·

The Impacts of Environmental Changes on Individual Behaviors in Developing Countries

DISSERTATION

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

By

Wei Chen

Graduate Program in Agricultural, Environmental and Development Economics

The Ohio State University

2019

Dissertation Committee:

H. Allen Klaiber, Advisor

Daniela A. Miteva

Sathya Gopalakrishnan

Copyrighted by

Wei Chen

2019

Abstract

This dissertation consists of three essays. All three essays explore how individuals make decisions in response to natural and man-made environmental changes in developing countries and how these individual behaviors lead to aggregate effects on the environment.

In Chapter 2, I estimate the causal effect of municipal road expansion on Vehicle

-Kilometers Traveled (VKT) in 103 Chinese cities while accounting for the potential for increased car adoption to affect VKT. A novel matching IV strategy is developed to address endogeneity concerns that complements previously used historical infrastructure instruments to provide time-varying identification in a panel data setting. I find that the estimated elasticity of

VKT with respect to road length is approximately 1.1, indicating that newly built urban roads lead to a more than proportional increase in total traffic. Given large ongoing infrastructure investment combined with more recently enacted traffic alleviation policies in many Chinese cities, this result provides important new information on the potential impacts of infrastructure investment on traffic.

In Chapter 3, I develop a novel compensating differential model of quality of life rankings with an agricultural sector. I introduce an additional farm income component into the household budget in the theoretical equilibrium system alongside housing and labor markets. I apply this model in a developing country, Indonesia, to examine the existence of compensating differentials and recover quality-of-life rankings for jurisdictions across the country at distinct time periods. To conduct the research, I use detailed household data from the fourth and fifth waves of the Indonesian Family Life Survey (IFLS) fielded in 2007 and 2014 across 11

ii provinces in Indonesia. I estimate implicit prices for various amenities based on hedonic equations of housing rents, non-farm wages and agricultural returns. The results indicate that compensating differentials exist across the country and in particular the impacts of amenities on agriculture and the non-farm labor market are different. Moreover, I calculate the quality-of-life rankings for 192 districts (regencies and cities) in the 11 Indonesian provinces in 2007 and 2014, respectively. The significant changes between these two time periods imply that the gap in quality of life across the country has narrowed.

In Chapter 4, I use citizen science eBird data to investigate the impact of deforestation on birdwatching ecotourism across Mexico. Utilizing detailed data on individual trips reported to eBird, I construct annual count data on visits to 1,810 Mexican municipalities from 2008 to 2016 to examine how changes in forests affect tourist visitation patterns. I find that an additional percentage point of deforestation reduces the probability of a municipality being visited by birders by 12.5% and decreases the number of birdwatching visit days by 29.4%. These results offer new insights into the impacts of deforestation on economic returns from the ecotourism industry and provide evidence that citizen science data perform well as a relatively untapped source of potentially valuable information on economic behavior in otherwise limited data settings.

iii Acknowledgements

I would like to sincerely thank many people without whom this dissertation would not have been possible. I would like to express my sincere gratitude to my advisor Dr. H. Allen Klaiber for his continuous support of my research throughout my doctoral program. As the old Chinese saying goes, “Once my teacher, forever my father.” I do respect him as much as I do my father and have been very grateful for his guidance and patience over the past five years. He provides me endless help and advice without reservation based on his immense knowledge and extensive professional experience. He always offers me numerous academic opportunities and financial resources even before I ask. He is not only a successful scholar, but also an extraordinary educator. I have heard countless miserable stories of Ph.D. students’ hard lives, but barely do I have empathy for them because my study experience is surprisingly enjoyable and exciting under Allen’s supervision. I could not have imagined having a better advisor and mentor during this five-year-long journey.

Besides my advisor, I would also like to thank the rest of my committee members who have constantly inspired and pushed me to improve my work. I am thankful to Dr. Daniela A.

Miteva for all the academic workshops she has enthusiastically organized, which significantly help improve my research skills. I will never forget her encouraging words that I was much better than I thought when I was filled with fear on the job market. My sincere thanks also go to

Dr. Sathya Gopalakrishnan for setting a high expectation for me, both in and out of the classroom.

She was never reluctant to spare her valuable time to write long emails to give me suggestions every time I made a presentation. In those days when I took job interviews, she was consistently available to support me through challenging times.

iv I am also very grateful to Dr. Jon Einar Flatnes and Dr. Tim Haab for their precious advice on this dissertation. I would especially like to thank Jian Chen, Shicong Xu, and Corinne

Bocci for their company and truly friendship throughout the doctoral program. I would like to thank my parents in China who always support my plans, even though they have no idea about what I have been doing in the United States. Finally, I deeply appreciate the opportunity to enroll in the Ohio State University that starts my life-changing and cross-cultural academic experience.

v Vita

2010...... Loudi No.1 Middle School, Loudi, Hunan, China

2014...... B.A. Economics, Peking University, Beijing, China

B.S. Mathematics and Applied Mathematics, Peking

University, Beijing, China

2016...... M.S. Agricultural, Environmental, and Development

Economics, The Ohio State University, Columbus,

Ohio

2014 to present...... Graduate Research Associate, Agricultural,

Environmental, and Development Economics, The

Ohio State University, Columbus, Ohio

Publications

Wolf, D., Chen, W., Gopalakrishnan, S.G., Haab, T. and Klaiber, H.A., “The economic impacts of harmful algal blooms and E.coli on recreational behavior in Lake Erie.” Land Economics, forthcoming.

Chen, W., Huang, Z. and Yi, Y., “Is there a structural change in the persistence of WTI-Brent oil price spreads in the post-2010 period?” Economic Modelling 50 (2015): 64-71.

Fields of Study

Agricultural, Environmental and Development Economics

vi Table of Contents

Abstract ...... ii

Acknowledgements ...... iv

Vita ...... vi

List of Tables ...... ix

List of Figures ...... xi

Chapter 1: Introduction ...... 1

Chapter 2: Does Road Expansion Induce Traffic? An Evaluation of Vehicle-Kilometers

Traveled in China ...... 5

2.1 Introduction ...... 5

2.2 Literature Review ...... 9

2.3 A Conceptual Model ...... 12

2.4 Data ...... 14

2.5 Empirical Strategy ...... 18

2.6 Estimation Results ...... 28

2.7 Discussion ...... 35

Chapter 3: Compensating Differentials of Rents, Wages and Agricultural Returns: the

Quality-of-life Among Indonesian Regencies and Cities ...... 37

3.1 Introduction ...... 37

3.2 Literature reviews ...... 39

3.3 Theoretical Models and Empirical Strategy ...... 45

vii 3.4 Data ...... 48

3.5 Results ...... 53

3.6 Discussion ...... 61

Chapter 4: The Impact of Deforestation on Ecotourism: a Birdwatching Example in Mexico 63

4.1 Introduction ...... 63

4.2 Data ...... 69

4.3 Empirical Models ...... 75

4.4 Results ...... 78

4.5 Discussion ...... 84

References ...... 87

Appendix A: Supplemental Materials ...... 105

viii List of Tables

Table 1. Summary statistics ...... 17

Table 2. Mean covariate differences between focal and matched cities ...... 26

Table 3. OLS estimates of road effect on VKT ...... 29

Table 4. 2SLS estimates of road effect on VKT ...... 30

Table 5. Robustness to matching specifications ...... 33

Table 6. Panel estimates of road effect on VKT with city fixed effects ...... 34

Table 7. Variables and data sources ...... 50

Table 8. Housing rent equation with amenities, 2014 ...... 54

Table 9. Wage equation with amenities, 2014 ...... 55

Table 10. Farm income equation with amenities, 2014 ...... 57

Table 11. Implicit prices of amenities, 2014 ...... 59

Table 12. Descriptive Statistics ...... 74

Table 13. Results from the Logit Model ...... 79

Table 14. Estimates from ZINB Model ...... 81

Table 15. Incidence Rate Ratio from ZINB Model ...... 83

Table 16. OLS estimates of road effect on VKT with varying fixed effects ...... 106

Table 17. 2SLS estimates of road effect on VKT with varying fixed effects ...... 107

Table 18. Robustness to lag specification ...... 108

Table 19. Mean covariate differences between focal and matched cities, by city ...... 109

Table 20. Housing rent equation with amenities, 2007 ...... 112

ix Table 21. Wage equation with amenities, 2007 ...... 113

Table 22. Farm income equation with amenities, 2007 ...... 115

Table 23. Implicit prices of amenities, 2007 ...... 117

Table 24. Quality of life rankings for 192 districts in Indonesia, 2014 and 2007 ...... 119

Table 25. Results from Probit Model ...... 124

Table 26. Estimates from Poisson and NB Models ...... 125

Table 27. Incidence Rate Ratio from Poisson and NB Models ...... 126

Table 28. Results from the Logit Model, with Deforestation Lagged by One Year ...... 127

Table 29. Estimates from ZINB Model, with Deforestation Lagged by One Year ...... 128

Table 30. Incidence Rate Ratio from ZINB Model, with Deforestation Lagged by One Year ... 129

x List of Figures

Figure 1. Differences in VKT and roads in 103 Chinese cities, 2011-2014 ...... 15

Figure 2. Example of matched cities for Changsha in 2014 ...... 26

Figure 3. 11 selected provinces in Indonesia ...... 49

Figure 4. Quality of life rankings for 192 districts in Indonesia, 2014 ...... 58

Figure 5. Aggregate eBird Reports of Day Trips to Mexico, 2008-2016 ...... 71

Figure 6. Forest cover (2000) and deforestation (2008-2016) based on Hansen et al. (2013) ..... 72

Figure 7. Quality of life rankings for 192 districts in Indonesia, 2007 ...... 118

xi Chapter 1: Introduction

Humans care about the environment, not because of the environment per se, but its influence on well-being and quality of life. To predict the consequence of environmental changes, it is important to understand the interaction between individuals and the environment. Environmental challenges are the focus of growing concerns across developing countries, resulting in policies targeted at relieving environmental pressures. At times, however, these policies may lead to unintended effects on the environment due to unforeseen responses of, which have the potential to undermine the effectiveness of these policies. Governments in a number of developing countries are also targeting improvements in local environmental conditions by investing in infrastructure and other amenities in order to raise residents’ quality of life. Nevertheless, the provision of these amenities, which changes the environment of local communities in a broad sense, impacts a number of households’ behaviors including residential location decisions, leading to migration and changes in market.

Motivated by the governments’ need for information on household behavior in order to aid in the design of effective environmental policies, this dissertation investigates individuals’ decision-making in response to natural and man-made environmental changes in developing countries and the resulting aggregate effects on the environment. Chapter 2 studies the impact of road expansion on traffic congestion in China. Chapter 3 evaluates the quality of life in a number of Indonesian districts based on the revealed values of local amenities and disamenities. Chapter

4 investigates the impact of deforestation on bird-watching ecotourism in Mexico.

1 Chapter 2 investigates the causal relationship between municipal road expansion and vehicle-kilometers traveled (VKT) in China, a developing country that has a rapidly increasing number of car ownership. Road infrastructure construction is often regarded as a preferred policy tool to mitigate traffic congestion. Nevertheless, changes in drivers’ behavior in response to new roadways may impede achieving the goal of reduced traffic congestion. It seems that new provision of road infrastructure can help dilute existing traffic, but at the same time newly-built roads are also likely to increase the demand to travel by car and therefore raise the overall level of traffic. To address potential endogeneity, I adopt a combination of traditional historical infrastructure instruments and a novel matching IV strategy. Focusing on 103 Chinese cities during 2011-2014, I find that the estimated road elasticity of VKT is approximately 1.1. This empirical result suggests that urban road expansion leads to a more than proportional increase in traffic. As a result, road construction does not appear to reduce congestion, and potentially exacerbates the problem. Although road infrastructure construction has long been viewed as a solution to road congestion, it might not be as effective as is expected since the newly-provided roads induce new traffic.

Chapter 3 analyzes compensating differentials and quality of life in Indonesia, a developing economy with a large agriculture sector. Given the increasing quantity of investment in infrastructure in developing countries which aims to raise local quality of life, it is important to investigate the potential feedback effects of this investment in how households respond to the newly-constructed infrastructure along with amenities that impact residential location choices.

Understanding the households’ decision-making process as it is affected by local amenities adds

2 important information for policymakers seeking to develop policies for local community development. In this chapter, I use data from Indonesia to recover quality-of-life rankings for jurisdictions across the country at distinct time periods and examine how quality of life rankings have changed over time. I extend Roback’s (1982) model by introducing an agriculture sector in the theoretical equilibrium system alongside housing and labor markets. To measure the quality of life, I estimate compensating differentials by utilizing detailed household data from the fourth and fifth waves of the Indonesian Family Life Survey (IFLS) fielded in 2007 and 2014 across 11 provinces in Indonesia. The results indicate that compensating differentials exist across the country. Moreover, the differentials of farm income and non-farm wages with respect to local amenities are different. Based on these differential estimates, I further calculate the quality-of-life rankings for 192 districts (regencies and cities) in the 11 provinces for 2007 and

2014, respectively. The results show that the rankings have significantly changed from 2007 to

2014.

Chapter 4 uses citizen science eBird data to explore how deforestation impacts the frequency of bird watching visits to ecotourism hotspots in Mexico. The ecological risk associated with deforestation in many developing countries brings negative economic effects to local communities which are often overlooked. Particularly in Mexico, bird biodiversity is vulnerable to habitat destruction which is a potential threat to the ecotourism industry, an important financial source for natural conservation. To examine how changes in forest cover affect birding visits, I use detailed data on individual trip reports submitted to eBird and construct annual count data on 1843 Mexican municipalities from 2008 to 2016. Adopting both

3 binary and count data models, I find that deforestation not only decreases the probability of visiting a municipality for a decision-maker, but also reduces the number of her visits conditional on choosing to visit a location. When there is an additional percentage point of deforestation, the probability of a municipality being visited by birders decreases by 12.5% and the number of visits decreases by 29.4%. These findings indicate negative impacts of deforestation on ecotourism industry and local economy.

4 Chapter 2: Does Road Expansion Induce Traffic? An Evaluation of Vehicle-Kilometers Traveled in China

2.1 Introduction

Traffic in cities is an inevitable consequence of urbanization and population growth. In many developing countries, this issue has taken on new prominence as congestion has snarled large cities and pollution and air quality has worsened. Increasing levels of traffic not only affect daily commute times in urban areas, but also raise the cost of commuting and increase inefficiencies in economic activities that depend on transportation. In addition, high traffic causes environmental challenges including reductions in air quality and increases in noise. For policymakers, constructing additional roads seems to be a simple solution and has been adopted by policy-makers in many countries such as the United States (Hansen et al. 1993), China and

India (Pucher et al. 2007). Intuitively, additional roads dilute existing traffic. However, road expansion also has the potential to induce new traffic which is likely to offset some of the anticipated benefits of urban road construction. Newly constructed roads facilitate car travels and create an incentive for individuals to choose vehicle commuting over alternative modes of transportation. When making decisions of road infrastructure investment, it is crucial to understand the magnitudes of induced new traffic relative to new roads provided.

This chapter investigates the effect of urban road expansion on the level of total traffic in urban areas of China. The research is motivated by recent challenges facing policy-makers in a number of developing countries experiencing increased traffic in urban areas. Given the rapid

5 economic growth and increasing population in China, it is not surprising that excessive amounts of road traffic have become a problem and are likely to grow in severity in the future. In fact, a number of Chinese cities have recently adopted car driving and/or buying restriction policies within municipal districts. At the same time, China is also investing a large amount of financial resources on roadway facilities such as urban roads. One goal of these construction plans is to ease traffic. Under these circumstances, it is paramount to predict drivers’ behavioral changes in response to road provision and determine the relationship between roads and traffic. In this chapter, I provide answers to this question.

Rapidly increasing road traffic, particularly in developing country contexts, also raises growing environmental concerns. Vehicle travel, generating air pollutant emissions such as carbon monoxide, oxides of nitrogen and particulate matter, is one of the major contributors to urban air pollution in China (Cai and Xie 2007). Evidence in China shows that increased road traffic significantly raises the level of air pollution (Fu and Gu 2017). In fact, the Chinese government has realized the key role of traffic when making policies targeted to mitigate air pollution. Recently, a number of areas have implemented direct traffic control policies to address this issue (He et al. 2016; Viard and Fu 2015) along with expansion of public transit systems (Li et al. 2019). In addition to pollution, traffic also is a primary contributor to noise in urban settings which has been shown to impact health and wellbeing of nearby residents (Pucher et al.

2007; Song et al. 2016). The strong correlation between traffic and environmental pollution makes understanding the potential for induced traffic resulting from road infrastructure construction a key policy input.

6 Focusing on road traffic in a developing context also raises novel questions when considering the role of car ownership changes in increasing traffic. Previous studies, which mainly focus on developed countries, pay limited attention to vehicle ownership decisions. In the

U.S., for example, more than 80% of Americans own a motor vehicle. As a result, vehicle ownership decisions are likely a relatively small component of evaluating travel decisions (i.e. duration, destination) and are often overlooked in the literature. Nevertheless, the decision to purchase a car is likely a primary consideration in developing countries such as China where existing levels of car ownership are much lower. Due to increasing income and improving roadway facilities, car ownership has recently expanded dramatically. The level of civil vehicle ownership in China has increased tenfold over the past 20 years and has recently been subject to policies seeking to curtail this growth. In this context, road expansion has impacts on traffic not only through changes in driving behavior for those who already own cars, but also through changes in travel mode decisions if road provision encourages vehicle ownership. It is an empirical question to see how road construction affects traffic via different channels.

This chapter focuses on the impact of urban road construction on vehicle-kilometers traveled (VKT) in a developing country context using panel data from 103 Chinese cities. I make three main contributions to the existing literature. First, I provide new insights into urban traffic and transportation infrastructure construction by showing that roadway expansion in China has led to a more than proportional increase in the total amount of traffic. This is the first comprehensive empirical analysis in China that links roadway infrastructure construction with traffic, using a novel dataset that contains panel data on traffic volume in 103 cities across China

7 from 2011 to 2014. Unlike previous research that focus on the expansion and traffic of interstate highways, this chapter focuses on municipal roadways within urban districts where traffic affects households’ daily lives directly and broadly. Not only is the issue important for city residents in terms of commuting time and living standards, but also for policy-makers who are faced with providing infrastructure enhancements to improve residents’ well-being. This chapter offers an added perspective to those who view road expansion as a strategy to ease traffic.

Second, this chapter explores changes in traffic in a context where individuals have options to choose from a number of potential travel modes. As Huo et al. (2012) points out, VKT in an area is influenced by vehicle stock level, which is an important variable apart from individual travel distance. While it is more difficult to see how car purchasing decisions affect

VKT in developed countries since most households already own an automobile, in developing countries such as China where vehicle ownership is still at low levels this is a central issue that must be confronted. There is substantial room for households to respond along this margin in response to roadway construction. I provide new insights into the importance of controlling for vehicle ownership in this context by analyzing the impact of infrastructure construction on traffic while controlling for growth in the level of car ownership.

The third contribution is the development of a novel matching IV strategy to address endogeneity and aid causal identification, which is combined with traditional instruments such as historical infrastructure plans. I combine the nearest neighbor matching strategy (Abadie and

Imbens 2002) and the idea of “other cities” instruments (Hausman 1996) to derive a new instrument for urban road infrastructure. Road construction in similar but distant cities is likely

8 to be correlated with local roads due to competitive effects among policy-makers but is unlikely to have a direct impact on local traffic. Empirical tests demonstrate the validity of the matching instruments and they are robust across different specifications. In addition, traditional instruments for roadway expansion are also adopted. Following the literature that considers historical roadway networks (Baum-Snow et al. 2015a; Baum-Snow et al. 2015b), I use the length of Chinese municipal roads in 1984 and highways in 1962 as complementary instruments.

As a whole, these instruments enable me to investigate the causal effect of road construction on

VKT. Finally, I validate these results using panel estimation techniques which directly account for time-invariant unobservable attributes specific to each municipality in my study area.

2.2 Literature Review

Downs (1962, 1992) proposes the “fundamental law of highway congestion”, which states that highways are always filled with traffic at their full capacity. Given increasing populations, longer commutes due to suburbanization and the limited amounts of public financing available for infrastructure, understanding how new infrastructure investments are utilized has emerged as an important policy and research objective. A number of studies have empirically examined the impact of new road construction on VKT; however, most of these studies are facility-specific

(Jorgensen 1947; Yagar 1973; Pells 1989). These studies focus on particular roads, and they generally find evidence that expanded roads induce new traffic. Far fewer studies investigate the problem in a larger area (Koppelman 1972; Dyett et al. 1980; Newman and Jeffrey 1989). In these studies, the elasticity of VKT with respect to roadway length in a geographic area is

9 estimated. The results are generally positive indicating that roadway expansion and construction induce additional traffic, although the magnitudes vary.

Duranton and Turner (2011) summarize the problem as a “fundamental law of road congestion” which they extend to general roadways including urban roads. They posit that newly built roads will spontaneously fill with newly generated traffic until the new roads reach a maximum level of capacity. In other words, VKT increases proportionally with respect to the length of roads when all roads are at full capacity. In their empirical analysis, they conduct a comprehensive investigation of the causal effect of lane kilometers of roads on VKT in the

United States. They consider a broad class of roadways in metropolitan areas across the country.

Their results confirm the fundamental law of road congestion implying that new road construction does not reduce the capacity utilization of roadways. Hsu and Zhang (2014) examine the relationship between roads and VKT for national expressways in Japan. They find that expressway expansion leads to a more than proportional increase in VKT, confirming the hypothesis that roadway infrastructure construction induces traffic.

An increasing amount of total traffic does not necessarily suggest road congestion.

However, the elasticity of VKT with respect to roads may imply a change in capacity utilization due to infrastructure construction. In the case of a unit elastic VKT, it is expected that congestion levels neither increase nor decrease as traffic grows proportional to the extent of new roadway expansion. These papers have revealed that road expansion results in more total traffic which can offset the positive effects of expansion on congestion relief. Nevertheless, few existing papers have discussed the importance of accounting for vehicle ownership decisions as a potential

10 confounder in identifying the impact of road construction on traffic levels, despite several related papers focusing on individual decisions of car purchase and use (Berkovec 1986; West 2004;

Bento et al. 2009). Most of the existing VKT literature has focused on developed country contexts where there are limited car ownership expansion possibilities relative to developing countries. As many households in developing countries are entering the car ownership market for the first time, I would expect there to be a greater scope for car ownership changes to confound the implications of road construction on VKT. In this setting, failure to control for these confounding impacts would bias the relationship between roadways and VKT in an upward direction.

In developing country contexts, there is a growing body of literature analyzing the impacts of transportation infrastructure construction in China. Baum-Snow et al. (2015a) investigates the influence of roads and railways on urban form in China. They find that these infrastructure investments displace population and the location of manufacturing jobs from central cities to surrounding areas. Similarly, Baum-Snow et al. (2015b) empirically demonstrates that road construction, which improves access to other broad markets, contributes to the regional growth of population, GDP and income in China. Yu et al. (2015) also finds that motorway construction in China leads to geographic concentration of economic activities. Few papers consider the consequences of road construction in relation to traffic, particularly within urban districts. To my knowledge, there has been no discussion surrounding the applicability of the relationship between road construction and VKT in China.

11 2.3 A Conceptual Model

To illustrate households’ behavioral changes in response to road construction, I develop a conceptual model of individual decision making linking the travel distance decision and travel mode decision. In this model, the decision maker is consuming VKT, denoted by Q, and leisure, denoted by L. She maximizes utility subject to a time constraint. The maximization problem is written as follows.

max 푄훼퐿1−훼 (1) 푄,퐿 푄 s. t. + 퐿 + 푡 = 푇 푣(푅, 퐶푎푟) VKT makes a positive contribution to her utility because travel enables individuals to reach a larger area or visit the same place more frequently. Utility is specified to be

Cobb-Douglas, where α, the elasticity of VKT, is positive but less than one. In the time constraint, travel speed 푣(푅, 퐶푎푟) is assumed to be a function of roads, denoted by R, and vehicle ownership is denoted by Car.1 In this specification, roads can facilitate travel as more roads that are available dilute traffic on each road, ceteris paribus. Owning cars (and thus choosing driving as the travel mode) also increases the travel speed, because in general it is faster than walking or riding bicycles and is more convenient than public transport due to a more flexible time schedule.

It is noticeable that general equilibrium effects on the aggregate VKT and speed have not been considered in this simple model regarding an individual. Hence, 푣(푅, 퐶푎푟) is an increasing

휕푣(푅,퐶푎푟) 휕푣(푅,퐶푎푟) function in R and Car, i.e. > 0 푎푛푑 > 0. Moreover, roads and cars are 휕푅 휕퐶푎푟 assumed complementary. The marginal effect of roads on travel speed rises with a car as roads

1 Here I treat Car as a continuous variable.

12 mainly serve automobiles. In turn, a larger quantity of roadway provision also enhances the

휕2푣(푅,퐶푎푟) 휕2푣(푅,퐶푎푟) practical value of cars. Therefore, I assume that > 0 ( > 0). In this case, 휕푅휕퐶푎푟 휕퐶푎푟휕푅 road expansion makes it more worthwhile to purchase a car and choose driving as the travel mode.

In the constraint, VKT or travel distance divided by the speed represents the time that the individual spends on travel. Apart from travel time and leisure, t is the time allocated to working, which is assumed to be fixed in this model. The sum of these three terms is restricted to be no greater than the total time T per day. Under this constraint, although the individual receives utility from the ability to travel greater distances, she loses utility from the associated travel time that competes for leisure time. Substituting the constraint into the utility function and taking the first order condition with respect to Q, the optimal VKT is obtained.

