopy biomass, vigor or stress (Tucker 1979). VCI as well indicates the vigor of vegetation cover. VCI is computed as a function of NDVI and it normalizes NDVI according to its variability over many years. Hence compared with NDVI, VCI is a consistent index for assessing crop condition and potential yield (Vogt et al, 2000). TCI is based on the fact that the brightness temperature of vegetation canopy rises with increasing water stress and hence it is an appropriate measure for tracking drought conditions. On the basis of these vegetation indices, sequential crop profiles can be developed during the growing season that show the progression of canopy emergence, maturation and senescence. Using these crop profiles, the crop yield performance can be assessed with considerable accuracy.

In this study two indices are used- rainfall (a meteorological variable) and VCI (a remotely sensed variable). A rainfall index can be used to write an insurance contract that effectively protects against crop losses due to drought or floods. However a disadvantage with a meteorological-based index in general is that the data are derived from a sparse network of weather stations with a poor spatial representation. This fails to measure spatial distribution relating to the variable of interest (e.g., crop losses, drought risk). On the other hand, satellite- based indices come with high spatial and temporal resolution essential for dynamic monitoring

50 of agronomic conditions and crop yield assessment. However, most satellite variables such as NDVI and VCI, are difficult to interpret in heterogeneous terrain (Vogt and Miemeyer, 1998).

3.5.2 How correlated is the Index with Crop yield losses? Figure 3.2 (a) shows the different phenological growth stages of rice during a growing season. Figure 3.2(b) shows how the pattern of the Pearson correlation coefficient between reflectance + − in channels 5 ( Ch5 ) and 7 ( Ch7 ) and ratio ( (ch7 ch5 ) (ch7 ch5 ) versus different growth stages of rice evolves over time. It is observed that the early stages of development (0-8 days) rice is negatively correlated with channels 5/7 or ratio. Correlation gradually increases as crops grow (8-13) and attains a maximum during flowering stage (20-30 days). However, with maturation and senesce setting in after 30 days, correlation decreases.

Some important implications can be drawn: (i) highest correlation tends to occur during critical yield determining stages such as flowering. Thus, by induction it should be the case that final yield is correlated with reflectance data (channels 5/7 and/or ratio). Hence high yield should be associated with high reflectance values and vice versa. (ii) By examining seasonal crop growth profiles over many seasons, one can identify critical times in crop-growth cycles essential to understanding the relationship between vegetation indices (NDVI, VCI/TCI) and crop yields. Such knowledge is necessary for developing yield distribution and loss prediction.

51 (a) Crop Growth Phases boothing End of tillering + Boothing Earing Flowering of ripening stages Initial earing

0 8 13 22 30 43 50 Growing days b) Correlation at Various Growth Stages

Ch5

Ch7

Correlation

Figure 3.2: Evolution of Correlation coefficient between channels 5 ( Ch5 ), and 7 ( Ch7 ) versus different growth stages of Rice Source: Adapted from Berg (1981).

52 3.5.3 Designing and Pricing an Index-based Contract This section presents a methodology for designing and pricing an index insurance contract for crop production. The general setup follows one suggested by Vedenov and Miranda (2001). In Chapter 5, we adapt this setup to compute empirical premium rates for insurance contract using two indices, rainfall and VCI.

Suppose an elementary contract pays an indemnity f (τ ) conditional on realization of the index τ according to the following schedule ⎧x if [τ ≤ λi* ]...... (i) ⎪ ⎪ i* −τ f (τ ) = x if [λi* ≤ τ < i* ]...... (ii) (38) ⎨ − λ * ⎪ (1 )i ⎪ * ⎩0 if [τ > i ]...... (iii) This means that the contract pays, whenever the index τ falls below a pre-defined trigger i*.. The three cases illustrate that for case: (i), maximum indemnity $x is paid since the index fall below

critical value λi* . For case (ii), an indemnity proportional to the difference between the index and trigger is paid for λi* ≤ τ < i* , and for case (iii) no indemnity is paid, if the realized index τ is above the trigger i*. The parameter 0 ≤ λ < 1is interpreted as the probability of ruin17 and signifies a catastrophe if value is closer to 1 and non-catastrophe otherwise. For extreme catastrophe, λ = 1 and the contract pays the maximum indemnity, if the index falls below the trigger level i* , but pays nothing otherwise.

To specify a particular contract we need to determine contract parameters [i* ,λ, x]. To ease our analysis it is necessary to standardize the contracts under consideration by stating that all contracts have an expected indemnity or pure premium of $1. This normalization could be convenient from a buyer’s perspective, since it allows him to view the amount of protection he can buy for $1 in pure premium. With this normalization, parameters for the standard contract must be chosen so that:

* ⎛ λii**−τ ⎞ = τ * λ = τ τ + i τ τ 1 Eτ fi ( ;i , , x) x⎜ h( )d h( )d ⎟ (39) ⎜ ∫∫0 λi* * ⎟ ⎝ (1− λ)i ⎠

53 where h(τ ) is the probability density function of the index τ . Once a premium and largest possible liability are known, a premium rate can be computed as a ratio of the two. In this case, as discussed above, the standard contract has a pure premium of $1 and maximum liability of $x, hence the premium rate becomes 1 π = .(40) x

Using equations (39) and (40), we can uniquely identify contract parameters [i* ,λ and x].

Next is to establish the relationship between index τ and farmer’s income. Assume the farmer’s income ω is subject to variations in index-related risk and can be expressed in the form of a stochastic relationship: ω = g(τ;ε ) (41) where ε represents the basis risk. Assume further that the farmer has target income, ω * that he is interested in protecting. This target income could be say 70% of expected income, 0.70ω , ~ where ω = Eg(τ;ε ) . Income below the target level is considered a loss L , hence ~ L = max[0,ω * − g(τ;ε )]

An additional parameter we need to determine is N, the number of standard contracts an individual farmer would purchase. Suppose an individual farmer buys N standard contracts, using equations (39) and (40), we can define a total loss function with N contracts as ~ L = max{0,ω * −[g(τ;ε ) + Nf (τ;π ,λ) − N]}.

Thus an optimal contract is defined by parameters [π * ,λ* , N * ] and the task boils down to how to determine these parameters. Vedenov and Miranda (2001) use a numerical method to solve an optimization problem:

1 ∞ 2 min E⎜⎛ [max{0,ω * −[g(τ;ε ) + Nf (τ;π ,λ) − N]}]2 h(τ )dτ ⎟⎞ (42) {N ,π ,λ} ε ⎝ ∫0 ⎠

In sum, the process of designing and pricing of a standard index contract for a representative farmer involves: (i) determining the distribution of the underlying index τ ; (ii) finding the

17 λ can be trivially computed by setting λi* = τ

54 relationship between the index and buyer’s risk exposure L, and (iii) finally solving for optimal number N and contract parameters via (42) or other means.

3.6 Data Requirements In addition to the survey data (discussed in Chapter 4) the following data were collected: (i) historical crop yield data by district (iii) remotely sensed data from NOAA’s AVHRR obtained from SADC remote sensing unit based in Harare. District-resident Agricultural Research and Extension officers who collaborate with Central Statistical Office (CSO) collect official crop yield data in Zimbabwe. These crop data are specifically used as input for national production forecasting purposes. The data are disaggregated by and districts and hence ideal for our purpose. In this study, crops of interests are maize and cotton. Maize is the main staple grown by over 80% of smallholders and accounts for about 60% of all cultivated land each year. Cotton is the main cash crop predominantly grown by smallholder farmers and approximately it accounts for about 18% of cultivated land each year. Inter-annual crop yield variation in smallholder sector is directly related to the amount of rainfall received per season since all production is rain- fed.

Climatic data primarily needed are 10-day cumulative rainfall obtainable from Drought Monitoring Center based in Harare. These data are available by different weather stations dotted throughout the country. Other ancillary data collected include SST-anomalies denoted as nino34. These data measure the influence of El-Nino and were obtained online from IRI (International Research Institute) website.

NDVI data were obtained from SADC remote sensing unit based in Harare. The data are received at dekadal intervals and geo-referenced to Zimbabwe’s 56 districts. In turn we use NDVI to compute weekly VCI using the formula shown in Table 2.2b. The VCI values were computed pixel by pixel for each week and then averaged to determine a district value. This mean value was taken to represent a designated crop area upon which insurance contract is based. Because the satellite data comes geo-referenced, geometrically and radiometrically corrected these are ideally suitable for the purpose of this study.

In sum, these data are used to estimate various empirical models described in chapters 4 and 5.

55 CHAPTER 4

EXPLORATORY DATA ANALYSIS

This chapter outlines the data collection methodologies, looks at descriptive statistics and explores the correlation between key variables, such as VCI, Niño34 versus crop yield (maize and cotton). The last part of the chapter looks at kernel density distribution of vegetative condition index variable and seasonal rainfall.

4.1 Data Collection Methods Given the objectives and hypothesis of the study as discussed in Chapter 1, field surveys based on the contingent valuation method (CVM) were conducted in Zimbabwe for the period of December 2003 to May 2004. The surveys covered nine identified sites located throughout Zimbabwe’s agro-ecological regions II-V (Table 1 and Appendix C).

The main data collection procedures involved two-stage sampling procedure, questionnaire pre- testing, focus group meetings and elicitation of a farmer’s willingness-to-pay. These procedures are briefly discussed below.

4.1.1 Sampling Method A two-stage cluster sampling procedure is used where the first stage involves a random selection of wards18 within an identified district. Three to four wards were selected in each district and a sample frame of farmers for each ward was obtained from the respective ward councilors. The second stage involved choosing a random selection of about 32 to 42 farmers per ward from a sample frame of about 300 households. Samples of about 100 to 125 farmers were selected from

18 Sub-units of a district with approximate size of a county. Often wards are subdivided into Villages

56 each identified district. In total, 1,125 households were drawn from the nine identified districts. A local resident enumerator was appointed to help in administering the questionnaire by holding face-to-face interviews with the farmers.

4.1.2 Focus Group Meetings and Pre-testing the Questionnaire Focus group meetings were held throughout the identified survey sites, except two areas (Chivi and Beitbridge). These meetings had five objectives: First, to explain to farmers and the local authority (councilors, headsmen and chiefs) the goals of the study. Second, in order to elicit meaningful WTP, farmers have to understand in simple terms the WTP concept. Third, using the participatory approach, farmers were engaged in interactive games demonstrating ‘with’ and ‘without’ situations, involving adopting seasonal forecasts and area-yield drought insurance. Here I was cautious to emphasize not only the novelty of these programs, but their disadvantages as well. This approach helped to avoid gross bias as farmers were left in a good position to make independent decisions regarding WTP for the two programs. The meetings were well advertised and oftentimes attracted large crowds that at times exceeded 500 farmers. Fourth, the focus group meetings were also appropriate platforms to obtain price bids essential for eliciting WTP both for seasonal forecasts and drought insurance (see sections 4.1.3 and 4.1.4 below).

Finally, we used the Focus group meetings to pre-test the designed questionnaire before the actual data collection began. Pre-testing the questionnaire was essential to identify and refine vague sections of the questionnaire. In essence, after pre-testing, we were able to correct for inconsistencies, fill informational gaps, improve ambiguous questions and re-word some vague sections. The pre-test was conducted by a group of nine enumerators, each drawn from the identified survey sites. In addition, pre-testing also provided a good opportunity to review each enumerator and to assess his/her enumeration ability. Those not good enough were dropped.

4.1.3 Eliciting Farmers’ WTP for Seasonal forecasts To elicit farmers WTP for seasonal forecasts, a hypothetical market situation is created where a farmer is envisaged receiving, but at a cost, ‘improved’ seasonal forecasts19 scaled down by a specific region. The advantage with improved forecasts is that they are hypothesized to give better prediction than the current broad El Niño-based forecasts. Therefore, farmers preferences for seasonal forecasts can be evaluated using the CVM approach on the presumption that credible, region-specific forecasts are provided. Hence six price bids were proposed that ranged

57 from Z$4,000 to 6,500. The first bid was equated to the price of a locally available newspaper and the other bids are arbitrary selected. These bids were ‘tested’ and modified during the Focus group meetings. The enumerator used a six-faced die to randomly assign one of the bids to each respondent.

4.14 Eliciting Farmers’ WTP for Drought Insurance Of prime concern is whether a drought insurance scheme suggested in this study will be feasible. Feasibility among other things depends on whether farmers posit strong demand for insurance services. In particular, provision of free food-aid, a practice that has been in place for many decades in Zimbabwe, may reduce the demand for insurance services. Hence, farmers’ demand for drought insurance is evaluated using the CVM approach. A farmer’s WTP for insurance services is elicited under two scenarios: first scenario denotes a case where food-aid is not available and second scenario denotes the opposite.

Six price bids were proposed that ranged from Z$4,800 to 17,000. Similar to the case discussed above, the enumerator used a six-faced die to randomly assign one of the bids to each respondent. These bids, determined during the focus group meeting, were based on current maize price charged in different rural areas20. In most rural areas maize price is normally quoted as units per 20-litre bucket. The first bid represents the lowest price, while the last bid represents the highest price of maize charged per 20-litre bucket across the nine sites. Because maize is a staple crop and under safety-first assumption, farmers would want to insure maize crop and cushion their households against transitory food insecurity.

4.2 Descriptive Analysis 4.2.1 Demographic Characteristics of Sampled Households Table 4.1 indicates that smallholder farmers have an average landholding of 6.5 acres per household. During the 2003/04 season land under cultivation averaged 4.8 acres per household. Of the sampled households only 20% received formal agricultural training21. More than 80% of the sampled households treat peasant farming as their main occupation and, in addition, most are resident in their respective areas. The level of literacy among the surveyed farmers was very high,

19 The location-specific forecasts are hereafter referred to as improved seasonal forecasts 20 The underlying assumption is that farmers are solely concerned in buying insurance against major crops, which in this case are maize and cotton. 21 Agricultural training of rural farmers is traditionally offered by the Department of Agricultural Research and Extension Services (AREX).

58 as only 8.9% indicated that they could not read nor write. Illiteracy was highest in Sanyati at 15.2%, followed by Beitbridge at 12.8%.

District N Total Cultivated Full time Resident Proportion of Illiteracy Household Farm area Occupation status (%) female headed level received size in 2003/04 (%) household (%) agricultural acres (acres) Training (%) Peasant Resident farmer

Chiweshe 125 4.8 2.4 86.4 98.4 0.29 5.6 13.6 Gutu 125 6.8 5.0 48.8 96.8 0.36 4.0 35.2 Sanyati 125 8.0 7.0 67.2 98.4 0.27 15.2 31.5 Chivi 125 6.4 4.6 99.2 99.2 0.59 5.6 32.0 Mt Darwin 125 7.4 5.2 90.4 100 0.19 8.8 4.0 Wedza 125 5.9 3.5 78.2 99.2 0.51 4.0 12.0 Hurungwe 125 9.4 7.3 85.6 100 0.45 1.8 4.0 Shamva 125 4.4 3.0 84.8 96 0.32 6.4 40.0 Beitbridge 125 5.2 5.1 96.8 98.4 0.44 12.8 3.2 Overall 6.5 4.8 82.9 98.5 38.0 8.9 19.5

Table 4 1: Demographic Statistics of Sampled Households

4.2.2. Basic Assets Ownership For most households, cattle are the primary source of draft power, while a disc plough is the most basic equipment for tilling the land. Farmers without these basic resources may not be able to engage in meaningful land cultivation. A working radio is another important asset essential for communicating seasonal forecasts, especially to the wide geographically distributed smallholder farmers. With less than 5% of the rural households having access to televisions, a radio remains the main media for communicating seasonal forecasts.

Table 4.2 summarizes ownership of these basic assets across surveyed regions. From the results (column 3) one can infer that about 25% of sampled households experienced draft power shortage during the 2003/04 season. In addition, one can as well infer (by column 4) more than 20% of the sampled households do not own a plough. Draft power shortage was acute in Sanyati, Chiweshe, Wedza and Mt Darwin where most households own on average three heads of cattle or less. For the 2003/04 farming season, a sizeable number of farmers (> 25%) resorted to hiring

59 draft power during 2003/04 season at an average cost of Z$90,300 (US$16.40) per acre. For an average landholding of 4.8 acres (as shown earlier), a typical household without any form of draft power would need to invest Z$430,000 for having his whole land tilled. This indeed could prove too expensive for most households, given average farm income of Z$245,000 (see Table 4.5)

District NR Basic Asset Ownership Hired draft power % owning a working radio No. of Cattle % % Households % hired Hiring cost owned owning owning a disc plough (Z$’000 cattle per acre) Chiweshe II 3 66.4 68.0 32.8 106 52.0 Gutu III 6 86.4 90.4 23.2 190 40.8 Sanyati IV 2 56.8 68.0 30.0 93 36.8 Chivi V 5 83.2 89.6 7.2 55 30.8 MtDarwin IV 3 61.6 65.6 25.6 73 47.2 Wedza III 4 70.4 79.2 26.4 81 26.4 Hurungwe II 7 92.8 99.2 15.2 101 76.8 Shamva II 5 72.8 74.4 19.2 42 67.2 Beitbridge V 12 79.2 98.4 3.2 72 66.4 Overall 5 74.4 78.9 21.1 90.3 50.2

Table 4.2: Household Basic Assets Ownership for the Season 2003/04

4.2.3 Maize and Cotton Yield In this section farmers provided yield estimates for maize and cotton based on historical yield realizations. To overcome the recall problem, farmers were asked to provide best yield realized during the last 5-10 years and conversely the worst yield for the same time period. From Table 4.3 the best maize yield averaged 0.66 t/acre whilst worst yield averaged 0.14t/acre. Compared across regions, the best maize yields were highest in NR II at 1.29 t/acre and lowest at 0.003 t/acre in NR V during worst years.

Cotton yield averages 2.32 bales22/acre during the best season and 0.89 bales/acre during the worst season. Unlike maize, cotton adapts well to semi-arid conditions and hence can thrive

22 On average, 1 bale weighs 200 kg.

60 even in NR IV. In Table 4.3 cotton is the major crop in areas such as Sanyati and Mt Darwin, which both are located in NR IV.

4.2.4 Sources of Farm Income Among crop enterprises, maize and cotton are the main source of income for most households except Beitbridge (Table 4.4). Cotton is the main source of income in areas such as Sanyati (44%), Mt Dawrin (66.4%) and Hurungwe (95.2%). Maize on the other hand is the main income of source in Hurungwe (99.2%), Gutu (54.4%), Chiweshe (30.4%) and Shamva (34.4%). For livestock, cattle and goats are the main sources of income as they overly contribute 27% and 18.8%, respectively, to household income. These enterprises are the main source of income for especially arid regions NR V e.g. Beitbridge and Chivi. Under other enterprises (last 3 columns in Table 4.4), selling of vegetables is also another important source of source.

