Health Insurance in Rural Cambodia: Impacts and Selection by Rachel A

Health Insurance in Rural Cambodia: Impacts and Selection

Rachel A. Polimeni

A dissertation submitted in partial satisfaction of the requirements for the degree of Doctor of Philosophy

Economics

in the

GRADUATE DIVISION of the UNIVERSITY OF CALIFORNIA, BERKELEY

Committee in charge: Professor Edward Miguel, Chair Professor David Levine Professor Ronald Lee

Fall 2011 Health Insurance in Rural Cambodia: Impacts and Selection

Rachel A. Polimeni 1

Abstract

Health Insurance in Rural Cambodia: Impacts and Selection

Rachel A. Polimeni

Doctor of Philosophy in Economics

University of California, Berkeley

Professor Edward Miguel, Chair

High health care expenditures following a health shock can lead to long-term economic consequences. Health insurance has the potential to avert economic di¢ culties following health shocks, increase health care utilization and improve health. However, adverse selection in health insurance markets may stop voluntary health insurance markets from providing protection to most consumers without substantial regulation and subsidization. If uninsured individuals forgo valuable health care due to lack of funds, health insurance can also increase health care utilization and improve health. These potential bene…ts of insurance have led many developing nations to consider health insurance as a policy tool. Yet, even in developed nations, there have been few studies to measure its e¤ectiveness. This dissertation consists of three chapters that evaluate the SKY Micro-health insurance program in rural Cambodia. In Chapter 1 I evaluate the health and economic e¤ects of the SKY insurance program on rural households using a randomized controlled trial. By randomizing the insurance premium we induce random variation in the likelihood of insurance take-up that allows us to estimate the causal e¤ects of health insurance on economic outcomes, health utilization, and health outcomes. We …nd that SKY insurance has the greatest impact on economic outcomes, as expected from an insurance program. For example, SKY decreased total health-care costs of serious health shocks by over 40%, and households with SKY had over one-third less debt and over 75% less health-related debt. SKY also changed health-seeking behavior, increasing use of (covered) public facilities and decreasing use of (uncovered) unregulated care. At the same time, SKY had no detectable impact on preventative care. As expected due to low statistical power, we did not …nd statistically signi…cant impacts on health. In Chapter 2 I study adverse selection into this insurance market. As part of this study I use the randomized experimental design to separate adverse selection from moral hazard. I test three implications of theories of adverse selection: that households joining are 2 more adversely selected based on characteristics observable at the baseline; that households that purchase insurance at a high price are more adversely selected on observables than those that purchase identical coverage at a lower price; and that households that purchase at the higher price will demonstrate more adverse selection in utilization than households purchasing coverage at a lower price even after holding constant baseline characteristics (“unobservable”selection). I …nd that households that purchase insurance have some characteristics consistent with higher expected health care utilization. Contrary to expectations, households paying a higher price do not demonstrate more selection on characteristics observable prior to insurance purchase. However, households that paid more for health insurance have substantially higher usage of both health centers and hospitals than households that received a discounted price, even when comparing households with similar observed baseline health. This result is consistent with substantial adverse selection based on factors we did not observe prior to insurance purchase. In Chapter 3 I go beyond adverse selection to examine several other factors that may be in‡uential in the purchase of SKY insurance. As insurance is a consumption- smoothing tool, risk-averse households may be more willing to purchase insurance. House- holds that can self-insure may be less likely to purchase insurance. Newer theories have hypothesized that budget constraints, present bias, or having little understanding of insurance may decrease the likelihood of buying insurance even for sick households. Age or gender bias may play into the decision, as may trust of Western medicine. These and other less-traditional type of selection factors may be particularly relevant in a developing country. Contrary to informational models, we …nd no evidence that risk averse households are more likely to purchase SKY, and instead …nd evidence of the opposite. Budget constraints, quality of health facilities, and age and gender of ill household members also in‡uence the decision to purchase insurance. i

For Allen ii

Contents

List of Figures v

List of Tables vi

1 Insuring Health or Insuring Wealth? An Experimental Evaluation of Health Insurance in Rural Cambodia 1 1.1 Introduction ...... 1 1.2 Previous Research ...... 2 1.3 The Setting ...... 5 1.3.1 Health care in Cambodia ...... 5 1.3.2 SKY Health Insurance ...... 5 1.4 Theory and Measurement ...... 6 1.4.1 Health seeking behavior ...... 6 1.4.2 Economic impacts ...... 7 1.4.3 Health Outcomes ...... 8 1.4.4 Trust in Providers and SKY ...... 9 1.5 Data and methodology ...... 9 1.5.1 Randomization of prices ...... 9 1.5.2 Estimation ...... 10 1.5.3 Data ...... 12 1.6 Results ...... 13 1.6.1 Tests of Experimental Design ...... 13 1.6.2 Summary statistics ...... 14 1.6.3 First Stage ...... 14 1.6.4 Health Seeking Behavior ...... 15 1.6.5 Economic E¤ects of Insurance ...... 17 1.6.6 Health Outcomes ...... 19 1.6.7 Trust in Providers and SKY ...... 19 1.7 Robustness Checks ...... 19 1.8 Conclusion ...... 20 1.9 Tables ...... 23 1.10 Figures ...... 34 1.A Supplementary Tables ...... 37 1.B Instrumental Variable Results Using Coupon as Instrument ...... 45 iii

2 Adverse Selection based on Observable and Unobservable Factors in Health Insurance 58 2.1 Introduction ...... 58 2.2 Previous Research ...... 60 2.3 Theory and Methods ...... 62 2.3.1 Selection on Observables ...... 62 2.3.2 Selection on Observables at High versus Low Price ...... 63 2.3.3 Selection on Unobservables ...... 63 2.4 The Setting ...... 64 2.4.1 Health Care in Cambodia ...... 64 2.4.2 SKY Health Insurance ...... 65 2.5 Randomization ...... 66 2.6 Data ...... 66 2.6.1 Household Survey ...... 66 2.6.2 SKY Administrative and Utilization Data ...... 67 2.6.3 Other Datasets ...... 67 2.6.4 Randomization ...... 67 2.7 Results ...... 68 2.7.1 Selection on Observables ...... 68 2.7.2 Selection on Observables by Price ...... 69 2.7.3 Selection on Unobservables ...... 69 2.7.4 Drop-out ...... 70 2.8 Robustness Checks ...... 71 2.8.1 Early versus Late Buyers ...... 71 2.8.2 Selection on Unobservables ...... 72 2.8.3 Behavioral Moral Hazard ...... 72 2.8.4 Hazard Rates by Price ...... 73 2.9 Financial Implications of Adverse Selection ...... 73 2.10 Conclusion ...... 74 2.11 Tables ...... 77 2.12 Figures ...... 84 2.A Supplementary Tables ...... 88 2.B Other Datasets ...... 103 2.B.1 Village Leader Interview ...... 103 2.B.2 Health Center Data Collection ...... 103 2.B.3 Village Meeting Data ...... 103 2.C Lucky Draw Implementation ...... 103 2.D Description of Variables, Adverse Selection ...... 104 2.D.1 Independent Variables ...... 105 2.D.2 Basic Covariates ...... 107 iv

3 Going Beyond Adverse Selection: Take-up of a Health Insurance Program in Rural Cambodia 110 3.1 Introduction ...... 110 3.2 The Setting ...... 111 3.2.1 Providers ...... 111 3.3 Literature Review ...... 112 3.3.1 Traditional Insurance Theory ...... 112 3.3.2 Recent Theory ...... 112 3.3.3 Developing Country Context ...... 113 3.3.4 Empirical Literature ...... 114 3.4 Speci…cation ...... 116 3.5 Data ...... 118 3.5.1 Household Survey ...... 118 3.5.2 SKY Administrative Data ...... 119 3.5.3 Village Leader Survey ...... 119 3.5.4 Health Center Survey ...... 119 3.5.5 Village Meeting Survey ...... 119 3.6 Background Results ...... 119 3.6.1 Summary Statistics ...... 119 3.6.2 Characteristics of Ill Members ...... 120 3.6.3 Qualitative Survey Responses ...... 121 3.7 Regression Results ...... 122 3.7.1 Traditional In‡uences on Take-up ...... 123 3.7.2 Other In‡uences on Take-up ...... 125 3.8 Robustness Tests ...... 126 3.8.1 Interview Lag and Delayed SKY Purchase ...... 127 3.8.2 Village Controls ...... 127 3.8.3 Wealth Interactions ...... 128 3.8.4 Coupon Status ...... 128 3.9 Conclusion ...... 129 3.10 Tables ...... 132 3.A Supplementary Tables ...... 141 3.B Theoretical Model ...... 144 3.C Description of Variables ...... 147

Bibliography 154 v

List of Figures

1.1 Timeline of Evaluation ...... 35 1.2 Proportion in SKY, by Months since Village Meeting and Coupon Type . . 36

2.1 E¤ect of Full Price (not steep discount) on Utilization, with and without baseline controls ...... 85 2.2 Proportion of SKY Members Using SKY-Covered Health Facilities, by Tenure inSKY ...... 86 2.3 Proportion of Households using SKY-covered Health Facilities for Care, by Premium and Tenure with SKY ...... 87 vi

List of Tables

1.1 Randomization Test ...... 24 1.2 First Stage Regression for Incident-level Outcomes, Round 1 and 2 Incidents Used...... 25 1.3 Health Utilization Following a Major Health Shock ...... 26 1.4 Provider Type, First Treatment after Major Health Incident ...... 27 1.5 Birth-Related Utilization ...... 28 1.6 Economic Impacts Following a Major Health Incident ...... 29 1.7 Method of Payment following a Major Health Incident ...... 30 1.8 Overall Economic Impacts on Households ...... 31 1.9 Health Impacts ...... 32 1.10 Trust in Providers and SKY ...... 33 1.11 Health Utilization after Major Health Incident - Ever Treated at Given Provider type ...... 38 1.12 General Health Utilization ...... 39 1.13 Overall Economic Impacts, Households with Health Incidents ...... 40 1.14 Instrumental Variables Regressions holding constant Round 1 Values . . . . 41 1.15 First Stage Regression for Individual-level Outcomes, Round 2 Data Used . 42 1.16 First Stage Regression for Household-Level Outcomes, Round 2 Data Used 43 1.17 First Stage Regression for Birth-Level Outcomes, Rounds 1 and 2 Data Used 44 1.18 IV using Coupon as Instrument: First Stage Regression for Incident-Level Outcomes, Rounds 1 and 2 Data Used ...... 46 1.19 IV using Coupon as Instrument: First Stage Regression for Individual-Level Outcomes, Round 2 Data Used ...... 47 1.20 IV Using Coupon as Instrument: First Stage Regression for Household-Level Outcomes, Round 2 Data Used ...... 48 1.21 IV using Coupon as Instrument: First Stage for Birth-Level Regressions, using R1 and R2 data ...... 49 1.22 IV Using Coupon as Instrument: Health Care Utilization following a Health Shock ...... 50 1.23 IV using Coupon as Instrument: Provider Type, First Treatment after a Major Health Incident ...... 51 1.24 IV using Coupon as Instrument: Birth-Related Outcomes ...... 52 1.25 IV using Coupon as Instrument: Economic Impacts following a Major Health Shock ...... 53 vii

1.26 IV using Coupon as Instrument: Method of Payment following a Major Health Incident ...... 54 1.27 IV using Coupon as Instrument: Overall Economic Impacts on Households 55 1.28 IV using Coupon as Instrument: Health Impacts ...... 56 1.29 IV using Coupon as Instrument: Trust in Providers and SKY ...... 57

2.1 Probit Regression of SKY Take-up on Baseline Characteristics ...... 78 2.2 Probit Regression of SKY Take-up on Baseline Characteristics Interacted with Price ...... 79 2.3 Summary Statistics, Buyers at Full versus Discounted Price ...... 80 2.4 E¤ects of Self-Selection on Utilization ...... 81 2.5 Cox Regression, Hazard of Dropping Coverage ...... 82 2.6 Financial Implications of Selection ...... 83 2.7 Research Sample ...... 89 2.8 Randomization Summary Statistics ...... 90 2.9 Summary Statistics for Selection on Observable Characteristics ...... 91 2.10 Summary Statistics for Early versus Late Buyers ...... 92 2.11 Autocorrelation of Health Expenses ...... 93 2.12 Robustness Check: Observable Selection using Full Sample (Early and Late Buyers) ...... 94 2.13 Robustness Check: Observable Selection using only Health Shocks lasting 7 or more days ...... 95 2.14 Robustness Check: Unobservable Selection using All Observable Covariates 96 2.15 Robustness Check: Unobservable Selection using All Covariates, and Indica- tor for Early Buyer ...... 97 2.16 Robustness Check: Unobservable Selection Controlling for Health Shocks 1-3 Months pre-Baseline ...... 98 2.17 Robustness Check: Unobservable Selection, No Oversampled Households . . 99 2.18 Robustness Check: Unobservable Selection, Tobit regression for Costs . . . 100 2.19 Robustness Check: Unobservable Selection, Only Shocks lasting more than 7days ...... 101 2.20 Robustness Check: Unobservable Selection, Inpatient Visits ...... 102 2.21 Independent Variables, Chapter 2 ...... 106 2.22 Basic Covariates used in Chapter 2 ...... 109

3.1 Treatment Behavior of Ill Households ...... 132 3.2 Summary of Hypotheses ...... 133 3.3 Summary Statistics (Traditional In‡uences on Take-up) ...... 134 3.4 Summary Statistics (Other Measures) ...... 135 3.5 Health and Utilization Regressed on Characteristics of Members ...... 136 3.6 Price Elasticity of Demand ...... 137 3.7 In‡uence of Traditional Selection Measures (Household Demographics, Clinic Characteristics, and Risk Characteristics) on SKY Purchase ...... 138 3.8 In‡uence of Self-Insurance Measures on SKY Purchase ...... 139 3.9 In‡uence of Other Selection Measures on SKY Purchase ...... 140 viii

3.10 SKY Purchase, by Wealth and Poor Health ...... 141 3.11 Robustness Checks ...... 142 3.12 Wealth/Health Interaction ...... 143 3.13 Baseline Survey Variables ...... 151 3.14 Village Leader Survey variables ...... 152 3.15 Health Center Survey variables ...... 153 3.16 Village Meeting Variables ...... 153 ix

Acknowledgments

I would like to thank my advisor, Edward Miguel, and my other committee members, Ronald Lee, and David Levine, for their support and advice over the course of the project. Their suggestions and guidance have made these papers in…nitely better. Thank you to my co-author David Levine for treating me as an equal despite having many more years of experience, and for giving me the opportunity to work on this project. Thank you to participants at the Berkeley Development Lunch, the Demography Brown Bag, the CERDI conference, USAID BASIS meetings, and other seminars, for their thoughtful comments, many of which have been incorporated into these papers. Special thanks to my colleagues and friends at U.C. Berkeley and Stanford, who, in addition to giving me useful research advice, also provided a supportive environment that kept me going through the years. Thank you to Patrick Allen for having endless patience for my last minute requests and for helping me to meet tight administrative deadlines; he could not have been a better graduate advisor. Thank you to AFD, USAID, and the Coleman Fung Foundation for their generous funding. Cooperation from GRET and SKY were essential in implementing this study. Thank you to the sta¤ at GRET for sharing their data and the …eld team at Domrei for their tireless data collection and cleaning. Jean-David Naudet and Jocelyne Delarue of AFD gave enormous guidance in the course of the evaluation and valuable feedback on these papers. Rachel Gardner and Francine Anene provided excellent research assistance. In addition, Rachel Gardner played an invaluable on-site role in the design of the evaluation, living in Cambodia for several months to observe village meetings and communicate with stakeholders, and contributed immensely to the project as a whole. Raj Arunachalam was an essential part of the early stages of this evaluation, contributing to the grant proposals, the evaluation design, and the creation of survey instruments. 1

Chapter 1

Insuring Health or Insuring Wealth? An Experimental Evaluation of Health Insurance in Rural Cambodia

With David Levine and Ian Ramage1

1.1 Introduction

Serious injuries and illnesses typically both increase medical expenses and reduce a family’shousehold income and home production (Wagsta¤ and Van Doorslaer 2003; Gertler, Levine, and Moretti 2003; Gertler and Gruber, 2002). “Each year, approximately 150 million people experience …nancial catastrophe, meaning they are obliged to spend on health care more than 40% of the income available to them after meeting their basic needs”(World Health Organization 2007). Poor households often forego high-value care, yet still often pay substantial sums for care of low quality (Das, Hammer, and Leonard 2008). High health care expenditures mean a short-term health shock can lead to debt, asset sales, and removal of children from school –creating long-term increases in poverty (Van Damme, Van Leemput, Por, Hardeman, and Meessen 2004; Annear 2006). Health insurance is designed to reduce economic di¢ culties following illness or injury. However, in developing countries few companies market health insurance to poor households (Sekhri and Savedo¤ 2005; Pauly, Zweifel, Sche• er, Preker, and Bassett 2006). Insurance companies do not target poor consumers for many reasons, ranging from their inconsistent incomes, which may lead to missed premium payments, to the relatively high transaction costs of servicing an inexpensive insurance policy. These problems are similar to those faced by the credit industry in developing countries, which led to the creation of micro-…nance. Micro-health-insurance agencies have followed the lead of micro-…nance and

1David Levine is a professor at Haas School of Business, University of California, and Ian Ramage is a Director of Domrei Research and Consulting in Phnom Pehn, Cambodia. 2 have started to o¤er insurance to this previously unserved population. Health insurance may also increase access to health care and, thus, improve health outcomes, especially if it reaches a poor population. The success of a micro-insurance program depends on its ability to improve economic and other outcomes while maintaining …nancial sustainability, or at the least assuring donors that their money is being spent in the most e¢ cient way possible. However, because health insurance is a relatively new product in developing countries, little is known about how best to design an insurance program to meet the needs of the poor. Unfortunately, rigorous evidence on the impact of insurance is scarce, and there are even fewer studies on the e¤ects of insurance in developing countries. One reason for the lack of evidence is that it is di¢ cult to …nd a valid group to compare with the insured. We cannot simply compare the outcomes of insured and uninsured households because health insurance status is typically strongly correlated with other household characteristics. For example, rich and well educated households typically have both better health (Asfaw 2003) and better health insurance coverage (Jutting 2004; Cameron and Trivedi 1991). Importantly, that correlation does not mean insurance improves health. At the same time, those in poor health may be more likely to purchase health insurance when it is o¤ered (Cutler and Reber 1998; Ellis 1989), but that correlation does not mean insurance worsens health. We evaluate the health and economic e¤ects of the SKY Micro-health insurance program on households in rural Cambodia using a randomized controlled trial. By randomizing the insurance premium we induce random variation in the likelihood of insurance take-up that allows us to estimate the causal e¤ects of health insurance on three main categories of outcome: Health care utilization, such as timely utilization of curative care and substitution to public facilities from private health centers and traditional medicine; Eco- nomic outcomes, such as out-of-pocket medical spending and new debt to pay for health care; and Health outcomes, such as frequency of major health shocks and stunting and wasting. We also investigate SKY’s impact on other outcomes such as opinion of public facilities and trust of the SKY program. SKY has the greatest impacts on economic outcomes, as expected from an insurance program. For example, SKY decreased total health-care costs of serious health shocks by over 40%, and households with SKY had over one-third less debt and over 75% less health-related debt. SKY also changed health-seeking behavior, increasing use of public facilities and decreasing use of unregulated care. At the same time, SKY had no detectable impact on preventative care. We did not …nd statistically signi…cant impacts on health, but the short time horizon of the study and the smaller sample size for these outcomes meant that, a priori, we did not expect to have su¢ cient statistical power to measure health impacts.

1.2 Previous Research

For the reasons noted above, rigorous evidence of the impacts of health insurance is rare. The small number of studies using randomization or natural experiments to establish causality typically …nd that health insurance increases health care utilization; in some cases 3 increased utilization also leads to detectable improvements in health.2 The RAND Health Insurance Experiment (from 1974 to 1982) in the United States is the only large-scale randomized experiment examining the e¤ects of health insurance on health and health care utilization to date. This experiment studied almost 4000 people in 2000 families. Some families were randomly assigned to a free care plan while others were assigned one of several plans that required varying co-payments. The study found that those assigned to a cost-sharing plan sought less treatment than those with full coverage (e.g. Lohr et al., 1986; Manning 1987). Forgone treatment for those with cost-sharing was primarily for preventive visits to doctors and “elective”care such as mental health treatment as opposed to emergency care (e.g., Keeler 1992). For most health outcomes there were no general health bene…ts from having more complete insurance (i.e., full coverage) (e.g. Brook, Ware, Rogers, Keeler, Davies, Donald, Goldberg, Lohr, Masthay, and Newhouse 1983). Health bene…ts were found, however, for individuals with poor vision and for persons with elevated blood pressure. Importantly, the improvement in high blood pressure led to a statistically signi…cant 10% reduction in mortality risk, apparently due to increased detection and treatment of high blood pressure among low-income households with free care (e.g., Keeler 1992). Several other studies examine changes in insurance eligibility rules, comparing outcomes for individuals who are just eligible to those who just missed the cut-o¤ for eligibility, or use other rigorous study designs. Across a variety of settings in the U.S. and Canada, expansions of health insurance coverage have consistently increased health care utilization (Fihn and Wicher 1988; Lurie, Ward, Shapiro, Gallego, Vaghaiwalla, and Brook 1986; Lurie et al. 1984; Currie and Gruber 1996; Currie and Gruber 1996; Currie and Gruber 1997; Lichtenberg 2002; Card, Dobkin, and Maestas 2007; Finkelstein 2005). Some studies …nd important improvements in health (e.g., Hanratty 1996; Currie and Gruber 1997), others modest or not statistically signi…cant improvements (e.g., Card, Dobkin, and Maestas 2007), and others evidence of no strong bene…ts (e.g., Finkelstein and McKnight 2008). Results are more mixed regarding the impact of health insurance on outcomes in poor nations. Most studies …nd a negative relationship between insurance coverage and out-of-pocket health expenditures (e.g., Jutting 2004, in Senegal; Jowett, Contoyannis, and Vinh 2003, in Vietnam; and Yip and Berman 2001, in Egypt). In contrast, Wagsta¤, et al., (2009) …nds that out-of-pocket spending is the same or even higher for the insured than the uninsured in China. They explain this surprising …nding as being a result of the institutional structure of health-care in China, which favors increased utilization and substitution toward more expensive services and treatments. Fewer studies look at health outcomes, though Wagsta¤ and Pradhan (2005) …nd that a national voluntary health insurance program in Vietnam is correlated with increased health care utilization and increased height-for-age and weight-for-age measures for children and with an increased (that is, healthier) BMI for adults. These studies in poor nations are useful, but are all subject to concerns that a very non-random group of people have health insurance. To our knowledge, no study of insurance in developing countries cleanly identi…es the causal relationship between health

2This literature review draws on Polimeni (2006) and Levine, Gardner, and Polimeni (2009) . 4 insurance and health spending, health care utilization or health outcomes. If health insurance increases utilization of e¤ective health care services, there is room for it to improve health in the poor area of Cambodia, where foregone care is an unfortunately common event (World Bank 2006). Past research has shown that the impacts of health insurance or changes in the price of health care on health are largest among the lowest income populations (e.g., in the RAND health insurance experiment in the US noted above, Manning 1987; and in the Indonesian Resource Mobilization Study, Dow, Gertly, Schoeni, Strauss, and Thomas 1997), though Wagsta¤ and Pradhan (2005) …nd smaller e¤ects of insurance for low-income households than for other households in Vietnam. While many studies have focused on the e¤ects of insurance on health and out of pocket health expenditures, health insurance can also in‡uence longer term economic outcomes. Health insurance may in‡uence a family’s long-term economic well-being by preventing families from selling productive assets or increasing child labor to cover medical expenses. Any increases in health can also lead to increases in productivity and income. For example, Thomas, et al., (2004) show that improving health via iron supplements has a signi…cant positive e¤ect on productivity for adults in Indonesia. Dow, et al. (1997) give evidence that higher prices for health care are associated with reduced labor force participation for women and lower wages for men in Indonesia. The study of the impact of insurance on health utilization also …ts into the emerg- ing literature on demand for health and health care services. Insurance will only have an impact on utilization of health care services if demand for health is somewhat elastic. If households utilize health care even at high prices, then lowering the marginal price of insurance should not increase utilization of care. On the other hand, because the SKY insurance program lowers the cost of public care as compared to other types of care, SKY may induce individuals to change health care provider (a stated goal of SKY). Several recent studies and literature surveys have examined elasticity of demand for health care services. In a recent literature review, Dupas (2011) concludes that demand for coverage of acute illness is relatively inelastic (e.g., Cohen, Dupas, and Schaner 2011, as referenced in Dupas 2011). Access to credit has not been found to increase utilization of health services, possibly because households insure against health risks through social networks (Townsend, 1994 and Robinson and Yeh, 2011, as referenced in Dupas 2011). Thus, we expect that SKY will not change percent utilizing health services following a major shock, although they may change provider type, as SKY only covers public providers. While households do not change utilization of health care services for some illness, they are often unable to cover the costs associated with major health shocks (Gertler 2002 and Fafchamps and Lund, 2003, as referenced in Dupas 2011). Families without access to credit may decrease investments in productive assets and otherwise jeopardize their future (Rosenzweig and Wolpin 1993 and Robinson and Yeh 2011, as referenced in Dupas 2011). While demand for treatment of acute illness is inelastic, demand for preventative services such as bednets, water treatment, and deworming products, has been found to be very price elastic (Kremer, Leino, Miguel, and Zwane 2011; Cohen and Dupas 2010; Kremer and Miguel, 2007; Abdul Lateef Jameel Poverty Action Lab 2011). A small decrease in cost produces a large increase in uptake. Thus, we may expect that SKY, by decreasing the 5 marginal price of preventative care, may have a large impact on the utilization of this care.

1.3 The Setting

1.3.1 Health care in Cambodia Cambodia is among the world’s poorest and least healthy nations. It ranks 188 out of 229 nations in GDP per capita, has the 38th highest infant mortality rate (of 224 countries with data), and the 46th lowest life expectancy (Central Intelligence Agency 2010). Cambodians rely on a mix of healthcare providers: public providers, private medical providers, private drug sellers (with and without pharmaceutical training), and traditional healers. Public facilities consist of local health centers, which provide basic care for every- day illnesses, Operational District Referral Hospitals, for illnesses requiring more involved treatment, and Provincial Hospitals, for care of more severe health shocks. Public facilities are subsidized by the Cambodian government or other organizations. However, public facilities have low utilization. According to the 2005 DHS, fewer than a quarter of those who sought treatment for illness or injury went to a public health facility. Private providers of varying capabilities are typically more popular than public ones, even when more expensive, because they often are more attentive to clients’needs, more available, visit patients in their homes, provide treatments patients prefer, and provide credit (Collins 2000; Annear 2006). At the same time, while households often utilize local private doctors and drug sellers for small health shocks, many visit public hospitals for surgery and other major health problems. The average rural household spends $9.60 per month on health care, of which $2.48 is spent on public health center and hospital visits (DHS 2005). Health shocks often contribute substantially to indebtedness and loss of land. For example, one study followed 72 households with a member who had su¤ered dengue fever following a 2004 outbreak in Cambodia. A year later, half the families still had outstanding health-related debt, with interest rates between 2.5% and 15% per month. Several of the 72 families had found it necessary to sell their land to pay their debt. (Van Damme, Van Leemput, Por, Hardeman, and Meessen 2004). Annear, et al. (2006) and Kenjiro (2005) found similarly high levels of indebtedness due to medical expenses.