푄 = 훼(푇 − 푡)푣(푅, 퐶푎푟) (2)

To further investigate the impact of road construction on VKT, I consider the total differential of the equation Q − 훼(푇 − 푡)푣(푅, 퐶푎푟) = 0 and the result is shown in equation (3) below. 휕푣(푅, 퐶푎푟) 휕푣(푅, 퐶푎푟) 푑푄 − 훼(푇 − 푡) 푑푅 − 훼(푇 − 푡) 푑퐶푎푟 = 0 (3) 휕푅 휕퐶푎푟 Rearranging I obtain 푑푄 휕푣(푅, 퐶푎푟) 휕푣(푅, 퐶푎푟) 푑퐶푎푟 = 훼(푇 − 푡) + 훼(푇 − 푡) (4) 푑푅 휕푅 휕퐶푎푟 푑푅 The marginal effect of roads on VKT is composed of two components. The first term depends on the contribution of roads to travel speed. It is the direct impact of roads on the

휕푣(푅,퐶푎푟) individual’s VKT decision. Based on the assumption that > 0, this direct effect is 휕푅 positive. In contrast, the second term illustrates the role of vehicle ownership where roads

13 influence VKT. The sign of this term depends on the contribution of car ownership to travel speed as well as the impact of road expansion on car ownership decisions. Recall that

휕푣(푅,퐶푎푟) 푑퐶푎푟 > 0 which means owning a car increases an individual’s travel speed. If > 0 휕퐶푎푟 푑푅 holds as well, the second term as a whole will be greater than zero. In other words, as long as a person’s propensity to purchase a car is positively correlated with the provision of roadway facilities, the indirect road effect on VKT through car ownership would also be positive.

This theoretical result reveals the different pathways through which road construction would affect VKT. Newly built roads directly stimulate car drivers’ demand for longer travel distance. More roadway facilities also generate demand for vehicles in general, which makes a positive contribution to traffic. In this chapter, I estimate the overall marginal effect of road construction on VKT while controlling for the potentially confounding impacts of changes in driving behaviors arising from changes in vehicle ownership.

2.4 Data

Traffic congestion is one of the major concerns for commuters in China. According to the Beijing

Transportation Research Center, in 2015 congestion (beyond moderate level) lasted for 3 hours on average per day in Beijing, while it was 1 hour and 55 minutes in 20142. Unfortunately, this level of congestion is not a unique issue in one or two large cities, but a widespread and growing problem across China. By 2015, there had been 13 major Chinese cities, including Beijing, where driving restrictions were applied to mitigate congestion, and an increasing number of

2 Beijing Transportation Research Center, "2015 Annual Report of Beijing Transportation"

14 cities are planning to take similar policies3. Meanwhile, the rapid construction of roadway infrastructure is also impressive especially in urban areas. As reported by the National Bureau of

Statistics of China (NBS), the total length of urban roads across the country increased from 130 thousand kilometers in 1995 to 365 thousand kilometers in 2015. The lengths of municipal roads in China have almost tripled over the past 20 years, and are still expanding rapidly.

(a) Change in vehicle-kilometers traveled (VKT), 2011-2014

8,000,000

6,000,000

4,000,000

2,000,000

0 Less0 than

(b) Change in kilometers of major urban roads, 2011-2014

2,000

1,600

1,200

800

400 0

Figure 1. Differences in VKT and roads in 103 Chinese cities, 2011-2014

3 Zhu J, Dai S (2014) Review on urban traffic demand management of vehicle restriction. Transportation Standardization 42.21: 33-39 (in Chinese)

15 I develop a novel panel dataset from 2011 to 2014 spanning urban districts of 103

Chinese prefecture-level cities and municipalities shown in Figure 1. These cities are distributed throughout China across 30 provincial administrative districts. The data are sourced from the

China Statistical Yearbook, the China Statistical Yearbook on Environment, the China City

Statistical Yearbook, and the China Statistical Yearbook for Regional Economy. Additional complementary data are obtained from each city’s statistical yearbook.

The dependent variable, VKT, is derived from the average traffic volume per hour reported in each city associated with urban roads. Traffic volume is defined as the number of cars passing any point on the urban road system in an hour. VKT on a road segment equals the traffic volume times the length of the segment. In my dataset, each city reports the average traffic volume weighted by the lengths of each road segment. Therefore, the total (hourly) VKT for each city can be derived as long as the total length of roads in the city is known.

Urban roads are defined as paved roads with a width of at least 3.5 meters within a city district. My data set contains the lengths of urban roads by central lines for vehicles. For car ownership, I consider the number of civil vehicles owned. All vehicles with a civil vehicle license plate are counted, including private vehicles. Non-car vehicles such as motorcycles and tractors are excluded in my data set. In addition, I use land area and population of the jurisdiction as controls for city characteristics. I also control for economic factors such as regional GDP and disposable income per capita of urban households. To further control for the effect of public transportation, I consider the numbers of buses and taxis in a city.

16 Table 1. Summary statistics

Year 2011 2012 2013 2014 Mean hourly VKT (1,000 km) 3614 3791 3653 3626 (7227) (6235) (5338) (4973) Mean average hourly traffic volume (car/h) 1945 1941 1877 1812 (911) (906) (920) (894) Mean kilometers of urban roads (km) 1511 1616 1643 1726 (1820) (1849) (1701) (1769) Mean number of civil vehicles owned (unit) 594318 675713 766128 877701 (628302) (686795) (756872) (838655) Mean land area of districts under city (sq.km) 2658 2658 2658 2658 (3331) (3331) (3331) (3331) Mean population (10,000 population) 234 238 243 249 (262) (264) (266) (278) Mean gross regional product (100 million yuan) 2108 2369 2631 2870 (3083) (3392) (3720) (4130) Mean urban household disposable income per capita (yuan) 22537 25517 27846 30313 (5337) (5846) (6435) (6982) Mean number of buses (unit) 2771 2881 3094 3175 (4194) (4287) (4523) (4583) Mean number of taxis (unit) 5991 6061 6231 6367 (9114) (9200) (9282) (9264) Number cities 103 103 103 103 Standard deviations in parentheses.

Descriptive statistics for all variables are presented in Table 1. Examining mean values, the length of urban roads and the level of car ownership increases year over year as one would expect. I also present the key variables of interest in Figure 2. Choropleth maps are drawn to show the changes in VKT and kilometers of urban roads in China from year 2011 to 2014. Each prefecture that an observed city is located in is colored and the darker values indicate higher growth (the whole prefectures are colored to display values clearly but my data only account

VKT and municipal roads within urban districts). To better examine the spatial distribution of the variables, I divide the country into 6 divisions based on geography and the level of economic development. From the resulting figures, I can see that the east coast and southwestern areas

17 experience relatively faster growth in urban road infrastructure and VKT compared to other areas.

Overall, the patterns of these variables provide intuition with respect to the positive relationship between VKT and roads.

2.5 Empirical Strategy

To empirically analyze how VKT is affected by road provision, I consider the following log-log model for empirical estimation.

ln(푄푖푡) = 퐴0 + 휌 ln(푅푖푡) + 퐴1 ln(푋푖푡) + 휃푖 + 휇푡 + 휖푖푡 (5)

In this reduced-form regression, i indexes cities and t indexes years. Q and R denote total VKT and total length of roads. X denotes observable characteristics at the municipal level. 휃푖 denotes time-invariant unobservable characteristics while 휇푡 denotes year fixed effects common across all cities.

My primary focus is the effect of roads on VKT. It is important to note that both roads and VKT take natural logarithm forms and thus the coefficient ρ represents elasticity of VKT with respect to the length of roads. My use of the term VKT refers to traffic. Higher VKT implies a greater amount of total travel distance by all vehicles in an area for a specified time period. In my empirical investigation, the first objective is to examine whether road expansion leads to an increase in the total VKT, i.e. ρ is positive and significantly different from zero. If this is the case, building more roads will increase the total amount of traffic, no matter how the increased traffic is allocated between existing and additional roads. When road expansion has a positive effect on

VKT, individuals change their behavior and drive longer distances in response to more roads

18 provided.

My second objective is to estimate the magnitude of the elasticity of VKT with respect to roads. An increasing amount of traffic is not necessarily associated with worsening congestion.

However, VKT relative to the length of road infrastructure may provide insights that allow me to infer congestion levels, assuming that all roads are identical, and the total traffic is evenly distributed among roads4. In a case where ρ = 1, VKT increases proportionally to road length.

Theoretically, newly built roads will be filled with a proportional amount of newly induced traffic. The extended road will be just as congested as already existing roadways and congestion levels will not be affected. Following this idea, I can compare ρ with one to infer the potential change in congestion levels. It is notable that these interpretations are under the assumption that the capacity of newly built roads is identical to that of existing roads. When the road quality is improving as is often the case in reality, an elastic VKT does not necessary imply that road infrastructure construction has no effect on congestion relief. However, it is a warning that newly induced traffic can offset the effects of road construction on mitigating congestion.

Instrumental variables and matching strategy

To estimate the causal effect of road construction on traffic, there are several potential identification concerns. Reverse causality is a concern if greater levels of traffic in a city indicate a higher demand for roads, which may stimulate construction of new roadways. However, in the

4 Some studies use speed to measure congestion (Couture et al. 2017). Indeed, this measure is more intuitive since speed is directly perceived by drivers and can be more easily interpreted. However, it is more difficult to monitor car speed on roads than traffic volume (a.k.a. flow rate) and I do not observe a nationwide record of the vehicle speed within urban areas in China. In addition, speed is just one factor of volume flow rate. As a comprehensive measure of traffic, volume flow rate is also influenced by car density on roads. See Tsekeris and Geroliminis (2013).

19 short term, this positive feedback effect may not be of significant concern due to the time it takes to adjust the provision of roads or to complete construction. Relative to the often multi-year length of construction projects, household travel decisions are significantly more responsive in short periods. As a result, attributing roads as a determinant of VKT is much more likely the case than vice versa over a short time period.

A second identification concern is that governments may invest more on infrastructure as a fiscal policy in cities that are less developed or that are faced with negative shocks to stimulate the local economy (Duranton and Turner 2012). Because VKT is positively correlated with local economic activities, this type of targeted investment could result in well-known endogeneity concerns. The negative feedback this would create would lead to an underestimation of road effects on VKT (Duranton and Turner 2011). To address this issue, I develop and adopt a novel instrumentation strategy.

Following the literature I first use historical roadway networks to instrument for current roads (Baum-Snow 2007; Duranton and Turner 2011; Baum-Snow et al. 2015a; Baum-Snow et al. 2015b). Baum-Snow (2007) suggests that the historical road networks could be an ideal instrument because current roads are built based on old roads while previous road plans are not affected by current economic activities. Therefore, the early road plan is relevant but exogenous, qualifying as a valid instrument for road construction. Duranton and Turner (2011) apply this strategy in the United States using exploration routes in 1835, railroads in 1898 and planned interstate highways in 1947 to instrument current interstate highways. Baum-Snow et al. (2015a) and Baum-Snow et al. (2015b) follow the same idea and adopt Chinese historical transportation

20 infrastructures in 1962 as instruments to address endogenous problem in terms of highway construction in China.

In this chapter, I adopt the footprint of 1984 Chinese municipal roads as an instrument for current urban roads, which is the earliest record I am able to find roadway maps covering all 103

Chinese cities used in my study. China did not begin market economic reforms until the 1980s.

Prior to this period, resources were generally not allocated by the market and road construction decisions were made using limited economic considerations (Baum-Snow et al. 2015a). Second, households in China did not begin to rapidly expand ownership of automobiles until the 21st century (Huo et al. 2012), reducing correlations between early roadways and present day traffic.

In addition to urban roads I also adopt the 1962 routes of Chinese highways as an additional instrument following Baum-Snow et al. (2015a) and Baum-Snow et al. (2015b). I digitize the

1962 map from SinoMaps Press (1962) and calculate the length of highways in each prefecture

(using current prefecture boundaries). It is noteworthy that highways and urban roads are two distinct types of transportation infrastructure designed to serve different purposes and are likely competing for funds and lands within a certain location. According to the design specifications for highway alignment issued by the Ministry of Communications of China, highway planning is supposed to avoid areas with high municipal road density. Therefore, I expect the correlation between the lengths of each type of infrastructure is negative.

For historical infrastructure to serve as valid instruments, the exclusion restriction requires that the instruments and the dependent variable be orthogonal conditional on control variables. As Duranton and Turner (2011) point out, population is a key control variable since

21 large cities are likely to receive more roads. Without population controls, old roads would not be exogenous because they were associated with local population levels at the time, which predicts current VKT. In this chapter, where I focus on developing countries, car ownership is of even greater concern because it is no longer highly coupled with population but evolves separately, serving as a potential link between historical roads and current traffic. Therefore, using appropriate control variables including both economic factors as well as car ownership measures is critical to ensure the exogeneity of old infrastructure as valid instruments.

A drawback of the historical roadway instrument is that it is usually cross sectional in nature. As a result, this instrument likely fails to provide identification for time varying additions of roadways, particularly in a panel data setting. In order to instrument panel data, as is the case in this chapter, I develop an additional time varying instrument building on the “other cities” instrumentation strategy developed by Hausman (1996) but shifting the context from prices to infrastructure. To address endogenous prices when estimating price elasticity of demand,

Hausman (1996) suggests that prices of the same goods in other cities can be an ideal instrument.

He argues that the goods in different cities share the same cost as a fraction of their prices, while the remaining differences in price arise from stochastic disturbances that are independent across cities. Therefore, prices in distant cities are correlated with local prices but do not affect local demand through other means. Using a similar argument, Autor et al. (2013) adopts Chinese imports of other countries as an instrument in their study on import competition in the U.S. The underlying assumption is that Chinese imports to the United States and other countries share a common within-industry component that is driven by comparative advantage and trade costs in

22 China.

I apply a similar strategy to form a time-varying instrument in this chapter using matching methods. I use the kilometers of urban roads in distant cities that are similar to but located away from the focal city of interest as an instrument. In China, investment in municipal facilities such as urban roads comes from several sources. Central and local government budgets each provide funding for urban infrastructure projects. Domestic banks also offer financial support for projects in the form of loans. In most cases these large banks are state-owned and managed by the Ministry of Finance and other central state-owned investment corporations. In

2014, the top five state-owned commercial banks had an 85.85% share in the domestic loan market.

Another source of funding for public investment is government bonds. Chinese local governments were not allowed to issue municipal bonds until 2011 when several pilot cities were certified to be nominal issuers but the repayment was implemented by the central government.

After 2014 local governments started to have limited autonomy in terms of municipal bonds but the quota under which local governments could issue bonds was strictly controlled by the central government. The rest financial sources include foreign investment and self-raised funds, i.e. local government non-budgetary revenue from selling the usufruct rights of land.

As a whole, sources used to finance local public projects can be classified into two categories: funding provided or allocated by the central government and funding managed by local governments. The former category includes financial allocation from central government budget, loans and bonds. According to the China Urban-Rural Construction Statistical Yearbook,

23 in 2014 these sources accounted for 30.34% of the national investment in urban service facilities.

Through these instruments, the central government decides how many resources each city can obtain and therefore different cities are competing with each other for limited funds. On the other hand, local funding for municipal infrastructure construction, i.e. local tax revenue and revenue from land markets, is affected by domestic trends in economic growth. From this perspective, investments on local infrastructure in different cities are likely to be positively correlated due to common economic trends.

For my "other cities" instrument, I focus on the competition among cities for central government funds as the mechanism through which roads in distant cities have explanatory power on local roads. The domestic trends that affect local public funds are endogenously correlated with traffic and I control for the positive correlation among roads in different cities by controlling for city economic characteristics. Having netted out local economic conditions, the remaining correlation among cities competing for local infrastructure construction leads to a negative correlation between local roads and roads in distant cities. Large quantities of urban road construction in other similar cities would suggest a low level of investment on local roadway systems in your own city.

In addition to funding, the central government also has substantial impacts on urban construction planning. For example, the overall urban planning of provincial capitals and other influential cities (most of the cities in my sample) requires approval by the State Council, where the urban sprawl in each city is strictly regulated by the central government. Therefore, centralized planning also contributes to competition among different cities for local

24 infrastructure.

Due to this competition, municipal roads in distant cities are relevant to local roads. This instrument, however, is exogenous to local traffic when holding constant the effect of local roads.

First, households’ travel decisions are unlikely to be directly affected by construction in a distant city. Urban roads in faraway locations do not influence local VKT in terms of commuting.

Second, road construction as a result of the competition among cities is likely correlated with traffic. Hence, roads in other cities can function as an instrument as they are correlated with the explanatory variable, local roads, but are uncorrelated with the error term in the explanatory equation. I also empirically test the exogeneity assumption given the over identified nature of my model.

To econometrically implement this instrument, it is important to define what constitutes appropriate “other cities” needed to derive the instrument. I adopt the nearest neighborhood matching strategy by Abadie and Imbens (2002) to find suitable city matches and use the kilometers of urban roads in the matched cities to instrument local urban roads. For each observation in the treated group, one or more observations in the control group are matched if the

“distance” between the matches is the shortest. The “distance” is measured by a specified norm between two vectors of covariates for observations. Details of implementation are explained in

Abadie et al. (2004).

In this chapter, I consider the Mahalanobis distance between cities based on a suite of city specific covariates listed in Table 2. For each city, I obtain the nearest 4 matched cities for each year of my data. To avoid spillover effects, I exclude cities located in the same province from

25 serving as matched cities ensuring that all matched cities are located outside of the province where the focal city is located in. The resulting matched cities are, on average, 1,074 kilometers away from the focal city. As a representative illustration, Figure 2 depicts a typical city Changsha and its 4 matched counterparts in 2014.

Table 2. Mean covariate differences between focal and matched cities

2011 2012 2013 2014 Pooled

Covariates Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD) Area 0.17 (0.81) 0.17 (0.81) 0.17 (0.81) 0.16 (0.81) 0.17 (0.81) Population 0.15 (0.72) 0.16 (0.72) 0.16 (0.70) 0.16 (0.72) 0.16 (0.71) GDP 0.13 (0.57) 0.14 (0.55) 0.14 (0.53) 0.14 (0.54) 0.14 (0.55) Income 0.07 (0.44) 0.06 (0.45) 0.07 (0.44) 0.07 (0.44) 0.07 (0.44) Every city is matched with 4 cities out of the province based on the listed covariates. In this table, all the covariates have been normalized to have zero mean and unit variance in each year. Average differences of the covariates within each matched pair are shown as well as the standard deviation of these differences.

Dalian

Qingdao Zibo

Changsha Fuzhou

0 500 1,000 km

Figure 2. Example of matched cities for Changsha in 2014

26 Having established matched cities for each city in my dataset, I next take the average of urban road kilometers in the 4 matched cities for each focal city. I consider a 5-year-lag of urban roads in the matched cities to further ensure exogeneity. In this way, the resulting matching instrument is “distant” from the traffic variable of interest in terms of political governance, geography and time while the matched cities are “close” to the focal cities regarding city characteristics. Table 2 provides measures of closeness for key covariates resulting from this matching exercise. Following Abadie and Imbens (2011), I normalize each covariate to have zero mean and unit variance in each year cross-section and report the mean of within-pair differences.

When each city (N in total) are matched with M cities, I obtain the following distance metric

푁 푀 ̅ (푚) ̅ 푋푖푡 − 푋푡 푋푖푡 − 푋푡 푚푒푎푛_푛표푟_푑𝑖푓푋,푡 = ∑ ∑ [ − ] /(푁 × 푀) (6) 푆푡 푆푡 푖=1 푚=1

∑푁 푋 ∑푁 (푋 −푋̅ )2 where 푋̅ = 푖=1 푖푡 and 푆 = √ 푖=1 푖푡 푡 . 푋(푚) is the covariate X in the mth matched city 푡 푁 푡 푁−1 푖푡 for city i in year t. The measures presented in Table 2 indicate high quality of my matches. As is expected, the average differences within matched pairs are very close to zero and small relative to the standard deviations of these differences. It suggests that my matching approach selects cities with very similar geographic, demographic and economic characteristics. Appendix Table

19 reports the mean of differences within pairs for each city.

The use of the matching process to derive a panel instrument for local municipal roads adds a key new source of identification to the standard historical instruments used in prior studies. Econometric tests are conducted to examine the validity of these instruments in Section

2.6.

27 2.6 Estimation Results

Out initial set of results uses naïve OLS models that will later serve as a benchmark to evaluate

2SLS estimation. The estimated elasticities of VKT for a series of OLS models are reported in

Table 3. All specifications include division-by-year fixed effects (Appendix Table 16 presents specifications with no fixed effects as well as division and year fixed effects). In column (1) I only include the length of urban roads which is my primary variable of interest. In this specification, I find that the estimated elasticity of VKT with respect to roads is large, exceeding

1 with an estimate of 1.211. In column (2) I add city characteristics such as land area, population, local GDP and disposable income per capita as additional controls. Having controlled for a number of local economic and city conditions, the elasticity of VKT decreases with an estimate of 1.109. Finally, in column (3) I further consider the level of car ownership as well as public transportation vehicles such as buses and taxis as additional controls that could influence VKT.

In this specification I account for drivers’ behavior changes in response to road expansion as a pathway to affect traffic. The estimated elasticity for this specification is further reduced and close to 1 with an estimate of 0.997, consistent with my expectations of a positive bias without including controls for car ownership.

2SLS estimates of the road effect on VKT

To address potential endogeneity concerns, I use a 2SLS framework that enables me to recover causal effects with the aforementioned instruments. The instruments I use include 1984 urban roads and 1962 highways as well as lagged roads in matched cities. For the matching instrument,

28 Table 3. OLS estimates of road effect on VKT

(1) (2) (3)

Variables ln(VKT) ln(VKT) ln(VKT) ln(Road km) 1.211*** 1.109*** 0.997*** (0.0348) (0.0729) (0.0739) ln(Car ownership number) 0.204*** (0.0556) ln(Land area) -0.0515 0.0747** (0.0315) (0.0330) ln(Population) 0.0717 -0.168** (0.0728) (0.0687) ln(GDP) 0.0919 -0.127 (0.0685) (0.0918) ln(Income) -0.185 -0.349* (0.159) (0.173) ln(Bus number) 0.185** (0.0740) ln(Taxi number) 0.122** (0.0474)

Division×Year FE (24) Yes Yes Yes Observations 409 409 409 R2 0.864 0.866 0.882 All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

I select 4 matched cities and calculate the average length of urban roads with a 5-year lag. I report results of both first and second stage estimates in Table 4.

Column (5) presents the first stage results of my preferred specification with a comprehensive control set and division-by-year fixed effects (for other fixed effect specifications, see Table 17 in the appendix). For old infrastructure instruments, the coefficient of urban roads in

1984 is positive while that of highways in 1962 is negative, confirming my hypotheses of historical-road based construction and road-highway competition. The coefficient of the matching instrument is negative, as I expect due to competition across cities for limited public funds. All three instruments are strongly statistically significant.

29 Table 4. 2SLS estimates of road effect on VKT

(1) (2) (3) (4) (5) (6)

Variables ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) 1.255*** 0.993*** 1.096***

(0.0375) (0.240) (0.217) ln(Car ownership number) 0.102 0.206*** (0.0621) (0.0543)

Instruments: ln(1984 Road km) 0.451*** 0.0915*** 0.0977**

(0.0546) (0.0310) (0.0400)

ln(1962 Highway km) -0.109 -0.0957*** -0.130***

(0.0683) (0.0353) (0.0241)

ln(5-year lagged road km 0.566*** -0.179*** -0.180***

from 4 matched cities) (0.0973) (0.0622) (0.0564) ln(Land area) 0.0829*** -0.0454 0.140*** 0.0642* (0.0310) (0.0319) (0.0257) (0.0356) ln(Population) 0.278*** 0.0915 0.160** -0.174*** (0.0622) (0.0781) (0.0727) (0.0624) ln(GDP) 0.608*** 0.172 0.457*** -0.179 (0.0435) (0.159) (0.0455) (0.117) ln(Income) 0.0312 -0.226 -0.109 -0.310* (0.178) (0.159) (0.198) (0.186) ln(Bus number) 0.249*** 0.158 (0.0884) (0.0962) ln(Taxi number) -0.108* 0.126*** (0.0583) (0.0454)

Division×Year FE (24) Yes Yes Yes Yes Yes Yes Observations 409 409 409 409 409 409 R2 0.751 0.863 0.898 0.865 0.908 0.881 First-stage statistic 458.48 7.85 12.15

Overidentification p-value 0.003 0.001 0.198

Odd columns present first stages and even columns present second stages. All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

To empirically assess the validity of my instruments, I conduct statistical tests to examine their relevance and exogeneity. For relevance, I look at the F test in the first stage for the joint significance of the coefficients. As is shown in column (6) of Table 4, the F statistic indicates that historical and matched roads are strong instruments in my preferred specification. For exogeneity,

30 adopting more than one instrument allows me to test using overidentifying restrictions. I conduct

Sargan’s χ2 test with the null hypothesis that the instruments are uncorrelated with the error term.

Results shows that the p-value is greater than 0.1 and I cannot reject the null of exogeneity at a

10% significance level. Through these tests, I have empirically demonstrated that historical and matched infrastructures are valid instruments in my model.

The second-stage regression estimates of my preferred specification in column (6) contain the elasticity of VKT with respect to roads. The estimated elasticity is 1.096, which is greater than the estimates from OLS models. The underestimation by OLS suggests a negative feedback as I discussed in Section 2.5 that infrastructure is likely to be allocated to cities experiencing negative shocks as a potential economic recovery effort. The resulting estimate is slightly greater than 1 and indicates that providing 1% more urban roads increases VKT by

1.096%. The elasticity above 1 is possible if earlier roads were not fully congested or the capacity of newly built roads are of higher quality due to better design or engineering improvements. This is very likely the case in China where the number of cars increases from a low starting point while roadway infrastructure experiences substantial investment.

When vehicle control variables are not included in the 2SLS model as shown in columns

(3) and (4), my instruments are no longer exogenous and do not pass the overidentification test.

This reveals the importance of controlling for vehicle ownership in order to exclude any non-road pathways through which the instruments would affect VKT. Without appropriate controls, my instruments are unsurprisingly unable to address potential endogeneity issues

31 associated with the road elasticity covariate5.