District Farming BEST SEASON WORST SEASON BEST SEASON WORST SEASON experience Maize yield Maize yield Cotton yield Cotton yield

(#years) Ave.Yld Harvest Ave.Yld Harvest Bales Harvest Bales Harvest (t/acre) area (t/acre) area (kg/acre) area (kg/acre) area (acres) (acres) (acres) (acres) 1 12 1.29 4.2 0.31 2.4 2.86 2.2 0.89 1.9 2 14 0.72 3.9 0.09 3.7 - - - - 3 16 0.37 4.6 0.03 4.3 2.02 4.7 0.44 3.4 4 15 0.54 3.4 0.10 2.9 3.0 2.0 1.65 1.7 5 7 0.17 3.9 0.05 3.1 1.88 3.2 0.73 2.6 6 15 0.49 4.9 0.14 3.4 - - - - 7 5 0.91 7.3 0.30 4.7 1.84 5.6 0.74 3.8 8 14 1.24 2.7 0.22 2.0 - - - - 9 9 0.21 4.2 0.003 3.7 - - - - Overall 10 0.66 4.3 0.14 3.4 2.32 3.54 0.89 2.68

Table 4.3: Average Cotton and Maize Yield during Best and Worst Seasons District: Chiweshe(1); Gutu(2); Sanyati(3); Chivi(4); Mt Darwin(5); Wedza(6); Hurungwe(7); Shamva(8); Beitbridge(9)

For the 2003/04 season farm income computed across all enterprises averaged Z$245,000 per household (Table 4.5). Livestock sales, especially cattle, generated the highest income that averaged Z$910,000 per household. In general livestock sales, especially cattle and goats, are the main income earners for most farmers located in arid regions IV and V. In Beitbridge, for

61 example, cattle averaged Z$3.0 million per household. Maize was second with income averaging Z$430,000 per household. Other notable sources of income include goat sales (Z$310,000), vegetable sales (Z$300,000) and cotton (Z$290,000).

District NR Crop enterprises Livestock enterprises Other enterprises (%) (%) (%) Mze Cotton Peanut Cattle Goats Chicken Beer Veg Remittance brewing selling Chiweshe II 30.4 4.8 - 8.0 12.8 12.8 3.2 34.4 1.6 Gutu III 54.4 - 22.4 16.8 12.5 - 40.0 61.6 28.8 Sanyati IV 7.2 44.0 - 10.4 - - - - - Chivi V 8.0 20.0 16.0 30.4 21.6 4.8 - 76.0 27.2 MtDarwin IV - 66.4 - 15.2 12.8 - 4.8 3.2 - Wedza III 21.6 2.4 4.8 10.4 7.2 8.8 3.2 45.6 9.6 Hurungwe II 99.2 95.2 - 83.2 24.8 - 1.6 92.8 - Shamva II 34.4 20.8 - 12.8 7.2 - - 25.6 2.4 Beitbridge V - - - 56.0 69.6 2.4 - 2.4 2.4 Overall 28.4 28.2 4.8 27.0 18.8 2.2 7.7 38.0 8.1

Table 4.4: Crops and Livestock as Source of Income

Average income per household Average (Z$’000) Income (Z$’000) District Crops Livestock Other Mze Cot P/ Bean Cattle Goat Chick Beer – Veg Remit nut brew selling 1 811 537 - - 723 139 12.8 3.2 - 217 244.3 2 551 - 22.4 10.4 605 328 - 40.0 273 - 18.98 3 299 215 - - 138 - - - - 212 86.4 4 162 163 16.0 - 209 69 4.8 - 59 - 68.28 5 - 290 - - 105 82 - 4.8 - 132 53.18 6 195 308 4.8 2.4 269 76 8.8 3.2 193 - 106.02 7 378 311 - - 453 50 - 1.6 499 - 169.26 8 330 - - - 372 53 - - 133 - 88.8 9 - - - - 2,400 595 2.4 - - - 299.74 Overall 428 293 4.8 1.4 910 306 2.2 7.7 302 199 245.41

Table 4.5: Average Farm Income during the 2002/03 season District: Chiweshe(1); Gutu(2); Sanyati(3); Chivi(4); Mt Darwin(5); Wedza(6); Hurungwe(7); Shamva(8); Beitbridge(9)

62 4.2.5 Local Maize, Private trader and GMB Prices Maize price charged by the local GMB depot averages Z$16,100 per 50 kg bag whilst price charged by a neighboring farmer averages Z$32,300 per 50 kg bag (Table 4.6). Hence, maize price almost double that charged by local GMB depots. Private traders by comparison are the most expensive, charging Z$44,300 per 50 kg bag.

District NR Maize price (Z$/50kg bag) Local GMB–depot Neighboring farmer Private trader Chiweshe II 22,500 31,750 - Gutu III 28,000 45,000 34,500 Sanyati IV 25,000 75,000 36,000 Chivi V 13,000 62,500 - Mt Darwin IV 8,500 16,370 11,775 Wedza III 28,000 32,340 - Hurungwe II 20,000 13,780 14,940 Shamva II 22,500 26,550 - Beitbridge V 8,450 65,000 56,250 Overall 16,100 32,300 44,300

Table 4.6: Comparison of Local Maize, and Private trader vs. GMB prices

4.2.6 Food-aid Distribution during 2003/04 Season About 75% of the sampled households received food-aid during the 2003/04 season (Table 4.7). The food-aid was provided mostly by non-governmental organizations, notably World Vision and Christian Care. On average each household received about 275.5 kg, which translates to about six 50kg bags of maize. Except in a few areas (Gutu, Sanyati and Wedza), food-aid was predominantly in the form of maize.

The main problem affecting the food-aid program, as cited by 36.9% of the households, was ‘inadequacy’ while corruption (12.9%), erratic supply (7.2%) and lack-of-targeting (6.7%) were considered less severe.

63 District NR % receive Type of Food-Aid Problems associated with Food-aid food-aid % Qty of Corruption Inadequacy Erratic Lack-of- received maize supply targeting maize received per household (kg) Chiweshe II 77.6 76.8 140 3.2 5.6 0.0 1.6 Gutu III 72.0 17.6 344 19.2 12.8 4.0 22.4 Sanyati IV 68.0 28.0 158.2 6.4 8.8 4.8 0.0 Chivi V 87.2 84.8 339.7 8.0 35.2 22.4 13.6 MtDarwin IV 89.6 88.8 386.6 5.6 79.2 5.0 5.6 Wedza III 90.4 15.2 25.9 1.6 8.0 0.0 4.0 Hurungwe II 72.8 72.0 237.0 26.4 32.8 24.0 4.8 Shamva II 60.0 60.0 143.6 4.0 56.0 4.8 8.8 Beitbridge V 64.8 61.6 704.9 41.6 94.0 0.0 0.0 Overall 75.8 56.1 275.5 12.9 36.9 7.2 6.7

Table 4.7: Food Aid Distribution during season of 2003/04

4.3 Seasonal Forecasts At the beginning of each farming season, the Department of Meteorological Services issues seasonal forecasts, predicting the likelihood of rainfall across the whole country. For the 2003/04 season the Department predicted ‘a late but normal’ season. Worth investigating is whether farmers are getting the right information and how the information is influencing farming decisions.

The results in Table 4.8 shows that 80% of the households received seasonal forecasts during the 2003/04 season. A majority (47.4%) got the information via the radio. The results also show a mere 2.1% received the forecasts via extension workers. Important to investigate is the issue of whether farmers got the correct information, as broadcast by the Department of Met Services during 2003/04 season. From the results it is observed that only 4.3% got the ‘rightful’ message; about 48.2% got the ‘wrong’ message as they interpreted it as either ‘normal to above-normal’ season and about 10% mistook it as a bad season. Perhaps the confusion stemmed from the ‘late but normal’ message as farmers could have wrongly interpreted it as ‘normal’ season. A small proportion misinterpreted it ‘below’ normal season.

64 District NR % get How was the forecasts message received? What was the forecast message for the season? forecast Radio TV Discus Ext. Above Nor Ave Below Late but s worker normal mal ave normal Chiweshe II 78.4 48.8 4.8 26.4 - 5.6 24.8 16.8 18.4 12.8 Gutu III 76.0 47.2 7.2 16.8 - 1.6 59.2 10.4 2.4 - Sanyati IV 98.4 30.2 5.6 30.4 12.8 43.2 48.0 4.0 2.4 0.8 Chivi V 73.6 38.4 4.8 28.0 2.4 14.4 20.0 16.0 15.2 2.4 MtDarwin IV 83.2 46.4 0.0 36.0 - 2.4 32.8 28.0 18.4 0.8 Wedza III 57.6 37.6 0.8 16.0 3.2 6.4 19.2 0.8 17.6 10.4 Hurungwe II 98.4 90.4 2.4 4.8 - 1.6 70.4 1.6 3.2 - Shamva II 62.4 40.0 10.4 12.8 - 20.8 4.0 20.8 11.2 8.0 Beitbridge V 96.1 47.2 4.1 25.7 - 20.2 42.0 33.6 4.2 - Overall 80.2 47.4 4.1 25.7 2.1 12.8 35.4 14.5 10.0 4.3

Table 4.8: Households who received Seasonal forecasts during 2003/04

4.3.1 Confidence with the forecasts There seems to be a credibility problem associated with the forecasts. From Table 4.9, 37.6% of sampled households indicated lack of confidence compared to 41.3% who showed outright confidence. Farmers seem to have no problems in understanding the forecasts, as 65.2% classified the forecasts as easy to understand. With respect to timing of seasonal forecasts, most (38.3%) indicated the month of August (i.e., two months before farming season starts) as the optimal time to receive forecasts.

65 District Optimal time to issue Confidence level Understanding level Forecasts (%) (%) (%) July Aug Sept 12341234 1 46.4 27.2 15.2 51.2 24.0 9.6 24.8 74.4 - - 2 33.6 26.4 5.6 - 32.8 32.8 4.8 - 28.8 32.8 5.6 3 19.2 39.2 20.8 61.6 24.0 12.0 - 58.4 32.0 7.2 - 4 11.2 40.8 20.0 - 39.2 28.8 5.6 - 54.4 20.2 4.8 5 15.2 35.2 23.2 - 34.4 39.2 - 19.2 39.2 20.0 2.4 6 7.2 44.0 21.6 - 33.6 35.2 10.4 16.0 56.0 8.0 - 7 14.4 54.4 19.2 78.4 16.8 - - 72.0 21.6 - - 8 20.8 31.2 30.4 - 36.0 40.8 6.4 4.8 26.4 51.2 6.4 9 27.2 71.2 - 45.6 40.3 - - 48.0 38.4 2.4 Overall 14.6 38.3 26.6 18.3 23.0 33.5 4.1 22.9 42.3 20.1 2.8

Table 4.9: Assessing Farmers’ level of Confidence with the forecasts Confidence level: very confident(1); confident(2); somewhat confident(3); not confident(4) Understanding level: very easy(1); easy(2); somewhat easy(3); not easy(4) District: Chiweshe(1); Gutu(2); Sanyati(3); Chivi(4); Mt Darwin(5); Wedza(6); Hurungwe(7); Shamva(8); Beitbridge(9)

4.3.2 Decision-making factors This section looks at factors considered by farmers when deciding what crop(s) to grow. From Table 4.10, a majority (48.5%) chose food security is the predominant decision factor. Second and third important factors are producer price (26.1%) and loan repayment (19.2%). Loan repayment seem to be an important decision factor, especially in cotton growing regions such as Sanyati (72%), Chivi (48%) and Mt Darwin (23.2%). Drought tolerance features as an important factor mostly in driest regions V such as Beitbridge (96%) and Chivi (30.4%). The use of seasonal forecasts as a decision-making factor is still very low at 12.8% in most areas. In some areas (Gutu Sanyati, Mt Darwin and Wedza) seasonal forecasts are not used as decision factor at all.

Results also show that input availability rather than price of inputs is a more important decision factor. Price of inputs and local selling price seem to be the least important decision factors. For most areas price of inputs (fertilizers, seed maize) is not a decision-making factor at all. This result is surprising, given the high inflation and general instability of farm input prices currently affecting the country.

66 District NR Decision Factors (%) Loan Producer Drought Local Seasonal Input Price Food repay price tolerance selling forecasts availability of security price inputs Chiweshe II - 8.0 - - 12.0 22.4 36.8 61.6 Gutu III 29.6 - - - - - 7.2 40.0 Sanyati IV 72.0 - 15.2 20.0 - - - 45.6 Chivi V 48.0 - 30.4 - 13.6 - - 43.2 Mt Darwin IV 23.2 83.2 - - - 56.0 - 43.3 Wedza III - 14.4 - - - 10.4 - 48.0 Hurungwe II - 92.0 - - 14.3 44.8 - 37.6 Shamva II - 37.2 - - 15.2 11.2 - 40.0 Beitbridge V - - 96.0 - 20.0 - - 77.6 Overall 19.2 26.1 15.7 2.9 12.8 16.1 4.9 48.5

Table 4.10: Decision-making Factors

4.4 Impact of Drought on Smallholder Farmers The impact of drought is often severe and long-lasting. This section looks at how drought impacts smallholder farmers. From Table 4.11, the majority of households (65.1%) indicated crop loss as the most severe impact they endure when drought occurs. Cattle loss comes second at 60.3% while food insecurity is third at 58.0%. Further, more than 30% indicated ‘hunger related diseases’ and ‘water scarcity’ as significant problems they suffer when drought occurs. About 25% reported ‘inability to send children to school’ as major problem whilst 10% reported ‘failure to repay loans’. The latter appears a key concern, especially for the cotton growers in such regions as Sanyati and Mt Darwin. Following economic reforms introduced in early 1990s (see Chapter 1, section 1.5), a new marketing system has emerged where cotton is grown under contractual agreement between buyers, consisting mainly of large firms and smallholder farmers. As part of the contract, the contracting firm provides loans to the farmer in the form of essential inputs (chemicals and seeds). In return, the farmer is under obligation to sell his produce to and only to the contracting firm and repay the loans as well.

When asked to rank, a majority (34.1%) picked food insecurity as most severe. Crop loss was ranked second at 19.6%, cattle loss third at 13.6% and water scarcity fourth at 12.4%.

67 District Loss impact (%) Crop Cattle Donke Goats Failure Unable to Unable Hunger- Water Loss loss y loss Loss to repay feed own to send related scarcity loans household children diseases to school increase Chiweshe 74.4 20.0 - - 3.2 81.6 20.0 28.0 28.0 Gutu 30.4 63.2 1.6 - 9.6 82.2 36.8 39.2 37.6 Sanyati 30.2 69.6 10.4 12.8 25.6 82.4 47.2 48.0 26.0 Chivi 24.0 52.0 3.2 6.4 - 89.6 56.0 20.8 30.4 MtDarwin 79.2 56.8 - 6.6 39.2 28.0 5.6 72.8 80.0 Wedza 83.2 28.0 - 4.0 - 40.0 9.6 9.6 21.6 Hurungwe 99.2 96.8 - 20.0 10.4 22.4 8.8 25.6 56.0 Shamva 65.6 60.8 3.2 2.4 5.6 94.6 28.0 38.4 64.8 Beitbridge 99.2 95.2 96.8 95.2 2.4 2.4 - - 5.6 Overall 65.1 60.3 12.8 16.4 10.8 58.0 23.6 31.5 38.9 Ranked 19.6 13.6 - - - 34.9 2.8 4.9 12.4 loss

Table 4.11: Impact of Drought on Smallholder Farmers

4.4.1 Drought coping strategies This section identifies coping strategies that farmers adopt in face of drought. From Table 4.12, more than 60% indicated ‘buying more grain’ as the most dominant strategy; second was ‘seeking off- farm employment’ at 49.8%, third was ‘selling of livestock’ (38.1%) and ‘food-aid’ was fourth at 28.7%. Less than 1% indicated ‘formal insurance’ as a drought coping strategy. The latter emphasizes the absence of formal insurance within the smallholder farm sector.

To gain further insights, farmers were asked to rank the strategies according to frequency and reliability of the coping measure. A majority ranked buying more grain (26.6%) as most reliable measure, while seek off-farm job and sell livestock was ranked second and third respectively. About 8.1% engage in barter in trade while 5% selected food-aid as a reliable strategy despite food-aid being free and in existence for many decades.

68 District Coping Strategies (%) 1 2 3 4 5 6 7 8 9 10 11* 1 83.2 8.0 52.2 4.2 - 16.8 - 8.0 10.4 19.2 - 2 60.0 38.4 57.6 4.0 13.6 31.2 - 13.6 19.2 52.0 - 3 76.8 47.2 63.2 11.2 5.6 48.8 3.2 8.0 36.0 2.4 - 4 74.4 16.8 38.4 - - 17.6 - 12.0 3.2 11.2 8.0 5 34.4 64.8 - - - 41.6 - 41.6 11.2 46.4 - 6 67.2 14.4 27.2 - - 20.8 - 18.4 4.8 - 28.0 7 53.6 89.6 32.0 - - 21.1 - - 7.2 27.2 - 8 8.0 27.2 74.4 3.2 16.8 - - 35.2 26.4 7.2 32.0 9 99.2 64.2 38.4 - - 60.3 - 41.6 1.6 - - Overall 61.9 38.1 49.8 2.8 4.4 28.7 0.9 19.8 13.3 18.6 8.4 Rank 26.6 17.2 18.8 5.2 2.5 2.3 8.1 12.9 measure

Table 4.12. Farmers’ Drought Coping Strategies Coping strategies: buy-more-grain(1); sell-livestock(2); seek-off-farm-job(3); stop-send-children-to- school(4);migrate-to-other-area(5); rely-on-food-aid(6); rely-on-formal-insurance(7); reduce consumption(8); borrow-grain-from-others(9); engage-in-barter-trade(10); other*(11) consists mainly vegetable selling. District: Chiweshe(1); Gutu(2); Sanyati(3); Chivi(4); Mt Darwin(5); Wedza(6); Hurungwe(7); Shamva(8); Beitbridge(9)

4.4.2 Drought Vulnerability Factors Many rural households are vulnerable to drought and in this section I seek to identify factors underlying this vulnerability. Table 4.13 shows a number of vulnerability factors identified across different regions. Overall about 70% of the households identified erratic rainfall as the principal vulnerability factor. In particular over 90% of households in arid region V (Chivi and Beitbridge) reported erratic rainfall as the key vulnerability factor. Another important vulnerability factor is poverty identified by 42.5% of sampled households. Other factors identified include lack of remittances (20.6%), poor crop choices (20.0%), lack of extension advice (17.2%), unavailability of credits and insurance (17.1%) and lack of climate forecasts (14.4%). Lack of seasonal forecasts is relatively more important in a few areas, such as Shamva (47.2%), Mt Darwin (28%) and Gutu (25.6%).

69 District Vulnerability Factors (%) Poor Poor Poor Lack Lack no Lack Erratic Location no crop gvt. ext. climate credits of rainfall assets choice Policy advice forecasts and remit to sell insuranc e 1 55.2 - 3.2 - - 25.6 5.6 58.4 - - 2 75.2 10.4 15.2 24.8 25.6 4.0 24.8 72.8 8.0 - 3 75.4 32.0 - 7.2 4.0 20.0 55.2 58.4 19.2 20.8 4 8.0 - - - - - 32.0 90.4 23.2 17.6 5 32.8 40.8 - 16.0 28.0 - - 96.0 28.0 19.2 6 14.4 5.6 - - 8.0 2.4 6.4 54.4 - - 7 12.0 29.6 - 62.4 4.0 48.0 12.8 67.2 28.0 4.0 8 32.8 28.8 - 25.6 47.2 50.4 24.8 23.2 - 39.2 9 76.8 29.6 - 15.2 12.8 - 23.2 98.4 4.8 16.0 Overal 42.5 20.0 2.8 17.2 14.4 17.1 20.6 68.8 12.9 14.3 Rank 9.8 - - 3.6 5.0 8.3 4.3 47.7 - -

Table 4.13: Drought vulnerability factors District: Chiweshe(1); Gutu(2); Sanyati(3); Chivi(4); Mt Darwin(5); Wedza(6); Hurungwe(7); Shamva(8); Beitbridge(9)

4.5 Measuring Correlation between Crop Yield vs. VCI and Rainfall From my discussion in Chapters 2 and 3, interest in remote sensing technology for modeling crop insurance stems mainly from the facts that first, it allows repeated temporal measurements necessary for dynamic monitoring of crop development during growing season, crop yield assessment/prediction and crop loss estimation. Second, it allows wide spatial coverage that may not be possible with ground-based field surveys. Third, remote sensing technology offers extra visual information essential for crop condition and yield assessment. Fourth, remote sensing allows early prediction and decision-making, and finally, the technology is potentially more cost- effective when compared with field-based methods.