1.3.2 SKY Health Insurance SKY health insurance was originally developed by the French NGO GRET as a response to high default rates among its micro-…nance borrowers due to illness. Since 1998 GRET has been experimenting with micro-insurance schemes by examining responses to di¤erent premiums and bene…ts. Historically, take-up of insurance has ranged from 2% in regions where insurance has been only recently introduced to 12% in the longest-served regions. While the SKY program targets the poor, it also is trying to avoid …nancial losses and become …nancially sustainable (without donor support) in the long term. Thus, the policy includes several terms that limit adverse selection. For example, SKY does not pay 6 for the delivery of babies within the …rst few months of joining. Also, insurance is purchased at the household-level, eliminating the possibility that households would purchase insurance for only very ill or frail members. Finally, SKY insurance does not cover long-term care of chronic diseases. (Government programs pay for the very expensive drugs for HIV/AIDS and tuberculosis.) At the time of the study SKY sold insurance at prices ranging from $0.50 per month for a single-person household to around $2.75 per month for a household with eight or more members. Households sign up for a six month cycle, paying for the …rst month’s coverage plus two reserve months up front. While a household can stop insurance payments at any time, failing to pay two consecutive months before the end of the six-month cycle results in the loss of one month of reserve. A household can join SKY at any time, but coverage will not begin until the start of the next calendar month. Households o¤ered insurance for the …rst time are o¤ered slightly lower premiums to encourage take-up. With their insurance, household members are entitled to free services and prescribed drugs at local public health centers and at public hospitals with a referral (SKY 2009).

1.4 Theory and Measurement

1.4.1 Health seeking behavior SKY health insurance lowers the cost of health care at public facilities. Thus, we expect that health insurance will increase health care utilization at public facilities, especially if households were seeking too little care prior to insurance purchase. We expect that most e¤ects of health insurance arise when someone has a serious illness or injury. At the same time, insured households may also increase preventative care. We measure both types of impacts, described below.

Health behavior following a health shock For health seeking behavior following a health shock, we focus on serious health incidents, which we de…ne as illnesses or injuries that lead to seven or more days of disability or death. On the one hand, by reducing the cost of care following a health shock, insurance can increase health-seeking behavior. On the other hand, if demand for health is relatively inelastic, as has been found in much of the recent health-demand literature, we may not see much increase in health care utilization, although insured households may shift away more costly private care towards SKY-covered care. We also measure reduction in foregone health care and reduction in delayed care. One of SKY’sprincipal goals is to reduce the share of families that forego necessary health care due to lack of funds. In our study, a sick household member is considered to have foregone care following an illness or injury if treatment was not sought, or was discontinued, due to cost. A concern in poor nations is that families delay treatment of illness due to costs. Thus, among serious incidents, we examine the e¤ect of insurance on the number of days until …rst treatment. More important for e¤ective treatment is that households are seeking quali…ed health care in a timely manner. Thus, we also measure time until they were treated at a hospital. 7

As noted above, health care in rural Cambodia is dominated by poorly trained informal doctors and drug sellers. SKY’s theory of success posits insured families will be less frequent users of ine¤ective informal care and unquali…ed private “doctors.” We proxy for those caregivers by looking at serious or costly incidents that used a drug seller, traditional healer (kru khmer), or private provider. Public health care providers are the only providers that are regulated by the Cam- bodian government. By partnering with only public facilities, SKY encourages utilization of these regulated facilities. To test this, we look at percentage of individuals visiting a public facility …rst or at all for care following a major health shock.

Other health seeking behavior We also analyze foregone care for households as a whole, whether or not they experienced a major health shock. To measure this, households are asked whether a member has ever foregone care due to lack of funds. Insurance may increase care following a health shock, but may also increase routine and preventative care. In general, having zero co-pay at public facilities may increase use of public health centers even in households without a major health shock. To test this, we examine use of a public provider in the three months prior to our household survey in households with or without a major health shock. While immunizations and some other forms of preventive care in Cambodia are already free, many Cambodians have little exposure to the public health facilities that provide and encourage preventive care. Thus, joining SKY (and using public facilities more) may increase preventive care. We test if SKY increases immunizations and modern contraception, and test whether SKY has any impact on birth-related outcomes such as ante- and postnatal care and location of birth.

1.4.2 Economic impacts The economic bene…ts of insurance require both that the health insurer pay after a serious injury or illness, and that the family reduce expenditures on expensive private providers. The net result is lower total out-of-pocket expenditures. Health care expenditures arise precisely when the family has lost productivity and often income from one or more adult. For example, if a patient is hospitalized, other household members typically must provide meals and other care for the patient, and may decrease labor supply to provide this care. The combination of low income and high expenditures can lead families to sell assets or take on debt. Market interest rates are high, so a loan often leads to asset sales at a later date. We hypothesize that when a serious health incident occurs, insurance will lower the rate of selling assets and of taking on debt to pay for care. We divide economic impact measures into two categories: economic consequences of individual health incidents, and overall economic impacts to a household.

Economic impacts following a health shock We use several outcomes to measure the impact of health insurance following a health shock. The goal of insurance is not focused on mean expenditures, but a substantial 8 reduction in the rate of very high expenditures. Thus, we look at economic behavior following only a major health shock, de…ned again as an illness or injury leading to death or an inability to carry out normal daily activities for seven or more days. To test whether SKY reduces out-of-pocket costs, we examine total out-of-pocket costs for health care (including transportation costs) following a major shock. Because insurance is most important for larger shocks, we also estimate whether insurance decreases the occurrence of costs exceeding 250 USD following a single incident (the top 10th percentile), or of costs exceeding 100 or 350 USD for a household (the top 35th and 10th percentiles). As mentioned above, to reduce out-of-pocket expenses, SKY must reduce the amount of money spent at expensive private providers. To test this, we look at the impact of SKY on large out-of-pocket costs paid for private care. To reduce out-of-pocket expenses, SKY must also pay for care following a health shock treated at a public facility. To test this, we measure how often SKY pays for care for insured households. If SKY lowers out-of-pocket expenses, households may be less likely to pay for care using costly means of payment. To test this, we examine how often health care expenses following a major health incident are covered by borrowing money, selling an asset, or raising money through extra work. If SKY increases health care or prompt utilization of quality health care, an ill individual may recover more quickly and may have fewer lost days of productive activity. We calculate the impact of SKY on the total number of days of missed activity for ill individuals.

Overall economic impacts on households If insurance is e¤ective, we expect insured families to be less likely to take on new loans due to health care costs and less likely to sell land and other assets. Above we describe our test for this outcome at the incident-level: we look at the percentage of major health incidents for which care is …nanced with a loan or asset sale. We also look at these measures at the household level: Out of all households, were insured households less likely to take out a loan or sell an asset in the past year due to health (not necessarily related to a major incident)? To increase precision we also run this analysis on the subsample of households that had a death or long-term disability during the year. If uninsured households sell productive assets or withdraw children from school to help pay for care, the result is that a short-term health shock can lower long term productivity and worsen long-term poverty (Van Damme, Van Leemput, Por, Hardeman, and Meessen 2004; Annear 2006; Jacoby and Skou…as 1997; Smith 2005; Dupas 2011). Conversely, if health insurance can avoid large out-of-pocket expenditures it may promote the accumulation of productive physical and human capital. Although this study was not designed to be large enough to measure such bene…ts unless they are very large, to test this we look at impact of SKY on productive assets and school enrollment.

1.4.3 Health Outcomes Prompt and appropriate curative care, avoidance of harmful care from unquali…ed providers, and increased preventative care will over time improve health. Unfortunately, it 9 takes an extremely large multi-year study to detect such e¤ects. Although this study was not designed to have much chance to measure such bene…ts, we measure how SKY insurance a¤ects objective measures of health such as frequency of major health shocks and children’s stunting and wasting.

1.4.4 Trust in Providers and SKY In addition to testing the health and economic outcomes of SKY members, we also test several other impacts of the SKY program. SKY typically selects relatively high-quality public sector providers and then works with them to improve quality. To the extent SKY is successful in both improving quality and increasing usage, SKY members will learn about the higher quality at public providers and increase their trust in these providers. SKY posits that their program provides good service to its members. If so, we expect that SKY members will observe this good service and increase their trust in SKY. We look at the impact of SKY insurance on the average of several measures of trust in SKY.

1.5 Data and methodology

Those who choose to purchase insurance typically di¤er markedly from those who decline insurance. To understand the causal e¤ects of insurance we implemented a randomized controlled trial that allows us to identify the impact of health insurance independently from all other factors that may a¤ect a household’s decision to take up insurance. No household was denied access to insurance. Rather, by subsidizing the premium of a randomly-selected group of households, we are able to estimate the e¤ect of insurance on households without substantially altering the existing SKY program.

1.5.1 Randomization of prices Our randomized experiment was carried out as the SKY program expanded to 245 villages from November 2007 to December 2008.3 The expansion took place in Takeo, Kandal, and Kampot provinces, all rural areas of Cambodia. When the SKY program …rst rolls out into a region, SKY holds a village meeting to describe the insurance product to prospective customers. The meetings are advertised ahead of time via loudspeaker announcements in each village.

3The analyses and data collection described in these papers is the result of a project that has been ongoing since 2006. While not discussed directly here, my role in the project included writing grant applications for the research (we received over one million dollars in funding through these research grants), designing the evaluation, and designing the survey instruments. To make the evaluation minimally intrusive to the operations of the SKY program, the …nal evaluation design was agreed upon after several discussions with GRET o¢ cers. Survey instruments were made culturally appropriate through discussions with our Cambodian-based research partners at Domrei Research and Consulting ("Domrei"), and several rounds of pilot testing. Randomization of coupon prices at the SKY village meetings and all data collection was carried out by the …eld team at Domrei. Throughout the project I visited Cambodia several times to choose our Cambodian-based partner (Dom- rei), discuss survey design, observe pilot tests, and to present evaluation results to SKY, Cambodian ministry o¢ cials, project and evaluation donors, and other stakeholders. 10

To randomize price of insurance, we implemented a Lucky Draw whose winners received a deeply discounted price: 5 months free insurance in the …rst 6-month cycle, with the option to renew for a second 6-month cycle with a coupon for 3 months free. At the start of each meeting, an Evaluation Representative recorded the name of one representative of each household in attendance, and throughout the meeting, recorded the names of those arriving late. SKY’s Field Coordinator introduced SKY in the typical fashion, explaining the product and to what it entitles the buyer. As the SKY Field Coordinator spoke about the product, the Evaluation Representative counted the number of households attending the meeting and determined the appropriate number of high and low coupons to be distributed. The number of 5 month coupons to be ra• ed o¤ was set equal to 20% of attendees for meetings of up to 60 households and equal to 12 for meetings of more than 60 households. The remaining households were entitled to a coupon for 1-month free o¤ of the …rst 6-month cycle. These high- and low-valued coupons, printed on colored heavy-weight paper, were placed into an opaque bag. At the end of the meeting, the Field Coordinator announced that there would be a “Lucky Draw”for coupons, and explained to what each coupon entitles the bearer. The Field Coordinator also explained that coupons could only be used by the family winning it. Next, the names from the attendance list were called o¤ one by one, and one representative from each family came to the front of the room to draw a coupon from the bag. High and low coupons were di¤erent colors, so that meeting attendees could see which type of coupon was drawn, but care was taken to ensure that coupon type could not be seen while drawing, and that high and low coupons could not be identi…ed by touch. The outcome for each draw was recorded next to the person’sname on the attendance sheet. All households winning a high coupon were selected to be part of our survey sample. Research …eld sta¤ also chose an equal number of low coupon households to be included in the survey sample. Low coupon households for the survey were chosen by picking every fourth household from the meeting roster until enough low coupon households had been chosen to equal the number of high coupon winners. Following the meeting, our sta¤ and the village chief drew village maps with the location of the families in our sample (that is, all the high-value coupon winners plus the low-value coupon winners that would also be surveyed). SKY Insurance Agents then visited these households to o¤er them health insurance. We encouraged members who received the steeply discounted o¤er to renew by o¤ering additional discounts after the initial 12 months had passed.

1.5.2 Estimation Intention to Treat The randomization of prices allows us to answer the question, “What is the e¤ect of o¤ering insurance at a deeply discounted price?”This result can be calculated by simply comparing average outcomes for households that did or did not receive a coupon for a large discount for SKY insurance. 11

Impact on the Insured (Treatment E¤ect on the Treated) We can also estimate the e¤ect of SKY insurance on households that purchased insurance due to the discount (the e¤ect of the Treatment on the Treated population). To estimate the e¤ect of insurance on the insured, we cannot simply compare outcomes of the insured to the uninsured. If we estimate how SKY predicts outcomes Y for household i at time t with ordinary least squares:

Yit = SKYit + "i (1.1) the estimated coe¢ cient OLS can have very large bias because SKY membership is endogenous. For example, if people with health problems purchase insurance more often, OLS could be strongly negative (that is, SKY predicts poor health), even if SKY insurance actually improves health. Thus, we instrument for SKY membership with the randomized treatment, with Ti = 1 for those o¤ered the steeply discounted price. Due to drop-out over time, SKY membership was higher a few months after a village meeting than several months later for those o¤ered the higher price. Thus, we also included as an instrument the o¤ered price interacted with the number of months since the village meeting (Monthsit):

SKYit = Ti + Monthsit + Monthsit Ti + uit (1.2) 1 2 3 Our survey collects data on major health shocks using respondent recall over the 12 month period immediately prior to the survey date. Thus, for incident-level outcomes, that is to say, outcomes that are a direct result of an individual health incident in month t, t is de…ned as the date of the incident, monthsit is de…ned as the number of months between the village meeting and time t, and the instrument Monthsit Ti is Monthsit multiplied by 1 if household i received a high coupon and 0 if the household received a low coupon. SKY status in month t, SKYit, is de…ned as a three-month average membership rate centered in month t, to account for imperfect recall of the timing of health incidents. Thus, SKYit can 1 2 take on the values 0, 3 ; 3 or 1. For example, for a health incident occurring t months after the village meeting, SKYit equals 1 if household i was insured in months t 1, t, and t + 1, but equals 1 if the household was insured in only time t 1. 3 Similarly, for birth outcomes, t is de…ned as the month of the birth, and monthsit as the number of months between the village meeting and time t. SKYit is again de…ned as a three-month average membership rate centered in month t. For all endogenous variables not related to a particular health incident or birth we de…ne Monthsit as the number of months between the village meeting and the date of the interview. For outcomes measured by behavior in the three months prior to the survey, such as having visited a public facility (for any reason, whether or not related to an illness), we de…ne SKYit as average membership in the 4 months prior to the survey (again, to account for imperfect recall). For outcomes that take time to accumulate such as health-related loans, SKYit is de…ned as the share of the year prior to the interview that the household was a SKY member. Finally, for variables that require only that the household be exposed to SKY, such as trust in SKY, SKYit =1 for households that had ever been a SKY member. The precise dating of membership never a¤ected results. 12

Using our randomized price as an instrument estimates the e¤ect of insurance on those households who purchase insurance due to the deeply discounted price. By de…ning SKYit at the time of an incident (or the other de…nitions, above) and including o¤er price interacted with months since the village meeting as an instrument, the “Treatment on the Treated” regression measures the impact of SKY on households that joined SKY and remained in SKY due to the large discount. For simplicity, we will often refer simply to the e¤ect of SKY on the “insured”and contrast it with the control group (those without a high-valued coupon), even though a small portion of the control group also purchased SKY. The causal e¤ect on this price-sensitive group is the local average treatment e¤ect (“LATE”; Imbens and Angrist, 1994). Unless the e¤ects of SKY are homogenous for all populations, the instrumental variables methodology does not allow the measurement of the impact of SKY coverage on households that would have bought SKY both with and without the large discount, or on households that choose not to buy insurance even at the largely discounted price. It is plausible the bene…ts of SKY are larger for the …rst group and smaller for the latter. As we also use months in SKY as an instrument, we are not measuring the impact of SKY on households that join SKY but immediately drop. The e¤ects of SKY may be lower for these households

1.5.3 Data Our analyses use a longitudinal household survey and SKY data on membership. We chose our sample size to have 80% power to detect a feasible and economically important reduction in several important outcome measures. For example, we expected to have 80% power to detect a 2.6 percentage point reduction in the percentage of households spending over $1.25 on health care in the previous four weeks (compared to the 10.1% mean in DHS 2005 data), or a 2.0 percentage point increase in the number of households using a public facility in the past four weeks (compared to the 5.1% utilizing public facilities in DHS 2005 data). Although we collected data on prenatal care, birth outcomes, anthropometric measures for children, and frequency of major illness or death, the evaluation was not designed to have statistical power to detect impacts on these measures. For example, using our sample, we calculated that we could detect a 3.5 percentage point decrease in the percentage of households reporting any illness in the last 4 weeks (compared to the baseline mean of 20.2% in DHS 2005 data). Using our actual survey measure of percent of individuals with an illness lasting more than 7 days, we have 80% power to detect a 2.6 percentage point decrease compared to the control of 10.2% reporting such an illness. Even with increases in utilization of public facilities, which may provide better care than unregulated treatment, we did not expect to see this level of change in the percentage reporting ill. For prenatal care, birth outcomes, and anthropometric measures, we have data on only a small portion of our sample, so it becomes even harder to detect changes in outcomes.

Household Survey Our main data source is a survey of over 5000 households. We rely largely on the follow-up survey, which took place 13 to 20 months after the initial SKY marketing 13 meetings. We also use some data from the …rst round survey administered one year prior to the follow-up, so 1 to 8 months after the village meetings. The surveys cover demographics, wealth, objective health measures, health care utilization and spending, assets and asset sales, savings, debt, trust of health care institutions, and so forth. We ask households to describe health utilization behavior following a major health shock, which we de…ne as a health incident causing a death, the inability to carry out usual household activities for seven or more days, or an incident causing an expense of over 100 USD. In most analyses we do not include behavior following a 100 USD health expense because households with SKY insurance would be less likely to fall into this category. In each village we interviewed all households that won the Lucky Draw (and were o¤ered the steeply discounted price) and an equal number of households o¤ered the regular price. We selected the control households by choosing every fourth non-winner from the village meeting attendance list, as described above. In total, our randomized sample consists of 2617 households o¤ered the deep discount and 2618 households o¤ered the regular price, of which we interviewed 2561 and 2548 households, respectively, in the baseline survey, and for which we have follow-up data for 2502 and 2506 households, respectively. Figure 1.1 summarizes the timeline and sample size of the evaluation. Because there was a delay between the …rst o¤er of insurance and the baseline survey, baseline survey results are not necessarily pre-insurance results. As a robustness check, we include “baseline” levels of some impact variables as controls. In that case, if insurance has already had an impact on households a few months after joining SKY, then the delay in the baseline will bias the estimated e¤ects of insurance downwards.

SKY Membership For each household that joins SKY, SKY records the date the household starts coverage, and (if not still a member) the date the household dropped out of SKY.

1.6 Results

1.6.1 Tests of Experimental Design Randomization Table 1.1 shows average characteristics of high and low coupon winners prior to the SKY meeting (for health events) or at the time of the …rst round survey. Of the thirty variables tested, only three show a statistically signi…cant di¤erence between high and low coupon at the 5% con…dence level. 14% of low coupon households have wealth level subjectively graded as “poor” by enumerators, while only 10% of high coupon households are rated as “poor”. Similarly, low coupon households are slightly more likely to live in a house made of palm, another measure of lower wealth. Other wealth indicators did not show signi…cant di¤erences. Households o¤ered a high coupon were also slightly less likely to be Khmer as opposed to another ethnicity: 94.6% versus 95.3% of high coupon households were Khmer. 14

We keep in mind these di¤erences when interpreting results, and for some variables, we test whether holding …rst round survey values constant impacts our results.

Analyzing serious health incidents We analyze a number of outcomes that measure behaviors following a major health incident, de…ned as an incident leading to missing seven days of usual activities (e.g., work) or a death. If insurance a¤ects the probability of a major incident, then for these measures we are no longer identifying the e¤ect of insurance solely using the randomized price. For example, suppose SKY induces an insured member to seek care for illness, and that seeking care means that the individual is unable to work for seven days or more. At the same time suppose that an uninsured household with the same illness would work through the illness. If this occurs, insured households will be counted as having a “serious” illness by our measure while the uninsured household would not. Behavior by the insured individual will be included in our measure, while that for the uninsured individual will not, causing bias in our results. One factor that helps to reduce this potential bias is that SKY does not greatly increase the incentive to spend a week at the hospital. Even with SKY insurance, hospital stays require family members be present to feed and provide some care for the patient. In addition, by the sixth day the marginal out-of-pocket cost of a hospital stay is zero even for the non-insured. SKY members may also be less likely to have a death than non-SKY members, although it is unlikely SKY would a¤ect death rates by much over such a short time. We believe that neither of these factors will have a meaningful e¤ect on the number of households from the insured and uninsured groups being classi…ed as having a serious incident by our measure. Consistent with our assumptions, the rates of serious incidents are almost identical in the high and low-coupon samples (Table 1.9). Among individuals in treatment and control households, there are almost identical numbers of deaths for the treatment group (those o¤ered the steeply discounted price) and the control group (0.007 average for both groups, no statistically signi…cant di¤erence). When we look at individuals with shocks requiring missed activity for 7 or more days, rates were also similar: 10.2% for both the treatment and control groups.

1.6.2 Summary statistics Summary statistics for each outcome, subdivided into Treatment and Control means, are presented in each outcome table. Comparing outcomes for the treatment and control group provides the intention to treat estimates of the e¤ects of distributing steep discounts.

1.6.3 First Stage Our instrumental variables methodology requires that SKY membership is strongly correlated with our instrument (i.e., the steeply discounted price plus time since the village 15 meeting). Figure 1.2 shows that this is in fact the case. Membership peaked at around 47% for treatments at month six and declined steadily over time. Membership for controls does not change much over time, slightly increasing to a peak of 3.3% at 20 months. Table 1.2 shows the …rst stage regression for incident level data. Recall that for the incident-level data, SKYit averages membership in the month of, prior to, and following the incident month t, and that months is de…ned as the number of months between the village meeting and month t. First stages for the other speci…cations are in Appendix Tables 1.15 through 1.17. All are similar to Table 1.2 and show similarly large e¤ects of the treatment on SKY membership and similarly strong statistically signi…cance.

1.6.4 Health Seeking Behavior Health seeking behavior following a health shock Here we present the impacts of health insurance on utilization following a serious health incident, which we de…ne as 7 or more days unable to perform usual activities due to health issues or a health incident resulting in a death. For the impact on forgone health care, our instrumental variables estimate is that those who purchased insurance due to the discount had a 3.2 percentage point reduction in discontinued treatment following a health incident compared to the control mean of 5.2%, but this di¤erence is not statistically signi…cant at conventional levels (Table 1.3, P = 0.19). We also examine the number of days until …rst treatment. Counter to expectations, insured individuals with a health shock have a longer delay before …rst treatment, and are less likely to receive care within a day (Table 1.3). However, this result may be due to the higher percentage of uninsured households receiving …rst treatment at drug-sellers (results below). More important is delay until e¤ective treatment. Thus, we also examine days until insured visited a “hospital,” where that term is best translated as “public hospital or private caregiver.” We top-coded this measure at 30 days, and coded those with no hospital or clinic visit as having a delay at the top-coded value of 30 days. We also measure the percent of individuals with incidents receiving care at a hospital within a day of the incident. There was no signi…cant di¤erence between baseline and those insured in either of these measures. Sources of care during a serious health care incident changed signi…cantly with insurance (Table 1.4). Speci…cally, SKY insurance doubled the odds that a serious incident’s …rst treatment was from a public health center. Among the control, almost half of serious incidents had their …rst source of care at a private provider, 14% at drug sellers, 16% at public hospitals and 14% at public health centers (NGOs and kru khmer traditional healers make up the rest). SKY reduced private providers as the …rst source of care by 11 percentage points (P<0.05), reduced drug sellers by 8 percentage points (P < 0.05) and increased public health centers by 18 percentage points (P< 0.001). Rates of …rst accessing public hospitals were not changed by economically or statistically signi…cant amounts. Many serious incidents have care from multiple providers. Rates of ever using each type of provider following a health shock also shifted in favor of health centers: among the control, 18% of households used a health center following a health shock, and this increased by 22 percentage points to 40% among SKY members after SKY purchase (P < 0.001). 16

The 9 percentage point decline in ever using a private provider (compared to the control with near two-thirds of all individuals with a shock) is marginally statistically signi…cant at the 7% level. (Appendix Table 1.11.)

Other health-seeking behavior At the household level, using instrumental variables, insured households that purchased due to the large discount were 1 percentage point less likely to forgo care compared to the control mean of 0.9% (essentially indicating that the insured had no forgone care) but this impact was not statistically signi…cant (Appendix Table 1.12). Respondents also were asked, “In the last three months, did you go to see a government doctor?”Inconsistent with SKY’stheory of change, SKY membership does not increase the share of respondents who report use of a public provider in the previous 3 months (Appendix Table 1.12). SKY also hoped to improve preventative care. The results on preventive care have low statistical power because of the smaller sample size of children (for immunization measures) and women of reproductive age (for birth outcomes and contraception). With that caution in mind, there is no detectable e¤ect on the proportion of children whose immunizations are up to date, or on the share of married women ages 16-45 using contraception or using modern contraception (Appendix Table 1.12). Table 1.5 presents SKY impacts on birth-related outcomes. On the one hand, the insured are no more likely to receive antenatal care in general, and there was no signi…cant impact on the percent receiving post-natal check-ups. On the other hand, the insured are much more likely to report having received at least one tetanus shot during pregnancy (P = 0.10, compared to the control mean of around 92.6%).4 Regardless of insurance, 99% of births had a trained birth attendant, midwife, or doctor present at the birth. Insured women were slightly more likely to give birth under the care of a trained birth attendant or doctor, and slightly less likely to give birth with a midwife, than were uninsured households, but these di¤erences are not statistically signi…cant at traditional levels. We do …nd some di¤erence in delivery location between insured and uninsured women. Women in insured households were 21 percentage points more likely to give birth in a public facility (the control mean is 59%), although given the small number of births the di¤erence is not statistically signi…cant. Pooling births at a any formal facility, insured women were 31 percentage points more likely to give birth in either a public or private facility (P = 0.06, control mean is 64%). Women not giving birth in public or private facilities gave birth either at home, the forest, or another location.