Robustness

In Table 5, I present several alternative strategies to create my matching instrument to examine whether the 2SLS estimates are sensitive to matching assumptions. First, I consider different numbers of matches when constructing the average length of urban road kilometers as an instrument. I use the mean of roads from 2 matched cities in columns (1) and (2) and 6 matched cities in columns (3) and (4). These instruments remain valid as the F statistics are high in the first stages and the overidentification tests cannot reject the null of exogeneity. The matching instruments are significant and negative in the first stage as expected regardless of how many matches are used. Estimates in the second stage are also robust. The estimated elasticities of

VKT with respective to roads are quantitatively and qualitatively similar to the prior estimates of

1.1, significant at a 1% level. These findings illustrate the robustness of my results to matching specification.

The second set of robustness results examines whether the geographic restrictions on city matches impact my results. Excluding matches within the same province cannot prevent neighboring cities from being matched if they are located close to the province borders. As a robustness check, I instead restrict all matches to be at least 200 kilometers away from each other and the results are presented in column (5) and (6) of Table 5. The instruments are still relevant and exogenous while the estimates are very similar to previous results.

5 I have also tried to include the car ownership number in the 2SLS model as an endogenous variable together with the road length. The results regarding roads highly resemble those in column (6) and my conclusion keeps the same.

32 Table 5. Robustness to matching specifications

(1) (2) (3) (4) (5) (6) Variables ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) 1.133*** 1.174*** 1.104***

(0.277) (0.224) (0.264) ln(Car ownership number) 0.0987 0.206*** 0.113* 0.207*** 0.100 0.206*** (0.0654) (0.0542) (0.0647) (0.0559) (0.0618) (0.0538)

Instruments: ln(1984 Road km) 0.0923** 0.0961** 0.0976**

(0.0428) (0.0402) (0.0429)

ln(1962 Highway km) -0.130*** -0.141*** -0.133***

(0.0258) (0.0253) (0.0249)

ln(5-year lagged road km -0.0889***

from 2 matched cities) (0.0314)

ln(5-year lagged road km -0.177***

from 6 matched cities) (0.0517)

ln(5-year lagged road km from -0.140***

4 matched cities 200 km away) (0.0473) ln(Land area) 0.140*** 0.0603 0.141*** 0.0560 0.137*** 0.0634 (0.0271) (0.0394) (0.0258) (0.0352) (0.0265) (0.0405) ln(Population) 0.0984 -0.176*** 0.151** -0.179*** 0.134* -0.175*** (0.0637) (0.0641) (0.0659) (0.0620) (0.0725) (0.0625) ln(GDP) 0.474*** -0.198 0.449*** -0.219* 0.471*** -0.183 (0.0451) (0.129) (0.0437) (0.113) (0.0457) (0.132) ln(Income) -0.298 -0.296 -0.0961 -0.280 -0.211 -0.307 (0.199) (0.215) (0.183) (0.199) (0.203) (0.200) ln(Bus number) 0.255*** 0.147 0.242*** 0.136 0.248*** 0.155 (0.0913) (0.113) (0.0886) (0.105) (0.0886) (0.103) ln(Taxi number) -0.105 0.128*** -0.103* 0.129*** -0.112* 0.126*** (0.0664) (0.0472) (0.0601) (0.0483) (0.0598) (0.0455)

Division×Year FE (24) Yes Yes Yes Yes Yes Yes Observations 409 409 409 409 409 409 R2 0.907 0.881 0.908 0.880 0.907 0.881 First-stage statistic 12.65 12.34 11.13

Overidentification p-value 0.212 0.220 0.196

The same regressions with different instruments are performed in all pairs of columns. Odd columns present first stages and even columns present second stages. All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

As a whole, I find that my estimates are robust across a number of different specifications.

The estimated elasticity of VKT with respect to road lengths ranges between 1.104 and 1.174,

33 indicating that urban road construction in Chinese cities results in a more than proportional increase in VKT. In Table 18 in the appendix I further explore robustness to changes in year lags of urban roads in the matched cities. Using either 1-year-lagged or current matching instruments, the results are virtually identical to those in previous models using 5-year-lagged instruments.

Table 6. Panel estimates of road effect on VKT with city fixed effects

(1) (2) (3)

Variables ln(VKT) ln(VKT) ln(VKT) ln(Road km) 0.877*** 1.102*** 1.087*** (0.134) (0.184) (0.187) ln(Car ownership number) -0.299 (0.224) ln(Population) -0.637 -0.638 (0.519) (0.505) ln(GDP) 0.400 0.421 (0.305) (0.328) ln(Income) -0.683** -0.345 (0.298) (0.295) ln(Bus number) 0.106 (0.139) ln(Taxi number) -0.00801 (0.132)

City FE (103) Yes Yes Yes Observations 409 409 409 R2 0.957 0.959 0.959 All regressions include a constant. Robust standard errors in parentheses (clustered by city). *** p < 0.01, ** p < 0.05, * p < 0.1.

The 2SLS estimation results rely on both spatial and temporal variation across and within cities. The historical infrastructure plan instruments are time-invariant and would be absorbed if I controlled city fixed effects. My final robustness results seek to isolate temporal variation for identification. To accomplish this, I consider a panel model with city fixed effects and the results are presented in Table 6. When city characteristics and vehicles ownership is controlled, the

34 estimated elasticity is 1.087 as is shown in column (3). This result is similar to my findings from

2SLS models, despite relying only on temporal variation within cities for identification.

2.7 Discussion

In this chapter, I examine the impact of urban road expansion on traffic in Chinese cities. The estimated elasticity of VKT with respect to road length is approximately 1.1, indicating that newly built urban roads would lead to a more than proportional increase in road traffic. A novel matching instrument is adopted to address possible endogeneity in addition to traditional instruments and the use of fixed effects. My findings reveal that it is important for policy-makers to pay attention to induced traffic when investing in road infrastructure. When draining a pool with hoses, an additional hose may help accelerate the process given that the total water volume is constant. However, it is more difficult to divert traffic using additional roads: the total amount of traffic increases because people drive more in response to new road provision. Road expansion may be effective to ease traffic if the total amount of traffic is constant or increasing due to other exogenous sources. Nevertheless, I find that the newly built roads themselves increase traffic levels.

My results demonstrate that the estimated elasticity of VKT with respect to roads is approximately 1.1. The elastic VKT suggests that newly built urban roads would be filled with more than a proportional amount of traffic. I am not able to predict the consequent congestion levels since more information on road quality over time such as additional width and capacity of roadways is needed. However, what I can conclude is that the large amount of induced traffic has

35 neutralized part of the positive effects of road expansion on congestion relief, if not all.

Admittedly, roadway construction has made tremendous contributions to urban growth and public welfare as is well documented in the literature. Despite these benefits, the induced traffic should be considered and measured especially when the construction is primarily aimed at relieving congestion.

Since I focus on city roads in urban areas, the results provide a cautionary warning to city planners that building additional roads to mitigate traffic congestion might not be as effective as expected. Some Chinese cities, such as Beijing, have utilized significant quantities of money investing in urban roadway infrastructure with a primary goal of reducing congestion. Not surprisingly given my results, the congestion reducing benefits of these policies have not fully materialized. In fact, alternative policies such as encouraging people to use public transport could potentially be a better method to relieve traffic congestion. Enhancement of underground capacity does not stimulate any VKT on roads. Provision of more buses and taxis is also not likely to induce traffic at the same level as road expansion. While there are potentially many other economic benefits to investing in roadway infrastructure, congestion reduction does not appear to be one of the main resulting benefits given the large amount of induced traffic.

36 Chapter 3: Compensating Differentials of Rents, Wages and Agricultural Returns: the Quality-of-life among Indonesian Regencies and Cities

3.1 Introduction

One of the most important goals of a society is to improve individual quality of life. It has long been challenging for the government and other interested parties to assess local residents’ quality of life, which is affected by numerous natural and man-made amenities. The difficulty is that these amenities, including many environmental goods and services, are often not traded in markets. Rarely are there explicit market prices that can be used to measure their monetary values. Despite the absence of markets, nevertheless, households make residential location decisions in part due to nearby amenities and consequently these nonmarket amenities are allocated among individuals through market mechanisms. The related markets for traded goods generate compensating differentials, providing information to reveal the value of nonmarket amenities (Mulligan and Carruthers, 2011; Powell and Shan, 2012; Halland and Greenman,

2015). This chapter develops a novel compensating differential model of quality of life rankings and introduces an explicit agricultural sector absent in the traditional quality of life literature. I use data from a developing country, Indonesia, and demonstrate the applicability of the model in this context. I recover quality-of-life rankings for jurisdictions across the country at distinct time periods, which allows me to examine how quality of life rankings have changed over time in response to government investment.

Originating from Roback (1982), the quality-of-life index is constructed as a sum of

37 amenities weighted by their implicit prices. The equilibrium model illustrates that the implicit prices for amenities is calculated using housing price and wage differentials which are estimated using hedonic methods. This framework provides a powerful tool to quantify the overall amenity level for residents and facilitates comparison in terms of quality of life among different places.

The result is beneficial for local households to help them understand their relative state of current living conditions as they make future residential choices. These rankings also yield relevant information for policy decisions regarding population, migration and public investment.

Using Roback’s (1982) model and its variants, many papers have demonstrated the existence of compensating differentials in the United States (Blomquist et al. 1988; Beeson and

Eberts 1989; Gyourko and Tracy 1991; Gabriel et al. 2003). In the developed country context, it is not surprising that the markets for tradeable goods such as houses and labor are tightly associated with local amenities where the observed prices of the goods and the unobserved values of the amenities are closely linked. If an apartment is more expensive than average, it is in part because the higher rent is due to higher levels of local amenities. In contrast, if there is a disamenity, anyone working in those locations is likely to require a higher wage to compensate for the negative ambient conditions. From these daily examples, however, it is fascinating to see that the key cause of the occurrence of compensating differentials is not the markets themselves, but humans’ inherent demand for “compensation”. The compensation can be achieved through markets, but not necessarily. Berger et al. (2008) found evidence of the existence of compensating differentials in a transition economy, Russia, where there remained many policy tools of the planned economy inherited from the former Soviet Union. An example is regional

38 wage coefficients -- a worker relocation policy that subsidizes citizens working in undesirable locations with extreme weather conditions. Other Soviet policies such as housing subsidies also functioned as tools to provide compensation in a similar way. It indicates that the concept of compensating differentials existed even if goods were not allocated purely through markets.

A natural following question is: are compensating differentials universal? To narrow down the question, this chapter is particularly focused on a developing country context. First, markets in developing countries may not be as well established as those in developed countries

(Freeman 2010; Marx et al. 2013). It is interesting to examine how compensating differentials, if any, work in a setting with rapidly expanding markets. Second, local economies in developing countries are often more dependent on agricultural production in the form of small farms (Hazell et al. 2010). Note that this form of agriculture in developing countries, where farmers do not productively take use of capital and consume most of their outputs, is different from commercial agriculture common in developed countries. These small farms fit the definition of a traditional

“subsistence” sector in contrast to a modern sector in Lewis’ (1954) classic model of dualism, a fundamental framework which is frequently used to analyze economic development in developing countries. The dualistic divide, supported by evidence of rural-urban splits in developing countries (Gollin 2014), not only stems from the natural differences between agriculture and industry in terms of production, but also the functional differences between them in a developing economy where the traditional agricultural sector serves as a substantial source of labor force for the industrial sector. Given the unique role of agriculture in developing countries, the effect of compensating differentials in the process of making residential location

39 decisions remains to be explored when both rural and urban households are considered.

Third, internal migration often increases with economic growth in developing countries.

There has been a trend of rural-urban migration in many developing countries such as China

(Zhao 1999; Meng and Zhao 2018), India (Haberfeld et al. 1999) and Indonesia (Lu 2010), as is illustrated in the Lewis model and many other theoretical models on labor migration (i.e. Todaro

1969; Lucas, Jr 2004). Empirical evidence has shown that rural households are motivated to move to urban areas for various reasons, such as to obtain higher income and better quality of life (Williams and Jobes 1990; Du et al. 2005). On the other hand, in some cases economic incentives and amenities in rural areas may also result in reluctance of farmers to migrate (Zhao

1999; Deller et al. 2001). Under these circumstances, the interactions among housing and labor markets, agricultural production and quality of life can help provide a comprehensive picture of how rural households make residential location decisions. Compensating differentials across different markets and sectors provide crucial information for policy-makers to understand internal migration in developing countries.

The unique characteristics in developing countries lead to a series of new questions that are relevant to the quality of life framework: Do compensating differentials exist in a developing country? Do compensating differentials exist in agricultural production? Do the agricultural compensating differentials differ from those in the non-farm labor market? What contributes to quality of life in a developing country context and how does quality of life change as a country develops?

To answer these questions, I introduce an agricultural sector into the compensating

40 differential framework to include rural households and rural areas. The new model distinguishes the farm income differential from the non-farm wage differential, allowing me to test the potential heterogeneity. In the empirical section, I apply the theoretical model to Indonesia using detailed household data from the Indonesian Family Life Survey (Strauss et al. 2016) fielded from late 1993 to early 2015. Illustrative of this data, the fifth wave contains information on

14870 households located across 13 provinces in Indonesia containing 83% of the total population. My results in this chapter are based on the fourth and fifth wave of IFLS in 2007 and

2014 and focused on the main islands in Indonesia including Sumatra, Java, Bali and West Nusa

Tenggara (11 provinces). The IFLS survey collects housing prices and wages with corresponding housing and individual characteristics as well as a few community-level facilities.

Complimentary data of amenities are sourced from various Indonesian government reports. I estimate implicit prices for amenities based on hedonic equations of housing rents, non-farm wages and agricultural returns. The results indicate that compensating differentials exist across the country and in particular the impacts of amenities on agriculture and the non-farm labor market diverge. Using these implicit prices, I further calculate quality-of-life indices and rank

192 districts (regencies and cities) in the 11 provinces in 2007 and 2014. The rankings in 2014 have changed significantly compared with 2007.

This chapter makes three main contributions to the literature. First, I introduce an expanded model of quality of life with an added agriculture sector in the theoretical equilibrium system alongside housing and labor markets. The explicit agricultural sector makes it possible to include rural areas directly, which are typically omitted in most previous studies on quality of life.

41 In addition, the model provides a framework to analyze the impact of amenities such as climate on agricultural production and how those impacts may differ from other industries. Given the large agricultural sector in many developing countries, including Indonesia, this innovation provides important new insights into quality of life rankings and changes in those rankings as countries develop and experience spatially targeted government investment.

Second, I conduct empirical research on quality of life adopting the new compensating differential framework in a developing country, Indonesia. Considering the distinctive features in developing countries, especially the dualistic social structure in the process of economic growth, my empirical work in Indonesia illuminates the understanding of households’ residential location choices in a developing economy context. The empirical results provide evidence of the existence of compensating differentials in a developing country. Furthermore, the effects of local amenities on agriculture and other industries are not necessary identical, resulting in heterogeneous compensating differentials. It reveals that households’ preferences towards amenities are mixed if looking from perspectives of different sectors.

Third, I further investigate the changes of quality-of-life rankings over a long time span in Indonesia. The changes in quality-of-life rankings reveal information on the effectiveness of investment in amenities at changing quality of life. For instance, policy-makers can evaluate the impact of providing more infrastructures by examining if it raises the local ranking. In addition, the pattern of ranking changes over time also shed light on the role of government investment in local amenities across space. If the government chooses to invest in developed regions with abundant amenities considering it a more efficient way to utilize resources, there will be an

42 increasing gap between local amenity levels in different districts. Consequently, the districts that are ranked high in terms of quality of life will continue to be of high rank. In contrast, if the government makes more investment in amenities in underdeveloped regions, quality-of-life indices will tend to converge. In this way, the pattern of quality-of-life ranking changes facilitates policy interpretation.

3.2 Literature reviews

In Roback’s (1982) seminal paper, she constructed a general equilibrium model which contains land and labor market to reveal households’ preferences for amenities. In this model, a number of identical decision makers choose to live in different places with different quantities of amenities and are faced with choosing locations given zero migration cost. In equilibrium, migration ensures that each individual achieves the same level of utility so that no one has an incentive to relocate. At this moment, an individual’s marginal willingness to pay for an amenity is equal to the partial differential of land rent minus that of wage with respect to the amenity:

푉푠 푑푟 푑푤 푗 = 푙푐 − . (7) 푉푤 푑s푗 푑s푗

The left hand side is the marginal willingness to pay for amenity 푠푗, or the implicit price. On the right hand side, 푙푐 is the residential land consumed, and r and w denote the rent and wage respectively. The interpretation of this relationship is quite intuitive: local land rents will be high if there is an amenity nearby. For a disamenity, local wages are high to compensate for any negative impacts, otherwise people will leave the location with disamenities and the local wages will rise until equilibrium is achieved. To summarize, households pay higher rents and receive

43 lower wages in order to enjoy nearby amenities. This is the key insight of compensating differentials and by using the formula above one can reveal the value of an amenity through land and labor markets.

Under this framework, Roback (1982) considers four indicators of natural amenities and disamenities in the United States: heating degree days, total snowfall, cloudy days and clear days.

She finds that the implicit prices of the first three items are negative while that of the last one positive. These signs are consistent with expectations of amenities and disamenities. Based on these implicit prices, she further calculates the quality-of-life indices for the 20 largest cities in the United States and obtains their rankings, thus demonstrating this model as a powerful tool to evaluate local amenity levels and compare the quality of life among different jurisdictions.

Since then, there have been a series of papers measuring the quality of life in the United

States considering an increasing number of amenities and urban areas (Blomquist et al. 1988;

Beeson and Eberts 1989; Gyourko and Tracy 1991; Gabriel et al. 2003). Blomquist et al. (1988) extend the framework by introducing housing markets instead of land markets to make the model easier to empirically implement. They take public services and social amenities into consideration and obtain a ranking of 253 urban counties in 1980 in the United States. Gyourko and Tracy (1991) employ several different specifications of the hedonic regressions regarding housing rents and wages. Gabriel et al. (2003) introduce local non-traded goods into the framework. They calculate two sets of rankings in 1981 and 1990 so that it is possible to investigate the evolution in the quality of life among United States cites.

Gabriel and Rosenthal (2004) introduce the concept "quality of business" by deriving the

44 implicit prices of amenities from the firms’ prospective. Chen and Rosenthal (2008) compare the quality of life and quality of business simultaneously in the United States and investigate the impacts of the two indices on migration decisions at individual levels. Berger et al. (2008) demonstrate the existence of compensating differentials in a transition economy, Russia. They estimate the implicit prices of local amenities based on household data in 39 cities and futher measure the quality of life in 953 cities. Bayer et al. (2009) point out the potential downward bias if migration cost is ignored and they integrate the migration cost into the hedonic equations.

Most of these works are undertaken in urban areas in developed countries, while few papers have paid attention to the quality of life in rural areas or developing countries. This chapter begins to fill this gap.

3.3 Theoretical Models and Empirical Strategy

The goal of this chapter is to introduce rural, agricultural areas into the framework of quality of life. To accomplish this, I consider households’ farm income in addition to non-farm wages, assuming there is no barrier to labor mobility across sectors in addition to zero migration cost.

Adding an agricultural sector to the original quality of life framework presents two primary challenges: 1) how to distinguish farming from non-farm activities of which both the marginal returns with respect to labor should equal each other in equilibrium; 2) once identified, how to combine two different income differentials to measure implicit prices for amenities.

푎 I introduce a non-constant returns to scale (non-CRS) farming return function 푓(퐿 , 푠푗) as an additional component of the household budget. This function is increasing in labor that is

45 푎 assigned to farming, 퐿 , and is also affected by amenities s푗 in location j. It has several desirable features. First, this specification assumes that farmers directly obtain returns from farming, which is common in developing countries lacking large-scale mechanized agricultural industries.

Second, it seems reasonable to assume farming to be diminishing returns to scale (DRS) in terms of labor input. In contrast, all firms in the original quality of life models are assumed to be CRS and workers earn fixed wages regardless of labor supply. Third, the non-CRS assumption fundamentally distinguishes farming returns from wages so that I am able to discuss their partial derivatives with respect to amenities separately. In equilibrium, the marginal returns of farming and non-farm activities with respect to labor must be equal. This is because individuals are free to choose to work in either industry or move from one to another, and thus any gap of marginal returns from labor between different industries will be eliminated. In this model, however, the return from farming is not proportional to labor and the partial derivative of it with respect to amenities does not necessarily equal that of return from non-farm working. This provides a key source of variation which is needed for empirical identification of each differential.

The utility-maximizing problem for households who farm for themselves is given by:

max 푢(푋, 퐻, s푗) (8) 푋,퐻,s푗

푛 푎 푠. 푡. 푋 + 푟(s푗)퐻 = 푤(s푗)퐿 + 푓(퐿 , 푠푗)

퐿푎 + 퐿푛 = 퐿̅ where the household allocates labor to farming, 퐿푎, and non-farm work, 퐿푛, while the total amount of available labor is fixed. Substituting the labor constraint into the budget constraint, I rearrange the FOCs to obtain

46 ∗ ∗ 푢푠 푑푟 푑푤 휕푓 푗 ∗ ̅ 푎∗ ∗ = 퐻 − (퐿 − 퐿 ) − (9) 푢푋 푑s푗 푑s푗 휕s푗 휕푓∗ 푤(s ) = . (10) 푗 휕퐿푎 The resulting equation (9) is quite similar to the original one in Roback (1982).

Consistent with the idea of quality of life, income differentials (RHS) are recovered from the implicit prices of amenities (LHS)6 since high income compensates low amenities. For income differentials, the positive sign of the housing differential illustrates that higher rents offset amenities of higher value. In contrast, the negative sign of the non-farm wage differential suggests that higher wages compensate for the presence of disamenities. This is also true for the farming return differential, though it is distinguished from the wage differential. Note that the wage is equal to the marginal farming return with respect to labor according to equation (10), but the wage and farming return can still be affected by amenities in different patterns.

Once obtaining the three differentials of housing rents, wages and farm returns with respect to an amenity, the implicit price for this amenity can be calculated using equation (9). To obtain the differentials, I estimate the following hedonic equations:

′ ′′ ln(푅푒푛푡푖푗) = 푎0 + 푎 퐴푖푗 + 푎 퐷푖푗 + ∑ 푎푘푠푘푗 + 휀푖푗 (11) 푘

′ ′′ ln(푊푎푔푒푖푗) = 푏0 + 푏 퐵푖푗 + 푏 퐷푖푗 + ∑ 푏푘푠푘푗 + 훿푖푗 (12) 푘

′ ′′ ln(퐹푎푟푚퐼푛푐표푚푒푖푗) = 푐0 + 푐 퐶푖푗 + 푐 퐷푖푗 + ∑ 푐푘푠푘푗 + 휖푖푗 (13) 푘

6 In Roback’s (1982) original model, the implicit prices are defined using partial derivatives of indirect utility functions. When agricultural production is not considered, the original definition of implicit prices is equivalent to the LHS in equation (9), which can be demonstrated through the envelope theorem.

47 where 푠푘푗 is the kth amenity in location j. 푎푘, 푏푘 and 푐푘 are the marginal percentage changes of housing rents, wages and farm income with respect to amenity k. 퐴푖푗, 퐵푖푗 and 퐶푖푗 are household or individual characteristics that may lead to idiosyncratic effects on the three prices.

퐷푖푗 are location controls for economy disequilibrium. To recover the compensating differentials,

I multiply those marginal percentage changes by the mean of rents and incomes. Applying equation (9) to combine the three differentials, I obtain the implicit price for each amenity.

Finally I calculate the quality of life for location j by summing all local amenities weighted by their implicit prices. The weighted sum reveals the monetary value of overall amenities in a given location.

푄푗 = ∑(푎푘̅푅푒푛푡̅̅̅̅̅̅ − 푏푘푊푎푔푒̅̅̅̅̅̅̅̅ − 푐푘퐹푎푟푚퐼푛푐표푚푒̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅)푠푘푗 (14) 푘

3.4 Data

The primary household level data used in this chapter is obtained from the Indonesian Family

Life Survey (Strauss et al. 2016) fielded from late 1993 to early 2015. I use the fourth and fifth wave of IFLS starting in 2007 and 2014, respectively. In 2014, 14870 households participated in the survey in 13 Indonesian provinces that contains 83% of the total population. I consider 11 provinces on the main islands including Sumatra, Java, Bali and West Nusa Tenggara (Figure 3).

The survey contains numerous questions about the household including living conditions, employment and farming business. In the section on housing, each household head reports the annual rents of their residential dwelling or the estimated rents if they own the property. The survey also records a number of housing characteristics based on the head of household’s

48 response and the recorder’s observation, such as the size and the number of rooms. In the section of employment, respondents report their salaries or net profits from their own business as well as their occupation. Traditional labor characteristics are also reported including ages, genders, marital status, education, etc. In the section of farm business, household heads report whether or not a family member cultivates a farm land, the net profit, the size of the farm land and the type of crops or livestock. These detailed data enable me to adopt hedonic methods to reveal partial differentials in the three market sectors. All individual characteristics serve as controls to mitigate idiosyncratic effects. See Table 7 for a comprehensive list of all the individual controls used in estimation.