But if VCI (or any index such as rainfall) is to be used as an index upon which insurance contracts can be drawn, it must fulfill these conditions: (i) be highly correlated with crop yield in such a way as to track the occurrence of insured losses; (ii) be easy to observe and measure on a regular basis, (iii) be able to predict crop yield with fair accuracy and at low cost, and (iii) not subject to manipulation by an individual farmer or insurer.

As discussed in detail in Chapter 2, NDVI measures vegetation vigor caused by chlorophyll activity and as a result it is highly correlated with photosynthetic activity in wilting and non-

70 wilted plant foliage. Thus, NDVI is a good predictor of plant canopy biomass, vigor or stress. VCI as well indicates the vigor of vegetation cover. VCI derived from NDVI, is a consistent index for assessing crop condition and potential yield. On the basis of these vegetation indices, sequential crop profiles can be developed during a growing season that show the progression of canopy emergence, maturation and senescence. Using these crop profiles, crop loss and yield performance can be assessed with a considerable degree of accuracy. Hence, an insurance contract can be drawn on the basis of vegetation indices. The following section analyses simple pair-wise correlation between maize and cotton yields versus VCI, a satellite variable derived from remotely sensed NDVI (see Chapter 2, section 2.2) and versus Nino34, a surrogate variable used for measuring the influence of El Nino. The section also illustrates profiles of VCI during the growing period (Oct-April) using three typical seasons classified as ‘best’, ‘normal’ and ‘worst’, according to average precipitation recorded.

4.5.1 Crop Yield vs. Drought-indices The results in Table 4.14 show the correlation between maize and cotton yield versus monthly VCI, Nino34 and regional precipitation. With respect to VCI (Table 4.14(a)), at the beginning of the farm season in December, correlation is very weak at -0.02 and –0.03 for maize and cotton, respectively. This is expected since during early season, most crops are still in their early stages of development and hence manifest little ‘greenness vigor’. However, as the season progresses, correlation improves remarkably and attains a maximum at 0.64 and 0.73 for maize and cotton, respectively. The correlation is at its peak during the months of March and/or February. By April, crops attain full maturation and senescence sets in and simultaneously the ‘greenness vigor’ diminishes, resulting in a gradual decline in correlation.

Apparently the highest correlation occurs during flowering/grain filling stage (i.e., February/March) which coincidentally is most crucial yield-determining stage. Hence it is logical to conclude that final crop yield must be correlated with VCI. If this relationship consistently holds, then VCI can be used to model yield losses with satisfactory precision.

71 National Maize National Cotton Smallholder Smallholder yield yield maize yield Cotton yield a)VC_variables vcDec 0.12 0.04 -0.2 -0.03 vcJan *0.47 0.36 *0.45 0.20 vcFeb *0.62 0.35 *0.63 *0.55 vcMar *0.58 0.36 *0.64 *0.73 vcApr 0.26 0.31 0.33 0.53 b) Nino34_variables n34Dec *-0.60 -0.25 *-0.56 -0.31 n34Jan *-0.61 -0.28 *-0.59 -0.33 n34Feb *-0.68 -0.28 *-0.64 *-0.35 n34Mar *-0.66 -0.30 *-0.62 -0.31 n34Apr *-0.58 *-0.37 *-0.56 -0.34 n34May *-0.50 *-0.40 *-0.53 -0.29

Table 4.14: Measuring Correlation between Crop yield vs. identified Indices for the period of 1980-2000 *Indicates statistical significance at α=.10 or smaller level

Results in Table 4.14(b) show that crop yield and Nino34 are negatively correlated. At the beginning of the season, correlation is low, it peaks during the months of February and March to attain a maximum at -0.68 and -0.35 for maize and cotton respectively. By April, when harvesting starts, correlation begins to decrease. These results imply that the El Nino is likely to exert its maximum impact during the months of February and March.

Knowledge relating to when maximum correlation occurs is important in three ways: it provides a basis for determining contract parameters (see Chapter 3), it allows yield to be predicted 2 to 3 months before harvesting begins and this is vital for planning purposes. Lastly, from the insurer perspective, this provides a gauge to assess the likelihood of insurable losses.

Figures 4.1 - 44 illustrate different VCI profiles during three distinct season viz. 1989/90, 1991/92 and 1992/93 all across natural regions II-V. The three seasons were classified good, average and bad according to precipitation received. Thus, 1992/93 season was classified the best with seasonal

72 precipitation greater than 750 mm, whilst 1991/92 season was classified the ‘worst’ with average precipitation of 315 mm and 1990/91 season was classified average season with average precipitation of 550 mm. What is observable in Figures 4.1-4.4 is a distinct decline in VCI during a bad year 1991/92 for the period January-April. Compared to good/average seasons, VCI was 20- 35% lower during drought year 1992/92. We also observe that during a good season, VCI attains well-spread peak during the month of January and February while for an average year it is a delayed peak that occurs around February and March. Figure 4.4 panel (b) is exceptional; the peak for a good and average year and depression for a bad year are all stacked around the month of January. These result emphasize three things: first, the sensitivity of VCI to drought-stressed conditions, second, the ability of VCI to track drought conditions especially during extreme seasons and third, the period January-February-March imply crucial yield determining stage during plant growth profile.

73 (a)

198990(average) 199192(drought) 199293(good)

0.9 0.8 0.7 0.6 0.5

VCI 0.4 0.3 0.2 0.1 0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

(b)

19890(average) 199192(drought) 199293(good)

0.9 0.8 0.7 0.6 0.5

VCI 0.4 0.3 0.2 0.1 0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

Figure 4.1: Illustrating VCI Profile in NR II during Good, Average and Bad Season for (a) district 1 and (b) district 7

74 (a)

198990 (average) 199192(drought) 199293 (good)

0.9 0.8 0.7 0.6 0.5

VCI 0.4 0.3 0.2 0.1 0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

(b)

198990 (average) 199192 (drought) 199293 (good)

0.8 0.7 0.6 0.5 0.4 VCI 0.3 0.2 0.1 0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

Figure 4.2: Illustrating VCI Profile in NR III during Good, Average and Bad Season for (a) district 3 and (b) district 6

75 (a)

198990 (average) 199192 (drought) 199293 (good)

0.7

0.6

0.5

0.4

VCI 0.3

0.2

0.1

0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

(b)

198990 (average) 199192 (drought) 199293 (good)

0.9 0.8 0.7 0.6 0.5

VCI 0.4 0.3 0.2 0.1 0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

Figure 4.3: Illustrating VCI Profile in NR IV during Good, Average and Bad Season for a) district 2 and (b) district 5

76 (a)

198990 (average) 199192 (drought) 199293 (good)

0.8

0.7

0.6 0.5

0.4 VCI 0.3

0.2 0.1

0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

(b)

198990 (average) 199192 (drought) 199293 (good)

0.8

0.7

0.6

0.5

0.4 VCI 0.3

0.2

0.1

0 Oct Nov Dec Jan Feb Mar Apr May Rain Season

Figure 4.4: Illustrating VCI Profile in NR V during Good, Average and Bad Season for a) district 4 and (b) district 9

77 4.6 Crop Yield loss Distribution Determining accurate premium rates requires an accurate measurement of crop yield risks or crop yield losses. But accurate measurement of crop yield losses in return depends on proper representation of the distribution of crop yields. The latter is no easy task. As discussed by Goodwin and Ker (1998), proper representation of crop yield distribution may be cumbersome for a number of reasons. First, it has been empirically observed that crop yield distributions are often negatively skewed. The skewness may be caused by biological constraints that limit plants from realizing their full potential yield. Second, environmental factors such as weather and pest damage hamper crop yield, resulting in lower yield being frequently observed. Third, when yield is disaggregated, say to a county/district level which is often the case in applied work, idiosyncrasies unique to a specific area or region may affect local yield distribution. Hence modeling crop yield may be very elusive.

A variety of approaches have been suggested in the literature (Goodwin, 1998, Gallagher, 1987) that represents yield distributions and measures yield risks. Two common methods include parametric versus non-parametric distributions. With respect to the former, a specific parametric distribution is selected a-priori and parameters of the distribution are estimated using observed yield data. For example, Gallagher (1987) used a gamma function to model the distribution of soybean yields. Nelson (1989) used a beta distribution to model corn yields. There are two disadvantages with the parametric approach: first, it requires a priori specification of a particular distribution, and, if a chosen distribution is incorrect, this results in misleading inferences. Second, parametric distributions do not allow bi-modality and hence using such functions may give inaccurate predictions. In view of these limitations, nonparametric methods to estimating yield distributions have been suggested.

Nonparametric density estimation techniques offer advantages in that: first, they do not assume a particular functional form for yield distribution, but rather allow the data to select the most appropriate representation of the yield distributions. Second, nonparametric density estimation techniques are fully flexible and hence can accommodate bi-modality. Lastly, these techniques can capture local idiosyncrasies in yield distribution that may not be possible with a parametric specification. But nonparametric density estimation techniques have their own limitations. For instance, they require a choice of kernel function and bandwidth.

78 This section considers nonparametric kernel density estimation for the variables VCI and rainfall. These variables are used as proxy variables to estimate crop yield densities and are eventually used to determine actuarially fair crop insurance premium rates for area-yield crop insurance (see Chapter 5).

4.6.1 Nonparametric density estimation When measuring yield risk for the purposes of rating an insurance contract, an insurer is interested in knowing the probability that a loss will occur and the expected indemnity. As an illustration, suppose a loss criterion is defined as 65% of mean μ. This means a loss is said to occur, if the yield falls below 65% of μ. Thus, one has to determine the probability that yield is less than 65μ. This probability is computed as the area under the probability density function between 0 and 65μ. Therefore, accurate measurement of yield density is needed to determine the loss likelihood.

Using the kernel approach (see Goodwin, 1998 for detail), each observation is surrounded by a symmetric weighting function K, which satisfies the following condition:

∞ ∫ K(t)dt = 1 −∞ and usually the weighting function will be a symmetric probability density function. Assuming n identically independently distributed observations of a univariate series = Y (Y1 ,,...... Yn ) , the kernel estimator of the density of Y is given by

1 n ⎛ y − Y ⎞ ∑ K⎜ i ⎟ . nh i=1 ⎝ h ⎠ The parameter h is a bandwidth that smoothes the density function. This parameter assigns weight to neighboring observations in constructing the density. Thus a larger bandwidth will smooth the density function, while a smaller bandwidth will result in a rough and irregular density. Of importance when estimating a nonparametric density function is how to choose an optimal bandwidth parameter. In the literature, Parzen (1962) derived an optimal bandwidth

parameter, which, assuming the unknown density is normal with variance σ 2 and a Gaussian kernel is used, is reduced to: = σ −(1/ 5) hopt 1.06 n . Using this optimality rule Silverman (1986) found that by setting

79 σ = min[]standard deviation, interquartile range/1.34 and reducing the scaling factor from 1.06 to 0.9 generally performed well for empirical work. It is Silverman’s rule-of-thumb method that is used to select the bandwidth parameters for the nonparametric density function shown below.

Figures 4.2-4.5 show nonparametric kernel density estimation for seasonal precipitation for the period of 1980-2000. The distributions are estimated by agro-regions (II-V) and display different characteristics. A common feature observed across the estimated distributions is that they exhibit bi-modality. The bi-modal distribution may underlie a mid-season dry spell that commonly occurs during the peak months of January and February. For regions II and III, the distributions are less uniform and often show multiple peaks compared to drier regions (IV, V). This result may imply that the risk of mid-season dry spell is more severe in regions II and III than in IV and V. For region V the bi-modal distribution tend to suggest extreme patterns of rainfall distribution, which despite being comparatively low is less erratic.

Figures 4.6-4.9 shows estimated kernel density functions for VCI for the period 1980-2000 across regions II-V. Figures 4.6 and 4.7 show bi-modal distribution while in contrast, Figures 4.8 and 4.9 do not exhibit this property except for the district of Chivi (Figure 4.9). As observed earlier, the bi-modal distribution tends to suggest the prevalence of mid-season dry spell whose impact is more felt in regions II and III than in counterpart regions IV and V. The results tend to make sense since for drier regions IV and V, highly drought resistant crop are commonly grown and often these crops are less susceptible to water stress. In contrast, crops grown in wetter regions (II and III) generally are high yield, but less drought-resistant and cannot withstand a prolonged period of water scarcity. In turn, these conditions are reflected by a reduced value of VCI.

80 Figure 4.5: Estimated Rainfall Kernel Density distribution for NR II

81 Figure 4.6: Estimated Rainfall Kernel Density distribution for NR III

82 Figure 4.7: Estimated Rainfall Kernel Density distribution for NR IV

83 Figure 4.8: Estimated Rainfall Kernel Density distribution for NR V

84 Figure 4.9: VCI Kernel Density distribution for NR II

85 Figure 4.10: VCI Kernel Density distribution for NR III

86 Figure 4.11: VCI Kernel Density Estimation for NR IV

87 Figure 4.12: VCI Kernel Density distribution for NR V

88 CHAPTER 5

DISCUSSION OF THE MAIN RESULTS

This chapter presents the main results of the study. The results are divided into three sections: the first section presents results on potential demand for improved seasonal forecasts using the WTP approach. The second section presents results pertaining to the potential demand for area- yield drought-index insurance using the WTP approach. The final section analyses the feasibility of offering index-based insurance. An indemnity function is designed based on remotely sensed VCI and rainfall index.

5.1 Smallholders’ Preferences for Seasonal Forecasts 5.1.1. Estimating a Random Utility Model with log-linear Income On the basis of the theoretical models and rationale discussed in Chapter 3 (sections 3.1 -3.2), random utility models, log-linear in income are estimated using logit and bivariate-probit model specifications. Responses to CV question as framed in section 3.2 are coded 1/0 for yes/no and estimated as dependent variable(s). The covariates on the other hand, include demographic variables (sex, farm-experience, education), resource-constraining factors (cattle-ownership, cultivated acres, working plough and radio) and regional-dummy effects. Table 5.1 shows descriptive information and definition of variables used in estimating various models. The loginc − bid income23 variable in particular, is defined as: log inc = log[ i i ] . For the random utility inci ∂(.) β model log-linear in income, the marginal utility of income becomes = which is ∂ (Inci ) Inci decreasing income for β>0.

23 The loginc parameter is used to recover parameter estimates of other covariates (see Haab and McConnell, 2001)

89 (a) Seasonal forecasts models Variable Description Mean Expected sign BID1(Z$) first bid 5,160 BID2(Z$) second bid 6,310 Rcvbid1 0/1 constructed dependent variable based on BID1 - Rcvbid2 0/1 constructed dependent variable based on BID2 - Acres average acre owned by a household 6.91 + CULTIACRE number of acres cultivated on average per season 5.01 + HHSEX sex of the head of responding household (1=male,0=female) 0.63 ? AGRICEDU did head of household received agricultural training? (1=yes,0=n0) 0.21 + WORKPLOU does your household own a working ox-drawn plough? (1=yes,0=no) 0.81 + WORKRADIO does your household own a working radio? (1=yes,0=no) 0.64 + LACKPOWER do you face severe draft power shortage? (1=yes, 0=no) 0.48 - FARMEXP do you have farming experience ranging 1-10 years? (1=yes,0=no) 0.36 ? D1 regional dummy variable if district is 1 - ? D2 regional dummy variable if district is 2 - ? D3 regional dummy variable if district is 3 - ? D4 regional dummy variable if district is 4 - ? D5 regional dummy variable if district is 5 - ? D6 regional dummy variable if district is 6 - ? D7 regional dummy variable if district is 7 - ? D8 regional dummy variable if district is 8 - ? D9 regional dummy variable if district is 9 - ? (b) Drought Insurance Models WFBID1($) response to first bid with food-aid available 9,200 WFBID2($) response to second bid with food-aid available 10,800 WOFBID11(Z$) response to first bid with food-aid not-available 10,100 WOFBID22(Z$) response to second bid with food-aid not-available111 11,285 Inc (Z$) household gross income 761,125 + Foodloss risk factor for drought-induced food-insecurity 0.48 + Loanloss risk factor for drought-induced farm-loan default 0.10 + Waterloss risk factor for drought-induced water-shortage 0.41 ? Rcvwfbd1 0/1 constructed dependent variable based on WFBID1 - Rcvwfbd2 0/1 constructed dependent variable based on WFBID2 - Rcvwofbd1 0/1 constructed dependent variable based on WOFBID11 - Rcvwofbd2 0/1 constructed dependent variable based on WOFBID22 -

(c ) SUR Models mzeprod smallholder maize production per district (kg) 20,993 cotprod smallholder cotton production per district (kg) 2,626 critVCI Vegetation-Index computed for the period Jan-Mar 0.64 + critN34 El-nino index measured for the period Dec-Mar 0.18 - cirtRf refers to critical rainfall measured during period Dec-Mar(mm) 124.8 + VCDec Vegetation index computed for the month of Dec 0.54 + VCJan Vegetation index computed for the month of Jan 0.65 + VCFeb Vegetation index computed for the month of Feb 0.69 + RFDec Rainfall measured during the month of Dec (mm) 138.1 + RFJan Rainfall measured during the month of Dec (mm) 177.8 + RFFeb Rainfall measured during the month of Dec (mm) 152.5 + RFMar Rainfall measured during the month of Dec (mm) 82.4 +

Table 5.1: Variable Definition and Description for (a) Seasonal forecasts (b) Drought Insurance and (c) SUR Models

90 Hence, log-linear specification provides marginal utility of income that is individual-specific and that varies across utility states as money income changes.

Another important parameter in which we are interested is the marginal effect. Marginal effects are computed by using an approach suggested by Haab etal (2001). Since both the logistic and normal distributions have median equal to zero, median WTP with respect to the error term for the bounded probit or logit model becomes:

− γ = − z j MD(WTPj ) y j (1 e )

− γ = − z j MD(WTPj ) y j (1 e ) where the latter equation shows the calculation of MD(WTP) computed at the mean of income ( y ) and covariates ( z ). The marginal effects on the MD(WTP) becomes easy to compute and thus, the effect of a change in one of an individual’s z on the median WTP becomes: ∂ MD(WTPj ) − γ = y (e zi k ).γ ∂ j k z kj and on re-arranging, this reduces to: ∂ MD(WTPj ) = (y − MD(WTP )).γ ∂ j j k z kj This equation measures the change in MD(WTP) for a marginal change in the covariates. The results on marginal effects are shown and discussed in the sections to follow.

The parameter estimates based on logit and bivariate-probit models are presented in Table 5.2. For the logit model, most variables (except agricedu and workplou) contain the theoretically expected signs (see last column Table 5.1). The variables, hhsex, workradio, lackpower, farmexp and some dummies (D1, D4,D6,D7 and D8) are statistically significant at α=0.10 level. For the bivariate model, the variables agricedu and farmexp are also significant at α=0.10 level.

Both models contain regional dummy variables (excluding the base case, D5) incorporated so as to capture fixed-effects due to regional differences. To justify the inclusion of these dummies, we conduct a likelihood ratio test where the null hypothesis equates all the dummies to zero against the alternative hypothesis. The general form of the likelihood ratio test statistic is: -2(log- likelihood value of the restricted model less the log-likelihood value from the unrestricted model). This test statistic is distributed Chi-square with degrees of freedom equal to the number of restrictions placed on the restricted model. With log-likelihood values of –268.65 and –294.77

91 for the unrestricted and restricted models, respectively, and degrees of freedon=8 and

χ 2 ()8 = 2.67 , we reject the null hypothesis at α=0.05 level and conclude that the regional dummy variables cause significant effect on the WTP for seasonal forecasts.