4The point estimate, taken literally, shows a 12 percentage point increase in reporting at least one tetanus shot; this e¤ect would lead to over 100% of SKY members having a tetanus shot. This anomaly is due to our choice of linear probability model coupled with sampling error. That is, if by chance a few high-coupon recent mothers who did not join SKY had a tetanus shot, our instrumental variable method will expand that sampling error to get the reported point estimate. 17

1.6.5 Economic E¤ects of Insurance Economic e¤ects following a health shock We analyze total out-of-pocket costs (Table 1.6), and then examine how households pay for costs of care (Table 1.7). To measure out-of-pocket costs, we top-coded each household’s total health care expenditures for serious (7 or more days unable to work) or fatal incidents at the 98th percentile (947 USD) to eliminate large outliers. We include both cost of treatment and cost of transport. The control mean cost for an incident is $103.81. The instrumental variable estimate is that households induced to purchase SKY due to the steep discount (who remained insured) paid $45.79 less in care and transport for a serious or fatal incident (P < .05, Table 1.6). Summing over all incidents in the last twelve months, we estimate that households that purchased SKY due to the deep discount paid $57.80 less in care and transport for these major incidents, compared to a control mean of $132.43 (P < 0.01). Importantly, much of this savings in out-of-pocket costs are due to lower rates of very high medical expenses. We cumulated out-of-pocket costs for each serious incident.5 While 11% of incidents in control households had health care costs of over $250, insurance decreased this percentage by 8.6 percentage points (P < 0.01).6 Moving to the household level (that is, cumulating across all incidents in the past year for a given household), insured households have 5.0 percentage points lower probability of spending over $350 (compared to control rate of 11.5%, P = 0.19), and 10.9 percentage points lower probability of spending over $100 following a shock (compared to the control rate of 38.2%, P < 0.10). SKY decreases costs in part by lowering the percentage of households paying for high-cost private visits, but the e¤ect is modest. The insured are 12.3 percentage points less likely to spend more than $5 at a private provider following a health shock compared to the control of 61.9% (P < 0.05), and 7.0 percentage points less likely to spend $150 compared to the control of 9.7% (P < 0.05). For private expenses, varying the cut-o¤ amount up to $1000 sometimes made the di¤erence insigni…cant, but the insured had lower private expenses than the uninsured in all but one (statistically insigni…cant) case. SKY also can reduce costs by paying for public care, but this will only be the case if they actually pay for care. Households induced to buy SKY with the large discount are 43.8 percentage points more likely than other households to have a treatment paid for by SKY insurance following a serious or fatal health shock (P < 0.001, Table 1.7). SKY households are also 9.2 percentage points less likely to sell assets following a shock (versus the control mean of 22.4%, P < 0.05, Table 1.7), 13.6 percentage points less likely to take out a loan with interest (versus the control mean of 19.6%, P < 0.01), and 6.4 percentage points less likely to take out a loan without interest (versus the control mean of 12.8%, P < 0.10), following a large health shock. SKY had no signi…cant impact on the use of extra work to pay for health care expenses.

5Results hold if we include households that did not have a death or missed 7 days, but spent over $100USD on care. 6We chose this cut-o¤ to correspond to the top 10th percentile of spending. We tested di¤erent cuto¤s under $250 and in all cases the IV regression showed that the insured had signi…cantly lower spending than the uninsured. Cuto¤s above $500 did not produce statistically signi…cant results. 18

In results not shown, while individuals with health shocks in insured households have an average of 1.9 fewer days lost due to illness (compared to the average control rate of 39.5 days ill), the di¤erence has very low statistical signi…cance (P = 0.82).

Overall economic impacts on households Separate from analyzing the costs of each incident, we examined economic outcomes of households. Consistent with insurance reducing out-of-pocket expenditures, households with SKY also have less debt. On average, insured households have $68 lower debt (P < 0.05), about one third of the mean for conrol households (Table 1.8). When we ask speci…cally about loans for health, insured families have $22 lower loans from health –compared to the control mean of $29 (P < 0.001). Also as we expect, the lower debt for SKY members shows up only in households with a serious health incident or death (Appendix Table 1.13). While SKY and non-SKY households with no serious incidents have lower debt than households with a serious incident, among those with no serious incident, debt is not especially lower for insured households (results not shown). Results were similar when we asked directly (in a di¤erent section of the survey) whether the household had more debt than the previous year due to health care costs or a birth. Households who bought insurance due to the high coupon were 7.7 percentage points less likely than control households (at 8.9%) to have such a loan (Table 1.8, P < 0.01). Looking at the impact of SKY on productive assets, the insured are less likely to report a reduction in land from the previous year, though the estimate is not statistically signi…cant (Table 1.8). When we focus on a reduction in farmland or village land because of health, we estimate that no SKY members sold land due to ill health; the IV point estimate shows that households that purchased SKY were 1.6 percentage points less likely to sell land for health reasons compared to the control mean of 1.1% (P = .051). For SKY donors, the hope is that over time health insurance promotes the accumulation of productive physical and human capital. (Recall this study was not designed to be large and long enough to be likely to measure such e¤ects.) Our IV results show that SKY members had substantially higher value of livestock ($96.9 higher, compared to the baseline mean of $540, P < .05, Table 1.8). There is no di¤erence in other asset classes: cash, gold, or non-farm businesses (not shown), or between Treated and Control groups as a whole.7 A wealth-index composed of the averaged z-scores of the value of cash, gold, animal, durable assets, and non-farm business shows a positive impact of SKY on wealth, but the e¤ect is not statistically signi…cant ( = .09, P = 0.13, Table 1.8).8 As expected, economic impacts on households with health incidents are generally larger than on households overall (Appendix Table 1.13).

7We test some outcomes holding baseline constant in Appendix Table 1.14. 8To create this index, we created z-scores for each of the …ve wealth values (gold, cash, animals, assets, business) by subtracting the overall mean of these variables and dividing by the standard deviation. The index is the average of these …ve zscores. This is similar to a procedure used by Kling (2007), except that that paper normalizes so that the mean and standard deviation of the index for control households is equal to zero. 19

Our instrumental variable estimate (Table 1.8) is that insured households have a 4.6 percentage point higher fraction of school-aged children enrolled in school versus the baseline mean of 83.1% (P = 0.14). While provocative, the higher enrollment is not driven by households with major health incidents (Appendix Table 1.13). Thus, this outcome is likely being driven by something other than SKY coverage. Future analyses will investigate this outcome in more detail.

1.6.6 Health Outcomes As mentioned previously, we did not …nd any di¤erence in the percentage of individuals in treatment versus control households with health shocks lasting 7 or more days or a death (Table 1.9). Although this study was not designed to have much chance to measure such ben- e…ts, we also measure how SKY insurance a¤ects objective measures of children’s health (BMI and height-and weight-for-age). Insurance had no detectable e¤ect on either measure (Table 1.9).

1.6.7 Trust in Providers and SKY SKY posited that increased exposure to the public sector (coupled with SKY’s selection of higher quality facilities and assistance to facilities) meant SKY members would raise their views of public doctors. To households visiting a public doctor in the three months prior to the Round 2 survey, we asked respondents their level of agreement with three statements regarding government doctors: “Government doctors are extremely thorough and careful”, “You have complete trust in government doctors”and “Government doctor’s medical skills are not as good as they should be”(reverse coded), each on a 1-5 scale. The mean was about 4 on each question, re‡ecting fairly good opinion of government doctors. SKY membership did not have a detectable e¤ect on trust in or con…dence in the skills of public-sector doctors (Table 1.10). This lack of improvement may be because there was no increase in usage of public facilities (for general care in the last three months), or perhaps because SKY did not increase quality. We measure views on SKY with agreement that SKY will pay, is honest, and is trustworthy (averaging scores in the three questions, each measured on a 1-5 scale). Con- sistent with SKY’s theory of change, our IV results show that SKY membership increases trust in SKY by 0.3 points on our scale, compared to a mean amongst the control of 3.4 (P < 0.001, Table 1.10). When we restrict the sample to those who have experienced a serious health incident, the e¤ect is larger (an increase of 0.42 points for insured households over the baseline mean of 3.4, P < 0.001, results not shown).

1.7 Robustness Checks

For many of the outcomes above, we ran tests on several sub-groups, for example, sometimes including only households with major shocks or without. In some instances we included health incidents not only that resulted in a death or seven or more day illness, but also those incidents that failed those criteria but on which more than 100USD was spent 20

on care. We also varied the cuto¤ for some economic outcomes, testing the percentage of incidents or households with expenditures above $5, $50, $100, etc. In most cases these changes did not a¤ect results, and when they did the di¤erence in outcome is mentioned above. Changes in our de…nition of SKYit in equation 1.2 also did not change general results. We also re-ran results using coupon status as an instrument for SKY purchase, rather than the interaction of coupon status and months since SKY. These results were very similar to the main results, and are presented in Appendix 1.B. Our randomization tests showed that high coupon households were slightly richer at the start of our study, meaning that some di¤erences in outcomes may have already been present before SKY. To test whether pre-SKY di¤erences were in‡uencing results, for a few variables we included the value of the variable at the time of the …rst round survey (Table 1.14). While statistical signi…cance decreased below the 5% level for some outcomes, the general results were the same. As noted above, because the …rst round survey was administered several months after the start of insurance, these results may be somewhat biased downwards. Future analyses will control for baseline characteristics for additional variables. To further analyze the data, forthcoming analyses will subdivide the data to test outcomes on various sub-populations. We will examine whether the e¤ects of SKY vary by age, gender, and wealth, or by whether the household member is permanently ill or disabled. Proximity to a higher quality public facility may in‡uence the impact of SKY, and we will examine this possibility as well.

1.8 Conclusion

SKY has several goals. First, it is trying to shift rural Cambodians from unregulated private providers and drug sellers to the public system. It appears to be successful in this regard. SKY also reduces expensive private care, though not by as much as we had anticipated. SKY aims to reduce delays prior to receiving quali…ed care. We do not …nd any reduction in delay prior to …rst care, but the uninsured often self-medicate from unquali…ed drug-sellers. Our measure is limited as we cannot distinguish delays prior to quali…ed providers. SKY’shope is that higher exposure to health messages at public health centers will increase preventive care such as immunizations and prenatal care. We do not …nd evidence of these e¤ects. Given that some forms of preventative care are already free (e.g., vaccines), it is perhaps not surprising that insurance did not increase this type of care. As in the general literature, it easier to detect changes in utilization than improvements in health. The sample size and timeframe of our study meant that we did not have statistical power to detect meaningful improvements in health. Thus, while we …nd no signi…cant impacts of SKY on health, we cannot draw any conclusions from this result. One the one hand, it is possible that SKY indeed has no impact on health: Treatment at public facilities is often a replacement for other types of care (private or drug sellers). Treatment at public facilities may not actually improve health compared to treatment at 21

other facilities, or if care is poor enough, may not improve health at all. On the other hand, even if public health centers are better than these other types of care and truly do improve health, they may not represent a big enough improvement in quality to cause a measurable di¤erence over our short time horizon and using our survey sample. While the above impacts focus on health and health care, health insurance is primarily designed to protect against economic loss. The e¤ects of SKY were typically larger on economic outcomes than on utilization. SKY reduced total medical expenses following a major health shock, and this reduction was largely due to lower rates of large expenses. SKY households also had lower accumulation of debt due to health problems, and were less likely to sell productive assets to pay for a large shock. Insured households were less likely to sell land to pay for a health issue, and had higher overall values of livestock than uninsured households. Our results suggest that most uninsured households will probably take on debt to pay for health care at some point in their lives.9 A substantial minority of those households will also sell productive assets such as land. SKY health insurance cuts the rates of these events by about a third. Importantly, the overall savings to insured households compare favorably with the cost of insurance for these households. On average, households pay 1.65 USD per month (taking into account average household size of SKY buyers), or 19.80 USD for a year of membership. Our calculations show a decrease in expenditures of 57.80 USD over the last 12 months for insured households (Table 1.6), and even higher reductions using uncensored results. Thus, assuming the value of SKY to a consumer equals averted out-of-pocket costs (ignoring any social cost, and any added or subtracted value of using a public facility over an alternative form of care or no care), the value outweighs the cost of insurance for the insured. If private care or self treatment via drug sellers is actually harmful, then our estimates of value to SKY members is an underestimate, as we are not including any value of averting private care. In addition, this calculation of bene…ts does not include any averted interest payments due to decreased loans for health care. Conversely, if public care is harmful, our estimates of bene…ts are an overestimate. As noted above, using our randomized price as an instrument estimates the e¤ect of insurance on the roughly third of households who purchase insurance due to the deeply discounted price. This price-sensitive group is relevant for business and public policy, as these customers are probably the most likely to purchase insurance if there were a greater subsidy, successful new marketing techniques, and so forth. At the same time, the e¤ects of insurance on this group are probably not representative of the e¤ects of insurance on the entire population. For example, a companion paper (Chapter 2) demonstrates substantially more self-selection among the 4% of the population who paid full price for SKY insurance than for the larger group who bought insurance only at a deeply discounted price. To the extent those who anticipate the greatest bene…ts of insurance buy insurance even at the full price, their bene…ts from insurance will be higher than our estimates. 9This back-of-the-envelope calculation assumes health care shocks are fairly independently distributed over time and ignores that dropout rates of SKY are high so that under current trends few households will be members of SKY for decades. 22

Conversely, those who decline insurance even with the steep discount may correctly expect low bene…ts, perhaps because they are unlikely to need health care or because they live far from high-quality public facilities. In that case, the never-buying group would have fewer bene…ts from insurance than our estimates. At the same time, those who decline insurance even with the steep discount may be the very poorest, those who do not understand insurance, or those who do not trust western medicine. If these groups could bene…t strongly from insurance, there are scenarios under which our estimated bene…ts will be below those expected for the never-buying group. Thus, it is di¢ cult to be sure how expansion to universal insurance would a¤ect this part the population. Using months since meeting as an additional instrument has similar caveats. House- holds that remain insured for a longer period of time are those who anticipate the largest bene…ts from SKY. In that sense, our estimates may overestimate the impacts on the population as a whole. In companion papers (Chapters 2 and 3) we …nd that SKY purchasers are not very di¤erent from those who decline on most observable factors such as education or risk aversion. At the same time, SKY members tend to have had more health problems prior to purchasing SKY, particularly if they paid the full price. We also provide evidence that SKY members paying the regular price have worse health in ways that we do not initially observe than SKY members buying with the deep discount provided by the high coupon. Speci…cally, holding constant measures of health observed at the baseline, SKY members who paid the regular price tend to use SKY facilities substantially more than those who purchased SKY with a high coupon. This gap in health care usage is predicted by theories of adverse selection. These results are relevant to our study, as it means that the SKY purchasers in‡uenced by the high coupon (the group of SKY members we analyze) have health and expected health care expenditures that are much more similar to others in their communities than are those who purchase SKY at the regular price. In addition to limitations of our identi…cation strategy, our measures all had limitations. For example, we did not measure the quality of private care. Thus, it is hard to tell if SKY increased e¤ective care, or simply replaced private with public care. As noted, the study was too small to detect changes in health along with several other longer-term outcomes. It bears repeating that “absence of evidence is not evidence of absence,” so it is possible that health insurance leads to long-term bene…ts for these outcomes. This study examines one insurer operating in a few regions of a single nation. We need more studies that rigorously evaluate micro-insurance and other innovations in health care …nancing. The low take-up of voluntary health insurance emphasizes the importance of other programs to increase access to health care for the rural poor (Bitran, Turbat, Meesen, and Van Damme 2011). SKY itself is managing one of Cambodia’s health equity funds, which provide free care for the rural poor. It is important to evaluate the impacts of health equity funds and other alternatives as a complement to this evaluation. 23

1.9 Tables 24

Offered Deep Offered Full Discount, Clustered Price, Mean Mean ttest Observ ations 2533 2536

Highest ranked wealth by enumerator 0.13 0.14 •0.98 Lowest ranked wealth by enumerator 0.14 0.10 3.96 ** Answered all literacy/numeracy questions correctly 0.15 0.15 0.13 Household Size 5.03 5.02 0.31 Education of health decision•maker (years) 4.61 4.72 •1.13 At least one household member with poor self• reported health 0.70 0.72 •1.15 At least one member over 65 0.25 0.26 •1.11 No child age 5 or under 0.55 0.57 •1.41 Household has a stunted or wasted child under age 6 0.16 0.15 0.88 All vaccines fulfilled for members under 6, 0 if no under 6, pre•mtg 0.27 0.25 0.96 Miss 7 or more days of work or death due to illness, 2 to 4 months pre•Meeting 0.07 0.07 0.07 Major health shock (*) and used health center for care (0 if no shock) 0.01 0.02 •0.97 Major health shock (*) and used hospital for care (0 if no shock) 0.02 0.02 0.22 Major health shock (*) and use private health care (0 if no shock) 0.05 0.05 •0.06 Ln of max days ill for a major health shock (*), pre meeting (0 if no shock) 0.22 0.23 •0.44 Major health shock (*) and spent 120,000 riel on care (USD30) (0 if no shock) 0.04 0.04 •0.34 Khmer household 0.953 0.946 2.00 * Ln of approximate value of animals, durables, and business (USD) 6.47 6.49 •0.64 Ln of approximate value of animals, durables, business, cash, and gold (USD) 6.68 6.74 •1.91

Area of farm land owned by household (hectares) 0.81 0.86 •1.05 Area of village land owned by household (hectares) 0.14 0.13 0.90 Household has at least one toilet 0.26 0.26 0.34 House made of palm 0.04 0.03 2.23 * Roof made of palm 0.05 0.04 1.40 Roof made of tin 0.37 0.38 •0.53 Roof made of tile 0.51 0.52 •0.66 House made of brick 0.03 0.03 •0.41 All variables are from the baseline survey. Sample is all high coupon households and all low coupon households in the randomized sample. Ttest clustered at village level. * Major shock includes all shocks causing 7 or more days of missed work or death.

Table 1.1: Randomization Test 25

Avg SKY Membership Prior, Post, Following Incident High Coupon 0.371*** (13.45)

Months Since Mtg 0.00227 •1.68

High Coupon Interaction With Months •0.00847** Since Mtg (•3.03)

Constant 0.0442*** •4.36 Observations 4009 Adjusted R 2 0.1502 F•Test 129.8

Table 1.2: First Stage Regression for Incident-level Outcomes, Round 1 and 2 Incidents Used 26 3887 3887 2429 2429 3887 IV N 1.19 2.451 IV T• •1.305 •2.488 •0.094 Statistic Impact on the Insured IV 1.628 (0.83) (0.06) (1.37) (0.07) (0.02) •0.032 •0.007 2.037* •0.143* Difference N 4207 4207 2749 2749 4207 1.413 2.181 •1.839 •1.785 •0.418 T•Statistic 0.49 (0.23) (0.02) (0.35) (0.02) (0.01) •0.013 •0.029 •0.008 0.505* Difference Intention to Treat 0.052 0.594 3.346 5.001 0.519 (0.18) (0.01) (0.23) (0.01) (0.01) Control Mean 0.04 3.851 5.491 0.511 (0.18) (0.02) (0.29) (0.02) (0.01) 0.565 Treatment Table 1.3: Health Utilization Following a Major Health Shock Stopped treatment because of no money until first treatment.Days Top•coded Never treated is 30 at 30 days. days. Percent receiving treatment on first of illnessday until hospital. Top•codedDays at 30 Never went to hospital as 30 days. days. Percent visiting hospital on first day of illness Foregone care Delayed Care Following a Major Health Shock All health incidents are for a death or 7 or more disabled. days Endogenous Variable: Average SKY status for months prior to, during, and post the incident. Instrument : months between incident and meeting, coupon status, and interaction between the two. incidents in Round until hospital uses 2 of data only Days collection. All other outcomes use incidents in Round 1 and Round 2. * p < 0.05, ** 0.01, *** 0.001 27 3887 3887 3887 3887 3887 3887 3887 3887 3887 IV N 0.88 •2.14 •3.56 0.016 4.333 3.532 IV T• •2.339 •0.036 •0.724 Statistic Impact on the Insured 0 IV 0.014 (0.04) (0.04) (0.05) (0.04) (0.05) (0.02) (0.01) (0.05) (0.02) •0.002 •0.011 •0.082* •0.113* 0.176*** 0.174*** •0.181*** Difference N 4207 4207 4207 4207 4207 4207 4207 4207 4207 T• 0.23 •3.56 1.038 4.011 3.527 •2.308 •2.025 •0.313 •0.521 Statistic 0.005 0.003 (0.01) (0.01) (0.01) (0.01) (0.02) (0.01) (0.00) (0.01) (0.01) •0.001 •0.002 •0.024* •0.031* 0.047*** 0.050*** •0.051*** Difference Intention to Treat 0.141 0.299 0.143 0.468 0.026 0.008 0.646 0.028 0.157 (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) Control Mean 0.16 0.188 0.349 0.118 0.437 0.032 0.008 0.595 0.025 (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) Treatment Table 1.4: Provider Type, First Treatment after Major Health Incident < 0.001 p *** < 0.01, p ** < 0.05, p Was the incident first treated at a public hospital? Was the incident first treated at a health center? Was the incident first treated at a public hospital or health center? Was the incident first treated at a drug seller? Was the incident first treated at a private doctor? Was the incident first treated with Kru Khmer? Was the incident first treated at an NGO? Was the incident first treated at a non•public place? Was the incident first treated at another place? All incidents for a death or 7 or more disabled. days Endogenous Variable: Average SKY status for months prior to, during, and post the incident Instrument : months between incident and meeting, coupon status, and interaction between the two * 28 436 337 337 436 436 337 310 337 IV N 0.34 1.20 1.88 0.51 1.631 0.693 IV T• •0.789 •1.009 Statistic Impact on the Insured IV 0.03 0.21 0.31 0.02 0.124 0.091 (0.09) (0.17) (0.17) (0.13) (0.11) (0.14) (0.05) (0.19) (0.08) •0.193 Difference N 436 337 337 337 436 310 337 436 T• 0.87 1.48 0.41 1.509 0.638 (0.76) •0.041 •0.972 Statistic 0.05 0.08 0.01 0.037 0.026 (0.03) (0.06) (0.05) (0.04) (0.04) (0.02) (0.05) (0.02) •0.033 •0.001 •0.052 Difference Intention to Treat 0.92 0.69 0.59 0.64 0.02 0.926 0.796 0.178 (0.02) (0.04) (0.04) (0.03) (0.03) (0.01) (0.04) (0.02) Control Mean 0.63 0.72 0.03 0.963 0.919 0.763 0.204 (0.02) (0.04) (0.04) (0.03) (0.03) (0.01) (0.04) (0.02) 0.639 Table 1.5: Birth-Related Utilization Treatment 1 2 2 1 2 2 1 Received at least one antenatal check•upReceived at least one antenatal Received at least one tetanus injection during pregnancy Gave birth in a public facility Gave birth in a public or private health facility a trained birth attendant Assisted at birth by Assisted at birth by a midwife a doctor Assisted at birth by Received at least one postnatal check•up Antenatal Care Antenatal Birth Post•Natal Care births listed in Round 2 survey. births in Round 1 and 2, except post•natal care which uses only Sample includes post•SKY status for months Average SKY prior to, during, and afterEndogenous variable: the birth Instrument: months since meeting, coupon status, and interaction of the two. 1: Includes most recent birth 3 or more months after the first start possible date. SKY 2: Using most recent birth after the first possible start date of SKY. 29 3887 2128 2128 3887 3887 3887 2128 IV N •2.4 •2.75 IV T• •2.384 •2.609 •1.305 •2.232 •1.846 Statistic Impact on the Insured IV •0.05 (0.03) (0.06) (0.04) (0.06) (0.03) •0.109 (19.20) (22.15) •0.123* •0.070* •0.086** •45.789* •57.804** Difference N 4207 2128 4207 4207 2128 2128 4207 •2.326 •2.555 •2.328 •2.328 •1.747 •1.081 •2.718 T•Statistic (5.76) (7.24) (0.01) (0.02) (0.01) (0.02) (0.01) •0.035 •0.014 •0.036* •0.020* •0.025** •13.404* •18.493* Difference Intention to Treat 0.11 0.382 0.115 0.619 0.097 (4.59) (5.73) (0.01) (0.02) (0.01) (0.01) (0.01) 132.43 103.811 Control Mean 0.084 0.347 0.101 0.583 0.076 (4.63) (5.33) (0.01) (0.02) (0.01) (0.01) (0.01) 90.407 113.94 Treatment 1 1 Table 1.6: Economic Impacts Following a Major Health Incident < 0.001 p *** < 0.01, p ** Total USD spent on care for a given incident a household on all Total USD spent on care by major health incidents in the last 12 months Share of incidents with total cost greater than 250USD Share of all households spending more than 100USD total on all major health incidents Share of all households spending more than 350USD total on all major health incidents Share of incidents with total cost greater than 5USD on a private provider Share of incidents with total cost greater than 150USD on a private provider < 0.05, Amount spent on care p Following a Major Health Shock All health incidents are for a death or 7 or more disabled. days variable, see text.Endogenous variable: Varies by Instrument : months between incident and meeting, coupon status, and interaction between the two. 1. Compressed to 98th percentile to remove outliers. * 30 IV N 3887 3887 3887 3887 3887 3887 3887 3887 IV T• •1.346 •2.615 •0.434 •0.902 •1.157 •1.977 •1.799 14.951 Statistic Impact on the Insured IV (0.03) (0.06) (0.03) (0.05) (0.03) (0.05) (0.04) (0.05) •0.077 •0.011 •0.044 •0.037 •0.064 •0.092* •0.136** 0.438*** Difference N 4207 4207 4207 4207 4207 4207 4207 4207 T• •1.47 •2.43 •0.087 •1.098 •1.117 •2.472 •2.028 12.329 Statistic •0.03 (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.02) •0.001 •0.016 •0.011 •0.032* •0.021* •0.035* 0.133*** Difference Intention to Treat 0.034 0.481 0.067 0.229 0.101 0.224 0.128 0.196 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Control Mean 0.09 0.16 0.167 0.457 0.066 0.213 0.191 0.107 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Treatment < 0.001 Table 1.7: Method of Payment following a Major Health Incident p *** < 0.01, p ** < 0.05, p Is sky used to pay for any of the used to pay Is sky treatments? for any of the Is cash used to pay treatments? for any used to pay Are savings of the treatments? of the for any pay Does family treatments? for any of the used to pay Is work treatments? for any of used to pay Are assets the treatments? used interest Are loans without for any of the treatments? to pay used to interest Are loans with of the treatments? pay for any All incidents for a death or 7 or more days disabled. Endogenous Variable: Average SKY status for months prior to, during, and post the incident two the between and interaction status, meeting, coupon and incident between Instrument : months * 31 4980 4980 4980 3528 4980 4980 4980 4980 IV N IV IV T• 2.096 1.524 1.462 •2.413 •1.314 •1.949 •3.558 •2.836 Statistic Impact on the Insured IV (6.27) (0.03) (0.03) (0.01) 0.087 (0.06) 0.046 (0.03) •0.035 •0.016 (28.37) (46.25) 96.945* •0.077** •68.469* •22.316*** Difference N 4980 3528 4980 4980 4980 4980 4980 4980 •2.29 0.917 0.824 1.091 •2.458 •1.485 •3.699 •2.932 T•Statistic (1.86) (0.01) (0.01) (0.00) 0.016 (0.02) 0.008 (0.01) (8.52) •0.012 14.797 •0.006* (13.56) •0.024** •20.937* •6.877*** Difference Intention to Treat (1.81) 0.089 (0.01) 0.093 (0.01) 0.011 (0.00) 0.023 (0.02) 0.831 (0.01) (10.07) 28.943 (18.03) 540.488 194.708 Control Mean (9.18) (1.49) 0.065 (0.01) 0.081 (0.01) 0.005 (0.00) 0.039 (0.02) 0.839 (0.01) 22.066 (17.67) 555.285 173.771 Treatment Table 1.8: Overall Economic Impacts on Households Amount borrowed intotal borrowed Amount health to related all Totalof loans value health to due year last than debt More birth a or reasons previous village the land than or farm Less year previous village the land than or farm Less reasons health to due year Total value of farm animals, USD, percentile 98th at compressed animal,gold, cash, for z•score Average value business and asset, in enrolled 6•17 ages children of Percent school Payment for care for Payment Capital Assets/Human Productive Ov erall Economic Impacts on Households two. the of interaction and status, meeting,coupon since months Instrument: * p < 0.05, p < 0.01, ** p < 0.001 *** 32 2206 2217 2207 IV N 24741 24560 •0.34 0.253 0.442 IV T• •0.641 •0.012 Statistic Impact on the Insured IV 0.001 0.071 (0.00) (0.16) (0.17) (0.13) (0.01) •0.007 •0.057 •0.001 Difference ` N 2221 2232 2222 24865 24684 •0.01 0.321 •0.079 •0.149 •0.114 T•Statistic 0.00 0.00 (0.00) (0.00) (0.05) (0.05) (0.04) •0.001 •0.008 •0.005 Difference Intention to Treat •0.69 0.007 0.102 (0.00) (0.00) (0.04) (0.03) (0.03) •1.385 •1.364 Table 1.9: Health Impacts Control Mean 0.007 0.102 (0.00) (0.00) (0.05) (0.04) (0.03) •1.386 •0.698 •1.369 Treatment Major Health ShocksMajor Percent of individuals who died in the last year Percent of individuals sick for 7 or more in the last year days Anthropometrics Length/height•for•age z•score BMI•for•age z•score Weight•for•age z•score see text. variable, by varies Endogenous variable: Instrument: months since meeting, coupon status, and interaction of the two. * p < 0.05, ** 0.01, *** 0.001 33 4929 1143 IV N IV IV T• 4.898 •1.405 Statistic Impact on the Insured IV (0.13) (0.06) •0.176 0.303*** Difference N 4929 1143 T• 5.396 •1.325 Statistic (0.04) (0.03) •0.054 0.147*** Difference Intention to Treat (0.03) (0.03) 3.976 3.411 Control Table 1.10: Trust in Providers and SKY Mean (0.03) (0.03) 3.923 3.558 Treatment < 0.001 < p *** < 0.01, < 1 p ** Trust of Public Doctors Public of Trust all over score (average questions) of heard (never inSKY Trust lowSKY trust) coded as < 0.05, < Trust of Public Providers Public of Trust of SKY Trust p Instrument: months since meeting, coupon status, and interaction of the two. the of interaction and status, coupon meeting, since months survey. the to Instrument: prior months three inthe publicprovider a visited who households only Includes 1: * 34