SUMATRA

JAVA BALI

NUSA TENGGARA

Figure 3. 11 selected provinces in Indonesia

49 Table 7. Variables and data sources Variable Description Source Housing rent equation variables Log of annual housing rent Log of Yearly housing rent (reported willingness to pay if self-owned, rupiah) IFLS4/5 Ownership 1 if the house is self-owned IFLS4/5 Single-unit&level 1 if the house is single-unit and single-level IFLS4/5 Size Size of the house (square meters) IFLS4/5 Room Number of rooms in this house IFLS4/5 Flooring 1 if the main flooring type is masonry IFLS4/5 Wall 1 if the main material used in the outer wall is masonry IFLS4/5 Electricity 1 if the house utilize electricity IFLS4/5 Pipe water 1 if the main source of water is pipe water IFLS4/5 Toilet 1 if the house has own toilet IFLS4/5 Garbage 1 if the house has trash can collected by sanitation service IFLS4/5 Ventilation 1 if the house has sufficient ventilation IFLS4/5 Yard 1 if the house has a moderately-sized yard IFLS4/5 Filth 1 if the house is surrounded by filth IFLS4/5 Wage equation variables Log of annual salary Log of salary and bonuses during the last year (rupiah) IFLS4/5 Log of annual hours Log of the total number of hours one works per year IFLS4/5 Years of schooling Highest year of school attended IFLS4/5 Experience Current age minus years of schooling minus the age starting school IFLS4/5 Experience squared Experience squared IFLS4/5 Ethnicity: Java 1 if ethinicity is Java IFLS4/5 Language: Indonesian 1 if language in daily life is Indonesian IFLS4/5 Language: Javanese 1 if language in daily life is Javanese IFLS4/5 Male 1 if one is male IFLS4/5 Married 1 if one is married IFLS4/5 Casual 1 if one is a casual worker IFLS4/5 Self-employed 1 if one is self-employed IFLS4/5 Occupation dummy variables 12 occupation dummies include academic professionals (omitted); non-academic professionals; IFLS4/5 officials and managers; clerks; sales workers; service workers; agricultural workers; production and related workers; craft workers; operators and assemblers; military specialists; students; others 50 Farm income equation variables Log of annual household farm income Log of the net profit generated by the farm business during last year (rupiah) IFLS4/5 Head: male 1 if the household head is male IFLS4/5 Head: age Age of the household head IFLS4/5 Head: married 1 if the houdsehold head is married IFLS4/5 Number of children Number of children (under 7) in the household IFLS4/5 Number of low-educated young males Number of young males (7-35) with schooling less than 6 years in the household IFLS4/5 Number of low-educated young females Number of young females (7-35) with schooling less than 6 years in the household IFLS4/5 Number of low-educated old males Number of old males (35 or older) with schooling less than 6 years in the household IFLS4/5 Number of low-educated old females Number of old females (35 or older) with schooling less than 6 years in the household IFLS4/5 Number of junior high school graduates Number of household members with schooling more than 6 years but no more than 9 years IFLS4/5 Number of senior high school graduates Number of household members with schooling more than 9 years but no more than 12 years IFLS4/5 Number of university graduates Number of household members with schooling more than 12 years IFLS4/5 Size of farm land Size of the farm land cultivated by the household (square meters) IFLS4/5 Crop dummy variables 8 crop dummies include tuber; nuts and beans; crops; vegetables; fruits; spice; rubber and wood; IFLS4/5 livestock; others (omitted) District-level Variables Temperature Annual average temperature, 2007/2014(°C) NOAA Precipitation Annual precipitation, 2007/2014(1,000 mm) CHIRPS PM2.5 Average PM2.5, 2007/2014(µg/m3) NASA Eruption Number of Volcanic eruptions, 2001-2007/2008-2014 DesInventar Flood Number of Floods, 2001-2007/2008-2014 DesInventar Forest Area of forest/district size, 2007/2014 (%) Hansen et al. (2013) Urbanization Area of urban land/district size, 2007/2014 (%) Liu et al. (2018) Population Total population (in millions), 2007/2013 BPS Primary schools Number of schools at primary level per 1,000 population, 2005/2011 PODES Doctors Number of doctors per 1,000 population, 2005/2011 PODES Hospitals Number of hospitals per 1,000 population, 2005/2011 PODES Immunization Immunization coverage for children under 5 years old (%), 2007/2014 SUSENAS Morbidity Morbidity rate (%), 2007/2013 SUSENAS Unemployment rate Number of people unemployed/people in labor force, 2007/2013 SAKERNAS

51 Apart from individual information, I obtain data on a number of natural and man-made amenities measured at the district level (the second level of Indonesian administrative divisions, regencies and cities). These variables are assumed to have impacts on quality of life and households’ residential location decisions, and are obtained from a variety of sources. Annual average temperature in 2014 is sourced from National Oceanic and Atmospheric Administration

(NOAA, Fan and Van den Dool 2008). Annual precipitation in 2014 is sourced from Climate

Hazards Group InfraRed Precipitation with Station data (CHIRPS, Funk et al. 2015). Average

PM2.5 in 2014 is sourced from National Aeronautics and Space Administration (NASA, Van

Donkelaar et al. 2015). I calculate the numbers of volcanic eruptions and floods from 2008-2014 based on the records in the United Nations disaster information system, DesInventar. I calculate the forest areain 2014 according to the database of Hansen et al. (2013) and I calculate the urbanization rate in 2014 according to the database of Liu et al. (2018). Total population in 2013 is sourced from Statistics Indonesia, known in Indonesia as BPS (Central Agency on Statistics).

Numbers of primary schools, doctors and hospitals per 1000 population are calculated from the

Village Potential Statistics (PODES) survey conducted by BPS in 2011. Percentage of

Immunization coverage for children under 5 years old (2013) and immunization rate (2014) are obtained from National Socioeconomic Survey (SUSENAS) conducted by BPS. In addition, I utilize unemployment rate from BPS Labor Survey (SAKERNAS) to control economy disequilibrium. All the district-level variables are also presented in Table 7.

52 3.5 Results

Table 8 shows the results of the housing hedonic regression using the 2014 sample of 11,699 households. All the standard individual variables of housing characteristics are significant and the signs are consistent with the literature, indicating that the household survey data is of high quality. The key results of interest are the housing rent differentials with respect to amenities in terms of percentage change. Based on the coefficients, housing rents tend to be low when the weather is very hot and there is excessive rain. Frequent volcanic eruptions are also associated with low housing prices. In contrast, high urbanization rates make positive contributions to the housing differential. Most of the signs are consistent with expectations, with only a few exceptions.7

Table 9 presents results for the non-farm labor market using the employment information from 10,610 workers in 2014. For instance, firms that are located in districts with severe air pollution are likely to offer higher wages to compensate the negative impact on life. As for districts with amenities such as primary schools, the local wages are likely to be low.

Using the data on 2,834 rural households’ information of family farm businesses in 2014, the differentials of farm income with respect to amenities are shown in Table 10 and are noticeably different from those of non-farm wages in Table 9. Specifically, excessive rainfall is detrimental to farming returns, which has no significant impact on non-farm wages. However, farmers do not gain a strong compensation effect for air pollution associated with farm income as

7 Interpretation of these differentials in isolation ignores the important role of labor and agricultural markets on the overall differentiation as the housing market is not the only channel to indicate compensation and the combination of all differentials is needed to obtain a comprehensive picture of the role of amenities on quality of life.

53 Table 8. Housing rent equation with amenities, 2014

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual housing rent 14.974 1.111 8.189 21.822 (Annual housing rent) 6,822,546 38,396,073 3,600 3,000,000,000 Housing characteristics Ownership 0.101*** (0.0241) 0.69 0.462 0 1 Single-unit&level -0.0900*** (0.0289) 0.739 0.439 0 1 Size 0.000447** (0.000204) 81.539 118.917 1 7,000 Room 0.0697*** (0.00605) 5.66 2.573 1 40 Flooring 0.185*** (0.0317) 0.687 0.464 0 1 Wall 0.229*** (0.0392) 0.848 0.359 0 1 Electricity 0.311*** (0.112) 0.993 0.081 0 1 Pipe water 0.0744** (0.0314) 0.282 0.45 0 1 Toilet 0.225*** (0.0367) 0.823 0.382 0 1 Garbage 0.351*** (0.0346) 0.39 0.488 0 1 Ventilation 0.0680** (0.0297) 0.858 0.35 0 1 Yard 0.0692*** (0.0227) 0.644 0.479 0 1 Filth -0.0458 (0.0287) 0.146 0.353 0 1 Amenities/Disamenities Temperature -0.111*** (0.0206) 26.859 1.783 20.931 30.058 Precipitation -0.103** (0.0509) 2.412 0.71 1.077 4.844 PM2.5 0.00728 (0.00866) 13.992 6.335 3.269 28.120 Eruption -0.113* (0.0587) 0.147 0.447 0 3 Flood 0.00108 (0.00122) 15.407 18.114 0 163 Forest -3.07e-05 (0.00218) 40.103 24.684 0.444 89.677 Urbanization 0.00491*** (0.00183) 23.752 35.632 0 100 Population 0.0117 (0.0490) 1.182 0.883 0.0436 5.202 Primary schools -0.755*** (0.200) 0.621 0.217 0.231 1.650 Doctors 0.405 (0.301) 0.22 0.181 0.0317 1.335 Hospitals -4.392 (4.307) 0.01 0.01 0 0.0592 Immunization 0.00657 (0.00860) 97.159 3.518 72.283 100 Morbidity -0.00244 (0.00331) 29.449 8.664 7.99 57.256 Disequilibrium Variable Unemployment rate -1.116 (1.449) 0.063 0.03 0.003 0.149

Observations 11,699 R-squared 0.303 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

54 Table 9. Wage equation with amenities, 2014

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual salary 16.029 1.507 8.700 21.416 (Annual salary) 21,337,375 39,258,114 6,000 2,000,000,000 Human capital characteristics Log of annual hours 0.640*** (0.0168) 7.224 1.064 0 9.336 Years of schooling 0.0990*** (0.00415) 9.868 4.362 0 21 Experience 0.0395*** (0.00399) 19.731 13.528 0 84 Experience squared -0.0504*** (0.00681) 5.723 7.238 0 70.56 Ethnicity: Java 0.0143 (0.0402) 0.472 0.499 0 1 Language: Indonesian 0.180*** (0.0344) 0.351 0.477 0 1 Language: Javanese -0.0735 (0.0486) 0.445 0.497 0 1 Male 0.362*** (0.0257) 0.634 0.482 0 1 Married 0.130*** (0.0321) 0.754 0.43 0 1 Casual -0.450*** (0.0377) 0.212 0.408 0 1 Self-employed -1.477*** (0.0637) Occupations 0.055 0.228 0 1 Non-academic professionals -0.229*** (0.0724) 0.025 0.157 0 1 Officials and managers 0.0871 (0.149) 0.089 0.285 0 1 Clerks -0.0589 (0.0680) 0.005 0.07 0 1 Sales workers -0.259*** (0.0703) 0.093 0.29 0 1 Service workers -0.390*** (0.0690) 0.102 0.303 0 1 Agricultural workers -0.396*** (0.0780) 0.169 0.375 0 1 Production and related -0.280*** (0.0680) 0.162 0.368 0 1 workersCraft workers -0.0724 (0.0748) 0.092 0.289 0 1 Operators and assemblers -0.255*** (0.0685) 0.045 0.208 0 1 Military specialists 0.404*** (0.0891) 0.211 0.408 0 1 Students -0.554 (0.695) 0.005 0.068 0 1 Others -0.141 (0.130) 0.0002 0.014 0 1 Amenities/Disamenities Temperature -0.0748*** (0.0212) 26.968 1.735 20.931 30.058 Precipitation -0.0556 (0.0551) 2.397 0.703 1.077 4.844 PM2.5 0.0270*** (0.00714) 14.255 6.296 3.269 28.120 Eruption -0.0985** (0.0380) 0.125 0.406 0 3 Flood 0.000488 (0.000872) 15.66 19.072 0 163 Forest -0.000274 (0.00186) 38.849 25.015 0.444 89.677 Urbanization -0.00147 (0.00123) 26.112 36.81 0 100 Population 0.0628** (0.0291) 1.199 0.874 0.0436 5.202 Primary schools -0.373** (0.148) 0.602 0.214 0.231 1.650 Doctors 0.380** (0.164) 0.23 0.184 0.0317 1.335 Hospitals -0.838 (3.407) 0.011 0.01 0 0.0592 Immunization 0.00349 (0.00614) 97.255 3.39 72.283 100 Morbidity -0.00486* (0.00251) 29.589 8.515 7.990 57.256

55 Disequilibrium Variable Unemployment rate -3.410*** (1.010) 0.063 0.031 0.0077 0.148808

Observations 10,610 R-squared 0.546 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

compared to off-farm labor. Urbanization has a negative impact on farming, but farmers would rather be close to such an amenity at the expense of lower farm income. Similar compensation occurs for forests and doctors.

Given estimates for the three differentials, I calculate the overall implicit price in 2014 for each amenity by combining them according to equation (9). Note that the calculation of marginal willingness to pay is at the household level. In my sample, the average number of workers per household is 2.66. Following Berger et al. (2008), I multiply the wage differential by this number to obtain the marginal change in household non-farm income. In addition, 31.95% of households in the sample cultivate farm lands. Since every decision maker is assumed identical in the model, I assume that 31.95% of the income of a typical household is from farming and the rest is from non-farm work. Therefore, I multiply the wage and farm income differentials by

68.05% and 31.95% respectively in the quality of life formula. As a whole, the implicit price is equal to the coefficient in Table 8 times the mean of housing rents, minus 68.05% times 2.66 times the coefficient in Table 9 times the mean of wages, minus 31.95% times the coefficient in

Table 10 times the mean of farm income.

Results for the implicit prices are shown in Column (4) in Table 11. These are the monetary measures for each amenity or disamenity in Indonesian Rupiahs (1,000,000 Indonesian

56 Table 10. Farm income equation with amenities, 2014

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual household farm 15.134 1.36 7.131 18.224 income(Annual household farm income) 7,818,082 10,102,699 1250 82,130,000 Farm household characteristics Head: male 0.387*** (0.108) 0.903 0.297 0 1 Head: age -0.000810 (0.00240) 48.731 14.166 19 93 Head: married 0.118 (0.0865) 0.889 0.314 0 1 Number of children -0.0370 (0.0421) 0.548 0.691 0 4 Number of low-educated young -0.0212 (0.0374) 0.363 0.604 0 4 malesNumber of low-educated young 0.0245 (0.0428) 0.38 0.629 0 4 femalesNumber of low-educated old males 0.0191 (0.0599) 0.539 0.561 0 3 Number of low-educated old females 0.0425 (0.0558) 0.63 0.599 0 3 Number of junior high school 0.0224 (0.0348) 0.664 0.811 0 5 graduatesNumber of senior high school 0.121*** (0.0334) 0.658 0.867 0 5 graduatesNumber of university graduates 0.144*** (0.0456) 0.224 0.58 0 5 Size of farm land 2.11e-05*** (5.23e-06) 7,346.275 15,058.04 1 400,000 Crop types Tuber -0.560* (0.288) 0.045 0.208 0 1 Nuts and beans -0.574* (0.322) 0.047 0.212 0 1 Crops -0.132 (0.150) 0.023 0.15 0 1 Vegetables 0.162 (0.205) 0.567 0.496 0 1 Fruits -0.348 (0.232) 0.04 0.196 0 1 Spice -0.0358 (0.155) 0.047 0.212 0 1 Rubber and wood 0.192 (0.150) 0.083 0.276 0 1 Livestock 0.163 (0.160) 0.082 0.275 0 1 Amenities/Disamenities Temperature -0.0982*** (0.0279) 26.645 1.978 22.026 30.055 Precipitation -0.164* (0.0948) 2.347 0.634 1.077 4.844 PM2.5 -0.0107 (0.0148) 11.12 4.891 3.269 28.120 Eruption 0.0740 (0.0738) 0.215 0.544 0 3 Flood 0.00373 (0.00577) 13.207 12.811 0 163 Forest -0.0119*** (0.00293) 49.976 21.383 0.444 88.408 Urbanization -0.0101*** (0.00350) 3.712 10.496 0 100 Population 0.103 (0.0721) 0.906 0.655 0.0495 5.202 Primary schools -0.218 (0.470) 0.763 0.183 0.252 1.352 Doctors -0.960* (0.566) 0.137 0.094 0.0329 0.773 Hospitals 20.94** (9.135) 0.007 0.006 0 0.059 Immunization 0.0138 (0.0117) 96.781 3.973 72.283 100 Morbidity -0.00405 (0.00504) 30.087 9.484 7.990 53.673 Disequilibrium Variable Unemployment rate -4.431** (2.187) 0.048 0.024 0.003 0.149

57 Observations 2,834 R-squared 0.164 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

Rupiahs are equal to 72.80 US dollars). Positive prices indicate amenities and negative prices disamenities. For example, air pollution and population are viewed as disamenities while temperature and urbanization are viewed as amenities. Implicit prices indicate the total differential of the effect of an amenity, and therefore significant results provide evidence of the existence of compensating differentials as a whole (Berger et al. 2008). The majority of the signs are consistent with my expectation, though some values imply households’ preferences in an unexpected direction.

Ranking

SUMATRA

Jakarta

BALI JAVA

NUSA TENGGARA

Figure 4. Quality of life rankings for 192 districts in Indonesia, 2014

58 Table 11. Implicit prices of amenities, 2014

(1) (2) (3) (4) Amenities Housing rent differential Wage differential Farm income differential Full implicit price Temperature -0.111*** -0.0748*** -0.0982*** 2375708*** (0.0206) (0.0212) (0.0279) (832334) Precipitation -0.103** -0.0556 -0.164* 1851128 (0.0509) (0.0551) (0.0948) (2170002) PM2.5 0.00728 0.0270*** -0.0107 -967017*** (0.00866) (0.00714) (0.0148) (284353) Eruption -0.113* -0.0985** 0.0740 2845533* (0.0587) (0.0380) (0.0738) (1531910) Flood 0.00108 0.000488 0.00373 -20812 (0.00122) (0.000872) (0.00577) (37575) Forest -3.07e-05 -0.000274 -0.0119*** 39991 (0.00218) (0.00186) (0.00293) (73562) Urbanization 0.00491*** -0.00147 -0.0101*** 115392** (0.00183) (0.00123) (0.00350) (49766) Population 0.0117 0.0628** 0.103 -2603638** (0.0490) (0.0291) (0.0721) (1187013) Primary schools -0.755*** -0.373** -0.218 9780518 (0.200) (0.148) (0.470) (5990732) Doctors 0.405 0.380** -0.960* -9501315 (0.301) (0.164) (0.566) (6824783) Hospitals -4.392 -0.838 20.94** -49901538 (4.307) (3.407) (9.135) (136738207) Immunization rate 0.00657 0.00349 0.0138 -124317 (0.00860) (0.00614) (0.0117) (246194) Morbidity rate -0.00244 -0.00486* -0.00405 181353* (0.00331) (0.00251) (0.00504) (100205)

Observations 11,699 10,610 2,834

R-squared 0.303 0.546 0.164

Housing rent, wage and farm income differentials are taken from Tables 8, 9 and 10 . The annual full implicit prices per household are evaluated at the means of housing rent, wage and farm income of the sample. 31.95% of households in the sample work for a farm business. For the remaining 68.05% of households, the average number of workers per household is 2.66. 1,000,000 Indonesian Rupiahs are equal to 72.80 US dollars. Robust standard errors in parentheses (clustered by districts). The standard errors on the full implicit prices are calculated from a linear combination of the standard errors in the housing rent, wage and farm income hedonic equations, assuming there is no covariance among the differentials. *** p<0.01, ** p<0.05, * p<0.1.

59 Finally I sum all of the amenities in each district weighted by the full implicit prices. As a result, I obtain the quality-of-life index for 192 Indonesian districts in 2014. It is noteworthy that although the index is still a monetary measure, it is not appropriate to directly interpret the magnitude of quality of life which depends on the unit and baseline of each type of amenity, and in the literature the index values are usually normalized to make them non-negative (See Berger et al. 2008). However, the rankings of quality of life, which is the focus in this chapter, provide relevant information to compare the level of overall amenities among different districts. The rankings in 2014 are shown in Figure 4. This figure shows that districts in the east of Java have a relatively higher level of quality of life. In contrast, the west of Java, especially areas around the capital Jakarta, suffers from a worse living environment.

Following the same procedure, I obtain the quality-of-life rankings for each Indonesian district in 2007 (See Appendix Tables 20-23 and Figure 7). The comparison of the rankings between the two periods is presented in Appendix Table 24. From 2007 to 2014, the rankings of quality of life have changed significantly. As the country develops, the level of local amenities in each district is not static and is very likely to be affected by government investment. As is shown in the maps, there is an obvious shift of high ranking districts from the west of Java to the east.

This finding suggests that the districts of lower quality of life in 2007 received additional investment and obtained relatively higher amenity levels over time, leading to higher rankings in

2014. This finding suggests that as Indonesia has experienced rapid development, the gap of quality of life across the country has narrowed. It would be interesting to further investigate why the gap has narrowed if we had data on government investment and analyzed its effect on quality

60 of life.

3.6 Discussion

In this chapter, I introduce an agricultural sector into the theoretical model of compensating differentials to calculate quality of life rankings. By adding a farming return component to the household budget, I manage to distinguish the farm income differential from the non-farm wage differential which provides a theoretical framework to test the effect of local amenities on different industries. The extended equilibrium model includes agricultural production, rural households and rural areas, which provides new insights into quality of life especially in most of developing countries.

Using household survey data in Indonesia, I apply the model to a developing economy with a large portion of small farm businesses. I find that compensating differentials exist across

Indonesia. In particular, compensating differentials exist in the agricultural sector. The farm income compensating differentials differ from non-farm wage differentials, suggesting that local amenities have heterogeneous impacts on agricultural production and non-farm labor market.

The quality-of-life rankings have also changed as the country develops. My findings indicate that the rapid development occurring in Indonesia has served to reduce the gap in quality of life across jurisdictions.

In addition to the overall quality of life rankings this chapter is focused on, the current work also provides a flexible framework for policy-makers to further analyze quality of life in specific aspects. Once the implicit prices are obtained from compensating differentials, one can

61 select relevant amenities to construct indices that reflect local environmental conditions on particular topics of interest. For example, based on estimated implicit prices for amenities that are related to natural resources and disasters, it is possible to construct multiple types of vulnerability indices to focus on topics such as climate change. The themed indices may help policymakers assess the effects of specific policies aimed at improving environmental conditions and quality of life, as many related policies have been implemented in Indonesia such as forest conservation and infrastructure investment (Koh and Ghazoul 2010; Gibson and Olivia 2010).

These indices could also be potentially used for predictions along with environmental evaluation.

62 Chapter 4: The Impact of Deforestation on Ecotourism: a Birdwatching Example in Mexico

4.1 Introduction

With the expansion of ecotourism across developing countries, significant research has focused on ecotourism as a way to conserve forests and contribute to economic development. The presence of forests offering aesthetic and recreational opportunities attracts tourists and revenue, yielding potentially considerable benefits to governments, firms, and local communities. Because ecotourism in many developing contexts depends on forests, ecotourism is often hailed as a way to promote sustainable development in an area (Ceballos-Lascurain 1996; Fennell and Weaver

2005; Cobbinah 2015). Given the increasingly important role of ecotourism in many economies,

(e.g., Naidoo and Adamowicz 2005a, 2005b; Hutchins 2007; Bhuiyan et al. 2011; Ramos and

Prideaux 2014; Hunt et al. 2015; Eshun and Tagoe-Darko 2015; Liang 2016; Snyman 2017), research is needed to better understand what characteristics of forest ecosystems affects individuals’ decisions on where and how much to recreate. Understanding the drivers of recreation and tourism choices provides important insights into the role of sustainable development in seeking to balance economic growth and the preservation of key ecological features making locations attractive for ecotourism.

A large number of studies in the natural sciences posit that significant deforestation and forest degradation result in low species richness and abundance (e.g., Fahrig 2003, 2017;

Laurance et al. 2014; Ochoa-Quintero et al. 2015; Giam 2017). On the other hand, the handful of

63 studies focusing on the impact of biodiversity (broadly defined) and ecotourism suggest tourists highly value biodiversity, which has a positive and significant impact on the economic profits derived from ecotourism (Naidoo and Adamowicz 2005b; Naidoo et al. 2011; Booth et al. 2011;

Kolstoe and Cameron 2017). This could be due to the presence of more species (e.g., Kolstoe and Cameron 2017) or more rare species (e.g., Booth et al. 2011) that increase the number of visitors as well as the distance tourists are willing to travel. While there is an abundance of theoretical and empirical work on the impact of deforestation on species richness and abundance, very little is known about how large-scale deforestation and forest degradation affect ecotourism behavior, particularly relevant in developing countries.

I focus on the impact of deforestation on birdwatching ecotourism in Mexico.

Birdwatching is becoming the fastest growing segment in the ecotourism market and generates significant tourist revenue as birders are generally well educated with above-average incomes

(Cordell and Herbert 2002; Şekercioğlu 2002; Lee et al. 2010; Steven et al. 2015; Kronenberg

2016; Ocampo-Penuela and Winton 2017). Birdwatching has long been a pastime in developed countries, such as the United States, where observing wild birds became popular in the late 19th century (Weidensaul, 2008). The United States Fish and Wildlife Service reported that in 2011

American birders’ total expenditure for birding was $40.94 billion, with $14.87 billion spent on trips (Carver 2013). According to the 2016 National Survey of Fishing, Hunting and Wildlife

Associated Recreation, 45.1 million Americans aged over 16 reported themselves as wild bird observers in 2016, while 36% of them (16.3 million) took birding trips away from home for an

64 average of 16 days a year.8

In developing countries, ecotourism is often adopted as a tool to simultaneously achieve goals of economic development and natural conservation. Birdwatching in particular has increasingly become popular as it is considered more sustainable than other forms of ecotourism

(Biggs et al. 2011; Li et al. 2013), i.e. birders are in general more concerned about natural conservation (Hvenegaard and Dearden 1998; Connell 2009). There is extensive literature focusing on birdwatching ecotourism in developing countries such as South Africa (Conradie and van Zyl 2013), Peru (Puhakka et al. 2011), Colombia (Ocampo-Penuela and Winton 2017),

Mexico (Cantú et al. 2011; Revollo-Fernández 2015), China (Li et al. 2013; Ma et al. 2013) and

Thailand (Hvenegaard and Dearden 1998).

Many birders, known as “listers”, keep track of their trips and the species they observe using online tools such as eBird, a checklist program launched by a large citizen-science project at the Cornell University Laboratory of Ornithology (Fink et al. 2017). This project is similar to many other volunteer environmental monitoring programs where citizens contribute to scientific research (Evans et al. 2005; Ricker et al. 2013; Luz et al. 2014; Stepenuck and Green 2015).