Important implications can be drawn from these results. First, the marginal effects results imply that for the logit model, the probability of a ‘yes’ response to WTP for seasonal forecasts increases by 0.06% if a household possesses a working radio. And for the bivariate model, this probability increases by 0.68%. Hence, a working radio is an important asset, if a household is to take full advantage of seasonal forecasts. Second, farm experience counts. The probability of responding ‘yes’ to WTP for seasonal forecasts increases by 0.16% and 1.55% for the logit and bivariate models respectively, for households with ten or less years of farming experience. Maybe households with less farm experience are keen to learn and seek efficient methods of avoiding drought risk and losses. Third, the lackpower variable captures how draft power shortage affects WTP for the forecasts. The marginal effects results indicate that for the logit and bivariate model, the probability of a ‘yes’ response to WTP for seasonal forecasts decreases by 0.05% and 0.39% respectively for a household without adequate draft power. Hiring draft/tractor power could be an alternative option, but only a handful can afford due to the high cost of hiring. Even if a farmer can hire power, it is unlikely for him to get power at such an opportune time that allows him to make full use of seasonal forecasts. What is frequently observed is those that farmers hire out draft power only do so when they have fully met their own tillage requirement. Often this may be too late for a hiring farmer to take full advantage of the seasonal forecasts. Lastly, results show that farmers with agricultural training are less willing to pay for seasonal forecasts. Marginal effects indicate that, for both the logit and probit model, the probability of answering ‘yes’ to CV question on WTP for seasonal forecasts decreases by –0.17% and –1.69% respectively, if the responding household received agricultural training (agricedu).

92 Logit Model Bivariate Model Variable Parameter Standard Marginal Parameter Standard Marginal Estimates error effects (%) Estimates error effects (%) Constant 1.835 *0.535 - 1.203 *0.310 - Cultiacre 0.023 0.041 0.01 0.013 0.024 0.05 Hhsex 0.397 *0.206 0.09 0.162 0.148 0.57 Agricedu -0.733 0.288 -0.17 -0.477 *0.182 -1.69 Workplou -0.305 0.333 -0.07 -0.211 0.208 -0.75 Workradio 0.276 *0.207 0.06 0.192 0.150 0.68 Lackpower -0.216 *0.102 -0.05 -0.111 0.153 -0.39 Farmexp 0.693 *0.264 0.16 0.438 *0.166 1.55 D1 -1.654 *0.542 - -1.039 *0.331 - D2 -0.296 0.466 - -0.169 0.271 - D3 -1.304 0.674 - -0.777 0.392 - D4 -0.911 *0.455 - -0.559 0.279 - D6 -1.588 *0.478 - -0.976 *0.279 - D7 2.283 *0.830 - 1.073 *0.453 - D8 -2.226 *0.490 - -1.370 *0.305 - D9 -0.464 0.469 - -0.260 0.284 -

Table 5.2: Parameter Estimates of Logit vs. Bivariate Models *Significant at 0.1 level or lower

5.1.2 Estimating WTP for Seasonal Forecasts Estimating parametric models from dichotomous choice CV responses allows us to understand how WTP responds to individual covariates as discussed above. But more importantly, estimated parametric models allow us to calculate WTP for the provision of non-market goods. Results in Tables 5.3(a) and (b) show the computed WTP for seasonal forecasts across districts and natural regions, respectively. For a single-bound (SB) model (or rather logit model) the estimated WTP ranged from Z$2,427 (district 8) to Z$4,676 (district 5). For double-bound (DB) model (or bivariate model), the estimated WTP ranged from Z$2,532 (district 8) to Z$4,225 (district 5). What is interesting to observe is the differential WTP pattern across districts. For households in wetter districts such as 1 and 8, WTP is consistently lower than those in drier districts 5 and 9. This picture is further illustrated in Table 5.3(b) where on average WTP is 36% and 30% lower in natural region II than in regions IV and V, respectively. Similarly, WTP for seasonal forecasts is 17% and 9.3% lower in region III than in regions IV and V respectively. Because the perceived drought risk is more ominous in drier regions (IV and V), households in these regions are willing to pay more for the provision of seasonal forecasts. The opposite is true for households in wet regions (II and III).

93 (a) WTP across Districts District24 Farm Region SB Model DB Models WTP(Z$) WTP(Z$)

1 II 3,183 2,931 2 IV 4,473 3,970 3 III 3,752 3,453 4 V 3,958 3,516 5 IV 4,676 4,225 6 III 3,739 3,424 8 II 2,427 2,532 9 V 4,414 3,954

(b) WTP across Natural Regions Natural SB Model DB Models Overall Regions Average II 2,805 2,732 2,769 III 3,746 3,439 3,593 IV 4,575 4,098 4,337 V 4,186 3,735 3,961

Table 5.3: Estimated WTP for Improved Seasonal Forecasts 1=Chiweshe; 2=Gutu; 3=Sanyati; 4=Chivi; 5=Mt Darwin; 6=Wedza; 7=Hurungwe; 8=Shamva; 9=Beitbridge

5.2 Assessing Smallholders’ Preferences for Drought Insurance As discussed in Chapter 3 (see section 3.3), crop insurance markets for smallholder farmers in Zimbabwe do not exist. The absence of formal insurance markets that could offer efficient risk sharing alternatives for smallholders amounts to market failure This market failure persists despite smallholders being risk averse and showing strong demands for insurance services (Binswanger, 1986). To characterize the demand for drought insurance under non-existence of formal markets, a non-market CV approach discussed earlier is used.

Assuming a world with and without food-aid, farmers are presented with a hypothetical insurance market and are asked to respond to CV questions eliciting WTP for area-based drought-index insurance (see appended questionnaire, A.3). Using this approach, we assessed smallholders’ preference for drought insurance in the presence of food-aid as discussed in the sections below.

94 5.2.1 .Estimated Parameter Coefficients using Discrete Choice Models Table 5.4(a), (b) and (c) presents estimated coefficients for the logit and bivariate25 models under with/without food-aid scenarios. The variables included in these models are defined in Table 5.1. In essence demographic variables (hhsex, agricedu), resource variables (acres, income) and regional dummies (D1-D9) are all similar to the discrete models discussed above. However, additional variables, collectively defined as risk factors (foodloss, loanloss, waterloss) are included in these models as well. Risk factors are dummy variables constructed so as capture the impact of drought on household food security (foodloss), loan repayment (loanloss) and household water consumption (waterloss). Of interest is to investigate how the risk factors affect a household’s WTP to pay for drought insurance.

Under with food-aid scenario and across all models, variables agricedu, loanloss, D4, D6 and D8 are statistically significant at α=0.10 level. Under without food-aid scenario variables loanloss, foodloss and dummy variables are significant, but in addition, risk-factor variable foodloss becomes significant (α=0.10). What is of interest is to compare how parameter coefficients shift between with and without food-aid regimes. The discussion below illustrates the impact of key variables on WTP for drought insurance.

First, look at the risk-factor variable, foodloss. Contrary to expectations foodloss variable carries a negative sign. For the logit model, under with and without food-aid scenarios, the marginal effects results imply that the probability that a household will answer ‘yes’ to a CV question on WTP for drought insurance decreases by -4.5% and –17.0% respectively, for a household facing foodloss risk. Extending the same analysis to bivariate model, the probability that a household will respond positively to CV question on WTP for drought insurance is likely to decrease by -6.4% and –20.4% under with and without food-aid, respectively. Hence, as the expectation of food loss, especially for the staple maize increases, households tend to seek less drought insurance protection. This result implies two things: first, food shortage expectation may negatively affect households to the extent of rendering them unable to pay for drought insurance. Because a significant portion of households derives incomes from crop sales, a severe drought will deprive

24 The proper district names are appended at end Table. District #7 was omitted because of inconsistency 25 For the bivariate model, only results obtained using the first bid are discussed since this is the conventional approach commonly used by researchers

95 them the means to pay for insurance. Second, because food losses occur predominantly in the form of staple maize typically grown for own consumption, households may be less inclined to insure a food-crop as opposed to cash-crop (see discussion below).

The marginal effects results on the loanloss variable (Table 5.4), show that, for the logit model under with food-aid situation, the probability that a farmer answers a ‘yes’ response to a CV question on WTP for drought insurance increases by 11.4% for households with access to loans. This probability increases to 21.7% under the without food-aid scenario. For the bivariate models, under with and without food-aid situations, marginal effects results on the loanloss variable show that the probability that a farmer answers a ‘yes’ response to a CV question on WTP for drought insurance increases by 13.9% and 23.2% respectively, for households with access to loans. This result leads to interesting implications: first, because most farmers do not have access to formal loans and credits due to lack of collateral security, drought insurance could substitute for a collateral security requirement. For a farmer who purchases insurance, his/her probability of defaulting on loan repayment is lowered. Banks view unsecured loans to insured farmers as more attractive than loans to uninsured farmers. Hence insurance, like collateral, increases the expected return of the loan. Second, by virtue of comparing foodloss versus loanloss risk factors, it appears like households are more willing to insure a cash- crop rather than food-crop.

Another variable with a significant impact on WTP for drought insurance is agricedu. For the logit model, the marginal effects imply that the probability that a farmer would respond ‘yes’ to CV question on WTP for drought insurance decreases by –9.7% for household who received agricultural training. However, the probability decreases to –7.3% under without food-aid situation. The bivariate model indicates a similar pattern, indicating that the probability decreases by, -11.9% and –8.3% under with and without food-aid, respectively.

96 (a) Logit Coefficient Estimates With Food-aid Without Food-aid Variable Parameter Std error t-value Marginal Parameter Std error t-values Marginal Effects (%) Effects (%) Constant 0.664 0.446 *3.168 - 0.806 0.501 *4.056 - ACRES -0.006 0.029 -0.451 -0.20 -0.020 0.032 -1.571 -1.06 FARMEXP 0.129 0.235 1.17 4.31 0.105 0.266 0.993 5.58 HHSEX -0.017 0.214 -0.172 -0.57 0.106 0.239 1.118 5.64 AGRICEDU -0.291 0.254 *-2.441 -9.72 -0.137 0.283 -1.224 -7.29 FOODLOSS -0.148 0.265 -1.189 -4.50 -0.320 0.311 *-2.595 -17.02 LOANLOSS 0.342 0.397 *1.834 11.4 0.407 0.583 *1.762 21.65 WATERLOSS -0.062 0.224 -0.589 -2.1 0.068 0.255 0.677 3.62 D1 -0.169 0.503 -0.717 - -0.085 0.563 -0.383 - D2 0.028 0.396 0.153 - 0.254 0.472 1.355 - D3 -0.184 0.677 -0.579 - -0.017 0.779 -0.053 - D4 -0.396 0.418 *-2.016 - -0.188 0.458 -1.034 - D5 -0.375 0.410 *-1.947 - -0.037 0.536 -0.176 - D6 -0.091 0.447 -0.432 - -0.275 0.477 -1.457 - D8 -0.638 0.482 *-2.814 - -0.345 0.516 *-1.688 - D9 0.229 0.429 1.135 - 0.142 0.500 0.714 -

(b) Bivariate estimates Constant 0.934 0.269 *3.088 - 1.027 0.306 4.03 - ACRES -0.006 0.018 -0.278 -0.18 -0.024 0.020 -1.404 -1.30 FARMEXP 0.182 0.146 1.104 5.41 0.111 0.165 0.813 6.00 HHSEX -0.043 0.129 -0.293 -1.28 0.123 0.149 0.989 6.64 AGRICEDU -0.400 0.155 *-2.298 -11.90 -0.153 0.165 -1.118 -8.26 FOODLOSS -0.214 0.162 -1.176 -6.36 -0.378 0.195 *-2.333 -20.42 LOANLOSS 0.466 0.233 *1.781 13.86 0.429 0.304 *1.694 23.17 WATERLOSS -0.116 0.137 -0.752 -3.45 0.090 0.157 0.685 4.86 D1 -0.247 0.314 -0.701 - -0.124 0.363 -0.411 - D2 0.007 0.244 0.026 - 0.231 0.288 0.966 - D3 -0.127 0.433 -0.261 - -0.050 0.510 -0.117 - D4 -0.630 0.250 *-2.241 - -0.318 0.283 -1.355 - D5 -0.500 0.247 *-1.799 - -0.101 0.315 -0.385 - D6 -0.140 0.280 -0.445 - -0.374 0.301 -1.494 - D8 -0.858 0.290 *-2.637 - -0.491 0.321 -*1.842 - D9 0.371 0.262 1.261 - 0.132 0.300 0.528 -

Table 5.4: Comparison of Single-bound vs. Double-bound Models

97 5.3. Impact of Seasonal Forecasts on Household WTP for Drought Insurance Following the comparative static presented in section 3.3.1, one objective of this study is to analyze how mitigation costs on seasonal forecasts impact WTP on drought insurance. To investigate this issue empirically, bids on WTP for seasonal forecasts are considered as mitigation costs r and incorporated as one of the covariates in the WTP function for drought insurance. Results in Table 5.5 help highlight the potential impact of mitigation cost under with and without food-aid scenarios. Under both with and without food-aid scenarios, the coefficient accompanying miticost variable has positive sign implying a positive relationship between WTP for drought insurance and WTP seasonal forecasts. Thus as WTP for seasonal forecasts increases, a farmers’ feels inclined to purchase insurance contract as well. These results tend to imply that if farmers have high expectations for a bad year, they may be more willing to invest in both seasonal forecasts and drought insurance. The converse could be true; as illustrated in section 3.3.1 where expenditures on mitigation could be zero for events of low probability. Hence, farmers may choose not to invest in drought mitigation, if the forecasts predict a good year and possibly he could lower the number of purchased insurance contracts.

Logit Model Estimated Coefficients With Food-aid Without Food-aid Variable Parameter Std error t-value Parameter Std error t-values Constant 3.193 0.497 0.432 -65.768 0.029 0.998 ACRES -0.246 0.027 0.890 -1.475 0.232 0.2549 FARMEXP -0.145 0.221 0.005 15.384 0.214 0.4316 HHSEX -1.771 0.211 0.741 15.737 0.255 0.5317 AGRICEDU -5.727 0.252 *5.402 -15.889 0.275 0.381 FOODLOSS -3.361 0.258 *1.781 -50.465 0.372 3.3134 LONLOSS 6.685 0.399 *2.944 -4.818 0.228 0.0165 WATLOSS -2.489 0.221 1.328 18.253 0.071 0.6264 MITICOST 2.361 0.067 12.896 30.919 0.511 18.6466 D1 -6.680 0.454 2.267 19.081 0.438 0.1368 D2 -0.228 0.360 0.004 94.101 0.683 4.5181 D3 0.214 0.639 0.001 74.010 0.436 1.1511 D4 -8.975 0.383 5.743 -22.768 0.442 0.2675 D5 -5.753 0.410 2.062 20.495 0.438 0.2109 D6 -4.075 0.391 1.137 -56.313 0.473 1.6201 D8 -10.831 0.464 5.707 -41.414 0.472 0.7502 D9 104.919 0.652 4.8516

Table 5.5. Impact of Mitigation costs on WTP for drought Insurance

98 5.4 WTP for Drought Insurance One of the pivotal objectives of the study is to investigate farmers’ WTP for drought insurance in the presence of food-aid. In section 3.3.2 we presented comparative statics that looked at the plausible impact of food-aid on the potential demand for drought insurance in a theoretical context. In this section empirical results to this effect are presented.

Tables 5.6 and 5.7 show the calculated WTP for drought insurance across districts and natural regions, respectively. First, under with food-aid scenario, results in Table 5.6(a) show that for the SB model, WTP range from Z$4,095 (district 8) to Z$6,861 (district 9), and from Z$3,926 (district 8) to Z$7,583 (district 9) for the DB.

Second, under without food-aid case (Table 5.6.b), WTP for SB models ranged from Z$5,417 for district 8 to Z$7,054 for district 9. For the DB models it ranged from Z$5,009 (district 8) to Z$7,431 (district 9).

a) With Food-Aid District Natural Single Bounded Double Bounded Model Region WTP(Z$) WTP(Z$)

1.Chiweshe II 5,417 5,678 2.Gutu IV 5,885 6,199 3.Sanyati III 6,187 6,831 4.Chivi V 4,111 3,979 5.MtDarwin IV 4,702 4,974 6.Wedza III 4,209 4,487 8.Shamva II 4,095 3,926 9.Beitbridge V 6,861 7,583 b) Without Food-Aid 1.Chiweshe II 6,028 6,325 2.Gutu IV 6,644 6,931 3.Sanyati III 6,757 6,713 4.Chivi V 5,591 5,660 5.MtDarwin IV 6,680 6,964 6.Wedza III 5,609 5,816 8.Shamva II 5,417 5,485 9.Beitbridge V 7,054 7,431

Table 5.6: Estimated WTP for Drought Insurance across Districts

99 What tends to be predominantly observed across models shown in Table 5.7 is that households located in drier regions of IV and V exhibit higher WTP for drought insurance than their counterparts in wetter districts of II and III. Under, with food-aid scenario, SB(WTP) for households in regions II and III averaged Z$4,756 and Z$5,198, respectively, compared to Z$5,294 and Z$5,486 for regions IV and V, respectively. The DB model showed a similar WTP pattern across regions. Under without food-aid scenario, both models show an increase in WTP ranging from Z$5,723 to Z$6,662 for SB model and from Z$5,905 to Z$6,948 for DB model.

With Food-aid Without Food-aid Natural SB(WTP) DB(WTP) SB(WTP) DB(WTP) Region II 4,756 4,802 5,723 5,905 III 5,198 5,659 6,183 6,265 IV 5,294 5,587 6,662 6,948 V 5,486 5,781 6,323 6,546

Table 5.7: Aggregated WTP for Drought Insurance across NR II-V

Results in Tables 5.8 (a) and (b) show percentage change in WTP for drought insurance across districts and natural regions in the presence of food-aid relative to the without food-aid case. Values accompanied by a negative sign indicate a decrease WTP and vice versa. The largest decrease of about 45% occurred in district 9, while the least decrease of 5.1% occurred in district 3.

Table 5.8(b) summarizes the change in WTP for drought insurance in the presence of food-aid across natural regions II-V. The largest decrease of 36.1% in WTP for drought insurance in the presence of food-aid is observed in region V. In contrast, WTP decrease by 10.6% in region II when food-aid is available. The results imply that the disincentive to purchase insurance in the presence of food-aid will be greatest in drier regions IV and V, and least in wetter regions II and III. Overall, WTP for drought insurance will decrease by 20.1% when food-aid is available. Thus, food-aid discourages farmers from seeking more efficient drought risk protection mechanisms, such as formal drought insurance.

100 a) Change in WTP across Districts (%) District WTP(SB) WTP(DB) Average Change 1 -10.1 -10.2 -10.2 2 -11.4 -10.6 -11.0 3 -8.4 1.8 -5.1 4 -26.5 -29.7 -28.1 5 -29.6 -28.6 -29.1 6 -25.0 -22.9 -24.0 8 -24.4 -28.4 -26.4 9 -41.9 -47.2 -44.6 b) Change in WTP across Natural Regions (%) Overall Change II -10.8 -10.4 -10.6 III -16.7 -10.5 -13.6 IV -20.5 -19.6 -20.1 V -34.2 -38.4 -36.2

Table 5.8: Change in WTP in the presence of food-aid relative to without food-aid case

5.5 Rating Index Insurance across Districts In Chapter 3 (section 3.5), a general methodology for designing and pricing index-based insurance was discussed. This section explores the possibility of implementing the methodology based on a remotely sensed vegetation index (VCI) and rainfall index. Our interest in using VCI (as discussed earlier) stems largely from remote sensing technology’s ability to provide repeated temporal measurements necessary for dynamic monitoring of crop growth, yield assessment and loss estimation. Further, VCI allows wide spatial coverage that may not be possible with meteorological-based indices (e.g., rainfall, soil moisture). In addition, using VCI does not require direct farm inspections nor individual field loss assessment and hence it’s likely to be inexpensive to administer. However, the VCI has its own disadvantages. VCI based insurance may be less be transparent, not easily understood by farmers and may prove difficult to interpret and sell to ordinary farmers. In light of these limitations and for comparison purposes, we use rainfall as an alternative index. Rainfall index is not only transparent to farmers, but also easy to interpret and understand.