1.10 Figures 35

Pilot testing to determine feasibility of randomization and necessary sample size (January – February 2007; 34 Village Meetings; Distribution of 325 five•month coupons, 748 one•month coupons)

Insurance Agent and Member Facilitator Qualitative Interviews: August 2007 (N = 26)

Phase 1 Village Meetings: Clinic survey: Phase 2 Village Meetings: November 2007 – May 2008 August • November September 2008 – December 2008 (N = 142 Villages, Distribution of (N = 103 Villages; Distribution of 1342 five•month coupons, 1342 2008 (N = 38) 1275 five•month coupons, 1276 one•month coupons selected at one•month coupons selected for random for control group. Village leader survey: control group) Maps of village households and October • December Maps of village households and location of health facilities and 2008 (N = 245) location of health facilities and workers workers

Phase 1 Baseline Survey: July • August 2008 Phase 2 Baseline Survey: (Interviewed 1305 five•month December 2008 coupon households, 1296 1•month (Interviewed 1256 five•month coupon households, plus 133 coupon households, 1252 1•month additional 1•month households not coupon households, plus 67 part of random sample (not used in additional 1•month households (not impact analysis)) used in impact analysis))

Village monographs: March • April 2009 (N = 7 villages, not part Phase 2 Round 2 Survey: Phase 2 Round 2 Survey: of impact evaluation) July • August 2009 December 2009 – January 2010 (Interviewed 1281 five•month (Interviewed 1221 five•month coupon households, 1282 1•month coupon households, 1224 1• coupon households plus 200 month coupon households, plus additional 1•month households not 72 additional 1•month part of random sample (not used in households not part of random impact analysis)) sample (not used in impact analysis))

Figure 1.1: Timeline of Evaluation 36

0.5

High Coupon Households 0.45

0.4

0.35

0.3

0.25

0.2

0.15

0.1

Low Coupon Households 0.05

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Months Since Meeting

Figure 1.2: Proportion in SKY, by Months since Village Meeting and Coupon Type 37

1.A Supplementary Tables 38 IV N IV 3887 3887 3887 3887 3887 3887 3887 3.56 IV T• IV 0.517 4.929 0.013 •2.117 •1.816 •0.597 Statistic Impact on the Insured Impact 0 IV (0.06) (0.04) (0.06) (0.04) (0.05) (0.03) (0.01) 0.029 •0.094 •0.018 •0.085* 0.219*** 0.200*** Difference N 4207 4207 4207 4207 4207 4207 4207 T• 1.052 4.565 3.704 •2.121 •1.929 •0.385 •0.439 Statistic (0.02) (0.01) (0.02) (0.01) (0.02) (0.01) (0.00) 0.017 •0.028 •0.003 •0.002 •0.026* 0.060*** 0.062*** Difference Intention to Treat Intention 0.18 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) 0.269 0.413 0.652 0.102 0.018 0.175 Control Mean 0.24 0.15 (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) 0.286 0.475 0.624 0.098 0.017 Treatment Table 1.11: Health Utilization after Major Health Incident - Ever Treated at Given Provider type Was the incident ever treated at a public hospital? Was the incident ever treated at a health center? Was the incident ever treated at a public hospital or health center? Was the incident ever treated at a drug seller? Was the incident ever treated at a private doctor? Was the incident ever treated with Kru Khmer? Was the incident ever treated at an NGO? disabled. days All incidents for a death or 7 more and post the incident SKY status for months prior to, during, Average Endogenous Variable: status, and interaction between the two : months between incident and meeting, coupon Instrument * p < 0.05, ** 0.01, *** 0.001 39 IV N 4980 2789 4980 3272 3272 IV T• IV 0.089 0.073 0.165 •1.145 •1.286 Statistic Impact on the Insured the on Impact IV 0.01 0.005 0.005 (0.01) (0.06) (0.06) (0.07) (0.06) •0.009 •0.078 Difference N 4980 4980 2805 3292 3292 0.17 •0.02 0.047 •1.137 •0.182 T•Statistic 0.00 0.002 0.001 (0.00) (0.01) (0.02) (0.02) (0.02) •0.003 •0.004 Difference Intention to Treat Intention 0.42 0.009 0.305 0.312 0.253 (0.00) (0.01) (0.02) (0.01) (0.01) Control Mean Table 1.12: General Health Utilization 0.006 0.307 0.417 0.311 0.253 (0.00) (0.01) (0.02) (0.01) (0.01) Treatment At least one household member did member household one least At in funds care due to lack of not get the last 12 months has visited a Household member three the last doctor in government months of up to date at time All shots 6 or under age children for survey Currently using contraception Currently using modern contraception Foregone care Foregone Care Preventative Other Health Seeking Behavior of age years under 6 and Age – Subpopulation Immunized 45 – 16 Age Women Married Subpopulation: Contraceptives sinceInstrument: months meeting, coupon status, and interaction of the two. p < 0.001 p < 0.01, *** p < 0.05, ** * 40 2128 1528 2128 2128 2128 2128 2128 2128 IV N 0.6 •2.19 •1.39 •3.15 3.047 2.077 IV T• •2.185 •3.233 Statistic Impact on the Insured IV 0.029 (0.05) (0.04) (0.02) (0.08) (0.05) •0.053 0.158* (41.07) (11.70) (62.44) •0.032* •0.155** •89.741* •36.853** 190.246** Difference N 2128 2128 1528 2128 2128 2128 2128 2128 T• 2.004 1.493 •2.426 •0.072 •1.427 •2.653 •3.472 •3.341 Statistic 0.04 (3.78) (0.02) (0.01) (0.01) (0.03) (0.02) •0.001 •0.018 (13.59) •20.414 40.905* •0.012** •32.951* •0.054*** •12.631*** Difference Intention to Treat 0.02 0.83 0.162 0.106 (3.22) (0.01) (0.01) (0.00) (0.02) (0.01) •0.047 49.409 (13.10) •20.994 234.609 484.266 Control Mean 0.109 0.088 0.008 0.829 (2.63) (0.01) (0.01) (0.00) (0.03) (0.01) •0.007 36.778 (11.18) •20.502 201.659 525.171 Treatment Table 1.13: Overall Economic Impacts, Households with Health Incidents Amount borrowed in total Total value of all loans related to health due to health More debt than last year reasons or a birth Less farm land than the or village previous year Less farm land than the or village due to health reasons previous year of farmTotal value animals, USD, compressed at 98th percentile Average z•score for cash, gold, animal, asset, and business value Percent of children ages 6•17 enrolled in school Payment for care Productive Assets/Human Capital Overal Economic Impacts on Households Instrument: months since meeting, coupon status, and interaction of the two. Sample: Incidents are All households with incidents. for a death or 7 disabled. more days p < 0.001 p < 0.01, *** * p < 0.05, ** 41

Impact on the Insured

IV Intercept IV Difference IV T• Statistic

Visited a health center following a health incident1 0.071 0.089** 3.21 (.007) (.028)

Total spent on care following a health incident1 47.418 •24.549* •2.15 (2.933) (11.392) Total spent on private care2,3 25.939 •11.145 •1.61 (1.765) (6.923) Spent more than 250USD total on care of health incident2 0.073 •0.046 •1.92 (.006) (.024) Total value of animals3 168.398 62.092 1.66 (10.649) (37.382) Total debt amount3 108.367 •47.679 •1.81 (6.941) (26.381) N = 4979 for all variables All outcomes are calculated at the household level. ` Round 1 survey levels of variables held constant in all regressions. Instrument: months since meeting, coupon status, and interaction of the two. 1. Health incident includes a death or incident with 7 or more days unable to perform daily actiities. 2: Health incident includes the above or one that cost over 100USD. 3. Compressed at 98th percentile

Table 1.14: Instrumental Variables Regressions holding constant Round 1 Values 42 •4.87 •1.33 24741 (•2.59) (•0.37) 125.38 0.1035 •0.0211 0.00491 0.490*** •0.0164* Sky Status Sky Last 4 Months SKY •7.35 •1.65 24741 (•4.17) (•0.60) 232.82 0.1939 •0.0297 0.00531 0.705*** •0.0251*** Percent Year in •4.12 •1.93 24741 (•0.15) (•0.58) 320.97 0.2366 •0.0363 0.00789 0.466*** •0.00105 Ever in SKY •2.28 •1.37 90.71 24741 Status 0.227* (•0.34) (•0.41) 0.0727 •0.0234 0.00505 •0.00212 Current SKY < 0.001 p *** < 0.01, Table 1.15: First Stage Regression for Individual-level Outcomes, Round 2 Data Used p 2 ** R < 0.05, p statistics in parentheses High Coupon Months Since Mtg High Coupon Interaction With Months Since Mtg Constant Observations Adjusted F•Test t * 43 4980 •4.32 •0.88 •0.16 (•1.77) 144.16 0.1002 0.00286 0.00803 0.375*** •0.00951 Sky Status Sky Last 4 Months SKY 4980 •6.52 •1.18 (•0.03) (•3.16) 244.86 0.1887 0.00335 0.579*** •0.00137 •0.0175** Percent Year in •0.6 4980 •3.47 •1.49 (•0.04) 337.05 0.2314 0.00408 0.00549 0.377*** •0.00232 Ever in SKY •1.5 4980 •0.92 •0.77 •0.12 0.125 Status 0.0061 109.01 0.0719 0.00398 0.00297 Current SKY < 0.001 p *** < 0.01, Table 1.16: First Stage Regression for Household-Level Outcomes, Round 2 Data Used p 2 ** R < 0.05, p statistics in parentheses High Coupon Months Since Mtg High Coupon Interaction With Months Since Mtg Constant Observations Adjusted F•Test t * 44

Avg SKY Membership Prior, Post, Following birth High Coupon 0.381*** •6.08

Months Since Mtg •0.00129 (•0.44)

High Coupon Interaction With Months Since •0.00919 Mtg (•1.36)

Constant 0.0709** •2.81 Observations 436 Adjusted R 2 0.1663 F•Test 23.77 t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

Table 1.17: First Stage Regression for Birth-Level Outcomes, Rounds 1 and 2 Data Used 45

1.B Instrumental Variable Results Using Coupon as Instru-

ment 46

Avg SKY Membership Prior, Post, Following Incident High Coupon 0.301*** (18.72)

Constant 0.0627*** (7.32) Observations 4028 Adjusted R 2 0.1461 F•T e st 350.61 t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

Table 1.18: IV using Coupon as Instrument: First Stage Regression for Incident-Level Outcomes, Rounds 1 and 2 Data Used 47 (8.48) 24741 (17.94) 0.1009 322.01 0.229*** 0.0572*** Sky Status Sky Last 4 Months 4 Last 570 SKY (9.18) 24741 (23.89) 0.1865 0.305*** 0.0550*** Percent Year in Year Percent 905 24741 (30.10) 0.2354 (11.20) 0.450*** 0.0895*** Ever in SKY Ever (8.41) Status 24741 (15.56) 0.0721 242.08 0.193*** 0.0570*** Current SKY Current < 0.001 p *** < 0.01, p 2 ** R < 0.05, statistics inparentheses statistics p High Coupon High Constant Observations Adjusted F•T e st t * Table 1.19: IV using Coupon as Instrument: First Stage Regression for Individual-Level Outcomes, Round 2 Data Used 48 4980 (9.02) (20.41) 0.0993 416.69 0.224*** 0.0535*** Sky Status Sky Last 4 Months 4 Last SKY 4980 (9.78) (26.10) 0.1848 681.31 0.301*** 0.0518*** Percent Year in Year Percent 4980 984.2 (31.37) 0.2302 (11.86) 0.442*** 0.0849*** Ever in SKY Ever 4980 (8.96) Status (17.69) 0.0709 312.92 0.189*** 0.0533*** Current SKY Current < 0.001 p *** < 0.01, p 2 ** R < 0.05, statistics inparentheses statistics p High Coupon High Constant Observations Adjusted F•T e st t * Table 1.20: IV Using Coupon as Instrument: First Stage Regression for Household-Level Outcomes, Round 2 Data Used 49 436 (8.33) (3.69) 69.45 0.1565 0.313*** Avg SKY Avg 0.0615*** Prior, Post, Prior, Membership Following birth Following < 0.001 p *** < 0.01, p 2 ** R < 0.05, statistics inparentheses statistics p High Coupon High Constant Observations Adjusted F•T e st t * Table 1.21: IV using Coupon as Instrument: First Stage for Birth-Level Regressions, using R1 and R2 data 50 3889 3889 2431 3889 2431 IV N IV IV T• 2.019 0.959 •1.519 •1.565 •0.327 Statistic IV Impact on the Insured (0.02) (0.81) (1.38) (0.06) (0.08) 1.326 •0.037 •0.087 1.631* •0.025 Difference N 4207 2749 4207 4207 2749 T• 2.181 1.413 •1.785 •0.418 •1.839 Statistic 0.49 (0.01) (0.23) (0.35) (0.02) (0.02) •0.029 •0.008 •0.013 0.505* Difference Intention to Treat 0.594 (0.01) 3.346 (0.18) 5.001 (0.23) (0.01) 0.519 (0.01) 0.052 Control Mean 0.04 (0.01) (0.18) (0.29) (0.02) (0.02) 3.851 5.491 0.511 0.565 Treatment Table 1.22: IV Using Coupon as Instrument: Health Care Utilization following a Health Shock Stopped treatment because of no money no of because treatment Stopped at Top•coded treatment. until first Days days. 30 is treated Never days. 30 30 at Top•coded until hospital. Days days. 30 as hospital to went Never days. day first on treatment receiving Percent illness of of day first on hospital visiting Percent illness Foregone care Foregone Care Delayed Following a Major Health Shock Health Major a Following disabled. days more or 7 or death a for incident. the are Allincidents healthpost and during, to, prior months for status SKY Average Variable: Endogenous 2. Round status. and 1 Coupon in Round : incidents Instrument use outcomes All other collection. data of 2 inRound incidents only uses until hospital Days 0.001 < p *** 0.01, < p ** 0.05, < p * 51 IV N 3889 3889 3889 3889 3889 3889 3889 3889 3889 IV T• 3.273 0.968 4.117 •3.194 •0.072 •2.011 •1.893 •0.134 •0.546 Statistic Impact on the Insured IV 0.017 (0.04) (0.04) (0.05) (0.04) (0.05) (0.02) (0.01) (0.05) (0.02) •0.003 •0.102 •0.001 •0.009 •0.076* 0.160** •0.163** 0.163*** Difference N 4207 4207 4207 4207 4207 4207 4207 4207 4207 T• 0.23 •3.56 4.011 3.527 1.038 •2.308 •2.025 •0.313 •0.521 Statistic 0.005 0.003 (0.01) (0.01) (0.01) (0.01) (0.02) (0.01) (0.00) (0.01) (0.01) •0.001 •0.002 •0.024* •0.031* 0.047*** 0.050*** •0.051*** Difference Intention to Treat 0.141 0.299 0.143 0.468 0.026 0.008 0.646 0.028 0.157 (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) Control Mean 0.16 0.188 0.349 0.118 0.437 0.032 0.008 0.595 0.025 (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.00) Treatment Was at a the incident first treated public hospital? Was at a the incident first treated health center? Was at a the incident first treated center? or health public hospital Wasa drug first treated at the incident seller? Was at a the incident first treated doctor? private Kru Was first treated with the incident Khmer? Wasan first treated at the incident NGO? Wasa non• first treated at the incident public place? Was first treated at the incident another place? All incidents for a death or 7 or more days disabled. Endogenous Variable: Average SKY status for months prior to, during, and post the incident status. : Coupon Instrument < 0.001 ** p < 0.01, *** * p < 0.05, Table 1.23: IV using Coupon as Instrument: Provider Type, First Treatment after a Major Health Incident 52 337 337 436 436 337 436 310 337 IV N 1.49 0.63 IV T• 0.88 •0.75 0.41 1.476 •0.965 •0.041 Statistic IV Impact on the Insured 0.16 0.02 0.121 0.259 0.083 (0.10) (0.18) (0.17) (0.13) (0.14) (0.05) (0.20) (0.08) •0.104 •0.191 •0.004 Difference N 337 337 337 436 436 310 337 436 T• 0.87 0.41 •0.76 1.478 0.638 1.509 •0.972 •0.041 Statistic 0.05 0.01 0.078 0.026 0.037 (0.03) (0.06) (0.05) (0.04) (0.04) (0.02) (0.05) (0.02) •0.033 •0.052 •0.001 Difference Intention to Treat 0.69 0.92 0.59 0.02 0.796 0.642 0.178 0.926 (0.02) (0.04) (0.04) (0.03) (0.03) (0.01) (0.04) (0.02) Control Mean 0.72 0.63 0.03 0.639 0.963 0.763 0.919 0.204 (0.02) (0.04) (0.04) (0.03) (0.03) (0.01) (0.04) (0.02) Treatment 2 1 2 Table 1.24: IV using Coupon as Instrument: Birth-Related Outcomes 2 1 1 2 Received at least one antenatal Received check•up at least one tetanus Received injection during pregnancy Gave birth in a public facility Gave birth in a public or private health facility Assisted at birth by a trained attendant midwife by a at birth Assisted doctor by a at birth Assisted at least one postnatal Received check•up Antenatal Care Antenatal Birth Post•Natal Care A birth is included in this sample if last birth is 3 months or more after SKY coverage. in 1st possible A birth is included 2 survey. births listed in Round post•natal care which only 1 and Round 2, except post•SKY births in Round Sample includes SKY status for Average months prior Endogenous variable: to, during, and post the birth Instrument: Coupon status. 1: Includes most recent birth 3 or more months after the first SKY start date. possible most recent birth after the first2: Using start date of SKY. possible 53 3889 2128 3889 3889 3889 2128 2128 IV N IV T• •2.001 •2.524 •2.277 •2.163 •2.001 •1.749 •1.077 Statistic IV Impact on the Insured (0.03) (0.06) (0.04) (0.06) (0.03) •0.107 •0.042 (19.14) (22.59) •0.071* •0.119* •0.058* •38.301* •57.011* Difference N 4207 2128 4207 4207 4207 2128 2128 •2.326 •2.555 •2.718 •2.328 •2.328 •1.747 •1.081 T•Statistic (5.76) (7.24) (0.01) (0.02) (0.01) (0.02) (0.01) •0.035 •0.014 •0.036* •0.020* •0.025** •13.404* •18.493* Difference Intention to Treat 0.11 0.382 0.115 0.619 0.097 (4.59) (5.73) (0.01) (0.02) (0.01) (0.01) (0.01) 132.43 103.811 Control Mean 0.084 0.347 0.101 0.583 0.076 (4.63) (5.33) (0.01) (0.02) (0.01) (0.01) (0.01) 90.407 113.94 Treatment < 0.001 p *** < 0.01, p 1 1 ** Table 1.25: IV using Coupon as Instrument: Economic Impacts following a Major Health Shock Total USD spent on care for a given incident Total USD spent on a household care by on all major health incidents in the last 12 months Share of incidents with total cost greater than 250USD Share of all households spending more than 100USD total on all major health incidents Share of all households spending more than 350USD total on all major health incidents Share of incidents with total cost greater than 5USD on a private provider Share of incidents with total cost greater than 150USD on a private provider < 0.05, Amount spent on care p Following a Major Health Shock All health incidents are for a death or 7 or more disabled. days variable, see text.Endogenous variable: Varies by Instrument : coupon status. 1. Compressed to 98th percentile to remove outliers. * 54 3889 3889 3889 3889 3889 3889 3889 3889 IV N IV IV T• 0.131 •1.112 •1.561 •1.211 •2.303 •2.257 •2.122 14.843 Statistic Impact on the Insured IV •0.09 (0.03) (0.06) 0.004 (0.03) (0.05) (0.03) (0.05) (0.04) (0.05) •0.054 •0.041 •0.119* •0.106* •0.075* 0.435*** Difference N 4207 4207 4207 4207 4207 4207 4207 4207 T• •1.47 •2.43 •0.087 •1.098 •1.117 •2.472 •2.028 12.329 Statistic •0.03 (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.02) •0.001 •0.016 •0.011 •0.032* •0.021* •0.035* 0.133*** Difference Intention to Treat 0.034 (0.01) 0.481 (0.01) 0.067 (0.01) 0.229 (0.01) (0.01) 0.224 (0.01) 0.128 (0.01) 0.196 (0.01) 0.101 Control Mean 0.16 0.09 0.167 (0.01) 0.457 (0.01) 0.066 (0.01) 0.213 (0.01) (0.01) 0.191 (0.01) 0.107 (0.01) (0.01) Treatment < 0.001 < p *** Table 1.26: IV using Coupon as Instrument: Method of Payment following a Major Health Incident < 0.01, < p ** < 0.05, < p Is sky used to pay for any of the treatments? the of any for pay to used sky Is treatments? the of any for pay to used cash Is the of any for pay to used savings Are treatments? treatments? the of any for familypay Does treatments? the of any for pay to used work Is the of any for pay to used assets Are treatments? any for pay to used interest without loans Are treatments? the of of any for pay to used interest with loans Are treatments? the disabled. days more or 7 or death a for Allincident incidents the post and during, to, prior months for status SKY Average Variable: Endogenous status. Coupon : Instrument * 55 IV N 4980 4980 3528 4980 4980 4980 4980 4980 IV T• •1.48 1.524 0.825 1.095 •2.268 •2.888 •2.425 •3.626 Statistic IV Impact on the Insured •0.04 0.087 0.027 (6.31) (0.03) (0.03) (0.01) (0.06) (0.03) 49.238 (28.73) (44.97) •0.019* •0.079** •69.668* •22.885*** Difference N 4980 4980 3528 4980 4980 4980 4980 4980 T• •2.29 0.917 0.824 1.091 •1.485 •2.932 •3.699 •2.458 Statistic 0.016 0.008 (8.52) (1.86) (0.01) (0.01) (0.00) (0.02) (0.01) •0.012 14.797 (13.56) •0.006* •0.024** •20.937* •6.877*** Difference Intention to Treat 0.089 0.093 0.011 0.023 0.831 (1.81) (0.01) (0.01) (0.00) (0.02) (0.01) 28.943 (10.07) (18.03) Control 540.488 194.708 Mean 0.065 0.081 0.005 0.039 0.839 (9.18) (1.49) (0.01) (0.01) (0.00) (0.02) (0.01) 22.066 (17.67) 555.285 173.771 Treatment Table 1.27: IV using Coupon as Instrument: Overall Economic Impacts on Households Amount borrowed in total Amount borrowed Total value of all loans related to health due to health year More debt than last reasons or a birth land than the Less farm or village year previous land than the Less farm or village due to health reasons year previous Total value of farm animals, USD, compressed at 98th percentile animal, for cash, gold, z•score Average asset, and business value Percent of children ages 6•17 enrolled in school Payment for care Payment Capital Productive Assets/Human Overal Economic Impacts on Households Instrument: Coupon status. * p < 0.05, ** 0.01, *** p < 0.001 56 2207 2206 2217 IV N 24741 24560 0.353 0.223 IV T• •0.114 •0.219 •0.015 Statistic IV Impact on the Insured 0.001 0.035 (0.00) (0.01) (0.16) (0.16) (0.13) •0.001 •0.036 •0.002 Difference ` N 2222 2221 2232 24865 24684 •0.01 0.321 •0.079 •0.149 •0.114 T•Statistic 0.00 0.00 (0.00) (0.00) (0.05) (0.05) (0.04) •0.001 •0.008 •0.005 Difference Intention to Treat •0.69 0.007 0.102 (0.00) (0.00) (0.04) (0.03) (0.03) •1.385 •1.364 Control Mean 0.007 0.102 (0.00) (0.00) (0.05) (0.04) (0.03) •1.386 •0.698 •1.369 Treatment Table 1.28: IV using Coupon as Instrument: Health Impacts Major Health ShocksMajor in the who died Percent of individuals last year Percent of sick individuals for 7 or in the last year more days Anthropometrics Length/height•for•age z•score BMI•for•age z•score Weight•for•age z•score see text. variable, by varies Endogenous variable: Instrument: Coupon status. * p < 0.05, ** 0.01, *** 0.001 57 1143 4929 IV N IV IV T• 4.898 •1.405 Statistic IV Impact on the Insured (0.13) (0.06) •0.176 0.303*** Difference N 1143 4929 T• 5.396 •1.325 Statistic (0.04) (0.03) •0.054 0.147*** Difference Intention to Treat (0.03) (0.03) 3.976 3.411 Control Mean 3.923 (0.03) 3.558 (0.03) Treatment Table 1.29: IV using Coupon as Instrument: Trust in Providers and SKY 1 < 0.001 < p *** < 0.01, < p ** Trust of Public Doctors (average Doctors Public of Trust all questions) over score SKY of heard (never inSKY Trust trust) low as coded < 0.05, < Trust of Public Providers Public of Trust of SKY Trust p Instrument: Coupon status. Coupon survey. the to Instrument: prior months three inthe publicprovider a visited who households only Includes 1: * 58