With the growing popularity of birding, an increasing number of listers are willing to travel long distances for new species that cannot be observed elsewhere. Estimates suggest the annual number of international trips worldwide with a main purpose of birdwatching is 3 million.9

According to American Birding Association’s survey, 49% of its members would like to travel

8 U.S. Department of the Interior, U.S. Fish and Wildlife Service, and U.S. Department of Commerce, U.S. Census Bureau. 2016 National Survey of Fishing, Hunting, and Wildlife-Associated Recreation. 9 Caribbean Tourism Organizations: Acorn Consulting Partnership, 2008. Bird Watching

65 outside the US to observe birds.10

Mexico is the country most visited by the US ecotourists.11 Largely influenced by the development of birdwatching activities in the US, observing birds as a pastime started to appear in Mexico from at least the beginning of the 20th century (Gómez de Silva and Alvarado Reyes

2010). Considered one of the mega-diverse countries in the world (Llorente-Bousquets and

Ocegueda 2008), Mexico is endowed with large areas of forests (Hendee et al. 2012). According to BirdLife International (del Hoyo and Collar 2014, 2016), Mexico is home to 10% of all bird species on the planet. Among these species, 10% are only observed in Mexico, making the country fifth in number of endemic bird species in the world. It is not surprising that Mexico is rapidly becoming one of the most popular destinations for birdwatching ecotourism in the world

(Revollo-Fernández 2015; Peterson and Navarro-Sigüenza 2016). It was estimated that in 2006

Mexico hosted more than 78 thousand birders resulting in expenditure of at least US $23.9 million (Cantú et al. 2011). In the eBird database, checklists that reported a birdwatching trip in

Mexico increased tenfold between 2008 and 2016, from 3,360 entries to 33,708, implying enormous growth in the number of birders visiting Mexico. However, Mexico is also experiencing widespread deforestation: the country lost 5.8% of its forests between 2000 and

2016 with a total loss area over 30,000 km2, which is equivalent to 1.52% of the total land area of the country (Hansen et al 2013; calculations by authors). The combination of increasing ecotourism and at-risk natural resources makes Mexico an ideal context to investigate the impact

10 American Birding Association, 1994. ABA Membership Survey 11 World Tourism Organization, 2002. The U.S. Ecotourism Market

66 of deforestation on birdwatching ecotourism.

To estimate the effects of deforestation on birdwatching visits in Mexico, I combine the citizen science data (the eBird database) with high-resolution satellite data of forest coverage

(Hansen et al. 2013). Using the eBird data, I construct annual visitation counts for 1810 Mexican municipalities 12 from 2008 to 2016 to examine how changes in forest conditions affect birdwatching visitation patterns. To empirically estimate this impact, I first consider binary probability models to examine the extensive margin of the role of deforestation in altering visitation patterns at a specific location. I then introduce count data models to investigate the effect of deforestation on the intensive margin of number of visits. My results indicate that deforestation not only decreases the probability of a municipality being visited, but also reduces the number of tourist visits. Specifically, my results indicate that when a municipality loses one percentage point of forested land, the number of birdwatching trips decreases by approximately

30%, ceteris paribus.

This chapter makes three primary contributions to the literature. First, I provide new insights into forest conservation and ecotourism in a developing country context. Although there have been many studies that evaluate the effects of ecotourism on forests, few focus on the contribution of state of the forest to ecotourism behavior. In contrast to several related papers that focus on a single birdwatching hotspot (e.g., Naidoo and Adamowicz 2005a) or sites in a small-scale region (e.g., Kolstoe and Cameron 2017), I conduct a comprehensive empirical investigation on deforestation and birdwatching tourism at the municipality level across Mexico

12 Municipalities are the second-level administrative divisions below states.

67 over a 9-year period.

Second, I demonstrate the potential role of citizen science data to conduct empirical research in settings where data availability from traditional sources is often lacking or prohibitively expensive to obtain across large areas. Recent research has suggested that social media data can be used as an alternative to traditional surveys to understand individual preferences for ecotourism, as there may be no significant differences between tourists’ stated preferences in surveys and revealed preferences from social media content (Hausmann et al.

2018). In a developing country context, there is a significant potential for citizen science data to play an important role in empirical research as a readily available, and relatively inexpensive, substitute for official data sources and on-the-ground data collection. Currently, only a handful studies, most of which are in developed country settings, have explored the potential of using citizen science data to inform about behaviors (e.g., Kolstoe & Carmeromn 2017; Arcanjo et al.

2016; Keeler et al. 2015; Hale 2018; Spalding et al. 2017; Donahue et al. 2018). My work shows that the eBird data contributed by citizen scientists performs well in Mexico yielding empirical results that are highly consistent with my prior expectations.

Finally, I provide new information on the potential economic impacts of deforestation on ecotourism across Mexico by adding new information on the potential scale of ecotourism revenue loss. My empirical results suggest that deforestation discourages birders from visiting a municipality in Mexico, reducing both the probability and the number of visits. The loss in ecotourism revenue is likely to affect not only local economies but also government tax income, further reducing the budget for conservation and leading to a potential negative feedback effect.

68 Thus, my results have important implications for the design of development and forest conservation polices in ecotourism-dependent communities seeking to enhance sustainable development.

4.2 Data

Curated by Cornell University in the US, the eBird dataset (Fink et al. 2017) is contributed by global birder volunteers who use an online checklist protocol to report their birding sightings.

Although the dataset was launched in 2002, the early eBird data was thin until 2008 when the number of participants and observations started to surge, coinciding with the mass market release of smart phones, including the iPhone in 2007, that enable birders to more easily report their sightings while in the field (Sullivan et al. 2009; Kolstoe and Cameron 2017). While eBird is a real-time online dataset, it also allows input of historic trips that occurred before a participant began using the reporting tool. To aid in identification of locations, birders are able to pick from a number of pre-defined “hotspots” which are frequent destinations and chosen by local “experts.”

In addition, birders are also able to input their own location data. To minimize inaccurate recall, I focus my use of the eBird records to birding visits occurring after 2008 as those are more likely to represent contemporaneous reports associated with visitation.

Each checklist submitted to eBird includes detailed information for a birding trip, such as location, time, duration of observation period, distance traveled as well as individual and species bird counts. Birders are asked to report the latitude and longitude of the observation location

(starting position if travelling), which I use to link ecotourism destinations to specific

69 municipalities in Mexico. I only consider coordinates on land and exclude sea tours, because I cannot unambiguously link the observations to municipalities.

Because the dates of the birding trips were also reported in these checklists, the eBird database allows me to separate observations occurring on different days. The duration of the actual birdwatching period as recorded in the checklists, which is usually several hours, never exceeds 24 hours. In other words, eBird users would submit a separate checklist each day with a corresponding location if they spent more than a single day partaking in bird watching activity.

As a result, a multiple-day trip would result in multiple checklists, enabling me to count each checklist as a day trip. If a birder reported several trips in different locations that happened within a day, I only consider the trip with the longest duration, the longest travel distance, or the largest number of birds observed, in that order. My resulting dataset for Mexico contains 64,572 day trips derived from checklists between 2008 and 2016 (Figure 5). Most trips occurred in the southern part of the country, as much of the north is arid and supports fewer bird species. I calculated the annual count of all day trips in each Mexican municipality and use this count as my primary measure of birdwatching ecotourism in the econometric analysis.

In addition to the eBird database, I also assemble data on a number of additional features that are likely to influence birder behavior. I use the Hansen datasets on percent tree cover and annual deforestation (Hansen et al. 2013) to recover information on forests. To convert the original percent tree cover in 2000 to percent forest cover in 2000, I use a cutoff of 25% to define forests (Sexton et al. 2015) and retain only deforestation events that occurred on forested cells in the base year of 2000. Using this data, I calculate annual forest loss area in each Mexican

70

Figure 5. Aggregate eBird Reports of Day Trips to Mexico, 2008-2016

municipality from 2008 to 2016 (Figure 6).13 As my main variable of interest, I define the annual deforestation rate as the ratio of deforested area to municipality size. Municipality areas are calculated based on the GADM database of global administrative areas (GADM 2009).

I also include a series of additional covariates that are likely to affect ecotourism. I divide these into groups including (1) forest and biomass attributes, (2) climate and weather, and (3) economic and infrastructure. Because the forest configuration is also likely to affect biodiversity,

I include covariates related to core and edge forests (Fahrig 2003; Fahrig et al 2017). Specifically,

I construct a ratio of forest edge to core forest using a reclassified version of the Hansen data 90

13 I do not consider reforestation due to the relatively low amount occurring during my study period and challenges in measuring reforestation from the Hansen data.

71 meters to define an edge; I calculate the forest edge and core areas using the Landscape

Fragmentation Tool (LFT) (Vogt et al. 2007). This variable is a measure of patch-level forest fragmentation and disturbance, which is known to affect patch-level biodiversity (Graham and

Blake, 2001), with more fragmented forests having higher values of the ratio. For my application,

I expect that more fragmented forests will have fewer core species and may have more introduced or invasive species that may not be of value to birders (Krebs 1972; Fahrig 2003).

However, forest edges also often provide easier viewing of species located in canopies that are difficult to observe when in the interior sections of forest tracts, making the overall effect an empirical question.

United States

Forest cover 2000 Deforestation 2008-2016 Belize

Guatemala 0 500 km Honduras

Figure 6. Forest cover (2000) and deforestation (2008-2016) based on Hansen et al. (2013)

72 In addition to forest edge ratio, I also calculate above ground woody biomass (average density per municipality) based on Baccini et al.’s (2012). This variable indicates the presence of old growth (i.e. high carbon value) forests vs younger less mature forests. I expect that older forests are likely to contain more rare species (Jenkins et al, 2013) making them more likely to attract birders, ceteris paribus.

To account for the impact of climate, I include measures of both precipitation and temperature. These are likely to play an important role in tourism behavior by affecting the demand for recreation as well as the presence and likelihood of observing certain bird species.

The temperature data are sourced from the Climate Prediction Center monthly global surface air temperature data and from National Oceanic and Atmospheric Administration (Fan and Van den

Dool 2008). I focus on the temperature in January and June respectively to distinguish seasonal effects. The precipitation data are obtained from Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS, Funk et al. 2015).

I construct control for economic and infrastructure factors from a variety of sources. At the municipality level, I control for GDP (in 2005, measured in US dollars) (INAFED, 2005), and population (in 2005) (INEGI, 2005). I also include the proximity to large urban areas as those are more likely to be influenced by urbanization, potentially discouraging birders from visiting. Specifically, I control for the Euclidean distance to the nearest major urban center which is defined as the geometric center of an urban area with population greater than 500,000 in 2000.

73 Using this population threshold14, I calculate distance to the nearest of one of 17 centers based on the map of urban areas from INEGI (INEGI 2000). For transportation infrastructure, I measure distance to the nearest airport using the Mexican airport map in 1998 from the North American

Transportation Atlas Data (Bureau of Transportation Statistics, 1998). I also measure the length of highways (in 2014) and railways (in 2000) (INEGI 2014, 2000). Finally, I identify whether there is an archaeological site or national protected area within a municipality based on the official maps from INEGI (INEGI 2000) and Mexico’s National Commission of Natural

Protected Areas (CONANP, 2018). I only consider national protected areas that were created before 2000 to exclude any endogenous factors caused by later deforestation and ecotourism.

Table 12. Descriptive Statistics Variables Mean SD Min Max Annual count of day trips (#) 3.84 25.43 0 848 Annual deforestation rate (%) 0.13 0.32 0 6.52 Municipality size (1,000 km2) 1.29 3.16 0.011 69.71 Forest area in 2000 (1,000 km2) 0.38 1.17 0 24.43 Edge to core ratio 3.64 9.24 0 167.50 Biomass density (Mg/ha) 46.62 41.71 0.039 185.49 Threatened bird species (#) 2.48 1.47 0 7.31 Temperature in January (ºF) 61.55 9.56 37 91 Temperature in July (ºF) 72.81 10.33 53 109 Precipitation (1,000 mm) 1.06 0.62 0.054 5.27 GDP (billion US dollars) 0.55 1.87 0.0013 22.86 Population (million) 0.05 0.14 0.00024 1.82 Distance to the nearest urban center (100 km) 1.56 1.39 0.015 7.52 Distance to the nearest airport (100 km) 0.47 0.27 0.0087 1.99 Length of highway (100 km) 0.95 1.28 0 14.45 Length of railway (100 km) 0.10 0.24 0 2.98 Archaeological site (0/1) 0.05 0.22 0 1 Protected area (0/1) 0.14 0.34 0 1

14 Alternatively, I also consider other population thresholds including 200,000 (41 centers) and 100,000 (67 centers) to calculate distance yielding qualitatively similar results. .

74 Table 12 provides a list of all the variables and their descriptive statistics. On average, each of the 1,810 Mexican municipalities in my sample received 3.84 day trips per year.

However, a significant number of municipalities, 77.59% of my municipality-by-year observations, experienced zero visits. Among the municipalities that received birdwatching visits, there is significant heterogeneity, with some having as high as 848 visits in a single year.

Deforestation was unevenly distributed across the sample as well. The maximum annual deforestation rate in a municipality was 6.52%, while the average was 0.13%. Moreover, 21.63% of the municipality-by-year observations did not experience any deforestation. These statistics are suggestive of the significant heterogeneity in my deforestation rate measure across municipalities and years which form the basis for my econometric analysis and identification.

4.3 Empirical Models

To empirically test the impact of deforestation on birdwatching ecotourism, I first estimate a binary probability model of a municipality attracting tourists. Using a logistic distribution, I calculate the probability of a municipality receiving at least 1 visit as: 1 푃(푦푖푡 = 1|풙푖푡) = ′ (15) 1 + 푒−풙푖푡휷 where 푦푖푡 is a binary variable equal to 1 if a municipality i received at least 1 visit in a year t,

풙푖푡 is a vector of the municipality characteristics that are likely affect the likelihood of receiving birding visits, and 휷 is a vector of coefficients.15 To aid in interpretation I report the average marginal effects of the estimated coefficients. I calculate the marginal effect at each observation

15 As a robustness check, I also estimate a probit model and report those results in the appendix.

75 and use the mean of the estimated marginal effects for each municipality characteristic to reveal how the characteristic affects the probability of the municipality being visited.

In addition to modeling the probability of receiving visits, I use count data models to measure the intensive margin impact of deforestation on the number of visits occurring in a municipality. Standard Poisson regression models may lead to inefficient estimates if the distribution of the count data is overdispersed (Gardner et al. 1995). Negative binomial (NB) regression models address overdispersion by introducing an idiosyncratic error term to capture unobserved characteristics (shown in the appendix). Given the large number of municipalities with no day trip visits, I also estimate zero-inflated negative binomial (ZINB) models (Greene

1994), where visitation is modeled as a two-stage process by expressing the count variable 푦푖푡 as a product of two latent variables

∗ 푦푖푡 = ℎ푖푡푦푖푡 (16) where ℎ푖푡 is a binary variable, with 0 indicating 푦푖푡 is a constant zero and 1 푦푖푡 follows a

∗ discrete distribution; 푦푖푡 follows a NB distribution. In the first stage, I use a logit model to capture the probability of 푦푖푡 being a constant zero count, i.e. the municipality is not in birders’ choice set. Thus, a positive coefficient indicates that the corresponding characteristic increases the probability of receiving no visits. The probability 푃(ℎ푖푡 = 0) in the first stage is denoted as

푞푖푡. Suppose ℎ푖푡 follows a logistic distribution, then 1 푞푖푡 = 푃(ℎ푖푡 = 0|풛푖푡) = ′ (17) 1 + 푒−풛푖푡휸 where 푞푖푡 is the probability that a municipality does not enter a birder’s choice set, and thus received zero birding visits; 풛푖푡 is a vector of municipality characteristics that affect the

76 probability of ℎ푖푡 being zero, and 휸 are coefficients.

When ℎ푖푡 = 1 with probability 1 − 푞푖푡, the second stage assumes 푦푖푡 follows a NB distribution. In this way, the model allows for 푦푖푡 to be zero as an outcome of the distribution, which is important in my application as some destinations are likely to receive zero visits despite being in birders’ choice sets. To estimate the model, the distribution of 푦 is given as follows:

∗ 푃(푦푖푡 = 0|풙푖푡, 풛푖푡) = 푃(ℎ푖푡 = 0|풛푖푡) + 푃(ℎ푖푡 = 1, 푦푖푡 = 0|풙푖푡, 풛푖푡) = 푞푖푡 + (1 − 푞푖푡)퐹(0) (18)

푃(푦푖푡 = 푘|풙푖푡, 풛푖푡) = (1 − 푞푖푡)퐹(푘), 푘 ≥ 1 (19)

∗ where 퐹(∙) is the probability distribution of 푦푖푡. Embedding the features of the NB model that capture omitted unobservable characteristics, the ZINB model places additional weight on the probability of observing zero counts through the added binary process to address excess zeros.

Consequently, the ZINB model is less likely to suffer from overdispersion, resulting in more efficient estimates.

The magnitudes of the coefficients in count data models cannot be as directly interpreted as those in the binary models. To interpret these coefficients, I calculate the incidence rate ratio

(IRR) of the estimates to illustrate the effects of each independent variable on the visitation counts. When the kth municipality characteristic increases by a unit, the expectation of the change in visit count relative to baseline is defined as the IRR for that characteristic. A value greater than 1 indicates an increase in visitation and a value less than 1 indicates a reduction in visitation.

To empirically estimate both binary models of visitation and count data models, I specify covariates as follows

77 ′ (1) (2) (푚) (푚+1) (퐾) 풙푖푡휷 = 훽0 + 훽1푥푖푡 + 훽2푥푖푡 + ⋯ + 훽푚푥푖푡 + 훽푚+1푥푖 + ⋯ + 훽퐾푥푖 + 흁풔 + 흉풕 (20)

(1) (2) (푚) where 푥푖푡 denotes the deforestation rate in municipality 𝑖 in year 푡, 푥푖푡 , ..., 푥푖푡 denote

(푚+1) (퐾) other time-varying covariates; 푥푖 , ..., 푥푖 are time-invariant. 훽0 is a constant, while

훽1, ..., 훽퐾 are coefficients on the independent variables. I also include state fixed effects, 흁풔, and year fixed effects, 흉풕 (s = 1, …, 32 and t = 2008, …, 2016). 휷 = {훽0, 훽1, … , 훽퐾, 흁풔, 흉풕} are the parameters to estimate in each of the binary or count data model mentioned above.

While this chapter is focused on the impacts of deforestation on ecotourism, there is a potential for ecotourism affecting deforestation simultaneously, and the resulting reverse causality would bias my estimates. If some forests do not attract sufficient visitors to generate revenue, they might have a higher risk of being deforested. On the other hand, prosperous tourism accompanied with infrastructure construction is also likely to accelerate deforestation.

To address these endogenous issues, I also consider specifications using lagged deforestation rate by one year, which is unlikely to be affected by birders’ visitations in the current year. The results are shown in Appendix Tables 28-30, highly consistent with my main results from no-lag models.

4.4 Results

Column (1) in Table 13 reports the estimated coefficients from the logit model. The negative coefficient on deforestation indicates that a municipality with increasing forest loss is significantly less likely to be visited by birders. In terms of the magnitude, the mean marginal effects in Column (2) suggest that on average birdwatching visitors are 12.5% less likely to visit

78 Table 13. Results from the Logit Model Probability of Visitation (1) (2) Variables Estimates Mean Marginal Effects Annual deforestation rate -1.001*** -0.125*** (0.113) (0.0137) Municipality size 0.0887*** 0.0111*** (0.0312) (0.00387) Forest area in 2000 0.212* 0.0265* (0.116) (0.0143) Edge to core ratio 0.00180 0.000226 (0.00444) (0.000558) Biomass density 0.00360* 0.000451* (0.00210) (0.000263) Threatened bird species 0.212*** 0.0266*** (0.0606) (0.00762) Temperature in January 0.0175 0.00220 (0.0196) (0.00242) Temperature in July -0.0219 -0.00275 (0.0205) (0.00253) Precipitation 0.0237 0.00297 (0.114) (0.0143) GDP 0.339** 0.0424** (0.144) (0.0182) Population -1.573 -0.197 (2.117) (0.266) Distance to the nearest urban center 0.0571 0.00715 (0.102) (0.0128) Distance to the nearest airport -1.202*** -0.151*** (0.287) (0.0350) Length of highway 0.423*** 0.0529*** (0.104) (0.0123) Length of railway 0.477** 0.0598** (0.223) (0.0284) Archaeological site 0.524** 0.0656** (0.236) (0.0292) Protected area 0.853*** 0.107*** (0.156) (0.0180)

Observations 16,254 16,254 Method Logit Logit State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

79 a municipality with an additional percentage point of deforestation. 16 The covariates on municipality size, forest area, biomass density, and the number of threatened bird species are all positively correlated with the probability of a municipality being visited. The results also indicate that birders value the level of economic development, transportation infrastructure and presence of other tourist attractions when making location choices. I see this by observing that birders are more likely to visit municipalities that have higher GDP, are closer to airports, and contain more highways, railways, archaeological sites, and national protected areas. These results are highly consistent with my expectations and economic theory providing my first evidence that the citizen science data collected by eBird appears to perform well in a revealed preference economic setting.

In Table 14, Columns (1) and (2) present the estimates of the first and second stages of

ZINB model, respectively. Consistent with the previous logit model, the positive coefficient of deforestation rate in Column (1) indicates that a municipality that no one chooses to visit is associated with a higher deforestation rate. The impacts of other municipality characteristics on birding tourism are also consistent. Note that the estimated signs in the first stage of the ZINB model are expected to be the opposite of those in the binary logit model (Table 13), because the dependent variables (probabilities) are defined differently.

Turning to the intensive margin, the estimated coefficient of deforestation rate in the second stage of the ZINB model is significant and negative, which suggests that visit counts in a municipality drop when the deforestation rate increases (Table 14 Column 2). Meanwhile, larger

16 Mean marginal effects from the probit model (in Appendix Table 25) are similar to the logit results.

80 Table 14. Estimates from ZINB Model (1) (2) Variables Constant zero Visit counts Annual deforestation rate 1.325*** -0.348* (0.282) (0.198) Municipality size -0.202*** 0.0392* (0.0483) (0.0219) Forest area in 2000 -0.741** 0.0798** (0.332) (0.0405) Edge to core ratio 0.0127 0.0344 (0.0124) (0.0350) Biomass density -0.00122 0.00428 (0.00322) (0.00278) Threatened bird species -0.229* 0.219*** (0.123) (0.0700) Temperature in January -0.0267 0.0165 (0.0345) (0.0263) Temperature in July 0.0152 -0.0460* (0.0345) (0.0239) Precipitation -0.363* -0.316** (0.205) (0.148) GDP -1.491 0.239** (2.250) (0.103) Population 6.769 -1.406 (20.77) (1.536) Distance to the nearest urban center 0.471*** 0.487*** (0.157) (0.135) Distance to the nearest airport 0.989** -1.192*** (0.393) (0.299) Length of highway -1.568*** 0.0907 (0.312) (0.0559) Length of railway -0.123 0.319 (0.508) (0.209) Archaeological site -0.729 0.273 (0.684) (0.329) Protected area -1.087*** 0.347** (0.271) (0.157) ln(α) 1.181*** (0.103) AIC 34632.53 Observations 16,290 16,290 Method ZINB ZINB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

81 municipalities, containing more forest area and hosting increased numbers of threatened bird species all lead to additional visit days. Municipalities with higher temperature in July and more precipitation receive fewer visitors, suggesting that hot weather in summer and rains are not favorable to bird sighting. GDP that reflects the level of local economic development contributes positively to the number of visits. In terms of locations, I find that birdwatching tourists prefer to visit locations farther from large urban centers, which confirms hypotheses that ecotourism activity is likely higher in more rural locations. On the other hand, locations that are closer to airports receive more visits, though the effects of other forms of transport infrastructures on the intensive margin of number of visits are not significant. This is intuitive if access is a key decision which makes these variables more likely to influence the extensive margin participation decision rather than the intensive margin. Locations with national protected areas also receive more visits.17

I report the IRR18 for the ZINB model in Table 15. Examining the IRR estimates, I see that the number of visits in a municipality decreases by 29.4% given an additional percentage point of deforestation. To place my results in context with deforestation in Mexico where the mean annual deforestation rate during my sample period is 0.13%, the number of visits would be expected to decrease by 4.42% per year in comparison with a baseline scenario of no deforestation. However, the significant heterogeneity in deforestation rate across Mexican

17 For comparison, I report results from Poisson and NB models in Appendix Tables 26 and 27. The Akaike information Criterion (AIC) of the ZINB model (in Table 14) is smaller than that of the NB model (in Appendix Table 26), suggesting there are excess zeros in my visit counts and the ZINB model is preferred. Moreover, the significant ln(α) in the ZINB model indicates overdispersion after the excess zero issue is addressed, showing that a NB process in the second stage fits my data better than a Poisson process. Even though these findings favor the ZINB model, the results in terms of deforestation across all the three count data specifications are qualitatively similar. 18 For example, an estimated coefficient of deforestation is -0.348 is equivalent to an IRR of 0.706 (= 푒−0.348).

82 Table 15. Incidence Rate Ratio from ZINB Model (1) (2) Variables Constant zero Visit counts Annual deforestation rate 3.764*** 0.706* (1.062) (0.140) Municipality size 0.817*** 1.040* (0.0395) (0.0228) Forest area in 2000 0.477** 1.083** (0.158) (0.0439) Edge to core ratio 1.013 1.035 (0.0126) (0.0363) Biomass density 0.999 1.004 (0.00322) (0.00279) Threatened bird species 0.795* 1.245*** (0.0978) (0.0872) Temperature in January 0.974 1.017 (0.0336) (0.0268) Temperature in July 1.015 0.955* (0.0351) (0.0229) Precipitation 0.696* 0.729** (0.142) (0.108) GDP 0.225 1.269** (0.507) (0.130) Population 870.4 0.245 (18,077) (0.377) Distance to the nearest urban center 1.601*** 1.628*** (0.251) (0.219) Distance to the nearest airport 2.687** 0.304*** (1.056) (0.0907) Length of highway 0.208*** 1.095 (0.0650) (0.0612) Length of railway 0.884 1.376 (0.449) (0.287) Archaeological site 0.482 1.313 (0.330) (0.433) Protected area 0.337*** 1.414** (0.0916) (0.222)

Observations 16,290 16,290 Method ZINB ZINB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

83 municipalities that ranges between 0% and 6.52% implies that some communities are likely to see significantly higher decreases in visitation which could exceed 90%. The interpretation for other characteristics are similar, where an IRR smaller than 1 indicates a negative percentage impact on visitation days.