For purposes of rating index insurance we need: first, an accurate measurement of crop yields risks, which in turn depends upon proper representation of yield distribution. By measuring yield risk, we are interested in determining the probability that a loss will occur (loss likelihood)

101 and the expected indemnity. This task was accomplished in Chapter 4, section 4.4 where nonparametric density functions were estimated for both VCI and rainfall index.

Second, we need to estimate yield forecasts for the crops (maize and cotton) we are interested in. We use seemingly unrelated regression equations (SUR). The general respective specifications of SUR for VCI and rainfall index are :

= β + β + β 2 + β + β + ε yit oi 1it 2it 3iVCI 4i Nino34 it

= α + α + α 2 + α + ε yit oi 1it 2it 3i Rain _ index it where i=1,…………..,J denote regions/districts, t=1,…………….,T denotes number of total number of observations and yit represents yield in district i for time period t. Further assume ε = ε ε ' = σ that E( i ) 0 and E( i j ) ij IT . The latter assumption underlies the possibility that the disturbances in different equations are mutually correlated. This assumption is plausible given that forces affecting crop yields such as drought are likely to be similar and correlated across space and time. For definition of other variables, see Table 5.1(c). Below we present the results based on trend analysis estimation of SUR coefficients and yield estimation.

5.5.1 Smallholder Cotton and Maize Yield Trend Analysis To capture technological changes in crop yields within the smallholder sector over time, we incorporate a linear (t) and quadratic trend (t2) into SUR models specified above. Table 5.9 and Figure 5.1 shows historical trends in maize and cotton production over a period of 20 years. For maize, there seems to be a weak upward trend with markedly increased volatility, especially after attaining national independence in 1980. Implications are that there has been sluggish technological change within the smallholder farm sector and hence yield levels have hardly changed from what they were at independence 1980. Further, wide fluctuations in maize yield observed especially after independence could imply an unstable policy environment (refer discussion Chapter 1, section 1.4). The lower panel of Figure 5.1 shows trend in cotton production. Besides sharp decreases that occurred in 1991/92 and 1995/96, cotton yield has maintained a strong upward trend. Following the agricultural policy reforms implemented by the government in early 1990s (see Chapter 1, section 1.4), cotton was fully decontrolled unlike maize. Currently, for major production regions

102 such as Hurungwe (district 7), Sanyati (district 3) and Mt Darwin (district 5) cotton is mostly grown under contractual-loan-cum system where, a contracting firm provides the farmer a loan in the form of essential inputs (seed, chemicals for insects and disease control, fertilizers). The contracted farmer on the other hand, is obligated to sell his produce to the farmer at an agreed-to price. This new farming system has plausibly introduced technical innovations possibly accounting for the upward trend.

Season Maize yield Cotton yield National Ave. (MT) (MT) Rainfall(mm) 198081 1000 45 731 198182 595 27 402 198283 285 33 310 198384 670 70 370 198485 1558 110 607 198586 1348 98 598 198687 628 83 333 198788 1609 137 607 198889 1188 123 412 198990 1262 103 538 199091 1019 138 450 199192 115 36 262 199293 1134 135 642 199394 1313 111 448 199495 399 56 372 199596 1687 158 609 199697 1453 198 643 199798 723 183 435 199899 845 188 681

Table 5.9 Trends in Smallholder Maize and Cotton Yield and Rainfall, 1980-99

103 National maize production

1800 1600

) 1400 1200 1000 800 600 Production (MT Production 400 200 0

70 74 76 80 86 90 92 96 98

1969 197172 1973 1975 197778 1979 198182 198384 1985 198788 1989 1991 199394 1995 1997 Seasons

National cotton production

250

200 )

150

100 Production (MT Production 50

2 8 80 778 586 9394 19697019717 197374197576197 1979 198182198384198 198788 19899019919219 19959619979 Seasons

Figure 5.1: National Maize and Cotton Production Trend by Smallholders (1969-1999)

104 As a step further to the trend analysis discussed above, we perform linear restriction tests so as objectively justify the inclusion of the trend variables into SUR models. Two linear tests are conducted: the first test aims to establish an appropriate model to adopt viz. a single-pooled regression model versus the system of SUR equations. Here the null hypothesis (Ho) is: β = β = = β 11 21 ...... j1 β = β = = β 12 22 ...... j2 ...... β = β = = β 1 j 2 j ...... jj With a computed F-statistic and degrees of freedom of F[24,143] =7.95 versus table-value of 1.61 we reject Ho at α=0.05. Hence we conclude that SUR model is a more appropriate model than the single-pooled regression model that can be used to account for cotton yield variation. A similar conclusion is drawn for maize yield.

The second test aims at establishing whether the inclusion of trend variables is statistically justifiable. For both maize and cotton yields, a linear restriction test is performed, such that Ho is β = β = 1 2 0 where these parameter coefficients are associated with the linear (t) and quadratic trend (t2) variables respectively. Using a restricted versus unrestricted SUR model we conducted a F-test where F[18,126]=4.67 versus table value of 1.75 and hence at α=0.05 we reject Ho and conclude that the inclusion of trend variable into SUR model for maize is statistically significant. Using a similar approach for cotton SUR model, we reject Ho as well and hence draw a similar conclusion.

Tables 5.10 and 5.11 show the estimated SUR parameter coefficients based on VCI and rainfall, respectively. Only VCI pertaining to critical yield-determining months of January, February and March (refer to our discussion section 3.6) were used. Similarly, weighted monthly rainfall for December, January, February and March were used. Results in Table 5.10 show that variables VCI and Nino34 are statistically significant at 5% level and, in addition, carry theoretically expected signs. For models pertaining to districts 4 and 7, trend variables, t and t2 were significant at the 5% level. While R2 is satisfactory for districts 8,7 and 4, it is very low for districts 1,3 and 5. Table 5.10 (b) show results based on maize and except for district 3, variables VCI and Nino34 carry the expected signs. However the Nino34 variable does not show strong significance as in the cotton model. R2 ranged from 0.11 to 0.68.

105 Results in Table 5.11 show the estimated SUR parameter coefficients based on weighted- monthly rainfall. The rainfall variable is strongly significant across most regions for both cotton and maize. For some regions (4,5 and 7) trend variables were significant at 5% level. For the cotton crop R2 was appreciably high for districts 4,5,7 and 8, but low for other regions. The maize crop showed a similar trend.

(a) Cotton

β β β 2 β β 2 district 0(constant) 1(t) 2 (t ) 3 (VCI) 4 (Nino34) R

1. -13.6 0.13 -0.002 28.6 -0.589 - (0.89*) (0.14) (0.01) (0.80*) (0.17*) 3 -1.59 0.003 -0.001 13.5 -0.017 - (0.89*) (0.18) (0.01) (0.31*) (0.22) 4 -3.4 0.54 -0.020 10.7 -0.26 0.78 (0.76) (0.15*) (0.01*) (0.59*) (0.18*) 5 -19.6 0.39 -0.004 33.8 -0.253 0.08 (1.78*) (0.35) (0.02) (0.94*) (0.42) 7 1.43 0.357 -0.010 7.59 -0.328 0.83 (0.49*) (0.08*) (0.04*) (0.43*) (0.10*) 8 -0.87 0.15 -0.002 8.82 -.349 0.74 (0.42*) (0.09*) (0.00) (0.10*) (0.11*)

(b) Maize

1 8.48 0.19 -0.008 0.926 -0.001 0.1 (0.82*) (0.18) (0.01) (0.00) (0.24) 2 6.76 -0.022 -0.001 5.43 -0.05 0.39 (0.41*) (0.09) (0.00) (0.00*) (0.12) 3 14.2 0.048 0.000 -8.18 -0.20 0.11 (1.5*) (0.32) (0.02) (0.01*) (0.45) 4 6.95 0.018 -0.002 4.57 -0.107 0.26 (0.65*) (0.14) (0.01) (0.00*) (0.19) 6 -0.11 -0.121 0.009 15.75 -0.207 0.14 (0.62) (0.13) (0.01) (0.00*) (0.19) 7 9.20 0.235 -0.008 1.28 -0.033 0.65 (0.36*) (0.08*) (0.00*) (0.00*) (0.11) 8 1.01 -0.012 0.005 9.13 -0.124 0.51 (0.53*) (0.11) (0.01) (0.00*) (0.16) 9 1.19 -0.006 0.012 8.22 -1.422 0.68 (0.65*) (0.5) (0.02) (0.00*) (0.29*)

Table 5.10: SUR on (a) Cotton and (b) Maize Yield on VCI with Quadratic Trend

106 (a) Cotton

2 2 district β0(constant) β1(t) β2 (t ) β3 (Rainfall) R

1. 7.65 -0.16 0.010 -0.004 - (0.69*) (0.15) (0.01) (0.00*) 3 4.79 0.200 -0.010 0.022 - (0.86*) (0.19) (0.01) (0.00*) 4 1.66 0.70 -0.025 0.004 0.58 (1.08) (0.23*) (0.01*) (0.00*) 5 -3.75 0.64 -0.020 0.040 0.60 (1.23*) (0.26) (0.01*) (0.00*) 7 6.56 0.309 -0.007 0.003 0.87 (0.35*) (0.08*) (0.04*) (0.00*) 8 2.92 0.206 -0.005 0.012 0.72 (0.47*) (0.10*) (0.00) (0.10*)

(b) Maize 1 7.54 2.61 -1.49 0.009 0.40 (0.67*) (1.87) (1.02) (0.00) 2 8.76 0.804 -0.272 0.003 0.33 (3.85*) (3.21) (0.65) (0.00*) 3 1.86 0.60 0.041 0.022 - (40.9) (20.4) (2.51) (0.60*) 4 -15.5 8.98 -0.828 0.828 0.04 (36.6*) (13.1) (1.16) (0.00*) 5 69.5 -19.8 1.49 0.018 0.55 (71.5) (19.9) (1.37) (0.00*) 6 -184.7 43.1 -2.42 0.018 0.28 (87.0*) (19.8*) (1.13*) (0.00*) 7 -114.04 23.3 -1.093 0.009 0.33 (72.5*) (13.9*) (0.67*) (0.00*) 8 16.3 -2.76 0.16 0.009 0.58 (98.7) (16.5) (0.69) (0.00*)

Table 5.11: SUR on Cotton and Maize Yield and Rainfall with Quadratic Trend

107 5.5.2 Yield Estimation The SUR regression parameter coefficients are used to estimate and generate ex-post forecasts for maize and cotton yield. The forecast series are detrended using a method suggested by Miranda26 (1997). For this method detrending factors are defined as: y = kT f kt yˆ kt where is ykt the observed yield in district k and year t.The corresponding predicted yield is given

by yˆ kt . The detrending factors are used to convert district level yields to specific year equivalence. In this case the detrended yields are converted to 1999-2000 yield equivalence. In turn, we use the detrended yield to construct representative agricultural portfolios of main crops viz. cotton and maize. An example of a representative portfolio is shown in Table 5.11 for Chiweshe districts. Local maize prices (discussed in Chapter 4) were assumed non random and fixed. For cotton we did not only assume non-randomness of prices, but uniform across districts. Weights assigned to each crop are calculated as a proportion of total cultivated land allocated for the production of each crop. Similar portfolios are developed for other districts [2 to 9]. Maize is the staple crop grown mostly for household consumption. Cotton, on the other hand, is grown mostly as a cash crop and grows well in drier regions, particularly region IV.

26 for details see Miranda AJAE .79 (February, 1997)

108 Year Revenue Maize Cotton (Z$)

Price per unit = 450 Price per unit = 337.5 Weight = 0.73 Weight = 0.09 Observed Predicted Detrended Observed Predicted Detrended

198081 11,522,037 19291.5 11115.9 34923.2 796.8 812.8,3 1638.5 198283 6,027,868 12218.0 13498.2 18214.6 504.1 577.0 1460.3 198384 13,163,070 31707.0 15973.2 39944.6 990.0 1218.3 1358.3 198485 15,507,350 42832.1 18311.9 47068.6 1217.4 1363.9 1491.9 198586 11,478,441 34168.5 19803.1 34720.8 1153.5 805.9 2392.1 198687 1,944,562 6361.2 22420.8 5709.3 898.3 660.5 2273.3 198788 3,026,017 10615.5 24191.2 8830.4 1675.3 679.1 4123.0 198889 1,846,497 6174.0 22694.3 5474.5 292.8 308.9 1584.1 198990 6,998,833 27903.0 26590.5 21116.4 690.0 564.3 2043.8 199091 7,348,139 30311.3 27471.9 22202.9 540.0 503.3 1793.2 199192 397,442 1620.0 28170.6 1157.2 132.0 387.5 569.4 199293 8,304,751 36998.0 29640.2 25118.4 817.0 777.7 1755.8 199394 7,287,327 33594.0 30590.9 22098.6 693.0 1259.6 919.6 199495 5,783,666 25656.0 29439.3 17537.1 434.0 969.4 748.3 199596 13,766,918 56119.0 27065.9 41723.6 1108.0 926.7 1998.5 199697 12,665,204 50013.0 26437.2 38068.2 3761.0 1194.8 5260.9 199798 12,488,473 46878.0 25011.5 37715.8 1660.0 852.8 3253.3 199899 13,159,285 42875.0 21650.5 39850.3 1517.4 1125.2 2254.0 199900 1,354,725 4050.0 20123.1 4050.0 800.0 1671.4 800.0

Table 5.12: Illustrating detrended Yield for District # 1 for the period, 1980-00

5.5.3 Rating Index-based Insurance To rate insurance contracts using VCI or rainfall index, we need to determine or rather provide values to primal parameters defined as set[x,λi,* i* ] and shown in Figure 5.2. From Figure 5.2, because we are assuming a standard contract with a unit liability, parameter x equals 1.

Parameter λi* defines the critical yield level below which full compensation is guaranteed. For VCI, this value is set at 0.36 and the value underlies an extreme drought year (McVicar and Jupp,

1998). The parameter i* is generally set as the average VCI value obtained during the critical crop growth period, such as the month of February. Hence in sum, unit liability contract is fully specified by the two critical values: 0.36 for λi* and 0.71 for i* . These parameters shall be referred to as deficit min and deficit max, respectively. The value of the deficit max i* , varies depending on region, while the value of deficit min remains fixed.

109 Loss cost

0 λi i* Index

Figure 5.2: Index-based Indemnity Schedule for a Standard Contract

Table 5.12 provides a summary of descriptive statistics for the indices VCI and rainfall across the nine identified districts. Notice that for drier districts such as 4 and 9, mean VCI values ranged from 0.50 to 0.64 compared to wetter districts such as 1 and 7 with mean VCI values ranging from 0.69 to 0.80. We also observe that the standard deviation for VCI is higher in drier districts (e.g., 9) than in wet districts such as 1. This implies that yield variability is higher in drier districts than wet districts. In addition, for most districts, the highest VCI is recorded during the month of February and may imply the most crucial yield determining stage.

110 District Descriptive VCI Rainfall Statistic Jan Feb Mar Apr Jan Feb Mar 1Min.55 .55 .62 .49 62.8 32.4 10.0 Max .78 .83 .81 .79 478.5 390.8 254.9 Mean .69 .76 .75 .67 227.7 219.8 119.6 Stdev .068 .074 .057 .068 105.8 106.2 78.0

2Min.42 .43 .49 .42 22.7 .30 3.6 Max .68 .68 .64 .62 319.7 341.1 119.9 Mean .54 .57 .56 .52 131.5 109.6 57.3 Stdev .068 .074 .057 .068 105.8 106.2 78.0

3Min.57 .58 .64 .51 49.9 3.5 9.7 Max .89 .89 .87 .81 387.4 265.0 204.0 Mean .77 .80 .78 .70 199.1 134.4 84.7 Stdev .078 .083 .070 .085 100.9 100.2 63.1

4Min.43 .45 .47 .42 22.7 .3 3.6 Max .77 .74 .79 .70 319.7 341.1 119.9 Mean .60 .64 .63 .59 131.5 109.6 57.3 Stdev .086 .086 .090 .082 93.3 93.9 41.6

5Min.57 .57 .55 .49 37.4 13.0 .70 Max .77 .79 .81 .77 452.8 394.6 287.3 Mean .66 .73 .73 .64 213.3 203.2 90.7 Stdev .063 .054 .055 .072 101.4 114.2 86.3

6Min.49 .53 .57 .43 26.5 7.1 12.6 Max .74 .74 .70 .66 415.9 374.8 205.1 Mean .61 .65 .64 .59 202.9 152.7 85.9 Stdev .062 .060 .043 .058 107.7 101.8 56.7

7Min.53 .58 .62 .55 67.6 60.5 29.1 Max .85 .87 .87 .83 365.5 427.6 228.9 Mean .75 .80 .78 .69 194.1 194.8 109.0 Stdev .077 .068 .065 .082 84.6 91.1 58.2

8Min.57 .60 .57 .57 78.5 .43.2 22.4 Max .77 .81 .79 .77 433.2 295.6 236.6 Mean .67 .73 .74 .66 230.5 179.7 104.3 Stdev .058 .062 .049 .057 95.1 91.0 68.4

9Min.34 .36 .36 .26 .0 2.2 3.7 Max .81 .74 .77 .68 294.4 394.2 119.9 Mean .54 .56 .54 .50 69.2 68.4 32.2 Stdev .134 .121 .124 .113 68.5 92.8 29.1

Table 5.13: Descriptive Statistics for VCI and Rainfall Index by Districts 1=Chiweshe; 2=Gutu; 3=Sanyati; 4=Chivi; 5=Mt Darwin; 6=Wedza; 7=Hurungwe; 8=Shamva; 9=Beitbridge

111 Rainfall drought index is constructed as a historical deficit from the seasonal mean computed over a 20- year period. For example if pth district has the largest precipitation deficit recorded as

70% below historical precipitation mean, then deficit min, λi* 27equals –0.70. In other words, if for any given season precipitation is 70% below the historical mean, then full compensation is paid.

Here we assume that precipitation is generated by a series that is stationary in mean and variance. The deficit max is fixed at 1.1, meaning if seasonal rainfall is 10% above average, there is no payment. This structure admits contracts that may be used to manage portfolio risk in regions that suffer losses primarily due to droughts and not losses due to excess rainfall. Though hedging against flood risk could be important too, our prime concern in this study is to insure losses due to drought risk.

To assess the viability and/or feasibility of index-based insurance, we seek a contract that maximizes the correlation, ρ between losses and indemnities, while simultaneously possessing reasonable pure premium rate, π which lies between 5-10%. High correlation is desirable, since it underlies the ability of index to ‘mimic’ the losses of the underlying variable of interest and hence minimizes basis risk. A higher pure premium rate could prove too expensive, especially to the rural poor, who essentially are the target for the insurance scheme.