Chapter 2

Adverse Selection based on Observable and Unobservable Factors in Health Insurance

2.1 Introduction

Health insurance can increase access to health care and decrease the potentially catastrophic e¤ects of large medical expenses. However, asymmetric information in health insurance markets may stop voluntary health insurance markets from providing protection to most consumers without substantial regulation or subsidization. Adverse selection is one type of asymmetric information that can raise an insurer’s costs and reduce coverage. The economic theory of insurance demand predicts that households that anticipate high health care costs are those that are most likely to be willing to purchase health insurance (e.g., Rothschild and Stiglitz 1976; Akerlof 1970). Volun- tary health insurance cannot be …nancially sustainable if adverse selection is severe because only the most costly patients would …nd it worthwhile to purchase insurance, and insurers’ premium levels would not be able to cover the high costs of care. Results of empirical studies of the importance of adverse selection remain mixed. In addition, in many empirical studies the e¤ects of adverse selection are di¢ cult to disentangle from that of moral hazard. Because adverse selection and moral hazard have very di¤erent implications for insurers and policy-makers, it is important to understand which aspects of asymmetric information are important in di¤erent settings. We utilize a randomized experimental design that allows us to separate adverse selection from moral hazard. The experiment and data were collected during the expansion of the SKY Micro-Health Insurance program in rural Cambodia. SKY partners with public health facilities and provides free care in exchange for a small (subsidized) monthly premium. As SKY expanded, we distributed coupons for deeply discounted insurance to a random subset of potential SKY customers, inducing a sharp increase in take-up rate. Sur- veys from over 5000 families provide information on health and health care utilization both before and after these households had the opportunity to purchase SKY. SKY administrative data provides further information on members’health care utilization. 59

We use these data to test three implications of theories of asymmetric information. Our …rst hypothesis (H1) is that those with high expected health care costs (based on factors observable to the econometrician such as past health care utilization) will be more likely to buy insurance. We call this selection on observables. Hypothesis 2 (H2) is that insurance buyers paying full price have higher predicted health care expenditures than buyers paying a deeply discounted price. Our …nal hypothesis (H3) is that among the insured, those who bought insurance at full price have higher health care utilization after they are insured than those who bought insurance at a deeply discounted price, holding constant observable factors. We call this selection on unobservables. Consistent with hypothesis 1 (adverse selection based on observables), while 67% of households that declined SKY report having at least one member in poor health, an even higher share –80% –of buyers report a member in poor health. Similarly, buyers are 2.3 percentage points (33%) more likely than decliners to report major health shock in the household during 3 months preceding the introduction of SKY. Contrary to hypothesis 1, households with young or with old members are not more likely to buy SKY, nor are those with a stunted or wasted child. Contrary to hypothesis 2 (worse adverse selection on observables at higher prices), we …nd only weak evidence that households that paid a higher price for insurance are more likely to report a member in poor health or to have experienced a serious health shock prior to the purchase of SKY than are households that paid a deeply discounted price for insurance. Consistent with hypothesis 3, households that paid more for health insurance have substantially higher usage of both health centers and hospitals than households that received a discounted price, even when comparing households with similar observed baseline health. This result is consistent with substantial adverse selection based on factors we did not observe at the baseline. Tests of our hypotheses are often di¢ cult in societies with well-established health insurance markets. First, moral hazard due to previous insurance coverage may have increased past health care utilization for some customers, so it is hard to tell whether prior health care utilization predicts purchase of insurance or if past utilization was due to being insured. In this study, no health insurance was available in these communities prior to SKY. In addition, in other settings high insurance prices often imply higher bene…ts such as lower copayments. In that case, moral hazard (in addition to adverse selection) can lead to higher health care utilization for those who pay high prices. In this study, bene…ts were identical regardless of price.1 Third, in other settings insurance prices may correlate with the quality of care or with factors the insurance company observes that are not observed by the econometrician. In this study, we randomized prices so they should not be correlated systematically with such factors. Studying health care in poor nations in particular is important because expanding voluntary health insurance is a popular policy prescription for the billions of poor people who lack a¤ordable access to health care. Most importantly, China has largely shifted to

1Households who pay more for insurance may have a desire to “get their money’s worth”. We discuss this possibility under robustness checks. 60 voluntary, government-subsidized health insurance for medical care coverage in rural areas (Wang 2007; People’s Daily Online 2008). Other developing nations such as Vietnam and Thailand also have rapidly expanded health insurance coverage (Vietnam Social Security 2010; Antos 2007). To understand the …nancial viability of voluntary insurance in these settings, it is important to study selection among poor populations. The success of voluntary health insurance in poor nations depends on the ability to improve health and economic outcomes while maintaining …nancially sustainable, or at the least assuring donors that their money is being spent in the most e¢ cient way possible. However, because health insurance is a relatively new product in developing countries, little is known about the risks and bene…ts of o¤ering this type of insurance, or how best to design an insurance program to meet the needs of the poor. This study provides evidence as to the extent of adverse selection in a developing country.

2.2 Previous Research

George Akerlof’sseminal article “The Market for Lemons”(Akerlof 1970) examines why health insurance companies do not raise their rates to match the risk of clients.2 Akerlof theorizes that individuals who are willing to pay the highest insurance premium are those people who expect the highest expected insurance payouts. For health insurance, poor health is a primary determinant of high expected insurance payouts. Individuals seeking insurance typically have more information on their health status than an insurance company. If an insurance company cannot identify health status, it can only charge a price based on observable characteristics. To be …nancially sustainable, that price must cover the expected health care costs for this group. Because healthier individuals have lower expected health care expenses, they may not be willing to pay the higher premium that is a consequence of risk pooling with less healthy individuals. Under certain assumptions, the healthiest people may drop insurance …rst, leading to a worse risk pool among the insured and higher prices. Higher prices, in turn, drive out the next most healthy until the market collapses altogether; this process is sometimes referred to as a "death spiral". There is an extensive empirical literature on the extent of adverse selection in insurance markets in developed countries. Studies are of various types, comparing individual characteristics by generosity of health plan, premium level, and choice of whether or not to remain uninsured (see Cutler and Zeckhauser 2000 for a review). Most studies have found evidence of adverse selection, although estimates of the magnitude vary. For example, several studies …nd that people with higher expected medical expenditures are more likely to buy insurance or choose health insurance with more generous bene…ts than those with lower expected medical expenditures (e.g., Cutler and Zeckhauser 1998; Cutler and Reber 1998). However, other studies …nd that adverse selection in health insurance and other insurance markets is minimal (e.g., Wolfe and Goddeeris 1991; Finkelstein and Poterba 2004) or non-existent (e.g. Finkelstein and McGarry 2006; Cardon and Hendel 2001; Cawley and Philipson 1999). In fact, there is even some recent evidence of positive selection into health insurance (e.g., Fang, Keane, and Silverman 2008).

2This literature review draws on Polimeni (2006) and Levine, Gardner, and Polimeni (2009) . 61

However, existing studies of asymmetric information often have trouble distin- guishing the e¤ects of adverse selection from moral hazard. For example, an observed cross sectional correlation between generous health insurance and high health care utilization could be because those who anticipate high health care utilization choose generous insurance (adverse selection) or because generous insurance induces more health care utilization (moral hazard). To focus on adverse selection when choosing among health care plans, Ellis (Ellis 1989) examines a change in health insurance options at a …nancial …rm. Out-of-pocket costs in the previous year among those selecting into the most generous (and highest premium) plan were 8.6 times higher than for those selecting into the least generous (and lowest premium) plan, a result strongly supporting adverse selection. Cutler and Reber (Cutler and Reber 1998) examine a similar natural experiment. In 1995 Harvard greatly increased the cost to most of its employees of its more generous health plans. 3,000 of its 10,000 employees were not subject to the price increase until a year later, providing a comparison group. Many employees who faced the higher prices switched to a lower-cost plan, but there was less switching among older employees and for those with high health care utilization in the past. There was enough of this adverse selection that by the third year after the reform the more generous plan was eliminated as an option for employees because it could no longer remain pro…table at a reasonable premium. The above studies are all from developed nations. There have been far fewer studies of selection in developing countries, partly because there are far fewer insurance markets in developing countries. On one hand, the theory of adverse selection implies potential customers with higher expected health care expenses will be more likely to buy health insurance in poor nations, just as in rich ones. On the other hand, there are several reasons to believe that clients in developing countries may behave di¤erently than what has been found in developing countries. For example, because insurance is a relatively new and unknown product, only those who are willing to take a risk on a new product may be willing to try it. More generally, Siegelman (2004) has argued there is plausibly much more adverse selection when people choose among insurance plans than when they choose whether or not to become insured. Non-experimental studies from developing countries have found enrollment to be more common in households with chronically sick members, consistent with adverse selection (e.g., Wagsta¤, Lindelow, Jun, Ling, and Juncheng 2009). These studies also typically …nd higher enrollment rates among wealthier households, potentially leading to positive selection because wealthier people also tend to be healthier (e.g., Wagsta¤, Lindelow, Jun, Ling, and Juncheng 2009; Wagsta¤ and Pradhan 2005; Jutting 2004; Lamiraud, Booysen, and Scheil- Adlung 2005). The current study is most similar in design to Ausubel (1999) and Karlan and Zinman (Karlan and Zinman 2009), who randomize o¤ers in credit markets. Ausubel (1999) …nds evidence for three hypothesized consequences of adverse selection in his study of who accepts credit card o¤ers, each corresponding to our three focal hypotheses. First, people who accept credit card o¤ers have worse observable credit characteristics (in his setting, credit histories) than those who reject the o¤ers. Second, he hypothesizes that people with worse credit histories are more likely to accept o¤ers with high interest rates than those 62

with good credit histories. Third, after controlling for observable characteristics of clients, those who accept o¤ers with high interest rates are more likely to default. Ausubel’s…nding of di¤erences in default rates may be due to adverse selection, but it may also be due to moral hazard: households with inferior credit o¤ers have more incentive to default. Ausubel argues that moral hazard is an unlikely explanation because default rates remain higher among households that accept the high interest rate credit card o¤ers even after interest rates equalize over time. Karlan and Zinman (2009) use a similar experiment but are able to disentangle the e¤ects of adverse selection and moral hazard by giving some borrowers who accepted a high interest rate a lower actual interest rate. They …nd only weak evidence of adverse selection. The research presented here adds to the literature in several ways. First, we present evidence on adverse selection in a developing country, while empirical studies have taken place for the most part in developed countries. Second, the above empirical studies have taken place in more traditional competitive markets, whereas the SKY program in Cambodia is the only health insurance option in the rural markets targeted. Because there is no plan choice, adverse selection may show up more starkly in this market, or, as Siegelman (2004) argued, there may be less adverse selection. In addition, in our setting, longitudinal data and ability to randomize prices make it possible to identify selection e¤ects in this dataset separately from moral hazard. Finally, unlike previous studies of health insurance, we measure selection based on both observable and unobservable factors.

2.3 Theory and Methods

2.3.1 Selection on Observables The core hypothesis of the theory of adverse selection is that customers with high expected future health care costs will be more likely to buy health insurance. We posit that expected future health care costs are higher for households with worse observable health Hij. This assumption implies our …rst hypothesis (H1): Adverse selection in joining based on observables: Customers with factors observable to the econometrician that predict high future health care costs (e.g., high past health care utilization) will be more likely to buy insurance than other households. We measure observable health as a vector including a member in self-reported poor health; a recent serious health shock, de…ned as an illness or injury that resulted in missing 7 or more days of normal activities, a death, or an illness resulting in an expense of more than 100USD; a child who is stunted or wasted; and a member who is under 6 or over 64 (groups with high health care utilization [DHS 2005]). Customers likely to use covered services (i.e., public health facilities) have higher expected costs. Thus, we also include in Hi an interaction of having had a serious health shock and having received treatment for it in a public facility. To test hypothesis 1 we estimate a probit predicting SKY membership for household i: 63

SKYi = F Hij + Dki + "i (2.1) 0 j Dk 1 j k X X @ A Here F ( ) is the probit function, Dki is a list of demographic and other control variables, and "i is an error term. We analyze the e¤ect of other variables on take-up in a companion paper (Chapter 3).

2.3.2 Selection on Observables at High versus Low Price At zero price economic theory suggests everyone takes insurance and there is no adverse selection. At a very high price, only those most likely to use expensive care will purchase insurance. More generally, Akerlof’s theory of adverse selection implies (H2): Adverse selection in joining based on observables and price: The e¤ects in Hypothesis 1 hold more strongly when insurance prices are high than when prices are low. We test hypothesis 2 by adding interactions of observable poor health and the price of insurance (Pi) to equation 2.2:

0 0 0 0 0 SKYi = F Pi + Hij + Pj Hij + Dki + " (2.2) 0 P0 j P j Dk i1 j k X X @ A Here Pi = 1 for consumers who were o¤ered the full price and Pj Hij is the interaction of full price and poor pre-SKY health, for the several measures of pre-SKY health, Hij (health prior to being o¤ered SKY). We use fewer health measures and covariates (Dki) in estimating equation (2) than in (1) because of the modest number of SKY buyers 0 who paid the full price. The implication of Hypothesis 2 is P j > 0 for all j > 0; that is, that poor health is a particularly strong predictor of SKY membership for those paying full price (Pi = 1).

2.3.3 Selection on Unobservables While the previous tests compare the health and health care utilization before being o¤ered insurance of those who later buy and decline insurance, the insurer cares about health care utilization after buying insurance. However, we cannot measure adverse selection by comparing utilization after being o¤ered insurance of those who join versus those who decline, as moral hazard as well as adverse selection can lead to high utilization. Our randomized prices provide a means to test for adverse selection, even when the selection occurs based on factors unobservable to the econometrician. As noted in hypothesis 2, economic theory suggests adverse selection will be more severe at high insurance prices than at low prices. As long as factors observed by the econometrician do not predict insurance uptake perfectly, we have H3: Adverse selection in utilization based on factors unobservable to the econometrician (but observed by households): Among the insured, those who bought insurance at a high price have higher health care utilization after they are insured than those who bought insurance at a lower price (holding constant observable factors). 64

The di¤erence in utilization between insurance buyers paying full versus the discounted price is equal to total selection due to the di¤erence in price. Any di¤erences in utilization that remain after controlling for characteristics that were observable at the baseline is plausibly due to selection based on unobservables. To test for selection on unobservables (Hypothesis 3), we run a probit regression among all SKY members in our sample, predicting post-SKY health care utilization as a function of insurance price and pre-SKY characteristics:

Yi = F P0 + Hij + Pj Hij + Dki + ui (2.3) 0 P0 j Pj Dk 1 j X X @ A Here, Yi is a measure of health care utilization for household i. We examine three measures of health care utilization in the three months following SKY purchase: an indicator variable equal to 1 for households with at least one health center visit; an indicator variable equal to 1 for households with at least one hospital visit; and the total of the list price of all services covered by SKY for all visits to public health centers and hospitals. The function F ( ) is probit for predicting the indicator variables, and OLS or tobit for predicting total costs. Hypothesis 3 posits households that pay full price for SKY (P = 1) will have higher utilization of health facilities following SKY purchase than households with similar

observables who paid a deeply discounted price: P0 > 0 . Because buyers paying di¤erent prices have identical health coverage, moral hazard due to di¤erent out-of-pocket costs cannot explain di¤erences in utilization post-SKY between buyers paying full and discounted prices. However, a more behavioral version of moral hazard is possible if households who purchase SKY with at full are more determined to “get their money’s worth,” a type of sunk cost e¤ect (Tversky and Kahneman, 1981). Thus, if we see that buyers paying full price have higher utilization than buyers paying a deeply discounted price, it may be due to a combination of adverse selection and this behavioral moral hazard. We present evidence on the importance of behavioral moral hazard under robustness checks, below.

2.4 The Setting

In this section we describe health care in Cambodia, SKY health insurance, and our randomization procedures.

2.4.1 Health Care in Cambodia Cambodia is among the world’s poorest and least healthy nations. It ranks 188 out of 229 nations in GDP per capita, has the 38th highest infant mortality rate (of 224 countries with data), and the 46th lowest life expectancy (Central Intelligence Agency 2010). Health shocks often contribute substantially to indebtedness and loss of land. For example, one study followed 72 households with a member who had su¤ered dengue fever following a 2004 outbreak. A year later, half the families still had outstanding health-related 65 debt, with interest rates between 2.5% and 15% per month. Several of the 72 families had found it necessary to sell their land to pay their debt (Van Damme, Van Leemput, Por, Hardeman, and Meessen 2004). Annear, et al., (2006) found similarly high levels of indebtedness due to medical expenses. Cambodians rely on a mix of healthcare providers: public providers, private medical providers, private drug sellers (with and without pharmaceutical training), and traditional healers. Public facilities consist of local health centers, which provide basic care for every- day illnesses, Operational District Referral Hospitals, for illnesses requiring more involved treatment, and Provincial Hospitals, for care of more severe health shocks. Public facilities are subsidized by the Cambodian government or other organizations. However, public facilities su¤er from low utilization rates. According to 2005 DHS estimates, of those who sought treatment for illness or injury, less than a quarter went to a public health facility, and instead preferred to visit private providers, including medical doctors, drug sellers, traditional healers and midwives. Private providers of varying capabilities are typically more popular than public ones, even when more expensive, because they often are more attentive to clients’ needs, more available, willing to visit patients in their homes, and willing to provide more of the treatments patients prefer. They are also usually willing to extend credit (Collins 2000; Annear 2006). However, even when households utilize local private doctors and drug sellers for small health shocks, many visit public hospitals for surgery and other major health problems.

2.4.2 SKY Health Insurance SKY health insurance was originally developed by the French NGO GRET as a response to high default rates among its micro…nance borrowers due to illness.3 Since 1998 GRET has been experimenting with micro-insurance schemes by examining responses to di¤erent premiums and bene…ts. Historically, take-up of insurance has ranged from 2% in regions where insurance has been only recently introduced to 12% in the longest-served regions. While the SKY program targets the poor, it also is trying to avoid …nancial losses and become …nancially sustainable (without donor support) in the long term. Thus, the policy includes several terms that limit adverse selection. For example, SKY insurance does not cover long-term care of chronic diseases. In addition, SKY does not pay for the delivery of babies within the …rst few months of joining. A government policy also reduces adverse selection: separate government programs pay for the very expensive drugs for HIV/AIDS and tuberculosis. Finally, insurance is purchased at the household-level, eliminating the possibility that households would purchase insurance for only very ill or frail members. At the time of the study, households were o¤ered insurance at a rate ranging from $.50 per household per month for a single-person household to around $2.75 per household per month for a household with eight or more members. Households sign up for a six month cycle, paying for the …rst month’scoverage plus two reserve months up front. While a household can stop insurance payments at any time, failing to pay two consecutive months

3SKY is a Khmer acronym of “Sokapheap Krousat Yeugn”, meaning “Health for Our Families.” 66

before the end of the six-month cycle results in the loss of one month of reserve. A household can join SKY at any time, but coverage will not begin until the start of the next calendar month. Households o¤ered insurance for the …rst time are o¤ered slightly lower premiums to encourage take-up. With their insurance, household members are entitled to free services and prescribed drugs at local public health centers and at public hospitals with a referral (SKY 2009).

2.5 Randomization

Our randomized experiment was carried out as the SKY program began a major expansion at the end of 2007. When the SKY program …rst rolls out into a region, SKY holds a village meeting to describe the insurance product to prospective customers. The meetings are advertised ahead of time via loudspeaker announcements in each village. Following the meeting, SKY insurance agents visit households to sell insurance to interested families. At the end of each meeting, SKY traditionally gives out coupons for one month of free insurance. To randomize price of insurance, we implemented a Lucky Draw for insurance coupons. Winners of the Lucky Draw received a coupon for a deep discount o¤ of the insurance premium, a “high” coupon, while others received the traditional one-month discount, a “low” coupon. The high coupon entitled the bearer to …ve months free out of the …rst six month insurance cycle, with the possibility of another three months discount if the household renewed for another six months. High coupons were meant to expire two months from the date of purchase. All households have the ability to purchase SKY at any time at full price (that is, …rst-time buyers can pay 5 months for 6 months’coverage).4

2.6 Data

Our analyses use a household survey; SKY administrative and health facility utilization data, and several other sources.

2.6.1 Household Survey Our main data source is a survey of over 5000 households. We rely largely on the baseline survey, which took place from two to nine months after the initial SKY marketing meetings. (Households could …rst start SKY coverage between one and two months after these meetings.) The baseline survey covered demographics, wealth, self-perceived and objective health measures, health care utilization and spending, assets and asset sales, savings, debt, health risk behaviors, willingness to take …nancial risks, trust of health care institutions, means of paying for large health expenses, and time preference. Appendix 2.D describes the measures used in this Chapter in more detail5. SKY meetings were held in Takeo, Kandal, and Kampot provinces from November 2007 to December 2008. For the baseline survey, we interviewed all Lucky Draw winners

4Details of the randomization of price are available in Appendix 2.C or in Chapter 1. 5Other measures are explored in Chapter 3 and decribed in the Chapter 3 Appendix. 67

(the 20% of the village meeting attendees able to purchase SKY at a deeply discounted price) and an equal number of households o¤ered the standard price (speci…cally, every fourth house on the village meeting attendance list that were o¤ered the standard price). To increase our sample of buyers who paid full price, we also interviewed all households with a low coupon that bought insurance. In total, our sample consists of 2537 households o¤ered the deep discount, 2534 households o¤ered full price and randomly included in the sample, and 196 over-sampled households who purchased insurance at full price. A summary of subjects and datasets is presented in Appendix Table 2.7.6

2.6.2 SKY Administrative and Utilization Data For each household that joins SKY, SKY records registration date, date the household starts coverage, and date the household drops out of SKY. We use this SKY administrative data to determine if and when each household from the village meeting purchased SKY insurance. To match our baseline data to the SKY database, for each village, we matched the name of household member in our survey to the names listed in the SKY database. SKY also collects utilization data whenever a member visits a health center or hospital with which it partners, including the list price of the services provided. We use these data to measure utilization of public health centers and hospitals in the months following SKY purchase, as well as the total cost of all visits paid for by SKY.

2.6.3 Other Datasets We also used data from a second round household survey, interviews with village leaders and measures of health center quality as control variables in some analyses (Appendix 2.B).

2.6.4 Randomization Appendix Table 2.8 shows average characteristics of high and low coupon winners at the baseline. To account for recall errors and to ensure we are looking at pre- randomization health events and behaviors, we look at health events taking place in the month before the SKY meeting and two months prior (for a total of three months). Of the thirty variables tested, only three show a statistically signi…cant di¤erence between high and low coupon at the 5% con…dence level. There is a statistically signi…cant but very small di¤erence in the percentage of households that are Khmer between the two groups (95.3% versus 94.6% in the full and discounted price groups, respectively). 14% of low coupon households have wealth level subjectively graded as “poor” by enumerators, while only 10% of high coupon households are rated as “poor”. Similarly, low coupon households are slightly more likely to live in a house made of palm, another measure of lower wealth. Other wealth indicators did not show signi…cant di¤erences. We control for wealth characteristics in our regressions and keep in mind this di¤erence when interpreting results. 6A timeline of the study is presented in Chapter 1. 68

2.7 Results

We test our three hypotheses regarding adverse selection: whether insurance buyers have worse observables prior to joining SKY, whether that adverse selection is more severe for those paying full price, and whether health care utilization is higher for those paying full price (conditional on observables). For additional evidence of selection, we compare utilization of households that drop SKY to that of households that remain insured.

2.7.1 Selection on Observables In this section we compare pre-SKY characteristics of households who bought insurance to those who did not. We focus on households that bought SKY within 63 days of the introductory marketing meeting. For these households, pre-meeting health shocks are more representative of health immediately before SKY purchase.7 Households that purchased SKY are 2.3 percentage points more likely to have had a major health shock in the two to four months prior to the SKY meeting than households that did not join SKY (9.4 percent of buyers versus 7.0 percent of decliners, P < 0.05, Appendix Table 2.9) and 12.7 percentage points more likely to report having a household member in poor health (80.0% versus 67.2%, P < 0.01). These results are consistent with Hypothesis 1, that SKY buyers are in worse health than decliners. In contrast to Hypothesis 1, SKY buyers and decliners are not signi…cantly di¤erent in other measures of high predicted health care expenditures such as presence of a stunted or wasted child, having children 5 and under, or having elderly household members. To investigate selection on observables further, we perform a probit estimation of membership as a function of pre-SKY health (as in equation 2.1). Consistent with Hypothesis 1, having a household member in self-reported poor health has a large e¤ect on insurance purchase. Around 22 percent of pooled low and high coupon households buy SKY insurance. A household is more than 12 percentage points more likely to buy SKY insurance if they have at least one household member in poor health at the baseline (Table 2.1, col. 1, 3, 4, 5, P < 0.001). A health shock pre-meeting leads to an 8.7 percentage point increase in the probability a household will buy SKY insurance (col. 2, P < 0.01). This variable is somewhat collinear with having a household member in poor health (col. 3), and the result is driven by households that used a public health center rather than a private facility for care (col. 4-5, joint results not signi…cant a traditional levels). Number of days ill and spending more than 30 USD on a shock does not have a signi…cant impact on take-up above and beyond having a major shock. As in the comparison of means, the other measures of high expected health care costs such as having a child or elder in the household and having a stunted or wasted child do not predict high SKY membership. We explore the e¤ects of other household characteristics on take-up in a companion paper (Chapter 3).

7Results are similar for the full sample (Table 2.12). 69

2.7.2 Selection on Observables by Price Hypothesis 2 posits that households that purchase SKY at full price (that is, with a low coupon) will have higher expected health care costs on average than households that bought insurance at a deeply discounted price. Table 2.2 presents results of a probit regression of SKY purchase on price and price interacted with health characteristics. We include only a limited number of health characteristics due to the small sample size of purchasers paying full price. Contrary to hypothesis 2, having a member in poor self-reported health does not raise SKY uptake statistically signi…cantly more among those o¤ered SKY at full price than among those o¤ered a deep discount, although the point estimate is positive for households in poor health that were o¤ered SKY at full price ( = 2.7%, SE = 3.3%, not signi…cant, col. 1). In contrast to hypothesis 2, there is no evidence a major health shock prior to the SKY meeting has more e¤ect on purchasing SKY among households o¤ered full price than among those o¤ered SKY at a deep discount ( = -3.8%, SE = -4.6, not signi…cant, col. 2).