4.5 Discussion

Using publicly available citizen science data, I examine the impact of deforestation on birdwatching ecotourism across Mexico. I show that deforestation has a negative impact on not only the probability that birders choose a municipality to visit, but also the number of trips taken.

My results indicate that when the deforestation rate increases by an additional percentage point, the probability of a municipality being chosen as a destination for birding decreases by 12.5% and the number of visits decreases by 29.4%. Considering that the average size of a Mexican municipality is 1290 km2, a percentage point of deforestation in an average municipality implies a loss of 12.9 km2 of forests. In other words, destruction of 12.9 km2 of forest cover reduces birding activities by 29.4%. Given such a large drop in visit counts, the impacts of deforestation on local ecotourism are far reaching.

The reduced number of visitation days due to deforestation is likely to result in significant revenue losses to local communities, firms, and governments. To illustrate the monetary magnitude of the economic impact for a single municipality, I take Xochimilco, a municipality in the Federal District of Mexico City as an example. Xochimilco contains 26.57 km2 of the Natural Protected Area of ejido of Xochimilco and San Gregorio Atlapulco which

84 provide an ideal habitat for migratory birds. Revollo-Fernández (2015) provides an estimate of the annual economic value of birdwatching in these locations at 8.75 million US dollars. A one percent increase in deforestation would result in a loss of 1.22 km2 of forests. This loss would lead to 29.4% fewer birdwatching visits. If I further assume that each birder’s daily expenditure is the same and that the economic impact is proportional to visit counts, this would result in a loss of 2.57 million US dollars.

In addition to direct losses of travel revenue, in the current information age, social media has been widely used as an advertising platform. As fewer tourists are present to share their experiences of visiting protected areas on platforms such as Instagram and Flickr (Hausmann et al. 2018), this form of word-of-mouth type advertising is likely to further imperil ecotourism in these areas over time. The reduction in ecotourism income affects not only the local economy and residents’ well-being, but also the fiscal budget for forest conservation (Sukserm et al. 2012;

Satyanarayana et al. 2012). With insufficient financial funding to prevent deforestation, the loss of forests is likely to continue.

My analysis also demonstrates the potential to use citizen science data in data-limited locations common in many developing countries. I show how citizen science data can be used to recover meaningful and intuitive results associate with economic decision-making. Given the large costs of acquiring data in many developing countries, this is a promising new venue to address issues related to conservation and ecotourism and develop more rigorous analysis to aid policymakers seeking to balance sustainable development and ecotourism. Using nationwide data for citizen science data on birdwatching data in Mexico, I find clear linkages between

85 patterns of deforestation and changes in ecotourism behavior.

86 References

[1] Abadie, A., Drukker, D., Herr, J.L., Imbens, G.W., 2004. Implementing matching

estimators for average treatment effects in Stata. Stata journal 4: 290-311.

[2] Abadie, A., Imbens, G.W., 2002. Simple and bias-corrected matching estimators for

average treatment effects.

[3] Abadie, A., Imbens, G.W., 2011. Bias-corrected matching estimators for average treatment

effects. Journal of Business & Economic Statistics 29.1: 1-11.

[4] Arcanjo, J.S., Luz, E.F., Fazenda, Á.L. and Ramos, F.M., 2016. Methods for evaluating

volunteers’ contributions in a deforestation detection citizen science project. Future

Generation Computer Systems, 56, pp.550-557.

[5] Autor, D.H., Dorn, D., Hanson, G.H., 2013. The China syndrome: Local labor market

effects of import competition in the United States. The American Economic Review 103.6:

2121-2168.

[6] Baccini, A.G.S.J., Goetz, S.J., Walker, W.S., Laporte, N.T., Sun, M., Sulla-Menashe, D.,

Hackler, J., Beck, P.S.A., Dubayah, R., Friedl, M.A. and Samanta, S., 2012. Estimated

carbon dioxide emissions from tropical deforestation improved by carbon-density maps.

Nature climate change, 2(3), p.182.

[7] Baum-Snow, N., 2007. Did highways cause suburbanization? The Quarterly Journal of

Economics: 775-805.

87 [8] Baum-Snow, N., Brandt, L., Henderson, J.V., Turner, M., Zhang, Q., 2015a. Roads,

railroads and decentralization of Chinese cities.

[9] Baum-Snow, N., Brandt, L., Henderson, J.V., Turner, M., Zhang, Q., 2015b. Transport

infrastructure, urban growth and market access in China.

[10] Bayer, P., Keohane, N. and Timmins, C., 2009. Migration and hedonic valuation: The case

of air quality. Journal of Environmental Economics and Management, 58(1), pp.1-14.

[11] Beeson, P.E. and Eberts, R.W., 1989. Identifying productivity and amenity effects in

interurban wage differentials. The Review of Economics and Statistics, pp.443-452.

[12] Bento, A.M., Goulder, L.H., Jacobsen, M.R., Von Haefen, R.H., 2009. Distributional and

efficiency impacts of increased US gasoline taxes. The American Economic Review 99.3:

667-699.

[13] Berger, M.C., Blomquist, G.C. and Peter, K.S., 2008. Compensating differentials in

emerging labor and housing markets: Estimates of quality of life in Russian cities. Journal

of Urban Economics, 63(1), pp.25-55.

[14] Berkovec, J., 1985. New car sales and used car stocks: A model of the automobile market.

The Rand Journal of Economics: 195-214.

[15] Bhuiyan, M.A.H., Siwar, C., Ismail, S.M. and Islam, R., 2011. Ecotourism development in

recreational forest areas. American Journal of Applied Sciences, 8(11), p.1116.

[16] Biggs, D., Turpie, J., Fabricius, C. and Spenceley, A., 2011. The value of avitourism for

conservation and job creation-An analysis from South Africa. Conservation and Society, 9,

pp.80-90.

88 [17] Blomquist, G.C., Berger, M.C. and Hoehn, J.P., 1988. New estimates of quality of life in

urban areas. The American Economic Review, pp.89-107.

[18] Booth, J.E., Gaston, K.J., Evans, K.L. and Armsworth, P.R., 2011. The value of species

rarity in biodiversity recreation: A birdwatching example. Biological Conservation,

144(11), pp.2728-2732.

[19] Bureau of Transportation Statistics, 1998. North American Transportation Atlas Data

(NORTAD). United States Department of Transportation.

[20] Cai, H. and Xie, S., 2007. Estimation of vehicular emission inventories in China from 1980

to 2005. Atmospheric Environment, 41(39), pp.8963-8979.

[21] Cameron, A.C. and Trivedi, P.K., 2013. Regression analysis of count data (Vol. 53), pp. 89.

Cambridge university press.

[22] Cantú, J.C., de Silva, H.G. and Sánchez, M.E., 2011. El dinero vuela: el valor económico

del ecoturismo de observación de aves. Defenders of Wildlife. Washington, DC, EUA.

[23] Carver, E. 2013. Birding in the United States: A Demographic and Economic Analysis. U.S.

Fish and Wildlife Service, Division of Economics.

[24] Ceballos-Lascurain, H., 1996. Tourism, ecotourism, and protected areas: The state of

nature-based tourism around the world and guidelines for its development, pp. 301. IUCN.

[25] Chen, Y. and Rosenthal, S.S., 2008. Local amenities and life-cycle migration: Do people

move for jobs or fun?. Journal of Urban Economics, 64(3), pp.519-537.

[26] Cobbinah, P.B., 2015. Contextualising the meaning of ecotourism. Tourism Management

Perspectives, 16, pp.179-189.

89 [27] Comisión Nacional de Áreas Naturales Protegidas (CONANP), 2018. Áreas Naturales

Protegidas Federales de la República Mexicana.

[28] Connell, J., 2009. Birdwatching, twitching and tourism: towards an Australian perspective.

Australian Geographer, 40(2), pp.203-217.

[29] Conradie, N. and van Zyl, C., 2013. Agreement of the international avitourist market to

ecotourism principles: a South African development perspective. African Journal of

Business Management, 7(30), pp.3013-3021.

[30] Cordell, H.K. and Herbert, N.G., 2002. The popularity of birding is still growing. Birding.

February 2002. pp 54-61.

[31] Couture, V., Duranton, G., Turner, M.A., 2017. Speed. Review of Economics and Statistics,

Forthcoming.

[32] del Hoyo, J. and Collar, N. J., 2014. HBW and BirdLife International Illustrated Checklist

of the Birds of the World. Volume 1: Non-passerines. Lynx Edicions and BirdLife

International, Barcelona, Spain and Cambridge, UK.

[33] del Hoyo, J., and Collar, N.J., 2016. HBW and BirdLife International Illustrated Checklist

of the Birds of the World. Volume 2: Passerines. Lynx Edicions and BirdLife International,

Barcelona, Spain and Cambridge, UK.

[34] Deller, S.C., Tsai, T.H., Marcouiller, D.W. and English, D.B., 2001. The role of amenities

and quality of life in rural economic growth. American Journal of Agricultural Economics,

83(2), pp.352-365.

90 [35] Donahue, M.L., Keeler, B.L., Wood, S.A., Fisher, D.M., Hamstead, Z.A. and McPhearson,

T., 2018. Using social media to understand drivers of urban park visitation in the Twin

Cities, MN. Landscape and Urban Planning, 175, pp.1-10.

[36] Downs, A., 1962. The law of peak-hour expressway congestion. Traffic Quarterly 16.3.

[37] Downs, A., 1992. Stuck in traffic: Coping with peak-hour traffic congestion. Brookings

Institution Press.

[38] Du, Y., Park, A. and Wang, S., 2005. Migration and rural poverty in China. Journal of

comparative economics, 33(4), pp.688-709.

[39] Duranton, G., Turner M.A., 2011. The fundamental law of road congestion: Evidence from

US cities. The American Economic Review: 2616-2652.

[40] Duranton, G., Turner, M.A., 2012. Urban growth and transportation. The Review of

Economic Studies 79.4: 1407-1440.

[41] Dyett, M.V., Fajans, M., King, M.C., 1980. The Land Use and Urban Development

Impacts of Beltways. No. DOT-P-30-80-38 Final Rpt.

[42] Eshun, G. and Tagoe-Darko, E., 2015. Ecotourism development in Ghana: A postcolonial

analysis. Development Southern Africa, 32(3), pp.392-406.

[43] Evans, C., Abrams, E., Reitsma, R., Roux, K., Salmonsen, L. and Marra, P.P., 2005. The

Neighborhood Nestwatch Program: Participant outcomes of a citizen‐science ecological

research project. Conservation Biology, 19(3), pp.589-594.

[44] Fahrig, L., 2003. Effects of habitat fragmentation on biodiversity. Annual review of

ecology, evolution, and systematics, 34(1), pp.487-515.

91 [45] Fahrig, L., 2017. Ecological responses to habitat fragmentation per se. Annual Review of

Ecology, Evolution, and Systematics, 48, pp.1-23.

[46] Fan, Y. and Van den Dool, H., 2008. A global monthly land surface air temperature analysis

for 1948–present. Journal of Geophysical Research: Atmospheres, 113(D1).

[47] Fennell, D. and Weaver, D., 2005. The ecotourium concept and tourism-conservation

symbiosis. Journal of sustainable tourism, 13(4), pp.373-390.

[48] Fink, D., Auer, T., Obregon, F., Hochachka, W.M., Iliff, M., Sullivan, B., Wood, C., Davies,

I and Kelling, S., 2017. The eBird Reference Dataset, Version 2016. Cornell Lab of

Ornithology, Ithaca, NY.

[49] Fisher-Vanden, K., Ho, M.S., 2007. How do market reforms affect China's responsiveness

to environmental policy? Journal of Development Economics 82.1: 200-233.

[50] Freeman, R.B., 2010. Labor regulations, unions, and social protection in developing

countries: market distortions or efficient institutions?. In Handbook of development

economics (Vol. 5, pp. 4657-4702). Elsevier.

[51] Fu, S. and Gu, Y., 2017. Highway toll and air pollution: Evidence from Chinese

cities. Journal of Environmental Economics and Management, 83, pp.32-49.

[52] Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G.,

Rowland, J., Harrison, L., Hoell, A. and Michaelsen, J., 2015. The climate hazards infrared

precipitation with stations -- a new environmental record for monitoring extremes.

Scientific data, 2, p.150066.

92 [53] Gabriel, S.A. and Rosenthal, S.S., 2004. Quality of the business environment versus

quality of life: do firms and households like the same cities?. Review of Economics and

Statistics, 86(1), pp.438-444.

[54] Gabriel, S.A., Mattey, J.P. and Wascher, W.L., 2003. Compensating differentials and

evolution in the quality-of-life among US states. Regional Science and Urban Economics,

33(5), pp.619-649.

[55] GADM, 2009. GADM database of Global Administrative Areas, version 1.0. Online url:

www.gadm.org.

[56] Gardner, W., Mulvey, E.P. and Shaw, E.C., 1995. Regression analyses of counts and rates:

Poisson, overdispersed Poisson, and negative binomial models. Psychological bulletin,

118(3), p.392.

[57] Giam, X., 2017. Global biodiversity loss from tropical deforestation. Proceedings of the

National Academy of Sciences, 114(23), pp.5775-5777.

[58] Gibson, J. and Olivia, S., 2010. The effect of infrastructure access and quality on non-farm

enterprises in rural Indonesia. World Development, 38(5), pp.717-726.

[59] Gollin, D., 2014. The Lewis model: A 60-year retrospective. Journal of Economic

Perspectives, 28(3), pp.71-88.

[60] Gómez de Silva, H. and Alvarado Reyes, E., 2010. Breve historia de la observación de aves

en México en el siglo XX y principios del siglo XXI. Huitzil, 11(1), pp.9-20.

93 [61] Graham, C.H. and Blake, J.G., 2001. Influence of patch‐and landscape‐level factors on bird

assemblages in a fragmented tropical landscape. Ecological Applications, 11(6),

pp.1709-1721.

[62] Graham, D.J., 2007. Variable returns to agglomeration and the effect of road traffic

congestion. Journal of Urban Economics, 62(1), 103-120.

[63] Greene, W.H., 1994. Accounting for excess zeros and sample selection in Poisson and

negative binomial regression models.

[64] Gyourko, J. and Tracy, J., 1991. The structure of local public finance and the quality of life.

Journal of political economy, 99(4), pp.774-806.

[65] Haberfeld, Y., Menaria, R.K., Sahoo, B.B. and Vyas, R.N., 1999. Seasonal migration of

rural labor in India. Population Research and Policy Review, 18(5), pp.471-487.

[66] Hale, B.W., 2018. Mapping potential environmental impacts from tourists using data from

social media: A case study in the Westfjords of Iceland. Environmental management, 62(3),

pp.446-457.

[67] Hall, M. and Greenman, E., 2015. The occupational cost of being illegal in the United

States: Legal status, job hazards, and compensating differentials. International Migration

Review, 49(2), pp.406-442.

[68] Hansen, M., Gillen, D., Dobbins, A., Huang, Y., Puvathingal, M., 1993. The air quality

impacts of urban highway capacity expansion: Traffic generation and land use change.

University of California Transportation Center.

94 [69] Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A.A., Tyukavina, A.,

Thau, D., Stehman, S.V., Goetz, S.J., Loveland, T.R. and Kommareddy, A., 2013.

High-resolution global maps of 21st-century forest cover change. science, 342(6160),

pp.850-853.

[70] Hausman, J.A., 1996. Valuation of new goods under perfect and imperfect competition, in:

The economics of new goods. University of Chicago Press 207-248.

[71] Hausmann, A., Toivonen, T., Slotow, R., Tenkanen, H., Moilanen, A., Heikinheimo, V. and

Di Minin, E., 2018. Social Media Data Can Be Used to Understand Tourists’ Preferences

for Nature - Based Experiences in Protected Areas. Conservation Letters, 11(1), p.e12343.

[72] Hazell, P., Poulton, C., Wiggins, S. and Dorward, A., 2010. The future of small farms:

trajectories and policy priorities. World Development, 38(10), pp.1349-1361.

[73] He, G., Fan, M. and Zhou, M., 2016. The effect of air pollution on mortality in China:

Evidence from the 2008 Beijing Olympic Games. Journal of Environmental Economics

and Management, 79, pp.18-39.

[74] Hendee, J.C., Dawson, C.P. and Sharpe, W.F., 2012. Introduction to Forests and Renewable

Resources: Eighth Edition, pp. 65. Waveland Press.

[75] Hoefle, S.W., 2016. Multi-functionality, juxtaposition and conflict in the Central Amazon:

Will tourism contribute to rural livelihoods and save the rainforest?. Journal of rural studies,

44, pp.24-36.

[76] Hsu, W. T., Zhang, H., 2014. The fundamental law of highway congestion revisited:

Evidence from national expressways in Japan. Journal of Urban Economics, 81, 65-76.

95 [77] Hunt, C.A., Durham, W.H., Driscoll, L. and Honey, M., 2015. Can ecotourism deliver real

economic, social, and environmental benefits? A study of the Osa Peninsula, Costa Rica.

Journal of Sustainable Tourism, 23(3), pp.339-357.

[78] Huo, H., Zhang, Q., He, K., Yao, Z., Wang, M., 2012. Vehicle-use intensity in China:

Current status and future trend. Energy Policy 43: 6-16.

[79] Hutchins, F., 2007. Footprints in the forest: Ecotourism and altered meanings in Ecuador's

upper Amazon. The Journal of Latin American and Caribbean Anthropology, 12(1),

pp.75-103.

[80] Hvenegaard, G.T. and Dearden, P., 1998. Ecotourism versus tourism in a Thai National

Park. Annals of tourism Research, 25(3), pp.700-720.

[81] Hymel, K., 2009. Does traffic congestion reduce employment growth?. Journal of Urban

Economics, 65(2), 127-135.

[82] INEGI, 2005. II Conteo de Población y Vivienda 2005.

[83] INEGI, 2014. Red Nacional de Caminos RNC 2014.

[84] Instituto Nacional de Estadística y Geografía (INEGI), 2000. Metadatos geoespaciales

digitales. Conjunto de datos vectoriales de la serie topográfica esc. 1:1 000 000.

[85] Instituto Nacional para el Federalismo y el Desarrollo Municipal (INAFED), 2005. Base de

datos de Finanzas PIB Municipal 2005.

[86] Jenkins, C.N., Pimm, S.L. and Joppa, L.N., 2013. Global patterns of terrestrial vertebrate

diversity and conservation. Proceedings of the National Academy of Sciences, 110(28),

pp.E2602-E2610.

96 [87] Jorgensen, R.E., 1948. Influence of expressways in diverting traffic from alternate routes

and in generating new traffic. In: Highway Research Board Proceedings, vol. 27.

[88] Keeler, B.L., Wood, S.A., Polasky, S., Kling, C., Filstrup, C.T. and Downing, J.A., 2015.

Recreational demand for clean water: evidence from geotagged photographs by visitors to

lakes. Frontiers in Ecology and the Environment, 13(2), pp.76-81.

[89] Koh, L.P. and Ghazoul, J., 2010. Spatially explicit scenario analysis for reconciling

agricultural expansion, forest protection, and carbon conservation in Indonesia.

Proceedings of the National Academy of Sciences, 107(24), pp.11140-11144.

[90] Kolstoe, S. and Cameron, T.A., 2017. The non-market value of birding sites and the

marginal value of additional species: biodiversity in a random utility model of site choice

by eBird members. Ecological economics, 137, pp.1-12.

[91] Koppelman, F.S., 1972. Preliminary study for development of a macro urban travel

demand model. No. DOT-P-5010.1.

[92] Krebs, C.J., 1972. The experimental analysis of distribution and abundance. Ecology. New

York: Harper and Row.

[93] Kronenberg, J., 2016. Birdwatchers’ wonderland? Prospects for the development of

birdwatching tourism in Poland. Journal of Ecotourism, 15(1), pp.78-94.

[94] Laurance, W.F., Sayer, J. and Cassman, K.G., 2014. Agricultural expansion and its impacts

on tropical nature. Trends in ecology & evolution, 29(2), pp.107-116.

97 [95] Lee, C.K., Lee, J.H., Kim, T.K. and Mjelde, J.W., 2010. Preferences and willingness to pay

for bird-watching tour and interpretive services using a choice experiment. Journal of

Sustainable Tourism, 18(5), pp.695-708.

[96] Lewis, W.A., 1954. Economic development with unlimited supplies of labour. The

manchester school, 22(2), pp.139-191.

[97] Li, F., Zhu, Q. and Yang, Z., 2013. Birding tourism development in Sichuan, China.

Tourism Economics, 19(2), pp.257-273.

[98] Li, S., Liu, Y., Purevjav, A. and Yang, L., 2019. Does subway expansion improve air

quality? Journal of Environmental Economics and Management, 96, pp. 213-235.

[99] Liang, H., 2016. Influence mechanism of development effect of forest ecotourism in China.

Forestry Studies, 64(1), pp.93-100.

[100] Liu, X., Hu, G., Chen, Y., Li, X., Xu, X., Li, S., Pei, F. and Wang, S., 2018. High-resolution

multi-temporal mapping of global urban land using Landsat images based on the Google

Earth Engine Platform. Remote Sensing of Environment, 209, pp.227-239.

[101] Llorente-Bousquets, J. and Ocegueda, S., 2008. Estado del conocimiento de la biota, en

Capital natural de México, vol. I: Conocimiento actual de la biodiversidad. Conabio,

México, 283-322.

[102] Lu, Y., 2010. Rural-urban migration and health: evidence from longitudinal data in

Indonesia. Social science & medicine, 70(3), pp.412-419.

[103] Lucas, Jr, R.E., 2004. Life earnings and rural-urban migration. Journal of political economy,

112(S1), pp.S29-S59.

98 [104] Luz, E.F., Correa, F.R., González, D.L., Grey, F. and Ramos, F.M., 2014. The

forestwatchers: a citizen cyberscience project for deforestation monitoring in the tropics.

Human Computation, 1(2), pp.137-145.

[105] Ma, Z., Cheng, Y., Wang, J. and Fu, X., 2013. The rapid development of birdwatching in

mainland China: a new force for bird study and conservation. Bird conservation

international, 23(2), pp.259-269.

[106] Marx, B., Stoker, T. and Suri, T., 2013. The economics of slums in the developing world.

Journal of Economic Perspectives, 27(4), pp.187-210.

[107] Meng, L. and Zhao, M.Q., 2018. Permanent and temporary rural–urban migration in China:

Evidence from field surveys. China Economic Review, 51, pp.228-239.

[108] Mulligan, G.F. and Carruthers, J.I., 2011. Amenities, quality of life, and regional

development. In Investigating quality of urban life (pp. 107-133). Springer, Dordrecht.

[109] Naidoo, R. and Adamowicz, W.L., 2005a. Economic benefits of biodiversity exceed costs

of conservation at an African rainforest reserve. Proceedings of the National Academy of

Sciences, 102(46), pp.16712-16716.

[110] Naidoo, R. and Adamowicz, W.L., 2005b. Biodiversity and nature-based tourism at forest

reserves in Uganda. Environment and Development Economics, 10(2), pp.159-178.

[111] Naidoo, R., Weaver, L.C., Stuart‐Hill, G. and Tagg, J., 2011. Effect of biodiversity on

economic benefits from communal lands in Namibia. Journal of Applied Ecology, 48(2),

pp.310-316.

99 [112] Newman, P.W.G., Kenworthy, J.R., 1989. Gasoline consumption and cities: a comparison

of US cities with a global survey. Journal of the American Planning Association 55, no. 1:

24-37.

[113] Nieuwenhuis, P., Beresford, A., Choi, A.K., 2012. Shipping or local production? CO2

impact of a strategic decision: an automotive industry case study. International Journal of

Production Economics 140.1: 138-148.

[114] Ocampo-Penuela, N. and Winton, R.S., 2017. Economic and conservation potential of

bird-watching tourism in postconflict Colombia. Tropical Conservation Science, 10,

p.1940082917733862.

[115] Ocampo-Penuela, N. and Winton, R.S., 2017. Economic and conservation potential of

bird-watching tourism in postconflict Colombia. Tropical Conservation Science, 10(1),

pp.1-6.

[116] Ochoa-Quintero, J.M., Gardner, T.A., Rosa, I., de Barros Ferraz, S.F. and Sutherland, W.J.,

2015. Thresholds of species loss in Amazonian deforestation frontier landscapes.

Conservation Biology, 29(2), pp.440-451.

[117] Pells, S.R., 1989. User response to new road capacity: a review of published evidence.

[118] Peterson, A.T. and Navarro-Sigüenza, A.G., 2016. Bird conservation and biodiversity

research in Mexico: status and priorities. Journal of Field Ornithology, 87(2), pp.121-132.

[119] Powell, D. and Shan, H., 2012. Income taxes, compensating differentials, and occupational

choice: How taxes distort the wage-amenity decision. American Economic Journal:

Economic Policy, 4(1), pp.224-47.

100 [120] Pucher, J., Peng, Z., Mittal, N., Zhu, Y., Korattyswaroopam, N., 2007. Urban transport

trends and policies in China and India: impacts of rapid economic growth. Transport

Reviews 27, no. 4: 379-410.

[121] Puhakka, L., Salo, M. and Sääksjärvi, I.E., 2011. Bird diversity, birdwatching tourism and

conservation in Peru: a geographic analysis. PLoS One, 6(11), p.e26786.

[122] Ramos, A.M. and Prideaux, B., 2014. Indigenous ecotourism in the Mayan rainforest of

Palenque: empowerment issues in sustainable development. Journal of Sustainable Tourism,

22(3), pp.461-479.

[123] Revollo-Fernández, D.A., 2015. Does money fly? The economic value of migratory

birdwatching in Xochimilco, Mexico. Modern Economy, 6(06), p.643.

[124] Ricker, B.A., Johnson, P.A. and Sieber, R.E., 2013. Tourism and environmental change in

Barbados: gathering citizen perspectives with volunteered geographic information (VGI).

Journal of Sustainable Tourism, 21(2), pp.212-228.

[125] Roback, J., 1982. “Wages, rents, and the quality of life.” Journal of Political Economy 90.6:

1257-1278.