While numerous approaches could be used to determine the optimal parameters (see discussion in Chapter 3), here we use a spreadsheet program to search for the desired optimal parameters at a specified loss criterion rate28. Once the user specifies the primal parameters (i.e., deficit min and deficit max) the program will: (a) compute the associated premium rate, b) compute the maximum correlation that can be achieved between portfolio losses and the optimal number of unit contracts, and c) graph portfolio losses versus the indemnity provided by the optimal number of contracts. This procedure allows us to search for a viable/feasible contract, meaning one that maximizes the correlation between losses and indemnities, while possessing reasonable pure premium rate of 5-10%. Thus, the results generated by the program includes: (i) the graph of losses versus indemnities paid when the optimal number of contracts is purchased and (ii) three

27 Parameter λ in this case has no apparent interpretation 28 Loss criterion defines a proportion at which an insured farmer is presumed to suffer a loss. For example for loss criterion rate of 0.75, the insured suffer a loss if his realizes a yield of 0.75 of normal yield.

112 important variables: the correlation ρ * between the contract indemnity and portfolio losses, the

optimal number of contracts to be purchased, N * , and pure premium rate for the contract,π * .

Table 5.14 shows the rating of area-yield insurance contracts based on remotely sensed VCI. Different loss criterion rates of 0.65, 0.75 and 0.90 are used across all districts. The primal parameters used to search for the optimal contract are shown in the last three columns. The districts 2, 4 and 9 showed highest correlation greater 0.80. The districts 1,3 and 6 show appreciable correlation ranging from 0.51-0.70. The remaining districts of 5, 7 and 8 showed low correlation ranging between 1-40%. With respect to the premium rates, most districts recorded acceptable rates within reasonable range 5-10% except district 9. The premium rate tends to be higher in drier regions (III, IV and V) than in wettest regions (II) except a few districts. The least number of contracts N (8,800) were recorded in the district 9 where premium rates were highest at 19.2% and hence too expensive for most households. In contract, largest number of contracts (601,700) was recorded in district 6 with lower premium rate of 2.50. This signifies that as premium rates decrease, insurance becomes cheaper and hence more contracts are bought.

District NR Contract Parameters

ρ # of Contracts Lambda Min Max (‘000) at different loss λ Deficit Deficit π at ( ) different loss rate rates .65 .75 .90 .65 .75 .90

1 II 4.79 0.58 0.58 0.60 114.4 138.5 174.6 0.54 0.36 0.72 2 IV 2.68 0.90 0.91 0.90 113.1 131.3 158.9 0.51 0.36 0.53 3 III 6.41 0.51 0.53 0.56 155.8 196.7 258.3 0.46 0.36 0.79 4 V 7.85 0.80 0.80 0.78 79.2 93.3 114.3 0.61 0.36 0.59 5 IV 5.30 0.29 0.32 0.30 31.7 38.9 49.5 0.51 0.36 0.71 6 III 2.50 0.66 0.62 0.59 371.1 463.6 601.7 0.60 0.36 0.60 7 II 6.32 0.09 0.09 0.12 85.1 108.9 166.1 0.46 0.36 0.79 8 II 2.22 0.30 0.32 0.32 49.1 58.1 71.5 0.53 0.36 0.68 9 V 19.18 0.84 0.86 0.88 8.8 10.2 12.3 0.71 0.36 0.51

Table 5. 14: Rating Area-yield Insurance Using VCI

113 The results in Table 5.15 shows the rating of area-yield insurance contracts across nine districts based on rainfall index. As the loss rate is increased from 0.75 to 0.90 while maintaining premium rate constant, we observe marginal increase in correlation across all districts. In addition as loss rate increases, the number of contracts increases by 15-20% across most districts. Thus as more and more coverage is provided at the same premium rate, more contracts are bought. Districts 2, 4,6, and 7 the rainfall index showed strong correlation that ranged from 0.70-90. The districts 8 and 9 showed modest correlation that ranged from 0.56-0.64 while districts 1 and 5 displayed lowest correlation ranging from 0.32-0.45. Low correlation is undesired, because this leads to a basis risk problem. Most districts recorded premium rates within reasonable range except district 9 with the highest rate of 19.6. This result corroborates our earlier observation in Table 5.14 and hence re-emphasizing the finding that index insurance may not be possible in district 9. The number of contracts N were lowest at 10,900 for district 9 and highest at about 1.2 million in district 7. The reason could be associated with high (expensive) and low (cheap) premium rates respectively.

Compared to VCI, the rainfall index showed better correlation for districts 7 and 8 whose correlation was improved from 0.12 and 0.32 to 0.82 and 0.60, respectively. However, for districts 1 and 9, VCI showed better correlation compared to the rainfall index. Both indices showed rather poor correlation for districts 3 and 5. For the remaining districts, the correlation was close.

District NR Contract Parameters

ρ # N Min Max at different loss Deficit Deficit π at different loss rate rates .75 .85 .90 .75 .85 .90

1 II 2.08 0.32 0.33 0.36 175.5 210.8 232.9 0.55 1.1 2 IV 4.24 0.87 0.89 0.90 84.1 97.8 104.6 0.50 1.1 3 III 7.47 0.39 0.41 0.45 178.8 215.0 234.0 0.85 1.1 4 V 7.66 0.83 0.85 0.85 113.9 133.6 143.4 0.60 1.1 5 IV 7.63 0.41 0.44 0.45 27.3 32.5 35.2 0.80 1.1 6 III 4.55 0.70 0.73 0.76 329.6 403.7 442.2 0.65 1.1 7 II 2.54 0.85 0.82 0.82 902.5 1,100.0 1,200.0 0.65 1.1 8 II 6.39 0.56 0.59 0.60 39.8 46.4 49.7 0.80 1.1 9 V 19.57 0.61 0.63 0.64 10.9 12.5 13.4 0.80 1.1

Table 5.15: Rating Area-yield Insurance Using Rainfall Index

114 The comparison of these indices is further illustrated in Figures 5.3-5.11 and Table 5.9. Figures 5.3 to 5.11 graphs the portfolio losses versus indemnities paid when the optimal number of contracts is purchased.

In Table 5.9, we classify drought into three categories: extreme, very bad and bad. This classification is arbitrarily based on average precipitation recorded during the growing seasons of 1980-2000. If rainfall is lower than 300 mm, the season is classified extreme, if in the range 301-350 it is classified very bad and if the range is 351-400 is classified as bad. From Table 5.9 we also observe concomitant decrease in yield for both maize and cotton during drought years. While it would be desirable to hedge against all forms of drought, practically this would be difficult given drought is a systemic risk. Hence the objective is to offer insurance protection against extreme and possibly very bad drought events. Therefore, in Figures 5.3-5.11 we are interested in an index which principally traces the extreme (1991/92), very bad (1982/83, 1986/87,1994/95) and bad (1988/89, 1997/98) events. A look at these figures indicates that both indices perfectly traced the extreme drought event of 1991/92. However, the indices missed drought events classified as bad and very bad. For the rainfall index, except for district 6, it was able to correctly trace the very bad drought events. However, the rainfall index failed to trace the bad events. For the VCI it was able to trace the very bad events almost accurately in seven out of nine districts. Unlike the rainfall index, VCI was also able to pick bad events in four out of nine districts. Hence, if the indices are used to hedge against extreme drought events only, basis risk is reduced to zero. Basis risk becomes an issue of grave concern, if index contract covers drought events of high to moderate (bad-very bad) intensity.

115 Season Maize yield Cotton yield National Ave. Classification (MT) (MT) Rainfall(mm) 198081 1000 45 731 198182 595 27 402 Bad 198283 285 33 310 Very bad 198384 670 70 370 Very bad 198485 1558 110 607 198586 1348 98 598 198687 628 83 333 Very bad 198788 1609 137 607 198889 1188 123 412 Bad 198990 1262 103 538 199091 1019 138 450 199192 115 36 262 Extreme 199293 1134 135 642 199394 1313 111 448 199495 399 56 372 Very bad 199596 1687 158 609 199697 1453 198 643 199798 723 183 435 199899 845 188 681

Table 5.16: Drought Classification

116 (a)

Por tf olio Los s Optimal Indemnity

7000.00 6000.00

5000.00 4000.00 3000.00

Loss Z$('000) 2000.00

1000.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Portfolio Loss Optimal Indemnity

6000.00

5000.00

4000.00

3000.00

2000.00 Loss Z$('000)

1000.00

0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

Figure 5.3: Portfolio loss vs. Optimal Indemnity for District #1 using (a) VCI and (b) Rainfall Index

117 (a)

Portf olio Loss Optimal Indemnity

4500.00 4000.00 3500.00 3000.00 2500.00 2000.00 1500.00 Loss Z$('000) 1000.00 500.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Portfolio loss Optimal Indemnity

Figure 5.4: Portfolio loss vs. Optimal Indemnity for District # 2 using (a) VCI and (b) Rainfall Index

118 (a)

Portfolio Loss Optimal Indemnity

10000.00 9000.00 8000.00 7000.00 6000.00 5000.00 4000.00

Loss Z$('000) 3000.00 2000.00 1000.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Por tf olio los s Optimal Indemnity

10000.00 9000.00 8000.00 7000.00 6000.00 5000.00 4000.00

Loss Z$('000) 3000.00 2000.00 1000.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

Figure 5.5: Portfolio loss vs. Optimal Indemnity for District # 3 using (a) VCI and (b) Rainfall Index

119 (a)

Portfolio Loss Optimal Indemnity

6000.00

5000.00

4000.00

3000.00

2000.00 Loss Z$('000)

1000.00

0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Portfolio loss Optimal Indemnity

7000.00

6000.00

5000.00

4000.00

3000.00

Loss Z$('000) 2000.00

1000.00

0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

Figure 5.6: Portfolio loss vs. Optimal Indemnity for District # 4 using (a) VCI and (b) Rainfall Index

120 (a)

Portfolio Loss Optimal Indemnity

1800.00 1600.00 1400.00 1200.00 1000.00 800.00 600.00 Loss Z$('000) 400.00 200.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Portfolio loss Optimal Indemnity

Figure 5.7: Portfolio loss vs. Optimal Indemnity for District # 5 using (a) VCI and (b) Rainfall Index

121 (a)

Portfolio Loss Optimal Indemnity

16000.00 14000.00 12000.00 10000.00 8000.00 6000.00 Loss Z$('000) 4000.00 2000.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Por tf olio los s Optimal Indemnity

Figure 5.8: Portfolio loss vs. Optimal Indemnity for District # 6 using (a) VCI and (b) Rainfall Index

122 (a)

Por tf olio los s Optimal Indemnity

(b)

Portfolio Loss Optimal Indemnity

16000.00 14000.00 12000.00 10000.00 8000.00 6000.00

Loss ZZZ$('000) 4000.00 2000.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

Figure 5.9: Portfolio Loss vs. Optimal Indemnity for District # 7 using (a) VCI and (b) Rainfall Index

123 (a)

Portfolio Loss Optimal Indemnity

(b)

Portfolio loss Optimal Indemnity

Figure 5.10: Portfolio Loss vs. Optimal Indemnity for District # 8 using (a) VCI and (b) Rainfall Index

124 (a)

Portf olio Loss Optimal Indemnity

1200.00

1000.00

800.00

600.00

400.00 Loss Z$('000)

200.00

0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

(b)

Portfolio loss Optimal Indemnity

1000.00 900.00 800.00 700.00 600.00 500.00 400.00

Loss Z$('000) 300.00 200.00 100.00 0.00 198081 198283 198384 198485 198586 198687 198788 198889 198990 199091 199192 199293 199394 199495 199596 199697 199798 199899 199900 Seasons

Figure 5.11: Portfolio Loss vs. Optimal Indemnity for District # 9 using (a) VCI and (b) Rainfall Index

125 CHAPTER 6

SUMMARY AND CONCLUSION

From Chapter 1, this study has three objectives. The first objective was to investigate whether a policy that encourages adoption of improved seasonal climate forecasts by smallholder farmers is a more efficient drought mitigation strategy. The second objective investigated the potential demand for area-yield drought-index insurance in the presence of free publicly provided food- aid. The third and final objective investigated the feasibility of offering index-based drought insurance to smallholder farmers on the basis of remotely sensed vegetation index, VCI versus rainfall index. Below we provide a summary and conclusion of each objective. Finally we discuss the weakness of the study and scope for further research.

Summary First objective: Do seasonal forecasts really matter for smallholders? Lack of basic farm resources, such as draft power and disc plough, tends to inhibit full utilization of seasonal forecasts by smallholder farmers. With more than 20% of households not owning enough draft power nor disc plough, such resource-constrained farmers fail to take full advantage of the information on seasonal forecasts. The alternative could be to hire draft or tractor power but, however, this is often too expensive for most farmers. Worse still, the hired power may not be available at the appropriate time that could enable the farmer to make maximum use of the forecasts. In addition, results on marginal effects indicate that the probability of getting a positive response to WTP for seasonal forecasts decreases by 0.39% for a household without adequate draft power.

Further, because a radio is the main media of communicating the forecasts and with half of the smallholders not owning a working radio, most farmers face the risk of either getting distorted message via ‘informal discussion’ or worse still not getting the information at all. The 2003/04 season is a case in hand where less than 5% of sampled households actually got the ‘right forecast’ message. The marginal effects indicate that the probability of WTP for seasonal forecasts

126 increases by 0.68% for a household owning a working radio. Hence, a working radio is an important asset if a farmer is to make full use of seasonal forecasts.

Seasonal forecasts suffer credibility problem as more than 40% households indicated outright lack of confidence with the forecasts. An improvement to current forecasts could be to offer region-scaled seasonal forecasts. Region-scaled forecasts are likely to be of immense value to the farmer, as opposed to current forecasts that encompass wide geographical regions, hence too general to be believed. In addition, little involvement of extension workers as a channel of communicating and interpreting seasonal forecasts tends to fuel the ineffectiveness of the forecasts.

Perhaps the most important results pertain to the estimated WTP for improved seasonal forecasts. The estimated WTP based on a single-bound model, ranged from Z$2,427 (district 8) to Z$4,676 (district 5). For a double-bound model, WTP ranged from Z$2,532 (district 8) to Z$4,225 (district 5). What was rather interesting to observe was the differential WTP pattern across districts. Households located in wet districts (e.g., 1 and 8) showed consistently lower WTP than those in drier districts (e.g., 5 and 9). When aggregated by natural regions, WTP for households whose districts are located in natural region II was 36.0% and 30.0% lower than those located in regions IV and V, respectively. A similar pattern was observed for households in natural region III whose WTP was 17% and 9.3% lower than in regions IV and V, respectively. Because the perceived drought risk is more ominous in drier districts located in regions IV and V, households in these regions are willing to pay more for the provision of improved seasonal forecasts.

Second objective: Does drought insurance really matter for smallholder farmers? Preliminary analysis showed that 75% of the sampled households received food-aid during the 2003/04 season. On average each household received about 275.5 kg which translates to about six 50kg bags of maize. Except for a few districts (2,3 and 6), food-aid was predominantly provided in the form of maize. Most households (36.9%) indicated inadequacy as the main problem affecting food-aid programs, while a few pointed corruption (12.9%), erratic supply (7.2%) and lack-of-targeting (6.7%) as key problems.

The impact of drought on smallholders’ economies is often severe and long lasting. The results identified those loss risk factors that severely impacts smallholders whenever drought occurs.

127 Most common risk factors picked by the households included crop loss (65.1%), cattle loss (60.3%) and food security loss (58.0%). Other important risk factors included hunger-related diseases (30%), water scarcity (28%), inability to send children to school (25%) and inability to repay loans (10%). A majority (34.1%) ranked food security loss as the most severe risk factor they face. Crop loss was ranked second at 19.6%, with cattle loss and water loss ranked third (13.2%) and fourth (12.4%), respectively.

To fend their households against the food insecurity risk factor, farmers have developed own coping strategies. More than 60% indicated buying more grain as the most dominant strategy. Other important strategies included seeking off-farm employment (49.8%), selling of livestock (38.1%) and relying on food-aid (28.7%). Less than 1% indicated formal insurance as a coping strategy, hence emphasizing absence of formal insurance within the smallholder farm sector. Upon ranking these strategies, majorities ranked buying more grain (26.6%) as the most reliable coping measure, while seek off-farm job and sell livestock were ranked second and third respectively. A mere 5% selected food-aid as a reliable strategy.

The result here have important implications; first, despite food-aid programs having been in existence for decades, most households view it as an unreliable drought coping strategy. In fact buying-more grain was ranked the most reliable strategy. Thus, second, if farmers are using their own savings to buy more grain so as cope with drought equally, farmers can be encouraged to buy insurance contracts with the potential to offer more effective protection especially against extreme drought events.

To gain more insights on potential demand for drought insurance by smallholders, WTP function is specified and estimated using single-bound logit and double-bound bivariate-probit models. The models were estimated under two scenarios with and without food-aid. Here our interests were centered on analyzing the impact of loss risk factors on WTP for drought insurance. A priori we anticipate that as the expectation of loss increases, farmers seek more protection in an effort to hedge against the risk threat. Results on the loanloss variable tend to agree with our a priori beliefs. For the single-bounded models, the probability of a ‘yes’ response to CV question on WTP for drought insurance increases by 34.2% for households with access to loans. The probability further increases to 40.7% in the absence of food-aid. This result leads to interesting implications: first, because most farmers do not have access to formal loans and credits due to lack of collateral security, drought insurance could substitute collateral security requirement.

128 For a farmer who purchases insurance, his/her probability of defaulting on the loan repayment is lowered. Banks view unsecured loans to insured farmers as more attractive than loans to uninsured farmers. Hence insurance, like collateral, increases the expected return of the loan. Second, by virtue of comparing foodloss versus loanloss risk factors, it appears like households are more willing to insure a cash- crop rather than food-crop.

Results on foodloss risk-factor variable did not conform to our a priori beliefs. In fact for the SB model, the probability that a farmer responds positively to the CV question on WTP for drought insurance decreases by -4.5% for a household facing food loss risk under with food-aid scenario. Under without food-aid case, this probability decreases by -17.0%. Extending the same analysis to bivariate models, the probability decreased from -6.4 and –20.4% under with and without food-aid, respectively. Hence, as the expectation of food loss especially the staple maize increases, households tend to seek less drought insurance protection. This result implies two things: first, food insecurity may negatively impact households to the extent of rendering them unable to pay for drought insurance. Because a significant portion of households derives incomes from crop sales, a severe drought will deprive them the means to pay for insurance. Second, because food losses occur predominantly in the form of staple maize, households may be less inclined to insure a food-crop as opposed to cash-crop.

Results on the loanloss variable (Table 5.4), show that, for the logit model, the probability of a ‘yes’ response to the CV question on WTP for drought insurance increases by 11.4% for households with access to loans. The probability further increases to 21.7% in the absence of food-aid. For the bivariate model, the probability of a ‘yes’ response to the WTP question increases from 13.9 (with food-aid) to 23.3% without food-aid. This result leads to the following implications: first, because most farmers do not have access to formal loans and credits due to lack of collateral security, drought insurance could substitute collateral security requirement. For a farmer who purchases insurance, his/her probability of defaulting on loan repayment is lowered. Banks view unsecured loans to insured farmers as more attractive than loans to uninsured farmers. Hence insurance, like collateral, increases the expected return of the loan. Second, by virtue of comparing foodloss versus loanloss risk factors, it appears like households are more willing to insure a cash- crop rather than food-crop.

One of the pivotal objectives of the study was to investigate farmers’ WTP for drought insurance in the presence of food-aid or rather how food-aid is likely to impact the potential demand for

129 drought insurance. Results showed that in the presence of food-aid, WTP for drought insurance decreases by more than 35% for households in regions IV and V, while it declined by less than 12% for households in regions II and III. The results imply that disincentive to purchase insurance in the presence of food-aid is greatest in drier regions IV and V and least in wet regions II and III. Overall demand for insurance is likely to decrease by more than 20% in the presence of food-aid. Thus, food-aid discourages farmers from seeking more efficient drought risk protection mechanisms, such as formal drought insurance.