2.7.3 Selection on Unobservables We did not …nd purchasers paying full price had worse observable health prior to buying SKY than purchasers paying the discounted price (in contrast to Hypothesis 2). However, if some adverse selection is based on characteristics that are unobservable to the researcher at the baseline, then SKY buyers who paid full price will use health facilities more than SKY buyers who purchased at a deep discount, even when holding constant baseline characteristics (Hypothesis 3). We examine all SKY buyers (not just early buyers) and measure health care costs based on the list cost (that is, the fees paid by uninsured patients to public facilities) of services provided to SKY members. Summary statistics are presented in Table 2.3. In the regressions with no covariates, buyers paying full price are 11.1 and 10.7 percentage points more likely to use a health center (P < 0.001) and hospital (P < 0.01), respectively, in the …rst 3 months after SKY purchase, than are buyers at the deeply discounted price (Table 2.4, col. 1 and 3).8 Buyers who paid full price also have 60% higher health care costs than buyers at a deep discount, on average (col. 5) .910 To test Hypothesis 3 we measure how much of this higher utilization remains after controlling for observed baseline characteristics (equation 2.3). Even after we condition on our pre-SKY health covariates, the price of insurance continues to predict increased likelihood of our three utilization measures (Table 2.4, col. 2, 4, 6). Estimated e¤ect sizes are only slightly lower (at about 20% for hospital use and total cost) than the unadjusted e¤ects, though none of the declines are statistically signi…cant (controlling for observables

8The summary statistics (Table 2.3) measure utilization in the …rst three months of SKY for all insured households, coding usage as zero in a month if a household dropped prior to that month (in which case we have no utilization data). In the regressions (Table 2.4) we include dummies for households that dropped in the …rst, second, or third month of SKY. 9For total cost, we take the log of cost in USD. The percent increases in spending by the households paying full price are calculated by comparing the exponentiation of the coe¢ cients on the price variable to the average cost of utilization by households that bought with a deep discount (Table 2.3). 10Results are similar when we use the Tobit functional form for columns 5 and 6 in Appendix Table 2.18 70

has almost no impact on the e¤ect size for health care use). Results are similar if we include all available covariates (Table 2.14, covariates listed in Appendix 2.D). Speci…cally, conditioning on baseline characteristics, buyers paying full price are 11.3 and 8.5 percentage points more likely to visit a SKY-partnered health center and hospital, respectively, in the …rst three months following SKY purchase (Table 2.4, P < 0.001, col.2, and P < 0.05, col. 4) than are buyers at a steep discount. Totaling health centers and hospital visits, buyers at full price use services costing about 56% more than buyers at a deep discount (col. 6). In short, even after controlling for all observed baseline characteristics, households that buy SKY at full price have higher utilization than those that buy at a deeply discounted price. Figure 2.1 summarizes our results from Table 2.4 graphically: Households paying more for insurance have higher utilization, and this result is barely decreased even after holding baseline variables constant. Baseline characteristics account for none of full price buyers’higher utilization of health facilities, 21.5% of their higher hospital utilization, and only 8.1% of their higher SKY-covered costs.

2.7.4 Drop-out Complementing selection into SKY, in this section we study self-selection among those who remain in SKY Dropout is highest among those who have not used SKY. For example, consider dropouts in months 6 (when the initial deep discount ran out for high-coupon buyers) to month 15 (when our dataset becomes too small to analyze). During these months, dropout averages 9% among those who have used SKY-covered services, but 15% among those that have not. (This calculation is complex because most households that remain in SKY shift into the “ever used SKY”category over time.) A hazard analysis of drop out on utilization in the prior month demonstrates this relationship as well. We analyzed the following equation to predict Dropoutit, the probability that a household i that is a member of SKY after t 1 months of membership will drop out in the next month:

1 HCi;t 1 + 2 HCi;t 2 + 3 Hospi;t 1 Dropoutit = f (2.4) + 4 Hospi;t 2 + 5 No_visiti;t 3 where HCi;t 1, HCi;t 2 , Hospi;t 1 and Hospi;t 2 are indicator variables for at least one visit to a health center (HC) or hospital (Hosp) in the 1 or 2 months prior to the given month, No_visiti;t 3 is an indicator variable equal to 1 if household i did not visit any public health facility between joining SKY and month t 3, and F ( ) is the Cox proportional hazard distribution. The hazard estimates (Table 2.5) are that households that utilized a health facility one additional time in the one or two months prior are 28.8 (P < 0.001) and 2.7 (not statistically signi…cant) percent less likely to drop SKY, respectively, in any given month. Similarly, a household that used a hospital one additional time in the one or two months prior are 48.3 and 42.1 percent less likely to drop SKY in any given month (P < 0.001). The low rates of dropout among those who use SKY facilities more will lead to an increasingly costly customer base if there is autocorrelation in health care utilization. Autocorrelation in utilization is, in fact, fairly high among SKY members. For example, 71 households that had a health center visit in months 1-3 were 36 percentage points more likely to visit health center in months 4-6 than were households that did not use a health center in months 1-3 (results not shown, 69.8 % versus 34.1%, P < .01). Further, the total list price of all health center and hospital care in months 1-3 is has a correlation of 0.41 with the cost of care starting in months 4, 5 or 6 (Appendix Table 2.11, col. 5). Autocorrelation in care coupled with the exit of households with low past utilization means SKY members’ average utilization will rise over time. Figure 2.2 shows this trend of rising utilization in hospital utilization, but not for health center utilization. The proportion using a hospital more than doubles over the …rst 18 months, while the proportion using a health center reduces by about a third. (Di¤erences over time are statistically signi…cant at the 5% level in 6 out of the …rst 8 months for health center utilization when comparing each month to month 18. Di¤erences are not signi…cant at traditional levels for hospital utilization.) When we look at costs of covered services (not shown), we …nd that covered health center costs decrease by 24%, from an average of 0.48USD per month in the …rst three months of coverage to 0.36USD per month in months 16 to 18 (P < 0.001), while hospital costs almost doubled, increasing from 0.37USD to 0.72USD in the same period (P < 0.001). Total average monthly SKY-covered costs (health center plus hospital) increased by around 27% over that period, from 0.85USD to 1.08USD. A second possible reason for increased adverse selection is that high coupon households paid only one of the …rst six months of insurance, but had to pay triple that price to purchase six additional months. Theory suggests this price increase will make insurance less attractive for households with modest expected health care expenditures, leading to an increase in adverse selection among buyers initially paying a deeply discounted price. Thus, renewing high coupon households will more closely resemble buyers who always paid the full price. Figure 2.3 graphs utilization for all households that are still members of SKY a given number of months after initial purchase. The …gure shows that while households that purchased at the full price initially have higher hospital utilization than those who purchased SKY with the deep discount, the di¤erence tapers o¤ by around the sixth month after SKY purchase (monthly di¤erences are not statistically signi…cant at the 5% level for any month). However, for health center use, households that purchased at the full price have consistently higher utilization than those that purchase at the discounted price (monthly di¤erences are signi…cant at P < 0.05 in months 2-5, 13, and 15).

2.8 Robustness Checks

In this section we present several robustness checks.

2.8.1 Early versus Late Buyers Households can purchase SKY insurance at any time. Also, although high-value coupons were intended to expire after 2 months, we observed a number of households using the coupons after that date. Thus, it is possible that some households waited to purchase SKY until they were more in need of health care. Thus, these late buyers may exhibit 72 high utilization immediately following SKY purchase that tapers o¤ shortly after. If low coupon households delay more than high coupon households, perhaps because some high coupon households believe they must use their coupon immediately, higher selection among low coupon as compared to high coupon households may be due to both higher price and timing of purchase. Some households do buy SKY after they have a health shock that needs care, and this pattern appears to be more common among households o¤ered SKY at full price (Appendix Table 2.10). Compared to early buyers at full price, late buyers who paid full price for insurance are less likely to report a member in poor health in the baseline survey (81% versus 86%, di¤erence not signi…cant at 10% level). On the one hand, this means household members in late-buying households are subjectively healthier, but on the other hand it may mean that their health care needs are less urgent, and thus these households wait to buy SKY until they experience an increase in utilization. Supporting this hypothesis, late buyers are twice as likely as early buyers to report a major health shock in the 3 months immediately following SKY purchase (P < 0.01). Also consistent (if less precisely), late buyers are 8 percentage points more likely to have a major health shock (t = 1.59), and 3 percentage points more likely to have used a hospital for care in the 3 months immediately preceding SKY purchase (t = 0.95). Results are not as clear for households o¤ered a deep discount that waited to purchase insurance. Like low coupon buyers, high coupon late buyers are somewhat less likely to have a member in poor health at the baseline. However, high coupon late buyers are not signi…cantly more likely to have reported a major health shock in the months immediately prior to or immediately following SKY purchase. Thus, unlike low coupon households that delayed purchase of SKY, high coupon late buyers may have delayed purchase for reasons other than waiting for a health shock to arise.

2.8.2 Selection on Unobservables We re-ran the tests of selection on unobservables including additional baseline covariates (Appendix Table 2.14); adding an indicator variable for the timing of SKY purchase (immediately following the SKY meeting versus later, Appendix Table 2.15), conditioning on data from 1 to 3 months pre-SKY instead of 2 to 4 months (Appendix Table 2.16); only including the randomly selected sample of households who were o¤ered the full price (dropping the over-sample of SKY buyers who were o¤ered full price, Appendix Table 2.17); and controlling for only pre-SKY shocks that caused 7 or more days of missed work (instead of also including a shock leading to a death or a 100USD expense, Appendix Table 2.19). We also switched the functional form from OLS to Tobit for predicting health care expenditures (Appendix Table 2.18). In general, results were very similar to those presented.

2.8.3 Behavioral Moral Hazard It is possible that buyers paying full price use more SKY services than those buying with a steep discount due to a sunk cost e¤ect, or “behavioral” moral hazard, whereby households that paid a higher price for insurance utilize services more to “get their money’sworth.”However, we do not believe this is the case. First, recent research has not 73 found evidence for this type of behavioral incentive to seek care. For example, in the case of treated bednets, Cohen and Dupas (2010) …nd no decreased usage of bednets for those receiving these nets for free, and Tarozzi, et al., (2011) found that households randomly chosen to receive free bednets used them even more than households paying for these nets. Second, the costs of health care include several costs that are not covered by SKY, including the opportunity cost of lost time and travel costs. Finally, while households may increase care for small illnesses to health centers to get their money’s worth, we found that low coupon households also have much higher utilization of hospital services, which typically require referral from a health center. To focus more speci…cally on very severe illnesses, we examined the subset of hospital visits with an overnight stay. Households purchasing SKY at full price are more than twice as likely to have an inpatient visit in the …rst three months after SKY purchase than households purchasing at the discounted price (approximately 12.8% versus 4.5%, P < 0.01, mean comparison not shown). The much higher utilization among those buying at full price remains after holding constant baseline characteristics (P < 0.05, Appendix Table 2.20, col. 2).

2.8.4 Hazard Rates by Price Results of the hazard rates on dropping out of SKY are similar for households paying either full or the deeply discounted price. One exception is that households that purchased at the high price are signi…cantly more likely to drop SKY, compared to households purchasing at the discounted price, if they have never used a health center or hospital through three months prior. Unlike households paying a lower price for insurance, households purchasing at the higher price were signi…cantly less likely to drop based on health center utilization in the two months prior, but this result does not seem meaningful when taking into account other small di¤erences in coe¢ cients (results not shown).

2.9 Financial Implications of Adverse Selection

Our results indicate that while there is some adverse selection based on characteristics observable at the baseline (e.g., past health care utilization and baseline health levels), most adverse selection was not based on baseline observable characteristics. How important is this observable and unobservable selection to SKY’sbottom line? The usual SKY insurance premium is set to cover the list price of services that SKY members use, what we will call “utilization costs.”(We have no data on administrative or sales expenses.) In our data, in the …rst six months after joining, the average utilization costs among all buyers who paid full price for SKY was $6.94 (Table 2.6, col. 1. Using back- of-the envelope calculations based on average household size and premium levels, average revenue per household was near $9.93 during that period, $2.99 above utilization costs per household. Our results showed that the large discount reduced self-selection of those with high utilization. The lowered price decreases average utilization costs from $6.94 (column 1) to 74

$5.29 (column 3) per person in the …rst 6 months of SKY. However, the steep price cut costs far more in revenue than it reduces utilization, so the steep discount leaves revenue $3.64 below utilization costs (column 3) - without considering administrative or sales costs. SKY does not wish to exclude any households from purchasing SKY, and in fact this would be counter to their mission. However, to demonstrate the size of adverse selection on SKY’sbottom line, suppose SKY excluded all households that reported “poor health”at the baseline, or that reported a large health shock prior to the baseline (where large shock is de…ned as above). For households that paid full price for SKY (Table 2.6, columns 1 and 2), these baseline measures are poor predictors of future health care utilization. Among buyers who purchased SKY at the full price (column 1), selling to only households that did not have observable poor health at the baseline actually increases average utilization costs (from $6.94 on average to $8.33 if they did not report poor health or to $7.58 if they reported no large health shock in the 3 months prior to purchase). This unexpected result may be due to the very low percentage of buyers at the full price that did not have a health shock or a member in poor health: only 16% of households that bought SKY at the full price did not have a member in poor health. While these households did not report poor health, our results on selection based on unobservables indicates that these buyers are apparently very adversely selected. If we look at only buyers that bought at full price soon after the Village Meeting (column 2), not selling to households with pre-SKY shocks lowers average monthly costs by around $0.40 ($6.23 to $5.82). Pre-SKY health does predict high utilization costs for households that purchased at the discounted price (columns 3 and 4). If SKY were to sell only to households that did not report a member in poor health, average costs over the …rst 6 months of SKY would be decreased from $5.29 to $3.60. However, there is little change in costs based on a major health shock in the 3 months pre-SKY. These results hold for the full sample (column 3) and when we look at early buyers only (column 4). Even if SKY were able to discriminate based on poor health, fewer than a quarter of buyers at the discounted price do not report a household member in poor health. In short, SKY would not be able to improve its …nancial outcomes substantially by avoiding self-selection based on factors we observed.

2.10 Conclusion

To study who buys health insurance, we randomized the price of SKY micro- health insurance and surveyed households close to the start of insurance purchase. Because insurance was not previously available, any di¤erences in baseline characteristics such as health care utilization were not in‡uenced by past insurance. Unlike many previous studies, the randomized price of insurance allowed us to eliminate the e¤ect of moral hazard even when looking at utilization of the insured after insurance purchase. We found mixed evidence of adverse selection on observable baseline characteristics (Hypothesis 1). Those who join SKY have had more past health shocks and are more likely to report a member in poor health.11 At the same time, SKY buyers are not more likely to

11Self-reported health was collected a few months after households joined SKY. Self-reports will underestimate adverse selection if SKY insurance improves health but will over-estimate adverse selection if SKY 75

have very young or very old members, who we believe would be more likely to be ill. There was little evidence that those paying the full price for insurance have more observable risk factors than those paying a deeply discounted price (Hypothesis 2). There was strong evidence that those who paying full price for insurance have more unobservable risk factors than those paying a deeply discounted price (Hypothesis 3). Speci…cally, those who bought insurance at the full price have much higher health care utilization than those who paid a lower price, and this gap in utilization by insurance price remains even when controlling for a number of observable baseline characteristics. While some households that paid full price may have been using more health care to “get their money’s worth” (a behavioral form of moral hazard), we have evidence that most of the e¤ect is due to adverse selection on unobservables. Speci…cally, the e¤ects were just as strong for hospital visits (which require a referral) –even overnight hospital visits – as for health center visits. Ignoring past health characteristics has little impact on SKY’s bottom line when we look only at households paying the full price for insurance. For these households, past health utilization and self reported health status had little predictive power on health utilization after purchasing SKY, but this result is probably due to the low proportion of these households purchasing SKY that did not have a member in poor health. For households paying the discounted price, discriminating against households with a member in poor self- reported health lowers total costs per household by around 32% in the …rst 6 months of SKY, but discriminating against households recent health shocks does little to lower health costs. The lowered price in general increases risk pooling and decreases average costs by 24%, but the steep decrease in revenue is not enough to cover these decreased costs. Finally, insured households that have low health care utilization are more likely to drop insurance coverage. Consisent with this result, households remaining with SKY have higher average hospital utilization over time. Drop-out of households with low utilization meant that by six months, average hospital utilization of remaining buyers who paid the deeply discounted price was equivalent to that of SKY members paying the full price. However, we did not …nd this convergence in health care utilization, and average health care utilization decreased for households who remained in SKY for longer periods of time. Future research will explore whether households with low utilization who dropped SKY overestimated their utilization needs, did not understand SKY insurance, or were dissatis…ed with SKY-partnered health care services. In short, our results imply that a lower insurance premium leads to a more diverse risk pool with lower average utilization, consistent with the economic theory of adverse selection. If health insurance is to address the problems of the global poor we must understand under what conditions insurance can be …nancially sustainable and can improve purchasers’lives. The …nancial viability of an insurance program depends in part on its ability to avoid adverse selection and excess utilization. SKY has a number of rules to reduce adverse selection. For example, the requirement that all members of a household must join SKY at once may explain why SKY households are not more likely to have children or elderly insurance raises members’awareness of health problems. 76

members than are SKY decliners.12 At the same time, SKY policies to limit adverse selection are limited. The result is that SKY faces a high-cost population of members, which reduces both its ability to attract new members and its …nancial viability. Protecting those in need is a major objective of donors and of policy-makers. Unfortunately, better protection against adverse selection, such as restrictions on coverage of pre-existing conditions, worsens insurance’sability to protect those in need. If households buying insurance are disproportionately those with the highest health care needs, subsidizing voluntary health insurance may be a cost-e¢ cient way to increase access to health care. Ongoing donor or government support is appropriate if health insurance both increases health and has other bene…ts such as reducing persistent poverty due to health-related debt. A companion paper examines the health care and economic impacts of SKY (Chapter 1). At the same time, without quality health care services, increases in utilization may make little di¤erence in health, so donors must weigh the value of increasing quality of care versus increasing access to care. In addition to its policy implications, our results provide evidence of adverse selection, separated from the often confounding e¤ects of moral hazard. More research is needed to deepen our understanding of insurance in developing countries, and to understand whether the lessons learned in our evaluation hold for insurance programs in other developing countries. It is important to understand the other characteristics that lead households to purchase health insurance in SKY and other programs. For example, if low quality health facilities or a lack of understanding of insurance are causing some households to remain uninsured, insurers may be able to attract a more diversi…ed risk pool by addressing these barriers. A companion paper (Chapter 3) explores these topics. It is also important to understand the dynamics of insurance. It is possible adverse selection will decline over time, for example, if insurance becomes more widespread. Longer- term studies of who buys and renews insurance policies are needed to address such questions.

12However, we do not …nd any evidence that even one-member households with an elderly member are more likely to purchase SKY (results not shown). Alternatively, households may be equally or more concerned with the care of working-aged members who contribute to household income. This possibility is explored in a related paper (Chapter 3). 77

2.11 Tables 78

(1) (2) (3) (4) (5) Offered Full Price (d) •0.362*** •0.361*** •0.362*** •0.362*** •0.369*** [0.0171] [0.0172] [0.0171] [0.0171] [0.0172]

At least one household member with poor 0.124*** 0.121*** 0.121*** 0.127*** self•reported health (d) [0.0125] [0.0125] [0.0125] [0.0131]

Major Health Shock (*) in the 2 to 4 0.0866** 0.0634* 0.103 0.12 months pre•Meeting (d) [0.0307] [0.0301] [0.108] [0.116]

Major health shock (*) and used health 0.0315 0.0212 center for care (0 if no shock) (d) [0.0563] [0.0578]

Major health shock (*) and used hospital •0.0313 •0.0171 for care (0 if no shock) (d) [0.0496] [0.0538]

Major health shock (*) and use private •0.027 •0.0451 health care (0 if no shock) (d) [0.0496] [0.0477]

Ln of max days ill for a major health •0.00865 •0.00632 shock (*), pre meeting (0 if no shock) [0.0196] [0.0210]

Major health shock (*) and spent 120,000 0.0285 0.0202 riel on care (USD30) (0 if no shock) (d) [0.0591] [0.0585]

At least one member over 64 (d) •0.02 [0.0154]

At least one member age 5 or under (d) •0.00207 [0.0163]

Household has a stunted or wasted child 0.0165 under age 6 (d) [0.0209]

Observations 4701 4701 4701 4701 4387 Pseudo R•squared 0.168 0.154 0.169 0.169 0.179

MarginalNotes: LHS effects; variable: Standard 1 if bought errors SKY,in brackets 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** p<0.01, LHS*** p<0.001. Variable: Marginal 1 if bought effects; SKY, Standard 0 if declined errors (SKY in brackets. Admin data)Robust standard errors clustered at the village level. (*) indicatesMajor health health shock shock means causing 7 days missed of missed daily work,activies death, for 7 or or 100USD more days, expense a death, in the or an2 to expense 4 months of pre•Meetingover Robust100USD. standard All data errors is from clustered the baseline at the villagesurvey. level. Wealth, household size and education are included in the Allregression other data but isnot from presented. the baseline Sample survey. is all SKY decliners and all SKY buyers who bought SKY within 63 days Wealth,of the Village household Meeting. size and(d) for education discrete arechange included of indicator in the regressionvariable from but 0not to presented.1. Sample is all SKY decliners and all SKY buyers who bought SKY within 63 days of the Village Meeting.

Table 2.1: Probit Regression of SKY Take-up on Baseline Characteristics 79

(1) (2) (3)

Offered Full Price (d) •0.381*** •0.358*** •0.381*** [0.0299] [0.0173] [0.0298]

At least one household member with poor 0.116*** 0.112*** self•reported health (d) [0.0163] [0.0162]

Full price * Poor self reported health 0.0268 0.0295 [0.0325] [0.0326]

Major Health Shock (*) in the 2 to 4 0.107** 0.0794* months pre•Meeting (d) [0.0389] [0.0375]

Full price * Major health shock •0.0379 •0.0304 [0.0458] [0.0456]

Observations 4701 4701 4701 Pseudo R•squared 0.168 0.154 0.169

MarginalNotes: LHS effects; variable: Standard 1 if bought errors SKY,in brackets 0 if declined (SKY Admin data). + p<0.10, * p<0.05, ** LHSp<0.01, Variable: *** p<0.001. 1 if bought Marginal SKY, effects; 0 if declined Standard (SKY errors Admin in brackets. data) Robust standard errors clustered (*)at theMajor village health level. shock (*) means indicates 7 days health of missed shock causingwork, death, missed or 100USDdaily activies expense for 7 in or the more 2 to days, 4 months a pre•Meeting Robustdeath, orstandard an expense errors of clustered over 100USD at the Allvillage data level.is from the baseline survey. Wealth, household size Pre•SKYand education utilization are included are for any in themajor regression shock (baseline but not presented.data). Sample is all SKY decliners and Allall SKYother buyers data is who from bought the baseline SKY within survey. 63 days of the Village Meeting. (d) for discrete change of Sampleindicator is variable all SKY from decliners 0 to 1. and all SKY buyers who bought SKY within 63 days of the Village Meeting. (d) for discrete change of dummy variable from 0 to 1

Table 2.2: Probit Regression of SKY Take-up on Baseline Characteristics Interacted with Price 80 * ** + ** ** 0.64 0.11 2.49 0.82 0.77 0.10 0.06 0.03 0.02 0.07 0.32 1262 Deep Discount Bought with 243 0.72 0.21 4.09 1.15 0.84 0.12 0.07 0.02 0.03 0.09 0.36 Price Bought at Full Table 2.3: Summary Statistics, Buyers at Full versus Discounted Price Major health shock (*) and use health center for care (0 if no shock) Major health shock (*) and spent 120,000 riel on care (USD30) (0 if no shock) Major health shock (*) and use hospital for care (0 if no shock) Major health shock (*) and use private health care (0 if no shock) Ln of max ill for a major days health shock (*), pre SKY start (0 if no shock) SKY paid for hospital visit in firstSKY 3 months of SKY Cost of care in first all SKY•paid 3 months of (USD) SKY Ln Cost of all SKY•paid care in first 3 months of (USD) SKY At least one household member with poor self•reported health Major Health Shock (*) in the 2 to 4 months pre•SKY SKY paid for health centerSKY visit in first 3 months of SKY Observations Notes: + p<0.10, * p<0.05, ** p<0.01, *** p<0.001, based on t•tests clustered (*) indicates at the village level. Sample buyers. is all SKY health shock causing missed activies for daily 7 or more a death, or days, an expense of over 100USD. Post•SKY = p < 0.001, based on ttests clustered = p < 0.01, *** at the village level. * = p < 0.05, ** data are fromutilization records; SKY for zeros averaged months in utilization 2 and 3 for households that dropped forZeros months averaged in utilization 2 and 3 for households in months that dropped SKY 2 or 3. in these months.SKY All other data is from buyers. Sample decliners the baseline survey. and all SKY is all SKY data are fromPost•SKY utilization records. SKY All other variables are from the baseline survey. 81

(1) (2) (3) (4) (5) (6) Use HC Use HC Use Hosp. Use Hosp. Total Cost Total Cost

Offered Full Price (d) 0.111*** 0.113*** 0.107** 0.0849* 0.414*** 0.329** [0.0321] [0.0328] [0.0342] [0.0359] [0.106] [0.110]

At least one household member with poor 0.161*** 0.0199 0.250*** self•reported health (d) [0.0344] [0.0253] [0.0595]

Major Health Shock (*) in the 2 0.0489 0.0372 •0.143 to 4 months pre•SKY (d) [0.152] [0.108] [0.354]

Major health shock (*) and use health 0.162* 0.0658 0.0816 center for care (0 if no shock) (d) [0.0770] [0.0835] [0.191]

Major health shock (*) and use hospital 0.115 •0.0592 •0.123 for care (0 if no shock) (d) [0.0885] [0.0552] [0.198]

Major health shock (*) and use private 0.0563 0.0286 0.12 health care (0 if no shock) (d) [0.105] [0.0872] [0.221]

Ln of max days ill for a major health •0.0452 •0.0131 0.0466 shock (*), pre SKY start (0 if no shock) [0.0408] [0.0257] [0.0878]

Major health shock (*) and spent 120,000 0.116 •0.0109 0.0875 riel on care (USD30) (0 if no shock) (d) [0.0787] [0.0639] [0.210]

At least one member over 64 (d) •0.0258 0.00533 •0.115+ [0.0315] [0.0227] [0.0638]

At least one member age 5 or under (d) 0.0852* •0.00221 0.111 [0.0347] [0.0265] [0.0774]

Household has a stunted or wasted child 0.0655 •0.0375 0.0853 under age 6 (d) [0.0454] [0.0284] [0.102]

Observations 1505 1255 1505 1255 1508 1255 Adjusted R•squared 0.024 0.045 Pseudo R•squared 0.009 0.048 0.017 0.029