[126] Satyanarayana, B., Bhanderi, P., Debry, M., Maniatis, D., Foré, F., Badgie, D., Jammeh, K.,

Vanwing, T., Farcy, C., Koedam, N. and Dahdouh-Guebas, F., 2012. A socio-ecological

assessment aiming at improved forest resource management and sustainable ecotourism

development in the mangroves of Tanbi Wetland National Park, The Gambia, West Africa.

Ambio, 41(5), pp.513-526.

101 [127] Şekercioğlu, Ç.H., 2002. Impacts of birdwatching on human and avian communities.

Environmental conservation, 29(3), pp.282-289.

[128] Sexton, J.O., Noojipady, P., Song, X.P., Feng, M., Song, D.X., Kim, D.H., Anand, A.,

Huang, C., Channan, S., Pimm, S.L. and Townshend, J.R., 2016. Conservation policy and

the measurement of forests. Nature Climate Change, 6(2), p.192.

[129] SinoMaps press, Map of the Peoples Republic of China, scale 1:4000000, as of 1962,

publication date February 1962, Zhong guo di zhi tu zhi yin chang.

[130] Snyman, S., 2017. The role of private sector ecotourism in local socio-economic

development in southern Africa. Journal of Ecotourism, 16(3), pp.247-268.

[131] Song, M., Zheng, W., Wang, Z., 2016. Environmental efficiency and energy consumption

of highway transportation systems in China. International Journal of Production Economics

181: 441-449.

[132] Spalding, M., Burke, L., Wood, S.A., Ashpole, J., Hutchison, J. and zu Ermgassen, P., 2017.

Mapping the global value and distribution of coral reef tourism. Marine Policy, 82,

pp.104-113.

[133] Stepenuck, K. and Green, L., 2015. Individual-and community-level impacts of volunteer

environmental monitoring: a synthesis of peer-reviewed literature. Ecology and society,

20(3).

[134] Steven, R., Morrison, C. and Castley, J.G., 2015. Birdwatching and avitourism: a global

review of research into its participant markets, distribution and impacts, highlighting future

102 research priorities to inform sustainable avitourism management. Journal of Sustainable

Tourism, 23(8-9), pp.1257-1276.

[135] Strauss, J., Witoelar, F. and Sikoki, B., 2016. “The Fifth Wave of the Indonesia Family Life

Survey (IFLS5): Overview and Field Report”. WR-1143/1-NIA/NICHD.

[136] Sukserm, T., Thiengkamol, N. and Thiengkamol, T., 2012. Development of the ecotourism

management model for forest park. Journal of the Social Sciences, 7(1), pp.95-99.

[137] Sullivan, B.L., Wood, C.L., Iliff, M.J., Bonney, R.E., Fink, D. and Kelling, S., 2009. eBird:

A citizen-based bird observation network in the biological sciences. Biological

Conservation, 142(10), pp.2282-2292.

[138] Todaro, M.P., 1969. A model of labor migration and urban unemployment in less

developed countries. The American economic review, 59(1), pp.138-148.

[139] Tsekeris, T., Geroliminis, N., 2013. City size, network structure and traffic congestion.

Journal of Urban Economics, 76, 1-14.

[140] Van Donkelaar, A., Martin, R.V., Brauer, M. and Boys, B.L., 2015. Global Annual PM2. 5

Grids from MODIS, MISR and SeaWiFS Aerosol Optical Depth (AOD), 1998-2012.

Palisades, NY: NASA Socioeconomic Data and Applications Center (SEDAC). doi, 10,

p.H4028PFS.

[141] Viard, V.B., and Fu, S., 2015. The effect of Beijing's driving restrictions on pollution and

economic activity. Journal of Public Economics 125: 98-115.

103 [142] Vogt, P., Riitters, K.H., Estreguil, C., Kozak, J., Wade, T.G. and Wickham, J.D., 2007.

Mapping spatial patterns with morphological image processing. Landscape ecology, 22(2),

pp.171-177.

[143] West, S.E., 2004. Distributional effects of alternative vehicle pollution control policies.

Journal of public Economics 88.3: 735-757.

[144] Williams, A.S. and Jobes, P.C., 1990. Economic and quality-of-life considerations in

urban-rural migration. Journal of Rural Studies, 6(2), pp.187-194.

[145] Yagar, S., 1973. Potential Demand Response to Improved Roadway Service. Traffic

Quarterly 27, no. 1.

[146] Yu, N., de Roo, G., de Jong, M., Storm, S., 2015. Does the expansion of a motorway

network lead to economic agglomeration? Evidence from China. Transport Policy

45 :218-227.

[147] Zhao, Y., 1999. Leaving the countryside: rural-to-urban migration decisions in China.

American Economic Review, 89(2), pp.281-286.

104

Appendix A: Supplemental Materials

105 Table 16. OLS estimates of road effect on VKT with varying fixed effects (1) (2)

Variables ln(VKT) ln(VKT) ln(Road km) 1.017*** 1.006*** (0.0693) (0.0731) ln(Car ownership number) 0.211*** 0.203*** (0.0506) (0.0547) ln(Land area) 0.0367 0.0761** (0.0267) (0.0325) ln(Population) -0.126* -0.175** (0.0650) (0.0664) ln(GDP) -0.0642 -0.136 (0.0921) (0.0903) ln(Income) -0.506*** -0.343** (0.121) (0.165) ln(Bus number) 0.202*** 0.186** (0.0625) (0.0720) ln(Taxi number) 0.0325 0.127** (0.0414) (0.0465)

Division FE (6) Yes

Year FE (4) Yes

Observations 409 409 R2 0.874 0.880 All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

106 Table 17. 2SLS estimates of road effect on VKT with varying fixed effects (1) (2) (3) (4) Variables ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) 1.070*** 1.086***

(0.153) (0.213) ln(Car ownership number) 0.128** 0.210*** 0.102* 0.205*** (0.0528) (0.0501) (0.0608) (0.0540)

Instruments: ln(1984 Road km) 0.149*** 0.0972**

(0.0416) (0.0391)

ln(1962 Highway km) -0.139*** -0.130***

(0.0212) (0.0233)

ln(5-year lagged road km -0.219*** -0.178***

from 4 matched cities) (0.0490) (0.0524) ln(Land area) 0.103*** 0.0349 0.141*** 0.0676* (0.0275) (0.0264) (0.0252) (0.0355) ln(Population) 0.182*** -0.131** 0.162** -0.179*** (0.0673) (0.0611) (0.0692) (0.0621) ln(GDP) 0.497*** -0.0947 0.450*** -0.177 (0.0529) (0.0972) (0.0446) (0.115) ln(Income) 0.0627 -0.498*** -0.0956 -0.313* (0.123) (0.110) (0.194) (0.179) ln(Bus number) 0.240*** 0.189*** 0.248*** 0.163* (0.0852) (0.0702) (0.0861) (0.0938) ln(Taxi number) -0.172*** 0.0368 -0.103* 0.130*** (0.0512) (0.0361) (0.0575) (0.0450)

Division FE (6) Yes Yes

Year FE (4) Yes Yes

Observations 409 409 409 409 R2 0.900 0.874 0.907 0.880 First-stage statistic 17.85 12.80

Overidentification p-value 0.251 0.188

The same regressions with different fixed effects are performed in all pairs of columns. Odd columns present first stages and even columns present second stages. All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

107 Table 18. Robustness to lag specification (1) (2) (3) (4)

Variables ln(Road km) ln(VKT) ln(Road km) ln(VKT) ln(Road km) 1.134*** 1.128***

(0.221) (0.220) ln(Car ownership number) 0.107 0.206*** 0.107 0.206*** (0.0658) (0.0552) (0.0651) (0.0553)

Instruments: ln(1984 Road km) 0.0943** 0.0953**

(0.0403) (0.0396)

ln(1962 Highway km) -0.129*** -0.130***

(0.0234) (0.0235)

ln(current road km -0.177***

from 4 matched cities) (0.0571)

ln(1-year lagged road km -0.197***

from 4 matched cities) (0.0578) ln(Land area) 0.138*** 0.0602* 0.136*** 0.0608* (0.0270) (0.0342) (0.0265) (0.0335) ln(Population) 0.169** -0.176*** 0.179** -0.176*** (0.0774) (0.0609) (0.0749) (0.0611) ln(GDP) 0.453*** -0.198 0.455*** -0.195 (0.0421) (0.122) (0.0427) (0.125) ln(Income) -0.125 -0.295 -0.0940 -0.298 (0.170) (0.194) (0.175) (0.191) ln(Bus number) 0.242*** 0.147 0.247*** 0.149 (0.0899) (0.0967) (0.0888) (0.0954) ln(Taxi number) -0.105* 0.128*** -0.112* 0.127*** (0.0592) (0.0460) (0.0586) (0.0456)

Division×Year FE (24) Yes Yes Yes Yes Observations 409 409 409 409 R2 0.908 0.881 0.909 0.881 First-stage statistic 11.25 12.60

Overidentification p-value 0.215 0.214

The same regressions with different instruments are performed in all pairs of columns. Odd columns present first stages and even columns present second stages. All regressions include a constant. Robust standard errors in parentheses (clustered by division×year). *** p < 0.01, ** p < 0.05, * p < 0.1.

108 Table 19. Mean covariate differences between focal and matched cities, by city Area Population GDP Income Cities Mean (SD) Mean (SD) Mean (SD) Mean (SD) 1 Beijing 2.08 (0.51) 1.50 (1.31) 1.44 (1.54) 0.31 (0.86) 2 Tianjin 0.62 (1.25) 0.50 (1.32) 0.47 (1.63) -0.42 (0.98) 3 Shijiazhuang 0.10 (0.29) 0.16 (0.33) 0.03 (0.26) -0.08 (0.26) 4 Tangshan 0.19 (0.30) 0.23 (0.31) 0.26 (0.29) 0.21 (0.21) 5 Handan -0.10 (0.12) -0.01 (0.21) -0.06 (0.19) -0.09 (0.21) 6 Baoding -0.11 (0.13) 0.01 (0.18) 0.04 (0.12) 0.02 (0.18) 7 Taiyuan 0.09 (0.23) 0.26 (0.17) 0.18 (0.16) 0.02 (0.27) 8 Datong 0.07 (0.14) 0.05 (0.12) 0.00 (0.11) -0.01 (0.19) 9 Yangquan -0.08 (0.12) -0.08 (0.09) -0.06 (0.06) -0.04 (0.11) 10 Changzhi -0.11 (0.09) -0.09 (0.09) -0.06 (0.05) 0.00 (0.08) 11 Linfen 0.01 (0.12) -0.05 (0.10) -0.09 (0.06) -0.10 (0.11) 12 Hohhot 0.20 (0.19) 0.01 (0.15) 0.11 (0.15) 0.12 (0.27) 13 Baotou -0.10 (0.29) -0.15 (0.31) 0.10 (0.47) 0.18 (0.60) 14 Shenyang 0.48 (0.19) 0.45 (0.50) 0.38 (0.58) 0.35 (0.41) 15 Dalian 0.18 (0.27) 0.27 (0.27) 0.47 (0.33) 0.24 (0.46) 16 Anshan -0.03 (0.16) 0.00 (0.22) 0.03 (0.20) 0.11 (0.21) 17 Fushun -0.02 (0.12) 0.03 (0.09) 0.07 (0.06) 0.07 (0.09) 18 Benxi 0.06 (0.09) 0.05 (0.10) 0.07 (0.07) 0.09 (0.14) 19 Jinzhou 0.02 (0.11) 0.03 (0.09) 0.02 (0.09) 0.06 (0.10) 20 Changchun 0.36 (0.38) 0.30 (0.28) 0.31 (0.23) 0.04 (0.27) 21 Jilin 0.32 (0.24) 0.26 (0.25) 0.17 (0.21) 0.04 (0.25) 22 Harbin 0.77 (0.30) 0.64 (0.17) 0.19 (0.23) -0.14 (0.21) 23 Shanghai -0.34 (1.22) 2.05 (1.03) 2.26 (1.22) 1.22 (0.82) 24 Nanjing 0.93 (0.29) 0.90 (0.31) 0.70 (0.38) 0.42 (0.42) 25 Wuxi -0.06 (0.21) 0.28 (0.27) 0.43 (0.35) 0.14 (0.44) 26 Xuzhou 0.29 (0.12) 0.36 (0.16) 0.30 (0.15) 0.05 (0.19) 27 Changzhou 0.06 (0.32) 0.24 (0.43) 0.37 (0.34) 0.22 (0.16) 28 Suzhou 0.48 (0.35) 0.27 (0.55) 0.58 (0.56) 0.74 (0.48) 29 Nantong -0.07 (0.14) 0.00 (0.20) 0.00 (0.24) 0.14 (0.28) 30 Yangzhou 0.09 (0.21) 0.24 (0.20) 0.23 (0.18) 0.03 (0.23) 31 Zhenjiang -0.11 (0.20) -0.09 (0.17) -0.09 (0.16) -0.12 (0.22) 32 Hangzhou 0.35 (0.55) 0.30 (0.52) 0.46 (0.40) 0.43 (0.29) 33 Ningbo 0.17 (0.22) 0.14 (0.21) 0.33 (0.29) 0.47 (0.37) 34 Wenzhou -0.04 (0.13) 0.01 (0.23) -0.14 (0.32) 0.14 (0.43) 35 Huzhou 0.08 (0.16) -0.01 (0.14) -0.19 (0.19) 0.02 (0.35) 36 Shaoxing 0.37 (0.26) -0.06 (0.34) -0.25 (0.39) 0.38 (0.42) 37 Hefei 0.07 (0.12) 0.25 (0.19) 0.43 (0.16) 0.28 (0.13) 38 Wuhu 0.19 (0.11) 0.16 (0.07) 0.17 (0.06) 0.17 (0.09) 39 Maanshan -0.17 (0.12) -0.12 (0.05) -0.12 (0.09) -0.19 (0.22) 40 Fuzhou -0.10 (0.24) 0.02 (0.27) 0.07 (0.19) 0.10 (0.2) 41 Xiamen -0.09 (0.20) 0.06 (0.20) 0.14 (0.40) 0.38 (0.33) 42 Quanzhou -0.13 (0.12) 0.00 (0.08) 0.00 (0.11) 0.01 (0.23)

109 Area Population GDP Income Cities Mean (SD) Mean (SD) Mean (SD) Mean (SD) 43 Nanchang -0.04 (0.10) 0.10 (0.21) 0.16 (0.24) 0.11 (0.22) 44 Jiujiang -0.06 (0.11) -0.02 (0.11) 0.02 (0.12) -0.04 (0.17) 45 Jinan 0.43 (0.06) 0.50 (0.07) 0.28 (0.25) 0.16 (0.51) 46 Qingdao 0.23 (0.36) 0.26 (0.38) 0.29 (0.43) 0.02 (0.45) 47 Zibo 0.27 (0.07) 0.18 (0.15) 0.07 (0.25) -0.02 (0.30) 48 Zaozhuang 0.21 (0.21) 0.09 (0.24) -0.14 (0.22) -0.06 (0.27) 49 Yantai 0.27 (0.17) 0.14 (0.15) 0.15 (0.15) -0.05 (0.27) 50 Weifang 0.32 (0.23) 0.08 (0.23) -0.06 (0.17) -0.08 (0.24) 51 Jinin 0.13 (0.21) -0.02 (0.24) -0.04 (0.17) 0.02 (0.16) 52 Taian 0.19 (0.24) 0.00 (0.24) -0.17 (0.17) -0.12 (0.19) 53 Zhengzhou -0.48 (0.23) 0.01 (0.19) -0.30 (0.55) -0.21 (0.65) 54 Kaifeng -0.12 (0.17) -0.16 (0.15) -0.12 (0.09) -0.15 (0.12) 55 Luoyang -0.06 (0.17) -0.08 (0.20) -0.16 (0.18) -0.06 (0.18) 56 Pingdingshan -0.18 (0.15) -0.05 (0.15) -0.04 (0.07) 0.00 (0.14) 57 Anyang -0.04 (0.12) 0.03 (0.13) -0.04 (0.06) -0.09 (0.19) 58 Jiaozuo -0.19 (0.15) -0.11 (0.13) -0.10 (0.07) -0.07 (0.11) 59 Sanmenxia -0.25 (0.10) -0.18 (0.06) -0.11 (0.06) -0.13 (0.16) 60 Wuhan -0.07 (0.30) 0.44 (0.57) 0.78 (0.50) 0.36 (0.3) 61 Yichang 0.31 (0.42) 0.05 (0.32) 0.11 (0.18) -0.27 (0.29) 62 Changsha -0.16 (0.2) 0.14 (0.24) 0.38 (0.30) 0.25 (0.31) 63 Zhuzhou -0.19 (0.22) -0.13 (0.18) -0.01 (0.13) 0.08 (0.15) 64 Xiangtan -0.01 (0.08) 0.02 (0.08) 0.08 (0.09) 0.08 (0.09) 65 Yueyang 0.12 (0.13) 0.13 (0.08) 0.14 (0.09) 0.04 (0.12) 66 Changde 0.21 (0.15) 0.20 (0.09) 0.16 (0.05) -0.11 (0.14) 67 Guangzhou -0.85 (1.13) -0.09 (1.38) 0.84 (1.53) 0.96 (0.74) 68 Shaoguan 0.02 (0.26) -0.09 (0.16) -0.09 (0.12) 0.01 (0.27) 69 Shenzhen -1.41 (1.10) -1.34 (1.55) 0.97 (1.64) 1.26 (0.82) 70 Zhuhai 0.17 (0.14) 0.00 (0.14) 0.10 (0.17) 0.24 (0.20) 71 Shantou -0.07 (0.31) 0.13 (0.43) -0.62 (0.38) -0.97 (0.50) 72 Zhanjiang 0.03 (0.09) 0.07 (0.11) 0.07 (0.09) -0.01 (0.10) 73 Nanning 0.85 (0.20) 0.48 (0.35) 0.21 (0.25) -0.15 (0.27) 74 Liuzhou 0.02 (0.13) 0.11 (0.08) 0.20 (0.06) 0.12 (0.15) 75 Guilin -0.04 (0.06) -0.02 (0.04) -0.02 (0.05) -0.03 (0.09) 76 Beihai -0.08 (0.11) -0.05 (0.08) -0.03 (0.07) -0.05 (0.13) 77 Haikou 0.14 (0.15) 0.09 (0.15) 0.06 (0.09) 0.13 (0.13) 78 Chongqing 6.62 (0.38) 5.30 (0.57) 1.62 (0.62) -0.34 (1.08) 79 Chengdu 0.06 (0.29) 0.61 (0.49) 0.53 (0.59) 0.35 (0.40) 80 Zigong 0.03 (0.19) 0.03 (0.16) -0.05 (0.10) -0.14 (0.09) 81 Luzhou 0.11 (0.14) 0.00 (0.15) -0.07 (0.11) -0.09 (0.12) 82 Deyang -0.04 (0.09) -0.07 (0.05) -0.04 (0.06) -0.04 (0.08) 83 Mianyang 0.02 (0.13) -0.02 (0.14) -0.02 (0.09) 0.02 (0.11) 84 Nanchong 0.25 (0.21) 0.06 (0.23) -0.18 (0.15) -0.36 (0.18) 85 Yibin 0.06 (0.11) -0.03 (0.13) -0.06 (0.08) -0.05 (0.13)

110 Area Population GDP Income Cities Mean (SD) Mean (SD) Mean (SD) Mean (SD) 86 Guiyang 0.13 (0.12) 0.17 (0.15) 0.14 (0.12) 0.05 (0.16) 87 Zunyi 0.08 (0.16) 0.02 (0.12) -0.03 (0.07) -0.12 (0.11) 88 Kunming 0.34 (0.18) 0.36 (0.32) 0.32 (0.26) 0.26 (0.24) 89 Qujing -0.02 (0.16) -0.06 (0.07) -0.04 (0.06) 0.01 (0.13) 90 Yuxi -0.08 (0.16) -0.15 (0.07) -0.04 (0.07) 0.04 (0.14) 91 Xi'an 0.41 (0.28) 0.30 (0.30) 0.06 (0.45) 0.73 (0.70) 92 Tongchuan -0.03 (0.28) -0.13 (0.17) -0.14 (0.11) -0.04 (0.14) 93 Baoji 0.14 (0.29) -0.14 (0.25) -0.20 (0.22) 0.01 (0.35) 94 Xianyang -0.24 (0.18) -0.16 (0.14) -0.11 (0.10) 0.07 (0.16) 95 Weinan 0.04 (0.18) -0.01 (0.13) -0.09 (0.07) -0.08 (0.08) 96 Yan'an 0.07 (0.20) -0.29 (0.32) -0.26 (0.27) 0.18 (0.42) 97 Lanzhou 0.08 (0.29) 0.20 (0.26) 0.14 (0.20) -0.05 (0.21) 98 Jinchang 0.03 (0.22) -0.22 (0.14) -0.08 (0.10) 0.09 (0.26) 99 Xining -0.10 (0.17) -0.03 (0.18) 0.00 (0.11) -0.14 (0.13) 100 Yinchuan 0.10 (0.25) 0.05 (0.16) 0.03 (0.11) -0.06 (0.09) 101 Shizuishan -0.01 (0.26) -0.14 (0.16) -0.08 (0.10) -0.10 (0.21) 102 Urumqi 1.56 (0.42) 0.50 (0.39) 0.33 (0.21) -0.56 (0.37) 103 Karamay 1.01 (0.42) -0.30 (0.38) -0.01 (0.21) 0.19 (0.39) Every city is matched with 4 cities out of the province based on the listed covariates. In this table, all the covariates have been normalized to have zero mean and unit variance in each year. Average differences of the covariates within each matched pair are shown (4 matched cities × 4 years = 16 pairs for each city) as well as the standard deviation of these differences.

111 Table 20. Housing rent equation with amenities, 2007

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual housing rent 14.158 1.16 7.601 21.311 (Annual housing rent) 3,725,404 29,087,408 2,000 180,000,0000 Housing characteristics Ownership 0.161*** (0.0305) 0.703 0.457 0 1 Single-unit&level -0.136*** (0.0299) 0.762 0.426 0 1 Size 0.000512* (0.000275) 78.171 118.796 3 8,000 Room 0.0556*** (0.0118) 5.482 2.814 1 99 Flooring 0.275*** (0.0357) 0.56 0.496 0 1 Wall 0.278*** (0.0290) 0.79 0.407 0 1 Electricity 0.231*** (0.0761) 0.964 0.186 0 1 Pipe water 0.140*** (0.0341) 0.271 0.445 0 1 Toilet 0.274*** (0.0329) 0.763 0.425 0 1 Garbage 0.358*** (0.0380) 0.335 0.472 0 1 Ventilation 0.0923** (0.0357) 0.845 0.361 0 1 Yard 0.0706*** (0.0242) 0.624 0.484 0 1 Filth -0.0874*** (0.0322) 0.183 0.386 0 1 Amenities/Disamenities Temperature -0.0802*** (0.0274) 26.638 1.438 20.896 28.753 Precipitation -0.0591 (0.0467) 2.472 0.719 1.246 4.459 PM2.5 0.0172* (0.00958) 11.582 5.15 2.324 22.036 Eruption -0.0449 (0.0830) 0.072 0.347 0 2 Flood 0.00737 (0.00593) 5.128 5.139 0 28 Forest 0.00352* (0.00205) 40.785 26.097 0.472 91.429 Urbanization 0.0103*** (0.00161) 24.33 36.535 0 100 Population -0.0168 (0.0612) 1.112 0.747 0.052 3.971 Primary schools -0.281* (0.168) 0.672 0.229 0.281 1.562 Doctors 0.0638 (0.247) 0.219 0.199 0.035 1.151 Hospitals -3.916 (3.420) 0.009 0.009 0 0.0574 Immunization -0.0154*** (0.00412) 95.738 4.321 68.491 100 Morbidity 0.00743** (0.00351) 31.373 7.175 14.673 46.542 Disequilibrium Variable Unemployment rate 0.131 (1.049) 0.096 0.044 0.0137 0.221

Observations 10,316 R-squared 0.366 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

112 Table 21. Wage equation with amenities, 2007

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual salary 15.382 1.41 2.303 20.752 (Annual salary) 10,373,382 21,181,208 10 1,029,249,984 Human capital characteristics Log of annual hours 0.609*** (0.0204) 7.323 0.906 0 10.146 Years of schooling 0.110*** (0.00468) 9.269 4.523 0 21 Experience 0.0474*** (0.00304) 18.949 13.449 0 78 Experience squared -0.0588*** (0.00505) 5.399 7.145 0 60.840 Ethnicity: Java 0.0119 (0.0380) 0.492 0.5 0 1 Language: Indonesian 0.0953** (0.0454) 0.302 0.459 0 1 Language: Javanese -0.00283 (0.0493) 0.434 0.496 0 1 Male 0.325*** (0.0313) 0.637 0.481 0 1 Married 0.156*** (0.0277) 0.721 0.449 0 1 Casual -0.423*** (0.0426) 0.224 0.417 0 1 Self-employed -1.426*** (0.0600) 0.057 0.231 0 1 Occupations Non-academic professionals -0.169** (0.0703) 0.026 0.159 0 1 Officials and managers 0.354** (0.158) 0.088 0.284 0 1 Clerks -0.0691 (0.0581) 0.003 0.053 0 1 Sales workers -0.352*** (0.0618) 0.074 0.261 0 1 Service workers -0.414*** (0.0660) 0.102 0.303 0 1 Agricultural workers -0.538*** (0.0782) 0.175 0.38 0 1 Production and related -0.395*** (0.0706) 0.189 0.391 0 1 workersCraft workers -0.251*** (0.0814) 0.095 0.293 0 1 Operators and assemblers -0.382*** (0.0710) 0.046 0.21 0 1 Military specialists 0.277*** (0.0851) 0.193 0.395 0 1 Students 0.284 (0.450) 0.007 0.081 0 1 Others 0.114 (0.143) 0.0004 0.019 0 1 Amenities/Disamenities Temperature -0.0295 (0.0264) 26.734 1.379 21.972 28.752 Precipitation -0.131*** (0.0425) 2.447 0.726 1.246 4.459 PM2.5 0.0150 (0.00939) 11.977 5.136 2.378 22.036 Eruption 0.0412 (0.0276) 0.06 0.318 0 2 Flood -0.000355 (0.00438) 5.235 5.2 0 28 Forest 0.000825 (0.00182) 38.441 25.964 0.472 91.429 Urbanization 0.000727 (0.00118) 28.384 38.563 0 100 Population 0.0517 (0.0336) 1.138 0.747 0.052 3.971 Primary schools -0.0520 (0.156) 0.638 0.218 0.281 1.518 Doctors 0.176 (0.143) 0.242 0.214 0.0350 1.151 Hospitals -0.570 (3.542) 0.01 0.009 0 0.0574 Immunization -0.0125** (0.00599) 96.207 3.768 68.491 100 Morbidity -0.00453 (0.00297) 31.414 7.161 14.673 46.542