Third objective: Would drought-index insurance be viable and/or feasible? To assess the viability and/or feasibility of the index-based insurance, we seek a contract that maximizes the correlation, ρ between losses and indemnities, while simultaneously possessing reasonable pure premium rate, π of 5-10%. High correlation is desirable, since its underlies the ability of the index to track losses of the underlying variable of interest. High correlation between index and actual losses of variable of interest minimizes basis risk. A pure premium rate within 5-10% range is actuarially sound and yet cheap enough to attract participation of the rural poor.

Except for 2/3 districts, both indices VCI and rainfall showed modest to high correlation that ranged from 0.60-0.90. Further, most districts had premium rates within the desired range of 5- 10% except district 9. In addition, both indices were able to perfectly trace extreme drought events but tended to miss on those events of lesser intensity. In as far as protecting against extreme drought events, VCI-based or rainfall-based contract could be sufficient. However for drought events of high-moderate intensity both indices tend to miss and basis risk becomes an issue of concern.

Conclusion Seasonal forecasts have potential to help the farmer as a drought mitigation tool, but the usefulness could be enhanced if possibly the forecasts are region-specific or region-scaled. Current forecasts tend to be too wide and too general to appeal to smallholders. Resource- constrained farmers, especially those, who lack draft power, disc plough and working radio, are severely handicapped and hence fail to use the forecasts to maximum benefit. If the current rural DDF (district development fund) tractor-program is expanded and timeously available it could go a long way towards addressing acute draft power shortages.

130 Food-aid discourages farmers from seeking more efficient drought risk protection mechanisms such as formal drought insurance. The disincentive to purchase insurance in the presence of food-aid was greatest in drier regions IV and V and least in wet regions II and III. Overall, demand for insurance decreased by more than 20% in the presence of food-aid. Furthermore, farmers, especially cotton growers who have access to formal loans and credits albeit limited, expressed a higher WTP for drought insurance in the absence of food-aid. This implies that since most farmers lack collateral security, they possibly view drought insurance substituting requirement for collateral security.

VCI showed appreciably high correlation coefficients sufficient to consistently track yield losses and quite comparable to the rainfall index. In addition, most districts showed reasonable premium rates actuarially sound and inexpensive enough to attract participation of the rural poor. In as far as hedging against extreme drought events, VCI-based contract could be sufficient. Basis risk becomes an issue if VCI and the indices are used to protect drought events of moderate intensity.

Further research may include use of satellite with high spatial and temporal resolution. While NOAA ‘s AVHRR has high temporal resolution, it has low spatial resolution that may not allow image specification. Combining LANDSAT, with a high spatial resolution, and AVHRR with high temporal resolution may do the trick. A better understanding of yield/VCI and yield/rainfall relationships for specific crops and/or region would also be useful. Finally, it may be useful to develop an index that combines satellite-based and meteorological-base information.

131 APPENDICES

Appendix A Drought-related Impact Sector affected Impact Economic • Loss from crop production: - Reduced crop yield - Reduced crop quality - Increase in insect infestation and plant disease • Loss from livestock production: - high/cost/unavailability of water for livestock - high/cost of feed for livestock - high livestock mortality rates - depletion of livestock for draft power - forced liquidation/marketing reduces price • Retards economic growth - Income loss to farmers - loss to industries directly dependent on agricultural production - cost to energy industry - increased energy demand and reduced supply because of drought-related power curtailment - decline in food production due to disrupted food supply - strain on financial institutions (greater credit risk) - revenue loss to government due reduced tax base Environmental • Damage to animal species and wildlife - lack of feed and drinking water - increased vulnerability to predation and disease • Wind and water erosion of soils • Damage to fish species • Damage to plant species

Social • Food shortages - Decreased nutritional level - Malnutrition - Famine - loss of human life • Food-aid emergencies - Inequity distribution - Costly to the government and taxpayer - Increased poverty - Decreased standard of living in rural areas - Reduced quality of life - Social unrest - Population migration (rural to urban areas) Source: Adapted from Wilhite, 1993

132 Appendix B

Mathematical Derivations Taking FOC with respect to α : −π * (1− Φ) ′ g U (.) = L(1− δ ) ls ′ Φ(1−π * ) U g nl −π * ′ g Φ U (.) ⇒ = L(1− δ ) ls − Φ ′ 1−π * 1 U (.) g nl Invert and simplify 1−π * ′ g 1 1− Φ U (.) ⇒ = nl −π * L(1− δ ) Φ U ′ (.) g ls ′ 1 1 1− Φ U (.) ⇒ 1− = nl π * L(1− δ ) Φ U ′ (.) g ls ′ 1 1 1− Φ U (.) 1 1 ~ ~ = 1− nl ⇒ = 1− ΦU π * L(1− δ ) Φ U ′ (.) π * L(1− δ ) g ls g where ~ 1− Φ ~ U′ (.) Φ = and U = nl Φ ′ U ls (.) ~ ~ 1 L(1− δ ) − ΦU * L(1− δ ) ⇒ = ⇒ π = ~ ~ π * L(1− δ ) g L(1− δ ) − ΦU g

∂ log(π ) ~ ~ ⇒ < for ΦU > L(1-δ ) ∂δ

133 Appendix C Existence of long-run equilibrium

From (37)

~ {(1− Φ)[(1− Φ) 2 λδσ 2 + ΦL(δ −1)]}2 V = L σ 2ψρ 2 L

~ ∂ (2(1− Φ) 2 λσ 2 + ΦL) (1− Φ)[(1− Φ) 2 λδσ 2 + ΦL(δ −1)] V = L L ∂δ σ 2ψρ 1 2 L 1424444444444 + 434444444444

~ ∂V ⇒ > 0 ∂δ

From (36)

~ 1 1 1 1 V = n e (α e ) 2 − n eα e [(1− Φ)(π e − ΦL − c) +γΦL] ψ σ 2 −φ −γ 2 + e − ρ L (1 ) [1 (n 1) ]

~ ∂V 1 1 2 1 1 1 1 ΦL = −n eα e [(1− Φ)(π e − ΦL − c) +γΦL]− n eα e ∂γ ψ σ 2 −φ −γ 3 + e − ρ ψ σ 2 −φ −γ 2 + e − ρ L (1 ) [1 (n 1) ] L (1 ) [1 (n 1) ]

~ ∂V 1 1 2 1 1 1 1 ΦL = − e α e [(1− Φ)(π e − ΦL − c) +γΦL]− n eα e ∂γ ψ σ 2 −φ −γ 3 + e − ρ ψ σ 2 −φ −γ 2 + e − ρ L (1 ) [1 (n 1) ] L (1 ) [1 (n 1) ] 14244444444444 + 4344444444444 142444444 + 43444444

~ ∂V ⇒ < 0 ∂γ

134 Appendix D QUESTIONNAIRE

Household Survey on WTP for Improved Seasonal Climate Forecasts and Drought Insurance Program for Smallholder Farmers in Zimbabwe Conducted by: Department of Agricultural Economics and Extension University of Zimbabwe in collaboration with Department of Agric. Environment and Development Economics, Ohio State University, USA Harare; Tel., 011-263-4-303311 Date:………………………………………………………………………………………... Questionnaire ID:…………………………………………………………………………... Enumerator:………………………………………………………………………………… Village:………………………………………… Ward…………………………………... District:…………………………………………………………………………………….. Province:……………………………………………………………………………………

INTRODUCTORY REMARKS: I am grateful to have you as one of my participants in this study exercise being conducted by the University of Zimbabwe, Department of Agricultural Economics, with the approval of Ministry of Agriculture and your local authority. The study is being conducted throughout rural sector where smallholder farmers like you are randomly surveyed. In brief the study aims to investigate the potential of two policy programs (a) a policy that calls for the adoption of improved seasonal forecasts as drought mitigation strategy and (b) a policy that advocates for drought insurance program as a risk management tool (whose details I discuss later).

GENERAL BACKGROUND: How much total farm land do you have? Acres…………………………………………………………………………………….. Hectares………………………………………………………………………………… Other (specify)………………………………………………………………………… Approximately how much land do you cultivate each year? Acres…………………………………………………………………………………….. Hectares………………………………………………………………………………… Other (specify)…………………………………………………………………………

Record 100: Demographic information Are you Sex Res/non Relation Number of dependents Occupation Highest Did you receive any the M/F resident to head of Under: Educ. agricultural training? HEAD of HH Level HH? Age<5 5

135 7=peasant farmer 8= other (specify)

1.MF= Master Farmer 2. DA=Diploma in Agriculture 3. AMFC=Advanced Master Farmer Certificate 4. LC=Livestock Certificate

Record 110: Inventory of Assets Owned: (a) Assets Which of these farm assets do you own? Is it working? Indicate by placing a tick Yes No Ox-plough Tractor Ox-cultivator Harrow Hoes Tractor Radio Television Other (specify)

Record 120: Draft Power Availability and Cost: (a) Cattle Do you If YES how many? ownership own cattle: ______YES NO (b) Draft power What type of draft Did you hire draft power for farming purposes this season 2003/2004? power did you use for this season Yes No 2003/04?

If YES what draft What was hiring cost $/UNIT? power did you hire? a)with fuel b)without fuel c)Total d) other cost ($) (specify) Ox-power Donkey-power Tractor power Other (specify)

Record 130: FOOD-AID Did your household receive FOOD-AID last year 2003? Yes______[if yes fill Table 13a below] No______[If NO check possible reason for not receiving?]

Not available in my area Did not meet recipient Chose not to participate New comer in the area Other (specify) selection criteria

If YES FILL the table below: What type of grain did Check approx. Length of period that your Estimate TOTAL quantity Did you pay anything you receive quantity received household received FOOD- of FOOD-AID received for in order to be a Per period AID the period indicated recipient? [DON’T ASK FARMER CALC YOURSELF] YES NO 1.Specify grain-type If YES how much per received month OR year 1. 5kg $/month OR $/year 2. 10kg

136 3. 20kg 4. 50kg (=1bag) 5. 5bags 6. 10bags 7. Other (specify)

2.Other benefits (specify) 1. 5kg $/month OR $/year 2. 10kg 3. 20kg 4. 50kg (=1bag) 5. 5bags 6. 10bags 7. Other (specify)

Record 140: Problems Associated with FOOD-AID In your opinion what are some of the PROBLEMS affecting the FOOD-AID program? Place a tick on the applicable: Corruption/ Inadequacy/ Erratic supply/ Lack of targeting/ Other (specify) Favoritism not enough Non consistent supply screening

Record 150: Historical Yield distribution For how many years have you been farming in this region? ______years. Given your farming experience I would like to know your yield during (a) best years and (b) worst years. If you can think of the past 5-10 years and to the best of your knowledge: how much did you harvest? Crop type Approx. harvest Approx. Size of (50 kg bags) harvested land (acres) a) Maize 1. BEST years 2. WORST years b) Cotton 1. BEST years 2. WORST years c) Tobacco 1. BEST years 2. WORST years d) Other (specify) 1. BEST years 2. WORST years

Record 160: Local Maize Selling Prices Indicate locally and currently prevailing maize selling prices per unit (e.g. 20-litre bucket) being charged by following: Local GMB depot price Local price by other farmers Private trader (from Other ($/unit) ($/unit) outside your area) sources(specify) ($/unit) ($/unit)

137 Record 170: Council and Cattle Dipping Fees In this record indicate whether the farmer is actually paying taxes OR is aware of ANY tax obligations he must pay either to the council or to a locally community-based organization by asking the following questions: RESPONSE Are you paying ANY Are you paying Dipping Are you a member of any Are you a member of taxes to the local fees for your cattle? locally available Burial ANY other locally council? society? available society?

YES NO If YES how $______/month $______/month $______/month $______/month much? $______/year $______/year $______/year $______/year

IMPROVED SEASONAL CLIMATE FORECASTS Record 200: Long-term Seasonal Forecasts The Department of National Met Services issues seasonal forecasts that predict the likely climatic outlook for a pending season. These forecasts are broadcast through the media (radios, TVs and newspapers) and are generally issued during the months of September or October. [Here emphasize seasonal forecasts as opposed to daily weather forecasts and throughout this section keep emphasizing this difference]

Are you receiving these forecasts? Yes______; NO______

What about this season did you receive the seasonal forecasts? YES______; NO ______;

If YES, how did you get the message for this season? Source Check ONLY ONE response 1. Local weather station 2. Agritex extension officer 3. Local paper 4. Farmer bulletin 5. Radio Program 6. TV program 7. Discussions/meeting 8. Other (specify)

Record 210: Forecast Message for the Season What is the forecasted message for this season? Message Check ONLY ONE response 1. Above normal (flood) season 2. Normal season 3. Average 4. Below normal (poor) season 5. Other (specify)

Record 220: Information Management This Record is a continuation of Record 200 and here we are seeking the response of both (a) farmers getting/hearing and (b) farmers NOT getting/hearing information about seasonal forecasts

(a) Farmer getting seasonal forecasts information You told me that you are obtaining information about seasonal forecasts, have you used this information to manage decisions related to your farm operations for this season?

YES ______[If YES, show how you used this information to change your crop decisions for this season by ticking the applicable in Table 220 below] NO______

138 Table 220a: Change Management Decision Management decision [Remember you DON’T Check the applicable (possibly MULTIPLE) have to read out these responses to the farmer] 1. Change crop variety 2. Change acreage 3. Change planting date 4. Change cultivar 5. Change fertilizers 6. Change chemicals 7. Change seed population 8. Other (specify)

(b) Farmer not getting forecasts information You told me above that you are NOT getting information about seasonal forecasts, let’s suppose you have received this information, would you use it? YES______[If YES which crop decisions would you have changed? (Tick the applicable in Table220a above)]

NO______[If NO why would you not use the information; (place a tick on possible reason Table 220b)]:

Table 220b. Reasons for not using Forecasts Forecasts not credible Lacks confidence Forecasts difficult to Other (specify) understand Reasons

Record 230: Confidence with the Forecasts NB: This Record applies only to farmers with YES response to Record 220(a) Please indicate how much CONFIDENCE you place in seasonal forecasts and how EASY is it to understand the forecasts issued by National Met Services when making production decisions.

Read out these responses to the farmer Check ONLY ONE response 1. very confident 2. confident 3. somewhat confident 4. not at all confident

Record 240: Understanding the Forecasts NB: This Record applies only to farmers with YES response to Record 220(a) Do you find these forecasts easy to understand and apply? Tick the applicable Read out these responses to the farmer Check ONLY ONE response 1. very easy 2. easy 3. somewhat easy 4. not at all easy

Record 250: Lead Time Making preparations for any pending new season requires TIME. For e.g. you need time to physically prepare your land, to purchase inputs such as fertilizers and crop seeds, to decide size of land to till and what crops to grow etc. Earlier as you may recall, you told me [Refer record 220(a) and (b)] how some of

139 your crop decisions are influenced by the availability of seasonal forecasts information provided by Dept. of Met. Now, how many months before the season starts would you consider most ideal for receiving information about seasonal forecasts?

Response Check ONLY ONE For what CROP variety or varieties? 1. 1 year 2. 9 months 3. 6 months 4. 3 months 5. 2 months 6. 1 month 7. <30 days 8. A few days before first rains 9. After receiving first rains 10. Other specify

Record 260: Optimal Time for Issuing Forecasts When do you think Dept. of Met should issue forecasts prior to the beginning of the season in order for the forecasts to be of MAXIMUM value to you as end-user:

Response Check ONLY ONE 1. January 2. February 3. March 4. April 5. May 6. June 7. July 8. August 9. September 10. October

Record 270: Production Decision Factors What factors do you consider when deciding what crop to grow each year. Producer price Drought Local price Seasonal Availability Price of Loan Food Other charged by depots tolerance charged by forecasts of inputs inputs repay security (specify) [GMB, Cotton, etc] neighbor concerns

Record 280(a): WTP for Seasonal Forecasts Suppose Dept. of Met Services is able to provide improved forecasts that predict likely seasonal outlook based on this region/district that you have CONFIDENCE in, would you be willing to use such information to make your crop production decisions; YES______; (if yes proceed to record 280a below) NO______; (if no proceed to record 280b below) Uncertain______(if uncertain proceed to record 280b below) Record 280(a): ! If YES, describe the following scenario to the farmer: Let’s suppose the Dept. of Met wants to provide these services to you in the form of Weather bulletin they will post directly to you as a subscriber each time before the season starts. This bulletin will provide information on improved

140 seasonal forecasts that can help you mitigate drought. Suppose the bulletin is written in your vernacular language shona, venda or ndebele and it’s price equals that of your local paper (herald, Kwayedza, or Umulimi). For the provision of such services; would you be WILLING TO PAY once per season: ⎧a.$4000 ______⎪ ⎪b.$4500 ______⎪c.$5000 ______Use a die to randomly assign one of these numbers to each respondent ⎨ ⎪d.$5500 ______⎪e.$6000 ______⎪ ⎩⎪ f .$6500 ______Depending on what initial PRICE BID (a-d) you offer the farmer and farmer’s response to it, ITERATE the bid UPWARDS if farmer’s response is YES or DOWNWARDS if response is NO. Continue asking/eliciting farmer’s WTP for HIGHER/LOWER amounts until he says NO/YES. Show this in Table below: Would you be Willing To Pay: Check with a (tick for Yes and X for No 1. 0 2. 500 3. 1000 4. 1500 5. 2000 6. 2500 7. 3000 8. 3500 9. 4000 10.4500 11.5000 12.5500

Record: 280(b). Probe Further ! If farmer’s response to initial question in Record 280(a) is NO/UNCERTAIN probe the farmer and get to know the ideal conditions the farmer expects so as induce him/her to use the MET forecasts or increase his/her certainty. ! Summarize the ideal conditions below: ______

! Now proceed to elicit the farmer’s WTP (similar to what you did in Record 280a) i.e. Let’s suppose the Dept. of Met wants to provide these services to you in the form of Weather bulletin that they post directly to you as a subscriber each time before the season starts. This bulletin will provide information on improved seasonal forecasts that can help you mitigate drought risk. For the provision of such services; would you be WILLING TO PAY once per season:

141 ⎧a.$4000 ______⎪ ⎪b.$4500 ______⎪c.$5000 ______Use a die to randomly assign one of these numbers to each respondent ⎨ ⎪d.$5500 ______⎪e.$6000 ______⎪ ⎩⎪ f .$6500 ______

Depending on what initial PRICE BID (a-d) you offer the farmer and farmer’s response to it, ITERATE the bid UPWARDS if farmer’s response is YES or DOWNWARDS if response is NO, continue asking/eliciting farmer’s WTP for HIGHER/LOWER amounts until he says NO/YES. Show this in Table below:

Would you be Willing To Pay: Check with a (tick for Yes and X for No 1. 0 2. 500 3. 1000 4. 1500 5. 2000 6. 2500 7. 3000 8. 3500 9. 4000 10.4500 11.5000 12.5500

DROUGHT-INDEXED INSURANCE PROGRAM Drought is one of the most devastating risks that you farmers face. For instance the season of 1991/92 is a good example. Now let’s discuss in detail how drought AFFECTS your household. Tell me, how does how does drought you affects your household?