Notes:Marginal + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Columns 1 and 2 (3 and 4): Indicator for use of a SKY• coveredRobust standard health center errors (hospital) clustered for at thethe firstvillage 3 months level. post SKY purchase; Columns 5 and 6: Ln of total cost (user fees, coveredRHS variables by SKY) are of from all SKY•covered baseline data. health center and hospital visits in the first 3 months post•SKY. Columns 1•4 use probit,Low coupon columns status 5•6 recordeduse OLS. at Marginal village meeting effects; afterStandard the Lucky errors Draw. in brackets. Robust standard errors clustered at the villageSKY status level. is (*)from indicates SKY administrative health shock data.causing missed daily activies for 7 or more days, a death, or an expense of over 100USD.LHS variables SKY use status SKY and utilization LHS variables data for use the SKY3 months data. post Coupon SKY statuspurchase. is recorded at the Village Meeting. All other data isLHS from cost the variables baseline (in survey. ln USD) Wealth, use OLS household regressions. size and education are included in the regression but not presented. SampleAll other isLHS all SKYvariables decliners use probitand all regressions. SKY buyers who bought SKY following the Village Meeting. (d) for discrete change of indicator variable from 0 to 1. Table 2.4: E¤ects of Self-Selection on Utilization 82

Pooled SKY data: Num. HC visits in 1 month(s) prior 0.712*** (0.035) SKY data: Num. HC visits in 2 month(s) prior 0.973 (0.023) SKY data: Num. Hosp. visits 1 month(s) prior 0.517*** (0.086) SKY data: Num. Hosp. visits 2 month(s) prior 0.579*** (0.076) Never used HC or Hosp. through month n•3 1.226* (0.111) Used HC but no Hosp. through month n•3 1.414*** (0.110) Offered Full Price 0.993 (0.102) N 17314 Exponentiated coefficients Notes:Sample * isp<0.05, all SKY ** householdsp<0.01, *** p<0.001.that were LHSmembers variable: anytime Hazard from of Jan. dropping 1, 2008 SKY. to Dec. Robust 31, 2009.standard errors Robust(in parentheses) standard errorsclustered clustered at the village at village level. level. Coupon status is recorded at the Village Meeting. All other data is from SKY records. Sample is all SKY buyers who were members anytime from Jan. 1, 2008 All data (drop•out and utilization) are from SKY records. to Dec. 31, 2009. (d) for discrete change of indicator variable from 0 to 1. * p<0.05, ** p<0.01, *** p<0.001

Table 2.5: Cox Regression, Hazard of Dropping Coverage 83 [4] 856 183 778 $1.65 $1.57 $1.67 $5.44 $3.08 $5.32 •$3.65 •$3.79 •$1.50 Early Buyers Large Discount, [3] 274 925 1197 $1.65 $1.60 $1.67 $5.29 $3.60 $5.48 •$3.81 •$3.64 •$2.00 Large Discount [2] 27 202 181 $9.79 $9.16 $9.80 $3.98 $6.23 $7.12 $5.82 $3.56 $2.05 Full Price, Full Early Buyers [1] 58 359 267 $9.93 $9.38 $9.91 $2.34 $6.94 $8.33 $7.58 $2.99 $1.05 Full Price Full Table 2.6: Financial Implications of Selection Number of Members, Month 1 Members, Month of Number Number of Members, Month 1 Members, Month of Number 1 Members, Month of Number All Buyers All months 6 Revenue, Average months 6 Revenue, Average months 6 Revenue, Average all Includes months 6 Margin, Potential purchase. SKY Average after months 6 first the for utilization data SKY on based household per costs Average finding Notes: by calculated utilizationis Average 2009. 31, December to 2008 1, January from inSKY enrolled households 6 the paidover premium anytheaverage household for averagethe months a utilization member that in is six each intaking of the thatby first month, calculated costs then are Revenues months. 6 these over utilizaton averaging month). ineach members of size household account into (taking months Average List Price of Services, 6 months 6 Services, of Price List Average months 6 Services, of Price List Average months 6 Services, of Price List Average Average Potential Margin, 6 months 6 Margin, Potential Average months 6 Margin, Potential Average Only Households w ithout "poor health" Only households w ithout a prior large shock 84

2.12 Figures 85

0.45 0.414 0.4

0.35 0.329

0.3

0.25

0.2

0.15 0.111 0.113 0.107 0.1 0.0849 Proportion Using or Log total Cost 0.05

0 No Controls*** Controls*** No Controls** Controls* No Controls*** Controls** Health Center Use, Proportion Hospital Use, Proportion Log (Total Cost Public Care)

Notes: Numbers presented are the coefficients from the regression of health center and hopsital utilization, and total cost, on having been offered the full price for insurance (no large coupon). Columns labeled "No Controls" use only an indicator variable for "Full Price" on the right hand side of the regression; these coefficients represent total adverse selection. Columns labeled "Controls" Use additional baseline controls; these coefficients represent hidden adverse selection (unobservable at the baseline). Stars represent significance of the coefficients in the regressions: * p < 005, ** p < •0.01, *** p < 0.001.

Figure 2.1: E¤ect of Full Price (not steep discount) on Utilization, with and without baseline controls 86

0.5 0.45 0.4 0.35 0.3 HC Use 0.25 Hosp Use 0.2 0.15 0.1

Proportion using Health Facility 0.05 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Months enrolled in SKY

Notes: Of households that are members for at least X months, the above chart shows the proportion that use a health center (HC) or hospital (Hosp) in that month. Thus, the number of SKY members is smaller in each subsequent month. Data is collected by SKY from health centers and hospitals. The chart includes utilization of any household who was a SKY member from January 2008 to December 2009.

Figure 2.2: Proportion of SKY Members Using SKY-Covered Health Facilities, by Tenure in SKY 87

0.6

0.5

0.4

0.3 HC Use, Deep Discount Hosp Use, Deep Discount 0.2 HC Use, Full Price Hosp Use, Full Price

Proportion using Health Facility Proportion using Health 0.1

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Months enrolled with SKY

Figure 2.3: Proportion of Households using SKY-covered Health Facilities for Care, by Premium and Tenure with SKY 88

2.A Supplementary Tables 89 528 196 195 5235 5267 5071 1754 1792 1597 1226 Total 68.0% purchase SKY following ever 348 N/A 986 N/A 2617 2537 2537 1334 1354 1354 52.6% Discount Offered Deep 420 438 243 180 196 240 195 2618 2730 2534 Price 15.4% Offered Full 1 Table 2.7: Research Sample Complete randomized baseline surveys, sample Randomized As Percent of Completed Baseline Bought 63 or more after days SKY meeting Complete oversample baseline surveys, Bought fewer than 63 days afterBought fewer than 63 SKY meeting days Oversample SKY buyers registered buyers afterSKY meeting date the Village Meeting. Not all of these households are members at once, as some drop out as others join. Notes: Households "Randomized at the Village Meeting" refers to the number of households chosen for our randomized survey sample. Of those, "Completed baseline surveys, randomized sample" refers to the number that completed the baseline survey. "Complete baseline surveys, oversample" refers to additional low coupon households that purchased SKY and that completed the baseline survey. SKY "Randomized" buyers are those that are chosen for our randomized sample and also purchased SKY. SKY "Oversample" buyers are those notwere part of the randomized sample but interviewed towere increase the number of households that bought SKY at the full price. SKY buyers can also be broken down into those that purchased soon after the Village Meeting (within 63 days) and a while after the Village Meeting (63 days or more). Households that SKY records indicate purchased SKY prior to the village meeting (around 43 households) are not included in the analyses. SKY take•up by full price households appears larger than the usual SKY take•up rate because we include any household that Randomized at the Village Meeting Complete Baseline Surveys, Total (withSKY Buyers SKYID, complete baseline) 90

Offered Deep Offered Full Discount, Clustered Price, Mean Mean ttest Observations 2534 2537 Highest ranked wealth by enumerator 0.13 0.14 •0.98 Lowest ranked wealth by enumerator 0.14 0.10 3.96 ** Answered all literacy/numeracy questions correctly 0.15 0.15 0.13 Household Size 5.03 5.02 0.31 Education of health decision•maker (years) 4.61 4.72 •1.13 At least one household member with poor self•reported 0.70 0.72 •1.15 At least one member over 64 0.25 0.26 •1.11 No child age 5 or under 0.55 0.57 •1.41 Household has a stunted or wasted child under age 6 0.16 0.15 0.88 All vaccines fulfilled for members under 6, 0 if no under 6, pre• SKY 0.27 0.25 0.96 Major health shock (*) in 2 to 4 months pre•Meeting 0.07 0.07 0.07 Major health shock (*) and used health center for care (0 if no shock) 0.01 0.02 •1.05 Major health shock (*) and used hospital for care (0 if no shock) 0.02 0.02 0.30 Major health shock (*) and use private health care (0 if no shock) 0.06 0.05 0.08 Ln of max days ill for a major health shock (*), pre meeting (0 if no shock) 0.24 0.25 •0.46 Major health shock (*) and spent 120,000 riel on care (USD30) (0 if no shock) 0.05 0.05 •0.26 Khmer household 0.953 0.946 2.00 * Ln of approximate value of animals, durables, and business (USD) 6.47 6.49 •0.64 Ln of approximate value of animals, durables, business, cash, and gold (USD) 6.68 6.74 •1.91 + Area of farm land owned by household (hectares) 0.81 0.86 •1.05 Area of village land owned by household (hectares) 0.14 0.13 0.90 Household has at least one toilet 0.26 0.26 0.34 House made of palm 0.04 0.03 2.23 * Roof made of palm 0.05 0.04 1.40 Roof made of tin 0.37 0.38 •0.53 Roof made of tile 0.51 0.52 •0.66 House made of brick 0.03 0.03 •0.41 AllNotes: variables + p<0.10, are from the* p<0.05, baseline **survey. p<0.01, *** p<0.001, based on t•tests clustered at the village level. (*) Sample is all high coupon households and all low coupon households in the randomized sample. indicates health shock causing missed daily activies for 7 or more days, a death, or an expense of over Ttest clustered at village level. 100USD. All data is from the baseline survey. Sample is all SKY decliners and all SKY buyers in the randomized sample.

Table 2.8: Randomization Summary Statistics 91 * * * * * ** 3575 0.248 0.559 0.022 0.158 0.070 0.015 0.053 0.232 0.045 0.672 Decliners 25.74% 1239 0.264 0.571 0.024 0.165 0.094 0.027 0.069 0.320 0.063 0.800 Buyers Pooled Households ** ** + ** ** ** 1183 0.265 0.571 0.055 0.011 0.016 0.045 0.188 0.035 0.154 0.651 Decliners 45.46% 986 0.272 0.578 0.089 0.029 0.023 0.064 0.314 0.062 0.162 0.787 Buyers Offered Deep Discount ** 2292 0.244 0.552 0.074 0.016 0.024 0.055 0.243 0.048 0.685 0.163 Decliners 9.48% 240 0.233 0.533 0.104 0.013 0.029 0.083 0.320 0.063 0.863 0.179 Offered Full Price Offered Full Buyers 1 Table 2.9: Summary Statistics for Selection on Observable Characteristics Major health shock (*) and used health center for care (0 if no shock) Major health shock (*) and used hospital for care (0 if no shock) Major health shock (*) and use care (0 if no shock)private health Ln of max ill for a major days health shock (*), pre meeting (0 if no shock) Major health shock (*) and spent 120,000 riel on care (USD30) (0 if no shock) Observations Percent Purchasing SKY Major Health Shock (*) in the 2 to 4 months pre•Meeting At least one household member with poor self•reported health At least one member over 64 No child age 5 or under Household has a stunted or wasted child under age 6 Notes: + p<0.10, * p<0.05, ** p<0.01, *** p<0.001, based on t•tests clustered (*) indicates at health shock the village level. causing missed activies for daily 7 or more a death, or days, an expense of over 100USD. Coupon status is recorded at the Village Meeting. All other data is from the baseline Sample survey. is all decliners plus households that63 bought SKY within of the meetingdays for which we have baseline data. (1) The low coupon percentage is the percentage coupon including all low households, which includes oversampled low coupon buyers. 92 N 240 179 175 ** 240 175 240 175 240 179 0.86 0.18 0.15 0.12 0.06 0.03 0.07 0.03 0.73 Early N 64 61 61 61 64 180 106 106 106 Bought at Full Price 0.33 0.20 0.08 0.03 0.07 0.06 0.70 Late N 986 + 0.81 679 958 + 0.28 986 679 986 679 986 958 0.79 0.12 0.14 0.11 0.06 0.04 0.04 0.03 0.63 Early N 69 69 69 348 304 134 134 134 304 0.74 0.19 0.08 0.11 0.06 0.03 0.07 0.02 0.67 Late Bought with Deep Discount Table 2.10: Summary Statistics for Early versus Late Buyers Visited health center in first 3 months post Baseline Data SKY, Major health shock (*) and use health center for care (0 if no shock) Visited hospital in first 3 months post Baseline Data SKY, Major health shock (*) and use hospital for care (0 if no shock) PRE•SKY/Pre•Meeting At least one household member with poor self•reported health POST•SKY Major Health Shock (*) in the 3 months post•SKY paid for hospital visit in firstSKY 3 months of SKY Notes: + p<0.10, * p<0.05, ** p<0.01, *** p<0.001, based on t•tests clustered (*) indicates at health shock the village level. causing activies missed daily for 7 or more a death, or days, an expense of over 100USD. Coupon status is recorded at the Village Meeting. Pre•meeting and pre•SKY variables are from the baseline Self•reported survey. poor health is reported at the time of the baseline which in some survey, cases is after initial SKY purchase. Pre•SKY variables are for the months immediately purchase. preceding Post•SKY SKY major shock variables are from All other the baseline survey. data post•SKY variables are from data. Sample means SKY of Early interviewed in the baseline bought survey. is all SKY buyers the within 63 days meeting, Late means all other buyers. Major Health Shock (*) in the 3 months pre•SKY SKY paid for health centerSKY visit in first 3 months of SKY * Major health shock as described in prevoius row. Pre•meeting and pre•SKY variables are from the baseline survey. Self•reported poor health is reported at the time of the baseline survey, which in some cases is a few months after initial SKY purchase. Major shock variables are for the 3 months preceding SKY purchase (not the village meeting). Post•SKY major shock variables are from the baseline survey. SKY data. from are variables post•SKY All other Early = bought within 63 days of the village meeting. 93

(1) (2) (3) (4) (5) Ln Cost Ln Cost Hosp, Ln Total HC Use Hosp Use HC, USD USD Cost, USD

SKY paid for health center visit in 0.359*** first 3 months of SKY (d) [0.0281]

SKY paid for hospital visit in first 3 0.182*** months of SKY (d) [0.0422]

Ln Cost of SKY•paid health center visits 0.602*** in first 3 months of SKY (USD) [0.0327]

Ln Cost of SKY•paid hospital visits in 0.139** first 3 months of SKY (USD) [0.0462]

Ln Cost of all SKY•paid care in first 3 0.412*** months of SKY (USD) [0.0412]

Observations 1261 1261 1280 1272 1271 Adjusted R•squared 0.34 0.013 0.127 Pseudo R•squared 0.09 0.028 Notes: Left•hand side variables use SKY utilization data for the 4 to 6 months post SKY purchase. CovariatesMarginal effects; use SKY Standard utilization errors data in forbrackets the 1 to 3 months post SKY purchase. Indicator variables are includedLHS cost (not variables shown) use to OLS,adjust other for households regressions that are dropped Probit. SKY in month 1 through 6. Cost variables useRobust OLS, standard all other errors regressions clustered use at Probit. the village + p<0.10, level. * p<0.05, ** p<0.01, *** p<0.001. Marginal effects; StandardCoupon status errors recorded in brackets. at village Robust meeting standard after errors the Luckyclustered Draw. at the village level. SKY status is from SKY statusadministrative is from SKYdata. administrative Coupon status data. recorded at the Village Meeting. Sample is all deep discount householdsLHS variables that use bought SKY utilizationSKY. (d) fordata discrete for the change4 to 6 months of indicator post SKYvariable purchase. from 0 to 1. RHS variables use SKY utilization data for the 1 to 3 months post SKY purchase.

Table 2.11: Autocorrelation of Health Expenses 94

(1) (2) (3) (4) (5) Offered Full Price (d) •0.379*** •0.376*** •0.379*** •0.379*** •0.384*** [0.0185] [0.0188] [0.0186] [0.0187] [0.0192]

At least one household member with poor 0.138*** 0.135*** 0.135*** 0.136*** self•reported health (d) [0.0140] [0.0141] [0.0141] [0.0149]

Major health shock (*) in 2 to 4 months 0.0761** 0.0528+ 0.139 0.142 pre•Meeting (d) [0.0286] [0.0289] [0.0968] [0.100]

Major health shock (*) and used health •0.0000219 •0.00463 center for care (0 if no shock) (d) [0.0585] [0.0599]

Major health shock (*) and used hospital •0.00649 0.00759 for care (0 if no shock) (d) [0.0544] [0.0588]

Major health shock (*) and use private •0.0503 •0.0664 health care (0 if no shock) (d) [0.0526] [0.0513]

Ln of max days ill for a major health •0.0188 •0.0148 shock (*), pre meeting (0 if no shock) [0.0190] [0.0198]

Major health shock (*) and spent 120,000 0.0365 0.0377 riel on care (USD30) (0 if no shock) (d) [0.0572] [0.0593]

At least one member over 64 (d) •0.0172 [0.0177]

At least one member age 5 or under (d) 0.00114 [0.0178]

Household has a stunted or wasted child 0.0061 under age 6 (d) [0.0228]

Observations 5229 5229 5229 5229 4871 Pseudo R•squared 0.141 0.129 0.142 0.142 0.147

Notes:Marginal LHS effects; variable: Standard 1 if bought errors SKY, in brackets 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** p<0.01,LHS Variable: *** p<0.001. 1 if bought Marginal SKY, effects;0 if declined Standard (SKY errors Admin in data)brackets. Robust standard errors clustered at theAll other village data level. is from (*) indicates the baseline health survey. shock causing missed daily activies for 7 or more days, a death, or an expense of over 100USD. All data is from the baseline survey. Wealth, household size and education are includedRobust standard in the regression errors clustered but not atpresented. the village Sample level. is all SKY decliners and buyers. (d) for discrete changeSample ofincludes indicator all variablebuyers that from purchased 0 to 1. SKY after the SKY meeting. (d) for discrete change of dummy variable from 0 to 1

Table 2.12: Robustness Check: Observable Selection using Full Sample (Early and Late Buyers) 95

(1) (2) (3) (4) (5) Offered Full Price (d) •0.362*** •0.361*** •0.362*** •0.362*** •0.368*** [0.0171] [0.0172] [0.0171] [0.0171] [0.0172]

At least one household member with poor 0.124*** 0.121*** 0.121*** 0.127*** self•reported health (d) [0.0125] [0.0125] [0.0125] [0.0131]

Miss 7 or more days of main activity due 0.0985** 0.0756* 0.0915 0.0837 to illness, 2 to 4 months pre•Meeting (d) [0.0348] [0.0343] [0.0701] [0.0688]

Major health shock (*) and used health 0.0307 0.025 center for care (0 if no shock) (d) [0.0544] [0.0560]

Major health shock (*) and used hospital •0.0244 •0.00884 for care (0 if no shock) (d) [0.0518] [0.0556]

Major health shock (*) and use private •0.019 •0.0301 health care (0 if no shock) (d) [0.0442] [0.0447]

Ln of max days ill for a major health •0.00784 •0.00221 shock (*), pre meeting (0 if no shock) [0.0170] [0.0176]

Major health shock (*) and spent 120,000 0.0461 0.0372 riel on care (USD30) (0 if no shock) (d) [0.0593] [0.0597]

At least one member over 64 (d) •0.0194 [0.0155]

At least one member age 5 or under (d) •0.00158 [0.0163]

Household has a stunted or wasted child 0.0167 under age 6 (d) [0.0208]

Observations 4701 4701 4701 4701 4387 Pseudo R•squared 0.168 0.154 0.169 0.169 0.179 MarginalNotes: LHS effects; variable: Standard 1 if bought errors SKY, in brackets 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** LHSp<0.01, Variable: *** p<0.001. 1 if bought Marginal SKY, effects;0 if declined Standard (SKY errors Admin in data)brackets. Robust standard errors clustered at (*)the Major village health level. shock (*) indicates means healthshock shockthat caused causing 7 days missed of misseddaily activies work in for the 7 or2 tomore 4 months days. pre•Meeting. All data is Robustfrom the standard baseline errors survey. clustered Wealth, at household the village size level. and education are included in the regression but not Allpresented. other data Sample is from is the all baselineSKY decliners survey. and all SKY buyers who bought SKY within 63 days of the SampleVillage Meeting. is all SKY (d) decliners for discrete and changeall SKY buyersof indicator who variablebought SKYfrom within 0 to 1. 63 days of the Village Meeting.

Table 2.13: Robustness Check: Observable Selection using only Health Shocks lasting 7 or more days 96

(1) (2) (3) Use HC Use Hosp Total Cost

Offered Full Price (d) 0.117*** 0.0733* 0.468*** [0.0344] [0.0357] [0.134]

At least one household member with poor 0.132*** 0.0158 0.372*** self•reported health (d) [0.0364] [0.0256] [0.102]

Major Health Shock (*) in the 2 0.0915 0.0334 •0.0993 to 4 months pre•SKY (d) [0.156] [0.127] [0.512]

Major health shock (*) and use health 0.167* 0.107 0.265 center for care (0 if no shock) (d) [0.0837] [0.0970] [0.272]

Major health shock (*) and use hospital 0.199** •0.0391 0.0987 for care (0 if no shock) (d) [0.0674] [0.0718] [0.261]

Major health shock (*) and use private 0.106 0.0855 0.429 health care (0 if no shock) (d) [0.109] [0.115] [0.312]

Ln of max days ill for a major health •0.0763+ •0.00551 0.0147 shock (*), pre SKY start (0 if no shock) [0.0454] [0.0277] [0.123]

Major health shock (*) and spent 120,000 0.0977 •0.0618 •0.173 riel on care (USD30) (0 if no shock) (d) [0.0996] [0.0543] [0.304]

At least one member over 64 (d) •0.0284 0.00983 •0.132 [0.0332] [0.0221] [0.0913]

At least one member age 5 or under (d) 0.0561 0.00233 0.127 [0.0511] [0.0375] [0.160]

Household has a stunted or wasted child 0.0199 •0.0374 0.0641 under age 6 (d) [0.0507] [0.0283] [0.146]

Observations 1199 1199 1199 Adjusted R•squared Pseudo R•squared 0.1 0.079 0.037

MarginalNotes: + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Column 1 (2): Indicator for Robustuse of a standard SKY•covered errors health clustered center at the(hospital) village for level. the first 3 months post SKY purchase; AdditionalColumn 3: covariatesLn of total notcost shown (user fees,(see Appendix).covered by SKY) of all SKY•covered health center and Clinichospital actual visits hours in the and first hygiene 3 months and post•SKY. inventory andColumns equipment 1•2 use score probit, are columnfrom the 3 Clinicuses OLS.Survey. TimeMarginal and effects; cost to Standardhealth center errors are in from brackets. the Village Robust Leader standard interview. errors clustered at the village Distancelevel. (*) toindicates health regionalhealth shock hospital causing is from missed interviews daily activieswith leaders for 7 at or the more Village days, Meeting. a death, or Allan otherexpense RHS of variablesover 100USD. are from SKY the status baseline and survey. LHS variables use SKY data. Coupon status is Lowrecorded coupon at the status Village recorded Meeting. at village All other meeting data isafter from the the Lucky baseline Draw. survey. Wealth, household SKYsize andstatus education, is from SKY and administrativeadditional variables data. as described in the Appendix, are included in the LHSregression variables but usenot presented.SKY utilization Sample data isfor all the SKY 3 months decliners post and SKY all purchase.SKY buyers who bought LHSSKY costfollowing variables the Village (in ln USD) Meeting. use OLS(d) for regressions. discrete change of indicator variable from 0 to 1.

Table 2.14: Robustness Check: Unobservable Selection using All Observable Covariates 97

(1) (2) (3) Use HC Use Hosp Total Cost

Offered Full Price (d) 0.113** 0.0742* 0.333** [0.0348] [0.0352] [0.112]

Joined 62 or fewer days after the •0.0567 0.0076 •0.0243 village meeting (d) [0.0467] [0.0269] [0.0900]

At least one household member with poor 0.136*** 0.0153 0.216** self•reported health (d) [0.0366] [0.0257] [0.0667]

Major Health Shock (*) in the 2 0.09 0.0352 •0.12 to 4 months pre•SKY (d) [0.155] [0.129] [0.401]

Major health shock (*) and use health 0.171* 0.106 0.178 center for care (0 if no shock) (d) [0.0823] [0.0972] [0.213]

Major health shock (*) and use hospital 0.204** •0.0395 •0.0408 for care (0 if no shock) (d) [0.0658] [0.0712] [0.205]

Major health shock (*) and use private 0.103 0.0865 0.357 health care (0 if no shock) (d) [0.109] [0.115] [0.236]

Ln of max days ill for a major health •0.0760+ •0.00575 0.0495 shock (*), pre SKY start (0 if no shock) [0.0452] [0.0280] [0.0925]

Major health shock (*) and spent 120,000 0.101 •0.0621 •0.225 riel on care (USD30) (0 if no shock) (d) [0.0987] [0.0540] [0.245]

At least one member over 64 (d) •0.0281 0.00975 •0.116+ [0.0331] [0.0220] [0.0659]

At least one member age 5 or under (d) 0.0535 0.00269 0.0544 [0.0513] [0.0376] [0.117]

Household has a stunted or wasted child 0.0224 •0.0375 0.0513 under age 6 (d) [0.0507] [0.0283] [0.113]

Observations 1199 1199 1199 Adjusted R•squared 0.05 Pseudo R•squared 0.102 0.08

Notes: + p<0.10, * p<0.05, ** p<0.01, *** p<0.001. LHS variables: Column 1 (2): Indicator for use of a SKY• Marginal effects; Standard errors in brackets covered health center (hospital) for the first 3 months post SKY purchase; Column 3: Ln of total cost (user Robust standard errors clustered at the village level. fees, covered by SKY) of all SKY•covered health center and hospital visits in the first 3 months post•SKY. Additional covariates not shown (see Appendix). Columns 1•2 use probit, column 3 uses OLS. Marginal effects; Standard errors in brackets. Robust Clinic actual hours and hygiene and inventory and equipment score are from the Clinic Survey. standard errors clustered at the village level. (*) indicates health shock causing missed daily activies for 7 Time and cost to health center are from the Village Leader interview. or more days, a death, or an expense of over 100USD. SKY status and LHS variables use SKY data. Distance to health regional hospital is from interviews with leaders at the Village Meeting. Coupon status is recorded at the Village Meeting. All other data is from the baseline survey. Wealth, All other RHS variables are from the baseline survey. household size and education, and additional variables as described in the Appendix, are included in the Low coupon status recorded at village meeting after the Lucky Draw. regression but not presented. Sample is all SKY decliners and all SKY buyers who bought SKY following SKY status is from SKY administrative data. the Village Meeting. (d) for discrete change of indicator variable from 0 to 1. This regression includes an LHS variables use SKY utilization data for the 3 months post SKY purchase. indicator variable for joining within 63 days of the Village Meeting. LHS cost variables (in ln USD) use OLS regressions.