113 Disequilibrium Variable Unemployment rate 2.350*** (0.726) 0.098 0.044 0.0137 0.221

Observations 8,426 R-squared 0.560 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

114 Table 22. Farm income equation with amenities, 2007

Main regression Coeff. Std. err. Mean Std. dev. Min Max Dependent variable Log of annual household farm 14.53 1.321 9.048 18.146 income(Annual household farm income) 4,383,950 6,693,436 8,500 76,000,000 Farm household characteristics Head: male 0.397*** (0.112) 0.889 0.314 0 1 Head: age 0.000695 (0.00191) 48.344 14.435 18 100 Head: married 0.221* (0.120) 0.88 0.325 0 1 Number of children 0.000547 (0.0348) 0.531 0.693 0 4 Number of low-educated young 0.0200 (0.0408) 0.47 0.688 0 5 malesNumber of low-educated young 0.0141 (0.0341) 0.486 0.697 0 4 femalesNumber of low-educated old males 0.0327 (0.0554) 0.585 0.556 0 3 Number of low-educated old females -0.00127 (0.0444) 0.706 0.594 0 3 Number of junior high school 0.0967*** (0.0324) 0.595 0.786 0 5 graduatesNumber of senior high school 0.241*** (0.0341) 0.511 0.8 0 5 graduatesNumber of university graduates 0.260*** (0.0600) 0.125 0.422 0 4 Size of farm land 6.81e-07 (4.84e-07) 8413.47 95455.699 5 5,000,000 Crop types Tuber -0.885*** (0.302) 0.037 0.189 0 1 Nuts and beans -1.228*** (0.220) 0.054 0.227 0 1 Crops -0.440*** (0.160) 0.015 0.12 0 1 Vegetables -0.202 (0.205) 0.655 0.476 0 1 Fruits -0.569** (0.225) 0.02 0.139 0 1 Spice -0.377** (0.158) 0.06 0.238 0 1 Rubber and wood 0.594*** (0.222) 0.066 0.249 0 1 Livestock -0.333 (0.249) 0.054 0.227 0 1 Amenities/Disamenities Temperature -0.0647 (0.0722) 26.279 1.552 21.971 28.609 Precipitation -0.459*** (0.150) 2.542 0.664 1.246 4.459 PM2.5 0.0418 (0.0292) 9.294 4.072 2.324 22.036 Eruption -0.0825 (0.104) 0.092 0.381 0 2 Flood -0.0189 (0.0138) 4.796 4.713 0 28 Forest 0.000311 (0.00610) 52.692 22.626 2.869 91.429 Urbanization -0.0136** (0.00527) 3.373 10.182 0 100 Population 0.0573 (0.117) 0.886 0.582 0.052 3.971 Primary schools 0.163 (0.328) 0.824 0.198 0.329 1.518 Doctors -0.933 (0.714) 0.125 0.092 0.0350 1.112 Hospitals 16.40* (9.258) 0.006 0.007 0 0.057 Immunization 0.0162 (0.0158) 94.683 5.301 68.491 100 Morbidity 0.0119 (0.00763) 32.1 7.289 16.507 46.542 Disequilibrium Variable Unemployment rate 3.993 (2.809) 0.073 0.031 0.0137 0.221

115 Observations 2,802 R-squared 0.156 Robust standard errors in parentheses (clustered by districts). The regression includes a constant. *** p<0.01, ** p<0.05, * p<0.1.

116 Table 23. Implicit prices of amenities, 2007

(1) (2) (3) (4) Amenities Housing rent differential Wage differential Farm income differential Full implicit price Temperature -0.0802*** -0.0295 -0.0647 344675 (0.0274) (0.0264) (0.0722) (513613) Precipitation -0.0591 -0.131*** -0.459*** 2886849*** (0.0467) (0.0425) (0.150) (839661) PM2.5 0.0172* 0.0150 0.0418 -276267 (0.00958) (0.00939) (0.0292) (183622) Eruption -0.0449 0.0412 -0.0825 -817003 (0.0830) (0.0276) (0.104) (618915) Flood 0.00737 -0.000355 -0.0189 61439 (0.00593) (0.00438) (0.0138) (86948) Forest 0.00352* 0.000825 0.000311 -2717 (0.00205) (0.00182) (0.00610) (35842) Urbanization 0.0103*** 0.000727 -0.0136** 44551* (0.00161) (0.00118) (0.00527) (24103) Population -0.0168 0.0517 0.0573 -1108516 (0.0612) (0.0336) (0.117) (688291) Primary schools -0.281* -0.0520 0.163 -313364 (0.168) (0.156) (0.328) (3006736) Doctors 0.0638 0.176 -0.933 -1703528 (0.247) (0.143) (0.714) (3002373) Hospitals -3.916 -0.570 16.40* -27634533 (3.420) (3.542) (9.258) (68571559) Immunization rate -0.0154*** -0.0125** 0.0162 151678 (0.00412) (0.00599) (0.0158) (115102) Morbidity rate 0.00743** -0.00453 0.0119 94933 (0.00351) (0.00297) (0.00763) (57980)

Observations 10,316 8,426 2,802 R-squared 0.366 0.560 0.156 Housing rent, wage and farm income differentials are taken from Tables 20, 21 and 22. The annual full implicit prices per household are evaluated at the means of housing rent, wage and farm income of the sample. 32.93% of households in the sample work for a farm business. For the remaining 67.07% of households, the average number of workers per household is 2.68. 1,000,000 Indonesian Rupiahs are equal to 72.80 US dollars. Robust standard errors in parentheses (clustered by districts). The standard errors on the full implicit prices are calculated from a linear combination of the standard errors in the housing rent, wage and farm income hedonic equations, assuming there is no covariance among the differentials. *** p<0.01, ** p<0.05, * p<0.1.

117

Ranking

SUMATRA

Jakarta

BALI JAVA

NUSA TENGGARA

Figure 7. Quality of life rankings for 192 districts in Indonesia, 2007

118 Table 24. Quality of life rankings for 192 districts in Indonesia, 2014 and 2007

Province Regency/City 2014 2007 Sumatera Barat Kepulauan Mentawai, Kab. 1 18 Jawa Timur Sampang, Kab. 2 180 Jawa Timur Pacitan, Kab. 3 39 Jawa Timur Blitar, Kab. 4 141 DI Yogyakarta Gunung Kidul, Kab. 5 47 Jawa Timur Sumenep, Kab. 6 179 Nusa Tenggara Barat Sumbawa Barat, Kab. 7 65 Jawa Timur Trenggalek, Kab. 8 105 Jawa Timur Pamekasan, Kab. 9 184 Jawa Timur Tulungagung, Kab. 10 85 Jawa Timur Bangkalan, Kab. 11 192 DI Yogyakarta Kulon Progo, Kab. 12 17 Jawa Timur Bondowoso, Kab. 13 124 Sumatera Barat Padang Pariaman, Kab. 14 31 Jawa Timur Probolinggo, Kab. 15 177 Jawa Timur Lumajang, Kab. 16 175 Nusa Tenggara Barat Bima, Kab. 17 86 Bali Karangasem, Kab. 18 162 Jawa Tengah Banjarnegara, Kab. 19 20 Jawa Tengah Wonogiri, Kab. 20 83 Jawa Tengah Purworejo, Kab. 21 63 Jawa Tengah Klaten, Kab. 22 174 Jawa Tengah Boyolali, Kab. 23 164 Bali Buleleng, Kab. 24 93 Jawa Timur Blitar, Kota 25 55 Jawa Tengah Kebumen, Kab. 26 7 Nusa Tenggara Barat Lombok Tengah, Kab. 27 125 Sumatera Utara Tapanuli Selatan, Kab. 28 112 Sumatera Utara Toba Samosir, Kab. 29 89 Sumatera Utara Tapanuli Tengah, Kab. 30 74 Jawa Timur Situbondo, Kab. 31 145 Jawa Timur Banyuwangi, Kab. 32 157 Jawa Tengah Wonosobo, Kab. 33 8 Sumatera Utara Mandailing Natal, Kab 34 71 Bali Jembrana, Kab. 35 81 Jawa Timur Lamongan, Kab. 36 176 Nusa Tenggara Barat Lombok Timur, Kab. 37 127 Jawa Timur Jombang, Kab. 38 121 Jawa Timur Ponorogo, Kab. 39 118 Bali Klungkung, Kab. 40 131 Jawa Tengah Kudus, Kab. 41 137 Nusa Tenggara Barat Dompu, Kab. 42 115

119 Province Regency/City 2014 2007 Jawa Tengah Jepara, Kab. 43 142 Jawa Tengah Tegal, Kota 44 90 Sumatera Barat Pariaman, Kota 45 2 Lampung Lampung Utara, Kab. 46 98 Nusa Tenggara Barat Sumbawa, Kab. 47 95 Bali Bangli, Kab. 48 136 Sumatera Selatan Lahat, Kab. 49 73 Sumatera Utara Sibolga, Kota 50 3 Bali Tabanan, Kab. 51 82 Sumatera Utara Karo, Kab. 52 114 Bali Badung, Kab. 53 97 Jawa Tengah Blora, Kab. 54 167 Jawa Tengah Brebes, Kab. 55 12 Nusa Tenggara Barat Lombok Barat, Kab. 56 126 Jawa Timur Ngawi, Kab. 57 159 DI Yogyakarta Yogyakarta, Kota 58 14 Jawa Tengah Cilacap, Kab. 59 13 Sumatera Utara Serdang Bedagai, Kab. 60 25 Jawa Barat Tasikmalaya, Kab. 61 54 Nusa Tenggara Barat Bima, Kota 62 160 Sumatera Utara Pakpak Bharat, Kab. 63 57 Jawa Tengah Kendal, Kab. 64 60 Jawa Timur Magetan, Kab. 65 146 Jawa Tengah Purbalingga, Kab. 66 4 Jawa Timur Gresik, Kab. 67 155 Jawa Timur Mojokerto, Kota 68 156 Sumatera Barat Limapuluh Kota, Kab 69 15 Jawa Timur Tuban, Kab. 70 190 Jawa Tengah Magelang, Kab. 71 153 Jawa Tengah Grobogan, Kab. 72 120 Sumatera Utara Tapanuli Utara, Kab. 73 147 Sumatera Barat Tanah Datar, Kab. 74 49 Jawa Tengah Rembang, Kab. 75 158 Sumatera Barat Solok Selatan, Kab. 76 72 Jawa Timur Kediri, Kab. 77 165 DI Yogyakarta Sleman, Kab. 78 144 Jawa Timur Malang, Kota 79 46 DI Yogyakarta Bantul, Kab. 80 34 Jawa Timur Malang, Kab. 81 151 Sumatera Selatan Ogan Ilir, Kab. 82 45 Jawa Barat Cirebon, Kota 83 16 Bali Gianyar, Kab. 84 123 Sumatera Barat Pesisir Selatan, Kab. 85 61 Lampung Way Kanan, Kab. 86 100

120 Province Regency/City 2014 2007 Sumatera Barat Pasaman Barat, Kab. 87 29 Lampung Tanggamus, Kab. 88 66 Lampung Lampung Tengah, Kab. 89 67 Jawa Timur Madiun, Kab. 90 148 Sumatera Selatan Ogan Komering Ulu Selatan, Kab. 91 132 Sumatera Utara Humbang Hasundutan, Kab. 92 116 Jawa Timur Batu, Kota 93 62 Nusa Tenggara Barat Mataram, Kota 94 38 Jawa Tengah Sragen, Kab. 95 138 Sumatera Selatan Ogan Komering Ilir, Kab. 96 107 Jawa Tengah Pati, Kab. 97 170 Lampung Tulang Bawang, Kab. 98 22 Sumatera Selatan Ogan Komering Ulu Timur, Kab. 99 111 Jawa Timur Mojokerto, Kab. 100 130 Lampung Lampung Barat, Kab. 101 122 Jawa Tengah Batang, Kab. 102 78 Jawa Tengah Banyumas, Kab. 103 6 Jawa Tengah Karanganyar, Kab. 104 51 Lampung Lampung Timur, Kab. 105 50 Jawa Timur Nganjuk, Kab. 106 96 Jawa Timur Pasuruan, Kab. 107 163 Jawa Timur Probolinggo, Kota 108 189 Sumatera Utara Asahan, Kab. 109 84 Jawa Timur Bojonegoro, Kab. 110 169 Jawa Barat Banjar, Kota 111 35 Jawa Tengah Semarang, Kab. 112 91 Sumatera Barat Pasaman, Kab 113 92 Jawa Tengah Salatiga, Kota 114 87 Sumatera Barat Solok, Kab. 115 40 Sumatera Barat Agam, Kab. 116 21 Lampung Lampung Selatan, Kab. 117 129 Sumatera Barat Sijunjung, Kab. 118 75 Sumatera Utara Samosir, Kab. 119 172 Sumatera Selatan Musi Rawas, Kab. 120 70 Jawa Timur Jember, Kab. 121 161 Sumatera Selatan Ogan Komering Ulu, Kab. 122 52 Sumatera Utara Simalungun, Kab. 123 109 Jawa Tengah Temanggung, Kab. 124 43 Sumatera Barat Sawahlunto, Kota 125 36 Jawa Barat Ciamis, Kab. 126 9 Bali Denpasar, Kota 127 134 Jawa Timur Kediri, Kota 128 171 Sumatera Utara Padang Sidempuan, Kota 129 113 Jawa Tengah Sukoharjo, Kab. 130 117

121 Province Regency/City 2014 2007 Jawa Tengah Pekalongan, Kota 131 53 Sumatera Utara Labuhan Batu, Kab. 132 128 Jawa Barat Kuningan, Kab. 133 11 Sumatera Utara Dairi, Kab. 134 102 Jawa Tengah Pekalongan, Kab. 135 32 Lampung Bandar Lampung, Kota 136 28 Lampung Metro, Kota 137 48 Jawa Tengah Demak, Kab. 138 143 Sumatera Selatan Banyuasin, Kab. 139 108 Jawa Tengah Tegal, Kab. 140 24 Sumatera Selatan Muara Enim, Kab. 141 110 Jawa Timur Madiun, Kota 142 188 Sumatera Utara Langkat, Kab. 143 42 Sumatera Selatan Prabumulih, Kota 144 64 Sumatera Barat Dharmasraya, Kab. 145 19 Jawa Barat Sukabumi, Kota 146 5 Jawa Tengah Pemalang, Kab. 147 58 Jawa Timur Sidoarjo, Kab. 148 166 Jawa Barat Tasikmalaya, Kota 149 10 Sumatera Selatan Musi Banyuasin, Kab. 150 77 Jawa Tengah Surakarta, Kota 151 59 Jawa Barat Majalengka, Kab. 152 56 Sumatera Barat Solok, Kota 153 26 Sumatera Utara Tanjung Balai, Kota 154 183 Sumatera Selatan Pagar Alam, Kota 155 149 Jawa Barat Cianjur, Kab. 156 80 Jawa Barat Sukabumi, Kab. 157 44 Sumatera Barat Padang, Kota 158 23 Sumatera Selatan Palembang, Kota 159 69 Jawa Tengah Magelang, Kota 160 88 Jawa Barat Garut, Kab. 161 173 Jawa Barat Depok, Kota 162 33 DKI Jakarta Jakarta Pusat, Kota 163 103 Jawa Barat Bogor, Kota 164 1 Sumatera Selatan Lubuklinggau, Kota 165 152 Sumatera Barat Payakumbuh, Kota 166 30 Jawa Tengah Semarang, Kota 167 140 Jawa Timur Surabaya, Kota 168 182 Jawa Barat Purwakarta, Kab. 169 41 Jawa Barat Sumedang, Kab. 170 37 DKI Jakarta Jakarta Utara, Kota 171 99 Jawa Timur Pasuruan, Kota 172 191 Jawa Barat Cimahi, Kota 173 27 Sumatera Utara Deli Serdang, Kab. 174 150

122 Province Regency/City 2014 2007 Jawa Barat Cirebon, Kab. 175 133 DKI Jakarta Jakarta Selatan, Kota 176 68 Sumatera Barat Padang Panjang, Kota 177 106 Jawa Barat Subang, Kab. 178 139 Jawa Barat Bekasi, Kota 179 104 DKI Jakarta Jakarta Barat, Kota 180 135 DKI Jakarta Jakarta Timur, Kota 181 76 Jawa Barat Bandung, Kota 182 101 Jawa Barat Indramayu, Kab. 183 186 Sumatera Utara Tebing Tinggi, Kota 184 178 Sumatera Barat Bukittinggi, Kota 185 154 Jawa Barat Bekasi, Kab. 186 168 Sumatera Utara Binjai, Kota 187 185 Jawa Barat Karawang, Kab. 188 181 Jawa Barat Bandung, Kab. 189 119 Sumatera Utara Pematang Siantar, Kota 190 79 Jawa Barat Bogor, Kab. 191 94 Sumatera Utara Medan, Kota 192 187

123 Table 25. Results from Probit Model Probability of Visitation (1) (2) Variables Estimates Mean Marginal Effects Annual deforestation rate -0.517*** -0.116*** (0.0690) (0.0148) Municipality size 0.0517*** 0.0116*** (0.0162) (0.00363) Forest area in 2000 0.118* 0.0264* (0.0671) (0.0149) Edge to core ratio 0.000938 0.000210 (0.00225) (0.000506) Biomass density 0.00185 0.000414 (0.00122) (0.000273) Threatened bird species 0.121*** 0.0271*** (0.0353) (0.00795) Temperature in January 0.0127 0.00284 (0.0117) (0.00258) Temperature in July -0.0145 -0.00324 (0.0120) (0.00266) Precipitation 0.00452 0.00101 (0.0585) (0.0131) GDP 0.173** 0.0387** (0.0840) (0.0189) Population -0.717 -0.161 (1.185) (0.266) Distance to the nearest urban center 0.0300 0.00672 (0.0591) (0.0132) Distance to the nearest airport -0.664*** -0.149*** (0.158) (0.0346) Length of highway 0.214*** 0.0479*** (0.0642) (0.0137) Length of railway 0.266** 0.0596** (0.116) (0.0264) Archaeological site 0.300** 0.0673** (0.143) (0.0317) Protected area 0.489*** 0.110*** (0.0919) (0.0193)

Observations 16,254 16,254 Method Probit Probit State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

124 Table 26. Estimates from Poisson and NB Models (1) (2) Variables Visit counts Visit counts Annual deforestation rate -0.488** -1.006*** (0.231) (0.259) Municipality size -0.0107 0.131*** (0.0343) (0.0299) Forest area in 2000 0.00680 0.0940 (0.0582) (0.155) Edge to core ratio 0.00110 0.0129 (0.0237) (0.0105) Biomass density 0.0112*** 0.00139 (0.00322) (0.00299) Threatened bird species 0.162** 0.350*** (0.0793) (0.0862) Temperature in January 0.0534* 0.0319 (0.0276) (0.0256) Temperature in July -0.0627*** -0.0578* (0.0222) (0.0299) Precipitation -0.293 -0.112 (0.219) (0.170) GDP 0.266*** 0.105 (0.0426) (0.189) Population -1.547** 1.142 (0.636) (2.905) Distance to the nearest urban center 0.469** 0.346** (0.209) (0.154) Distance to the nearest airport -1.722*** -1.883*** (0.652) (0.341) Length of highway 0.187*** 0.554*** (0.0716) (0.151) Length of railway 0.0615 0.276 (0.255) (0.392) Archaeological site 0.777** 0.298 (0.328) (0.240) Protected area 0.708*** 0.786*** (0.187) (0.193) ln(α) 1.985*** (0.104) AIC 168230.1 36271.28 Observations 16,290 16,290 Method Poisson NB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

125 Table 27. Incidence Rate Ratio from Poisson and NB Models (1) (2) Variables Visit counts Visit counts Annual deforestation rate 0.614** 0.366*** (0.142) (0.0949) Municipality size 0.989 1.140*** (0.0339) (0.0341) Forest area in 2000 1.007 1.099 (0.0585) (0.171) Edge to core ratio 1.001 1.013 (0.0237) (0.0106) Biomass density 1.011*** 1.001 (0.00326) (0.00299) Threatened bird species 1.176** 1.419*** (0.0932) (0.122) Temperature in January 1.055* 1.032 (0.0291) (0.0264) Temperature in July 0.939*** 0.944* (0.0208) (0.0282) Precipitation 0.746 0.894 (0.164) (0.152) GDP 1.305*** 1.111 (0.0556) (0.210) Population 0.213** 3.134 (0.135) (9.104) Distance to the nearest urban center 1.599** 1.413** (0.334) (0.217) Distance to the nearest airport 0.179*** 0.152*** (0.116) (0.0519) Length of highway 1.205*** 1.741*** (0.0863) (0.263) Length of railway 1.063 1.318 (0.271) (0.517) Archaeological site 2.175** 1.347 (0.714) (0.324) Protected area 2.031*** 2.195*** (0.381) (0.424)

Observations 16,290 16,290 Method Poisson NB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

126 Table 28. Results from the Logit Model, with Deforestation Lagged by One Year Probability of Visitation (1) (2) Variables Estimates Mean Marginal Effects Annual deforestation rate, lagged by one year -0.944*** -0.123*** (0.161) (0.0203) Municipality size 0.0925*** 0.0121*** (0.0359) (0.00464) Forest area in 2000 0.222* 0.0290* (0.120) (0.0155) Edge to core ratio 0.00156 0.000203 (0.00465) (0.000608) Biomass density 0.00294 0.000383 (0.00203) (0.000267) Threatened bird species 0.198*** 0.0259*** (0.0620) (0.00809) Temperature in January 0.0103 0.00134 (0.0189) (0.00245) Temperature in July -0.0172 -0.00224 (0.0203) (0.00262) Precipitation 0.0724 0.00945 (0.119) (0.0155) GDP 0.401*** 0.0524*** (0.137) (0.0181) Population -2.034 -0.266 (2.108) (0.277) Distance to the nearest urban center 0.0388 0.00506 (0.105) (0.0137) Distance to the nearest airport -1.165*** -0.152*** (0.278) (0.0355) Length of highway 0.441*** 0.0575*** (0.103) (0.0126) Length of railway 0.403* 0.0526* (0.231) (0.0306) Archaeological site 0.545** 0.0711** (0.233) (0.0299) Protected area 0.860*** 0.112*** (0.159) (0.0190)

Observations 14,448 14,448 Method Logit Logit State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

127 Table 29. Estimates from ZINB Model, with Deforestation Lagged by One Year (1) (2) Variables Constant zero Visit counts Annual deforestation rate, lagged by one year 1.188*** -0.417* (0.282) (0.228) Municipality size -0.219*** 0.0388* (0.0518) (0.0219) Forest area in 2000 -0.694** 0.0774* (0.297) (0.0442) Edge to core ratio 0.0120 0.0336 (0.0116) (0.0321) Biomass density -0.000353 0.00398 (0.00312) (0.00274) Threatened bird species -0.216* 0.229*** (0.126) (0.0703) Temperature in January -0.0187 0.0104 (0.0320) (0.0273) Temperature in July 0.0113 -0.0394 (0.0318) (0.0270) Precipitation -0.412** -0.310** (0.210) (0.149) GDP -1.606 0.218** (1.924) (0.101) Population 7.411 -1.080 (19.18) (1.540) Distance to the nearest urban center 0.478*** 0.491*** (0.160) (0.142) Distance to the nearest airport 0.848** -1.236*** (0.387) (0.295) Length of highway -1.545*** 0.0992* (0.298) (0.0597) Length of railway 0.000682 0.280 (0.533) (0.200) Archaeological site -0.766 0.279 (0.678) (0.306) Protected area -1.080*** 0.364** (0.286) (0.152) ln(α) 1.165*** (0.0995) AIC 32627.09 Observations 14,480 14,480 Method ZINB ZINB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

128 Table 30. Incidence Rate Ratio from ZINB Model, with Deforestation Lagged by One Year (1) (2) Variables Constant zero Visit counts Annual deforestation rate, lagged by one year 3.279*** 0.659* (0.926) (0.150) Municipality size 0.803*** 1.040* (0.0416) (0.0227) Forest area in 2000 0.499** 1.080* (0.148) (0.0478) Edge to core ratio 1.012 1.034 (0.0117) (0.0332) Biomass density 1.000 1.004 (0.00312) (0.00275) Threatened bird species 0.806* 1.257*** (0.102) (0.0884) Temperature in January 0.982 1.010 (0.0314) (0.0275) Temperature in July 1.011 0.961 (0.0321) (0.0260) Precipitation 0.662** 0.733** (0.139) (0.109) GDP 0.201 1.244** (0.386) (0.126) Population 1,654 0.340 (31,717) (0.523) Distance to the nearest urban center 1.613*** 1.634*** (0.258) (0.232) Distance to the nearest airport 2.334** 0.291*** (0.904) (0.0856) Length of highway 0.213*** 1.104* (0.0635) (0.0659) Length of railway 1.001 1.323 (0.533) (0.264) Archaeological site 0.465 1.322 (0.315) (0.405) Protected area 0.340*** 1.439** (0.0972) (0.219)

Observations 14,480 14,480 Method ZINB ZINB State FE (32) Yes Yes Year FE (9) Yes Yes All regressions include a constant. Robust standard errors in parentheses (clustered by states). *** p<0.01, ** p<0.05, * p<0.1.

129