Record 300: Impact of Drought on Household Don’t read out these responses to the farmer: Place a TICK to the Ask farmer to SINGLE out the MOST Let the farmer list and tick the applicable applicable SEVERE factor [only ONE] 1. loss of crops 2. loss of cattle 3. loss of donkeys 4. loss of sheep 5. loss of goats 6. unable to pay loans 7. unable to feed my household 8. unable to send children to school 9. hunger related disease 10. water scarcity 11. other (specify)

142 Record 310: Household Coping Strategies You have just told me [refer to Record 300] how drought affects your household, tell me what MEASURES you take to mitigate the impact/effects of drought on your household? Don’t read out these responses to the farmer: Place a tick to the Ask farmer to SINGLE out the MOST applicable FREQUENT measure he takes whenever drought occurs[only ONE] 1. buy more grain 2. sell livestock to mitigate impact 3. engage in off-farm job 4. stop sending children to school 5. migrate to other regions with plenty of food 6. rely on food-aid 7. rely on my crop insurance 8. reduce consumption 9. borrow grain from other farmers 10. engage in barter trade 11. other (specify)

Record 320: Vulnerability to drought You have told me how drought AFFECTS your household and what MEASURES you take to reduce the effects of drought on your household. Now tell me in your opinion WHY your household is vulnerable to the problem of drought? Don’t read out these responses to the farmer Single out what you consider the MOST significant factor [only ONE] 1. poverty 2. poor choice of crop types 3. poor government policy 4. lack of extension advice 5. lack of climate information 6. lack of credit and insurance services 7. lack of remittances 8. erratic rainfall 9. geographical location 10. no livestock assets to sell 11. other (specify)

143 Record 330: WTP for Drought Insurance One way that could help protect your household against drought is by purchasing DROUGHT INSURANCE. Are you aware of what INSURANCE is and how it operates? YES______;NO______;

Definition of Insurance Insurance is a program where you protect yourself against future possible disasters such as drought. You protect yourself by paying a small amount against a coverage you select, you desire and you consider adequate. In the event of drought occurring then you are assured/guaranteed of ‘compensation’ based on your coverage you initially declared.

How Insurance works The small amount you pay is called the PREMIUM and this safe-guards you against the future drought perils. The amount you select and you consider adequate to meet the subsistence needs of your household is called the COVERAGE. The return you get in the event of drought actually occurring is called INDEMINTY. But remember the higher the COVERAGE the higher the INDEMNITY and the higher the PREMIUM you will pay.

The organization that provides the insurance program is called the INSURER and your household that seeks insurance protection is called the INSURED. INDEMNITY is paid if and only if the insured peril occurs which is drought in this case. For e.g. if locusts or hailstorm destroys your crop such perils are NOT covered by your drought insurance. In this case INDEMINTY is paid based on the occurrence of drought, indexed by area or region and monitored during the rainy season by the INSURER. The payment of indemnity could be 100% or less depending on the severity of loss for that particular season as assessed by the INSURER. Your indemnity is NOT paid to you on the basis of yield losses you experience as an individual household but is rather paid on the basis of average yield realized within in a specifically identified AREA in which a particular farmer is geographically located. Such a scheme is likely to be less affected by problems of moral hazards [laziness, corruptive tendencies etc]. Therefore ALL insured farmers from within the same AREA will be charged the same premium and difference arises depending on the COVERAGE level that each farmer chooses.

Do you understand the program so far? YES_____; NO______

Benefits of drought insurance With drought insurance you are able to control and self-manage risks associated with climate. To emphasize the benefits are paid IF AND ONLY IF the insured peril occurs and in that case you are PROTECTED. Conversely if drought does NOT occur you will NOT receive compensation. Further, the premium payments you pay are not accumulative. The bottom-line is an insured farmer is PROTECTED and is likely to suffer fewer losses than an uninsured farmer who is NOT PROTECTED.

Case (a) If such a scheme is offered ! If food-aid were NOT available; Will you be willing to participate in the DROUGHT INSURANCE program I discussed above? YES______; NO______.

If YES, would you be WILLING TO PAY $______in order to receive a guarantee of at least ONE BAG (50kg) in event that drought occurs

144 ⎧a.$4800 ______⎪ ⎪b.$6000 ______⎪c.$8000 ______Use a die to randomly assign one of these numbers to each respondent ⎨ ⎪d.$10000 ______⎪e.$12000 ______⎪ ⎩⎪ f .$17000 ______Depending on what initial PRICE BID (a-d) you offer the farmer and farmer’s response to it, ITERATE the bid UPWARDS if farmer’s response is YES or DOWNWARDS if response is NO, continue asking/eliciting farmer’s WTP for HIGHER/LOWER amounts until he says NO/YES. Show this in Table below:

Premium WTP (for coverage of 1bag) Check with a tick for YES and X for NO. 1. $1000 2. $2000 3. $3000 4. $4800 5. $6000 6. $8000 7. $10000 8. $12000 9. $13000 10.$15000 11.$17000

1 bag= 50kg; 1 ton = 22 bags, maize is the main staple

Case (b): If such a scheme is offered ! If food-aid is AVAILABLE (as is the case now); Will you be willing to participate in the DROUGHT INSURANCE program I discussed above?

YES______; NO______.

If YES, would you be WILLING TO PAY $______[Here use the same figure as you selected in case (a) above] in order to receive a guarantee of at least ONE BAG (50kg) in the event that drought occurs: ⎧a.$4800 ______⎪ ⎪b.$6000 ______⎪c.$8000 ______Use a die to randomly assign one of these numbers to each respondent ⎨ ⎪d.$10000 ______⎪e.$12000 ______⎪ ⎩⎪ f .$17000 ______Depending on what initial PRICE BID (a-d) you offer the farmer and farmer’s response to it, ITERATE the bid UPWARDS if farmer’s response is YES or DOWNWARDS if response is NO, continue asking/eliciting farmer’s WTP for HIGHER/LOWER amounts until he says NO/YES. Show this in Table below:

145 Premium WTP (for coverage of 1bag) Check with a tick for YES and X for NO. 1. $1000 2. $2000 3. $3000 4. $4800 5. $6000 6. $8000 7. $10000 8. $12000 9. $13000 10.$15000 11.$17000

1 bag= 50kg; 1 ton = 22 bags, maize is the main staple

FARM INCOME ESTIMATION Record 400: FARM INCOMES AND COSTS Please indicate to the best of your ability your level of farm and non-farm income for the past year (2002/2003)

Source of income a) CROP SALES Quantity Price per unit Total 1. maize 2. cotton 3. sunflower 4. white sorghum 5. red sorghum 6. peanuts 7. tobacco 8. other (specify) b) LIVESTOCK SALES 1. cattle 2. sheep 3. goats 4. other (specify) c) OTHER 1. beer brewing 2. vegetables 3. carpentry 4. remittances 5. other (specify)

Record 410: Input Costs Please give information related to farming costs you incurred during the last season, 2002/2003 INPUT Type Quantity Cost per unit Total a) Seed Purchase

146 1. maize seed 2. tobacco seed 3. other (specify) b) Fertilizer/chemical purchases 1. Ammonium nitrate 2. Compound D 3. Other Fertilizers c) Chemical/Insecticides

d) Labor 1. tillage 2. planting 3. weeding 4. harvesting 5. other (specify)

Vote of Thanks: On behalf of the Department of Agricultural Economics, University of Zimbabwe in collaboration with the Department of Agricultural Environmental and Development Economics, Ohio State University, I would like to thank you very much for participating in this survey exercise.

147 Appendix E

Agro-ecological Zones of Zimbabwe

Farmland Natural Region-I Natural Region-IIA Natural Region-IIB Natural Region-III Natural Region-IV Natural Region-V

148 BIBLIOGRAPHY

Adesina, A.A., and J.H. Sanders. 1991. Peasant farmer behavior and cereal technologies: Stochastic programming analysis in Niger. Agricultural Economics, 5, 21-38

Asfaw, D. 1989. Drought preparedness in Ethiopia. Unpublished paper. Drought Management and Preparedness Training Seminar for Eastern and Southern Africa. Gaborone, Botswana.

Benedetti, R., and P. Rossini, 1993. On the use of NDVI profiles as a tool for agricultural statistics: The case study of wheat yield estimate and forecast in Emilia Romagna, Remote Sensing of Environment, 45:311-326.

Berg, A. 1981: General results of a European remote sensing project (AGRESTE) agronomic considerations and needs of society. Chapter in Application of remote sensing to agricultural production and forecasting, eds. A. Berg. Commission of European Communities Publ.

Binswanger, H.P. ‘Risk Aversion, Collateral Requirements, and the Markets for Credit and Insurance in Rural Areas in Rural Areas’. In Hazell, P.B.R., Pomeda, C. and Valdes, A. (eds), Crop insurance for Agricultural Development: Issues and Experience. Baltimore: John Hopkins University Press, pp. 67-86.

Cane, M.A., G. Eshel, and R.W. Buckland. 1994. Forecasting Zimbabwean maize yields using easten equatorial Pacific sea-surface temperatures. Nature (London) 370:204-205.

Cane, M.A. 2001. Understanding and Predicting the World’s Climate System. Impacts of El-niño and Climate Variability on Agricluture, 63, 1-20

Chimbwanda, W. 2002. Former Deputy General Manager with AFC (Personal communication)

Chakravarti, J.S. 1920: Agricultural insurance: a practical scheme suited to Indian Conditions. Government Press, Bangalore.

Doraiswamy, P.C., Moulin, S., P. W. Cook, and A. Sten. 2003. Crop yield assessment from remote sensing. Photogrammetric engineering and remote sensing., Vol. 69, No. 6, 665-674

Dijk, A.V. 1986. A crop condition and crop yield estimation method based on NOAA/AVHRR satellite data. Unpublished PH.D dissertation, University of Missouri- Columbia

149 Duncan, J and Myers, R.J. Crop Insurance under Catastrophic Risk. Amer. J. Agr. Econ .82(4) (November 2000):842-855

Eastman, J.R, and A. Anyamba. Prototypical Patterns of ENSO-Related Drought and Drought Precusors in Southern Africa: The thirteenth Pecora Symposium Proceedings, Soux Falls, South Dakota, August 20-22, 1996.

Easterling, W. E, and J. W. Mjelde. 1987. The importance of seasonal climate prediction lead time in agricultural decision making.Agric. and Forest Meteor., 40:37-50

Gallagher, P. “U.S. Soybean Yields: Estimation and Forecasting with Nonsymmetric Disturbances.” Amer. J. Agric. Econ. 69(November 1987):798-803.

Glantz, M.H., 1994: Forecasting El-niño: science’s gift to the 21st century. Ecodecision, 4, 78-81.

Gautam, M., P.B.R. Hazell and H. Alderman. 1994. Rural demand for drought insurance. Policy research working paper: World Bank

Gudger, M. 1991. Crop Insurance: Failure of the public sector and the rise of the private sector. In Risk and agriculture: Proceedings of the tenth agricultural sector symposium, D. Holden, P.Hazell and A. Pritchard, eds. Washington D.C.: World Bank

Goward, S.A., C.J. Tucker, and D.Dye, 1985. North American vegetation patterns observed with the NOAA-7 AVHRR, Vegetatio. 64:3-14

Goodwin, B.K., and A.P. Ker, “Nonparametric Estimation of Crop Yield Distributions: Implications for Rating Group-Risk (GRP) Crop Insurance Contracts.” Amer. J. Agr. Econ. 80(February 1998):139-53

Haab, T.C., and T,E. McConnell 2000. Valuing Environmental and Natural Resources. Edward Elgar Pres

Hanemann, M., J. Loomis., and B. Kanninen. 1991: Statistical efficiency of double- bounded dichotomous choice contingent valuation. American Journal of Agricultural Economics 73:1256-1263

Hazell, P.B.R. 1992. The appropriate role of agricultural insurance in developing countries. Journal of International Development 4: 567-581.

Hazell, P.B.R., C. Pomareda, and A. Valdes. 1986. Crop insurance for agricultural development: Issues and experience. Baltimore: The John Hopkins University Pres.

Heal, G, and Y. Lin. 1996. El-niño and Agriculture in Zimbabwe. Paper presented at Columbia University ..

Jensen, J.R. 1996. Introductory Digital Image Processing; A Remote Sensing Perspective. 2nd ed.,

150 New Jersey, Prentice Hall.

Johnson, G.E., V.R. Achutuni., S. Thiruvengadachari., and F.N. Kogan. 1993: The role of NOAA satellite data in drought early warning and monitoring: selected case studies. Chapter in: Drought assessment, management and planning: Theory and case studies, eds. D.A. Wilhite. Kluwer Academic Publ.

Ker, A.P and B.K Goodwin. “Nonparametric Estimation of Crop Insurance Rates Revisited.” Amer .J .Agr. Econ. 83(May 2000): 463-478.

Kelly, M and Kleffner, A.E. Optimal Loss Mitigation and Contract Design. The Journal of Risk and Insurance, 2003, Vol.70, No. 1, 53-72

Kogan, F.N. 1998: A typical pattern of vegetation conditions in southern Africa during El-nino years detected from AVHRR data using three-channel numerical index. Int. J. Remote Sensing, vol. 19, No. 18, 3689-3695

Kogan, F.N., and J. Sullivan. 1993: Development of a global drought watch system using NOAA/AVHRR data, Adv. Space Res. 13, 219-222

Labus, M.P., G.A. Nielsen., R.L. Lawrence., and R. Engel. 2002: Wheat yield estimates using multi-temporal NDVI satellite imagery. Int. J. Remote Sensing, Vol. 23, No. 20, 4169- 4180.

Lecomte D., and W. Thiao, 1995: SST anomalies in the equatorial Pacific and their relationship to corn yield in Southern Africa. Presented at the International Symposium on African Drought, Trieste, Italy.

Lillesand, T.M., and R.W. Kiefer. 2000. Remote Sensing and Image Interpretation. 4th ed., New York, John Wiley and Sons, Inc.

Maas, S.J., 1988. Using satellite to improve model estimates of crop yield, Agronomy Journal, 80:655-662

Matlon, P. 1991. Farmer risk management strategies: “The case of West Africa Semi-arid Tropics”. In D Holden, P. Hazell and A. Pritchard (eds.): Risk in Agriculture: Proceedings of the tenth agriculture sector symposium, 1990, The World Bank, Washington, D.C.

Mason, S.J. 1996. Regional manifestations of climate variability in Southern Africa. Paper presented at “Workshop on reducing climate-related vulnerability in Southern Africa”., Victoria Falls, Zimbabwe

Mason, S.J., and M.R. Jury. 1997: Climatic variability and change over Southern Africa: a reflection on the underlying process. Progress in Physical Geography 21, 23-50

151 McVicar, T.R., and D.L.B. Jupp. 1998: The current and potential operational uses of remote sensing to aid decisions on drought exceptional circumstances in Australia:a Review. Agricultural systems, Vol. 57, No. 3, 399-468

Ministry of Lands and Agriculture. 2000, Statistical Bulletin, Harare, Zimbabwe.

Ministry of Lands and Agriculture, 1995/96, Policy Statement, Harare, Zimbabwe.

Miranda, M.J. and J.W. Glauber. 1997. Systemic risk, reinsurance, and the failure of crop insurance markets. American Journal of Agricultural Economics 79:206-215

Mishra, P.K. 1996. Agricultural risk, insurance and income: A study of the impact and design of India’s comprehensive crop insurance scheme. Brookfield: Avebury Press

Mitchell, R. and R. Carson. 1989. Using Surveys to Value Public Goods. Resources for the Future, Washington, D.C.

Mjelde, J.W., S.T. Sonka., B.L. Dixon, and P.J. Lamb: Valuing Forecast Characteristics in a Dynamic Agricultural Production System. Amer.J.Agric.Econ.70(1988):674-84

Mjelde, J.W., J.B. Penson., C.J. Nixon. 2000. Dynamic aspects of the impact of the use of perfect climate forecasts in the corn belt region. Journal of Applied Meteorology. 39:67-79

Nelson, C.H., and P.V. Preckel. “The Conditional beta Distribution As a Stochastic Production Function.” Amer. J. Agr. Econ. 71(May 1989): 370-78.

NOAA (National Oceanic and Atmospheric Administration), 1997: Climatic conditions associated with the 1997-1998 El-nino: Impact and Outlook, Special Climate Summary, 97/3, 13, Washington D.C.

Parzen, E. “On Estimation of Probability Density and Mode.” Ann. Math. Statist. 33(September 1962):1065-76

Peters, A.J., Walter-Shea, E. A., L. Jii., A. Vina., M. Hayes., and M. D. Svoboda. 2002. Drought monitoring with NDVI-based standardized vegetation index. Photogrammetric engineering and remote sensing., Vol. 68, No. 1, 71-75

Phillips J.G., E. Makaudze and L. Unganai. 2001. Current and Potential Use of Climate Forecasts for Resource-Poor Farmers in Zimbabwe. Impacts of El-niño and Climate Variability on Agricluture, 63, 1-20

Odumodu L.O. 1983. Rainfall distribution, variability and probability in Plateau State, Nigeria, Journal of Climatology., vol. 3, 385-393.

152 Quiggin, J. 1994: The optimal design of crop insurance. Chapter in: Economics of Agricultural Crop Insurance: Theory and Evidence, eds. Hueth, D.L., and W.H. Furtan. Kluwer Academic Publ.

Roberts, R.A.J., and W.J.A. Dick. 1991: Strategies for crop insurance planning. Food and Agriculture Organization of the United Nations, Rome.

Ropelewski, C.F., and M.s. Halpert. 1987. Global and regional scale precipitation patterns associated with El-nino/Southern Oscillation. J. Climate, 115, 8, 1606-1626.

Sakurai, T., M. Gautam, T. Reardon, P. Hazell and H. Alderman. 1994. Potential demand for drought insurance in Burkina Faso. Mimeo., Agricultural Policies Division, Agriculture and Natural Resources Department, The World Bank, Washington, D.C.

Sen, A. 1976. Poverty: and ordinal approach to measurement. Econometrica, 44:219-31.

Silverman, B.W. Density Estimation for Statistics and Data Analysis. New York: Chapman and Hall, 1986.

Solow, A.R., R.F. Adams, K.J. Bryant, D.M. Legler, J.J. O’Brien, B.A. McCarl, W. Nayda, R. Weiher. 1998. The value of improved ENSO predictions of U.S. agriculture. Climate Change 39: 47-60.

Skees, J., P. Hazell, and M. Miranda. 1999. New Approaches to Crop Insurance in Developing Countries. Environment and Production Technology Division, vol. 55.

Skees, J.R., J.R. Black, and B.J. Barnett. 1997. Designing and rating an area yield crop insurance contract. American Journal of Agricultural Economics 79:430-438

Swinton, S.M. 1988. Drought Survival Tactics of Subsistence Farmers in Niger. Human Ecology, vol. 16, No. 2.

Tannehill, I.R. (1947) Drought: Its Causes and Effects, Princeton, NJ: Princeton University Press.

Tucker, C.J. 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote sensing of the environment 8: 127-50.

Unganai, L.S., and F.N. Kogan. 1998. Drought monitoring and crop yield estimation in Southern Africa from AVHRR data. Remote Sens. Environ. 63, 219-232.

Vedenov, D. V. and M.J. Miranda. 2001. Rainfall Insurance for Midwest Crop Production. Selected paper for AAEA annual meeting.

Vogt, J.V., S. Niemeyer., F. Somma., I. Beaudin., and V.V. Viau. 2000: Drought monitoring from space. Chapter in: Drought and Drought Mitigation in Europe, eds. Vogt, J.V., and F. Somma. Kluwers Publ.

Walker, T.S. and Jodha, N.S. ‘How small farm households adapt to risk’. In Hazell, P.B.R.,

154 Pomeda, C. and Valdes, A. (eds), Crop insurance for Agricultural Development: Issues and Experience. Baltimore: John Hopkins University Press, pp. 17-34. Wilhite, D. A., and M.H Glantz, 1985. Understanding the drought phenomenon: The role of definitions. Water International 10:111-20.

WMO (World Meteorological Organization). 1995: The global climate system review: Climate system monitoring, June 1991-Nov 1993, WMO, No. 819 150

Zeller, M and M. Sharma. 1998. Rural finance and poverty alleviation. IFPRI publication: Food policy statement No.27

155