Table 2.15: Robustness Check: Unobservable Selection using All Covariates, and Indicator for Early Buyer 98

(1) (2) (3) Use HC Use Hosp Total Cost

Offered Full Price (d) 0.114*** 0.0842* 0.312** [0.0339] [0.0364] [0.109]

At least one household member with poor 0.180*** 0.0173 0.272*** self•reported health (d) [0.0354] [0.0261] [0.0606]

Major Health Shock (*) in the 3 0.0216 •0.00395 •0.0864 months pre•SKY (d) [0.175] [0.0903] [0.311]

Major health shock (*) and use health 0.198** •0.0561 0.0329 center for care (0 if no shock) (d) [0.0688] [0.0371] [0.169]

Major health shock (*) and use hospital 0.137+ •0.0615 •0.255 for care (0 if no shock) (d) [0.0828] [0.0385] [0.169]

Major health shock (*) and use private 0.173* •0.02 0.122 health care (0 if no shock) (d) [0.0850] [0.0600] [0.208]

Ln of max days ill for a major health •0.0521 0.0287 0.0531 shock (*), pre SKY start(0 if no shock) [0.0433] [0.0251] [0.0817]

Major health shock (*) and spent 120,000 0.0237 0.0047 0.106 riel on care (USD30) (0 if no shock) (d) [0.0873] [0.0627] [0.185]

At least one member over 64 (d) •0.0289 0.00586 •0.123+ [0.0323] [0.0232] [0.0646]

At least one member age 5 or under (d) 0.0875* 0.000386 0.104 [0.0358] [0.0268] [0.0793]

Household has a stunted or wasted child 0.0601 •0.0358 0.0918 under age 6 (d) [0.0473] [0.0289] [0.102]

Observations 1215 1215 1215 Adjusted R•squared 0.045 Pseudo R•squared 0.055 0.028

MarginalNotes: + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Column 1 (2): Indicator for Robustuse of a standard SKY•covered errors health clustered center at the(hospital) village for level. the first 3 months post SKY purchase; RHSColumn variables 3: Ln of are total from cost baseline (user fees, data. covered by SKY) of all SKY•covered health center and Lowhospital coupon visits status in the recorded first 3 months at village post•SKY. meeting Columns after the 1•2Lucky use Draw. probit, column 3 uses OLS. SKYMarginal status effects; is from Standard SKY administrative errors in brackets. data. Robust standard errors clustered at the village LHSlevel. variables (*) indicates use SKYhealth utilization shock causing data for missed the 3 months daily activies post SKY for 7 purchase. or more days, a death, or LHSan expense cost variables of over (in100USD. ln USD) SKYuse OLSstatus regressions. and LHS variables use SKY data. Coupon status Allis recorded other LHS at variablesthe Village use Meeting. probit regressions. All other data is from the baseline survey. Wealth, RHShousehold dummies size areand included education, to adjustand additional for households variables that as dropped described SKY in thein month Appendix, 1, 2, orare 3. Sampleincluded is in all the households regression that but buynot presented.SKY for the Samplefirst time is after all SKY the declinersvillage meeting. and all SKY buyers who(d) for bought discrete SKY change following of dummythe Village variable Meeting. from (d)0 to for 1 discrete change of indicator variable +from p<0.10, 0 to 1.* p<0.05, These regressions** p<0.01, *** control p<0.001 for health shocks 1 to 3 months prior to joining SKY rather than 2 to 4 months prior to joining SKY.

Table 2.16: Robustness Check: Unobservable Selection Controlling for Health Shocks 1-3 Months pre-Baseline 99

(1) (2) (3) (4) (5) (6) Use Use Total Total Use HC Use HC Hosp. Hosp. Cost Cost

Offered Full Price (d) 0.0563 0.109* 0.108* 0.0925+ 0.304** 0.313* [0.0427] [0.0449] [0.0429] [0.0476] [0.114] [0.135]

At least one household member with poor 0.155*** 0.0183 0.252*** self•reported health (d) [0.0359] [0.0267] [0.0613]

Major Health Shock (*) in the 2 0.00618 0.146 •0.041 to 4 months pre•SKY (d) [0.183] [0.174] [0.444]

Major health shock (*) and use health 0.171* 0.0356 0.0473 center for care (0 if no shock) (d) [0.0802] [0.0765] [0.205]

Major health shock (*) and use hospital 0.151+ •0.0644 •0.0472 for care (0 if no shock) (d) [0.0893] [0.0513] [0.222]

Major health shock (*) and use private 0.0494 0.0171 0.168 health care (0 if no shock) (d) [0.114] [0.0867] [0.234]

Ln of max days ill for a major health •0.0451 •0.0254 0.0115 shock (*), pre SKY start (0 if no shock) [0.0441] [0.0281] [0.102]

Major health shock (*) and spent 120,000 0.150+ •0.0171 0.0466 riel on care (USD30) (0 if no shock) (d) [0.0776] [0.0626] [0.231]

At least one member over 64 (d) •0.0124 0.00777 •0.0867 [0.0332] [0.0227] [0.0623]

At least one member age 5 or under (d) 0.0926* 1.8E•05 0.101 [0.0385] [0.0273] [0.0847]

Household has a stunted or wasted child 0.0757 •0.0518* 0.0596 under age 6 (d) [0.0483] [0.0256] [0.101]

Observations 1380 1141 1380 1141 1383 1141 Adjusted R•squared 0.007 0.034 Pseudo R•squared 0.001 0.046 0.01 0.03

Notes: + p<0.10, * p<0.05, ** p<0.01, *** p<0.001. LHS variables: Columns 1 and 2 (3 and 4): Indicator Marginal effects; Standard errors in brackets for use of a SKY•covered health center (hospital) for the first 3 months post SKY purchase; Columns 5 Robust standard errors clustered at the village level. and 6: Ln of total cost (user fees, covered by SKY) of all SKY•covered health center and hospital visits in RHS variables are from baseline data. the first 3 months post•SKY. Columns 1•4 use probit, columns 5•6 use OLS. Marginal effects; Low coupon status recorded at village meeting after the Lucky Draw. Standard errors in brackets. Robust standard errors clustered at the village level. (*) indicates health SKY status is from SKY administrative data. shock causing missed daily activies for 7 or more days, a death, or an expense of over 100USD. SKY LHS variables use SKY utilization data for the 3 months post SKY purchase. status and LHS variables use SKY data. Coupon status is recorded at the Village Meeting. All other LHS cost variables (in ln USD) use OLS regressions. data is from the baseline survey. Wealth, household size and education are included in the regression All other LHS variables use probit regressions. but not presented. Sample is all SKY decliners and all SKY buyers who bought SKY following the Village RHS dummies are included to adjust for households that dropped SKY in month 1, 2, or 3. Meeting, only if they are part of the randomized sample (no oversampled low coupon buyers included). Sample(d) for discrete includes change households of indicator randomized variable into from the 0 study, to 1. but does not include oversampled buyers. Sample includes households that buy SKY for the first time after the village meeting

Table 2.17: Robustness Check: Unobservable Selection, No Oversampled Households 100

(1) (2) Total Cost, Tobit Total Cost, Tobit

Offered Full Price 0.565*** 0.458*** [0.130] [0.132]

At least one household member with poor 0.447*** self•reported health [0.0958]

Major Health Shock (*) in the 2 •0.161 to 4 months pre•SKY (d) [0.455]

Major health shock (*) and use health 0.174 center for care (0 if no shock) [0.238]

Major health shock (*) and use hospital •0.0409 for care (0 if no shock) [0.265]

Major health shock (*) and use private 0.126 health care (0 if no shock) [0.300]

Ln of max days ill for a major health 0.0307 shock (*), pre SKY start (0 if no shock) [0.113]

Major health shock (*) and spent 120,000 0.209 riel on care (USD30) (0 if no shock) [0.274]

At least one member over 64 •0.126 [0.0888]

At least one member age 5 or under 0.201+ [0.108]

Household has a stunted or wasted child 0.136 under age 6 [0.130]

Observations 1508 1255 Adjusted R•squared Pseudo R•squared 0.009 0.024 MarginalNotes: + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Ln of total cost (user fees, Tobitcovered regressions, by SKY) of standard all SKY•covered errors are health clustered center at andthe villagehospital level visits but in not the robust). first 3 months post• RHSSKY, variablesusing Tobit are regressions. from baseline Marginal data. effects; Standard errors in brackets. Robust standard Lowerrors coupon clustered status at therecorded village atlevel. village (*) meetingindicates after health the shock Lucky causingDraw. missed daily activies for 7 or SKYmore status days, isa death,from SKY or an administrative expense of over data. 100USD. SKY status and LHS variables use SKY data. LHSCoupon variables status useis recorded SKY utilization at the Village data for Meeting. the 3 months All other post data SKY is purchase. from the baseline survey. RHSWealth, dummies household are includedsize and toeducation adjust for are households included in that the dropped regression SKY but in not month presented. 1, 2, or 3.Sample is Sampleall SKY declinersis all households and all SKY that buybuyers SKY who for boughtthe first SKY time following after the thevillage Village meeting. Meeting, only if they are(d) partfor discrete of the randomized change of dummy sample variable(no oversampled from 0 to low1 coupon buyers included). (d) for discrete +change p<0.10, of *indicator p<0.05, variable** p<0.01, from *** 0p<0.001 to 1.

Table 2.18: Robustness Check: Unobservable Selection, Tobit regression for Costs 101

(1) (2) (3) (4) (5) (6) Use Use Total Total Use HC Use HC Hosp. Hosp. Cost Cost

Offered Full Price (d) 0.111*** 0.114*** 0.107** 0.0860* 0.414*** 0.328** [0.0321] [0.0328] [0.0342] [0.0360] [0.106] [0.109]

At least one household member with poor 0.163*** 0.0198 0.253*** self•reported health (d) [0.0344] [0.0252] [0.0592]

Miss 7 or more days of main activity due 0.190* 0.0645 0.211 to illness, 2 to 4 months pre•SKY (d) [0.0904] [0.132] [0.337]

Major health shock (*) and use health 0.141+ 0.0624 0.0186 center for care (0 if no shock) (d) [0.0820] [0.0810] [0.184]

Major health shock (*) and use hospital 0.107 •0.0619 •0.165 for care (0 if no shock) (d) [0.0899] [0.0541] [0.190]

Major health shock (*) and use private •0.0082 0.0188 0.00855 health care (0 if no shock) (d) [0.111] [0.0877] [0.239]

Ln of max days ill for a major health •0.0761* •0.0177 •0.0189 shock (*), pre SKY start (0 if no shock) [0.0361] [0.0284] [0.0841]

Major health shock (*) and spent 120,000 0.150* 0.00028 0.116 riel on care (USD30) (0 if no shock) (d) [0.0746] [0.0699] [0.207]

At least one member over 64 (d) •0.0251 0.00606 •0.115+ [0.0316] [0.0228] [0.0631]

At least one member age 5 or under (d) 0.0846* •0.0025 0.109 [0.0347] [0.0264] [0.0774]

Household has a stunted or wasted child 0.068 •0.0371 0.0894 under age 6 (d) [0.0455] [0.0284] [0.101]

Observations 1505 1255 1505 1255 1508 1255 Adjusted R•squared 0.024 0.045 Pseudo R•squared 0.009 0.05 0.017 0.029

Notes:Marginal + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Columns 1 and 2 (3 and 4): Indicator forRobust use ofstandard a SKY•covered errors clustered health center at the (hospital)village level. for the first 3 months post SKY purchase; Columns 5 andRHS 6: variables Ln of total are cost from (user baseline fees, data.covered by SKY) of all SKY•covered health center and hospital visits in theLow first coupon 3 months status post•SKY. recorded at Columns village meeting 1•4 use after probit, the columns Lucky Draw. 5•6 use OLS. Marginal effects; StandardSKY status errors is from in brackets.SKY administrative Robust standard data. errors clustered at the village level. (*) indicates major healthLHS variables shock causing use SKY missed utilization daily data activities for the for 3 7months or more post days. SKY SKY purchase. status and LHS variables use SKY data.LHS cost Coupon variables status (in is ln recorded USD) use at OLSthe Village regressions. Meeting. All other data is from the baseline survey. Wealth,All other householdLHS variables size useand probit education regressions. are included in the regression but not presented. Sample is all SKYRHS declinersdummies andare allincluded SKY buyers to adjust who for bought households SKY following that dropped the Village SKY in Meeting. month 1, (d) 2, orfor 3. discrete changeSample ofis allindicator households variable that from buy 0 SKY to 1. for the first time after the village meeting.

Table 2.19: Robustness Check: Unobservable Selection, Only Shocks lasting more than 7 days 102

(1) (2) Inpatient Visit Inpatient Visit

Offered Full Price (d) 0.0780** 0.0571* [0.0257] [0.0236]

At least one household member with poor 0.0303* self•reported health (d) [0.0120]

Major Health Shock (*) in the 2 0.0141 to 4 months pre•SKY (d) [0.0669]

Major health shock (*) and use health 0.0264 center for care (0 if no shock) (d) [0.0517]

Major health shock (*) and use hospital •0.0281 for care (0 if no shock) (d) [0.0220]

Major health shock (*) and use private •0.00273 health care (0 if no shock) (d) [0.0398]

Ln of max days ill for a major health 0.00131 shock (*), pre SKY start (0 if no shock) [0.0141]

Major health shock (*) and spent 120,000 •0.00699 riel on care (USD30) (0 if no shock) (d) [0.0292]

At least one member over 64 (d) •0.0118 [0.0130]

At least one member age 5 or under (d) 0.0167 [0.0175]

Household has a stunted or wasted child •0.00531 under age 6 (d) [0.0187]

Observations 1487 1239 Pseudo R•squared 0.027 0.055

MarginalNotes: + p<0.10,effects; Standard* p<0.05, errors** p<0.01, in brackets *** p<0.001. LHS variables: Indicator for an inpatient Probitvisit at regression,a SKY•covered robust hospital standard in the errors first clustered3 months atpost the SKY village purchase, level. using probit RHSregression. variables Marginal are from effects; the baseline Standard survey. errors in brackets. Robust standard errors clustered Lowat the coupon village statuslevel. recorded(*) indicates at village health meetingshock causing after the missed Lucky daily Draw. activies for 7 or more SKYdays, status a death, is from or an SKY expense administrative of over 100USD. data. SKY status and LHS variables use SKY LHSdata. variables Coupon usestatus SKY is recordedutilization at data the forVillage the 3 Meeting. months post All other SKY datapurchase. is from the baseline LHSsurvey. cost Wealth, variables household (in ln USD) size use and OLS education regressions. are included in the regression but not Allpresented. other LHS Sample variables is all use SKY probit decliners regressions. and all SKY buyers who bought SKY following the RHSVillage dummies Meeting. are (d) included for discrete to adjust change for ofhouseholds indicator variablethat dropped from 0SKY to 1. in month 1, 2, or 3. Sample is all households that buy SKY for the first time after the village meeting.

Table 2.20: Robustness Check: Unobservable Selection, Inpatient Visits 103

2.B Other Datasets

2.B.1 Village Leader Interview In each village we interviewed the village chief or another village leader in order to collect general village-level information, including travel time and cost to the nearest public health center; recent village-level shocks (drought, ‡ood, epidemics, etc.); availability of lending institutions; and price and availability of paid transportation. Selected variables from this survey are used only as controls in the unobserved selection analysis, and are described in the Appendix to Chapter 3.

2.B.2 Health Center Data Collection Households may be more likely to purchase SKY if the quality of the local public health clinic with which SKY partners is of good quality. To measure this, we administered a simple survey of health clinics in areas covered by our study. To minimize data collection costs, the health center survey focuses on observations by SKY member facilitators. SKY hires member facilitators to be present at health facilities to facilitate treatment for SKY members and manage client complaints and questions as needed. Member facilitators typically work mornings at one particular Health Center. The survey consists of checklists of operating hours, drug supply, cleanliness, and equipment supply. In the current paper, these variables are used only as controls in the selection on unobservables analysis. These variables are listed in the Appendix to Chapter 3.

2.B.3 Village Meeting Data At the end of each village meeting, our …eld team spoke to a village leader to collect village-level data relevant to our study. The Appendix to Chapter 3 lists variables created from responses to these interviews. In the results presented in this paper, we use these variables only as controls in the analysis of selection on unobservables.

2.C Lucky Draw Implementation

To implement the Lucky Draw, attendance was taken at the beginning of each meeting, and names of people who arrived late to the meeting were added to the roster as they arrived. We collected one name for each household represented at the meeting, so that households with more than one member attending were not counted more than once. While a SKY representative conducted the meeting, sta¤ from our …eld team counted out the number of high and low coupons to be distributed to the meeting attendees. The number of high coupons was set equal to 20% of households up to a cap of 12 per meeting. The high coupons were put into a bag along with enough low coupons to cover all other households attending the meeting. At the end of the meeting, the research team’s …eld coordinator announced that there would be a ra• e where the prize is a large-valued coupon for insurance and explained 104 the rules of the coupon. Family names were called o¤ one by one from the roster. When a name was called o¤, a representative from the household came to the front of the room and pulled a coupon from the bag. High coupons were brightly colored so that everyone could see what coupon was drawn after the fact but care was taken to ensure people could not see coupons as they were drawing. As coupons were drawn, the names of households who received a high coupon were recorded so that coupons could not be traded and used by other households.

2.D Description of Variables, Adverse Selection 105

2.D.1 Independent Variables The following variables were used on the left hand side of the main speci…cations. 106

Independent Variable Description Name Purchase SKY 1 if purchase SKY, 0 if not (observable selection regression) Health Center Use, …rst 3 1 if any household member months used a SKY-partnered health center in the …rst 3 months after insurance purchase, 0 if not (SKY data) (unobservable selection regression) Hospital Use, …rst 3 months 1 if any household member used a SKY-partnered hospital in the …rst 3 months after insurance purchase, 0 if not (SKY data) (unobservable selection regression) Cost of Health Center and Log of total cost of visits to Hospital Visits, …rst 3 months a SKY-partnered health center or hospital in the …rst 3 months after insurance purchase, by any household member (SKY data) (unobservable selection regression)

Table 2.21: Independent Variables, Chapter 2 107

2.D.2 Basic Covariates

The following is a description of the basic controls used in regressions measuring adverse selection. Additional controls used in robustness tests can be found in the Chapter 3 Appendix.

Variable Questionnaire Question Description Name Subjective How healthy is each household 1 if respondent describes poor health member? (Excellent health, health of any household good health, poor health). member as “poor” Primary respondent to questionnaire gives subjective response for all household members. Major health Three questions: In the last 1 if respondent answers “yes” shock, 2-4 year, were there any health to any of these three health months pre- problems in your household questions, AND the month meeting. (We that made someone unable to of the health shock was 2-4 use shocks work or go to school for one months prior to the date of the 1-3 months week or more? In the last year SKY meeting. pre-meeting as did anyone in your household a robustness pass away? In the last year check.) did anyone in your household spend more than 400,000 riel ($100 USD) on a single health problem? Visited a [If household member expe- 1 if, following a major health health center rienced major shock in 2- shock in the 2-4 months pre- for a major 4 months pre-meeting:] Did meeting, a household member health shock, [sick member] seek treatment visited a public health center 2-4 months for this health problem? If yes, for …rst or subsequent treat- pre-meeting where? [Respondent chose ment “Health center”] Visited a pub- [If household member expe- 1 if, following a major health lic hospital rienced major shock in 2- shock in the 2-4 months pre- for a major 4 months pre-meeting:] Did meeting, a household member health shock, [sick member] seek treatment visited a public health center 2-4 months pre for this health problem? If yes, for …rst or subsequent treat- meeting where? [Respondent chose ment “public hospital”] 108

Variable Questionnaire Question Description Name Visited a pri- [If household member expe- 1 if, following a major health vate facility rienced major shock in 2- shock in the 2-4 months pre- for a major 4 months pre-meeting:] Did meeting, a household member health shock, [sick member] seek treatment visited a private doctor for 2-4 months for this health problem? If …rst or subsequent treatment pre-meeting yes, where? [Respondent chose “private doctor (village or town)”] Spent more [If household member(s) Sum of treatment costs at any than $30 USD treated for a major shock in 2- facility (including traditional (120,000 riel) 4 months pre-SKY], what was healers, drug sellers, etc.) for for a major the cost of treating the health all household members experi- health shock, problem (at any facility)? encing a major shock in the 2-4 2-4 months months pre-SKY pre-meeting Max number [If household member expe- Maximum days ill for any sin- of days ill, rienced major shock in 2- gle health problem for any 2-4 months 4 months pre-meeting:] How household member in the 2-4 pre-meeting many days was he/she unable months before the SKY meet- do his/her usual activities being. cause of this health problem? Household has Height, age, and weight mea- 1 if household has a child that a stunted or sured for all children age 5 and is stunted or wasted (zscore wasted child under for height-for-age or weight- for-height is less than -2) according to WHO growth standards, 0 otherwise (including if household has no child age 5 or under) Household has Date of birth of each house- 1 if any household member is a member age hold member age 65 or older 65 or older Household has Date of birth of each house- 1 if any household member is at least one hold member age 5 or under, 0 otherwise member age 5 or under Household size Household roster: Name of Number of household mem- people who usually sleep here bers listed in the household (slept in the house 5 out of the roster 7 nights immediately preceding the interview) 109

Variable Questionnaire Question Description Name Poor house- Enumerator subjective 1 if enumerator rates household wealth ranking: poor- hold as poor est/medium/better o¤ Better-o¤ Enumerator subjective 1 if enumerator rates house- household wealth ranking: poor- hold as better-o¤ est/medium/better o¤ Respondent is (Round 2 survey) Four lit- 1 if respondent answers all lit- literate and nu- eracy/numeracy questions: eracy and numeracy questions merate Draw a line from each picture correctly to the correct word; Write the name of the village, commune and district where you live; Write the correct number of objects in the pictures, and what the object is; Tell me what time it is (picture of a clock shown) Education Who makes the decisions Education from 1 to 13 (13 of health about healthcare in your = tertiary education). If re- decision-maker family? What is the highest spondent decides with another (years) grade this person completed? household member, use max- What is the highest grade you imum education of the two completed? members. Indicator variables for 0 years or 1 to 4 years used in regressions. Table 2.22: Basic Covariates used in Chapter 2 110

Chapter 3

Going Beyond Adverse Selection: Take-up of a Health Insurance Program in Rural Cambodia

3.1 Introduction

Voluntary health insurance has gained popularity recently as a potential health policy tool for poor nations. To understand both the e¤ects of insurance and its …nancial viability, we must understand who purchases voluntary health insurance. Standard economic theory predicts that households that anticipate high insurance costs are those that are willing to purchase health insurance (e.g., Rothschild and Stiglitz 1976; Akerlof 1970). Strong adverse selection may make it di¢ cult for private voluntary insurance to survive. More optimistically, standard economic theory also predicts risk-averse households will also value insurance. If such households are also very cautious, insurance may ‡ourish in the market and grow to cover a substantial share of the population. Several other factors may also lead a household to purchase insurance, such as the quality (real or perceived) of the health facilities connected with the health insurance, convenience of visiting a covered facility, or budget constraints. At the same time, if consumers do not understand or trust insurers, private voluntary insurance will not provide an e¤ective safety net. To understand more about the factors leading a household to purchase insurance, we study the characteristics of households that purchase or abstain from purchasing SKY health insurance in rural Cambodia. SKY partners with public health facilities and provides free care in exchange for a small (subsidized) monthly premium. We utilize data from two survey waves from over 5000 households who attended a marketing meeting for SKY insurance. These surveys asked about health and health utilization behaviors when households were …rst o¤ered insurance and again one year later. The extent of adverse selection in this insurance market is explored in Chapter 2. This chapter goes beyond adverse selection and examines other factors that may induce households to purchase or abstain from purchasing insurance. The richness of our data allows us to provide empirical evidence on a wide variety of measures of selection into a 111 health insurance program. We also provide some of the …rst evidence on selection into a health insurance program in the context of a developing country, in which incentives to buy insurance may di¤er somewhat from a developed country. Contrary to standard economic theory, we …nd no evidence that risk averse households are more likely to purchase SKY, and instead …nd evidence of the opposite. Budget constraints, low quality of health facilities, and age and gender of ill household members also in‡uence the decision to purchase insurance.

3.2 The Setting

A description of the Cambodian setting and the SKY program can be found in Chapters 1 and 2. Below I describe in more detail the choice of health care providers in Cambodia.

3.2.1 Providers Cambodians rely on a mix of healthcare providers: public providers, private medical healthcare providers, private drug sellers (with and without pharmaceutical training) and traditional healers.1 SKY partners only with public providers, because these are the only providers regulated by the government. Thus, a household’s preference for and perceptions of providers may be an important determinant of take-up in this market. Cambodia’s public health system has three basic levels of healthcare facilities: Provincial-level hospitals, Operational District (OD) Referral Hospitals, and community Health Centers. The highest level of public care available within a province is at the Provin- cial Hospital in each province. Provinces are divided into several operational districts— a division speci…c to the healthcare system (that is, they di¤er from sub-provincial political administrative districts). Each OD has one (usually small) district-level Referral Hospital and an average of eleven Health Centers. Each Health Center, in turn, serves several villages and around 13,500 people on average. (Based on data from the Cambodian Government 2004.) Public facilities su¤er from low utilization rates. According to 2005 DHS estimates, of those who sought treatment for illness or injury, fewer than a quarter went to a public health facility (Table 3.1). Even a smaller percentage of second and third treatments were sought at public facilities. Typical complaints about public facilities in Cambodia include having to engage in costly and time-consuming travel to reach facilities (rather than seeking treatment nearby or even receiving home care visits from private providers), personnel absence from public health facilities, long waiting times at facilities, frequent shortages of medicines, unpredictable costs and poor health-worker attitudes toward patients (for example, scolding or belittling of patients) (Collins 2000; Annear 2006). Private providers of varying capabilities are typically more popular than public ones, even when more expensive, because they often are more attentive to clients’needs, more available, willing to visit patients in their homes, and willing to provide more of the treatments patients prefer. They are also usually willing to extend credit of various types

1This description of Cambodian health care providers draws on Levine and Gardner (2008). 112 to clients (Collins 2000; Annear 2006). However, private facilities are not regulated and may provide inappropriate care such as improper drug prescriptions (Fort, Ravenholt, and Stanley 1998) and high rates of unnecessary (and sometimes unsafe) injections (Vong, Perz, Sok, Som, Goldstein, Hutin, and Tulloch 2005). In addition, local private providers are usually not capable of treating more serious illnesses. For those, patients must seek care at public hospitals. Private hospitals are also available but are very costly, and are often used by only the wealthiest individuals. Self-medication through purchase from local uncerti…ed drug sellers is quite common. Pharmacists and other drug sellers are often situated near local public markets and so are more conveniently located than most public health centers. In addition, they are usually cheaper than a clinic and are willing to provide any medicine requested by customers. Thus drug sellers of various types are usually the …rst (and often the only) place rural Cambodians seek treatment for their illnesses (DHS 2005). Though traditional healers and midwives also serve clients in both urban and rural areas, they are less common and are responsible for less than two percent of care sought (DHS 2005).

3.3 Literature Review

We break down hypotheses into traditional measures and more modern extensions to these measures (Table 3.2). The dividing line is inherently arbitrary.

3.3.1 Traditional Insurance Theory The standard economic theory of the demand for insurance predicts that insurance markets will su¤er adverse selection, which occurs when less healthy people or people who are more risky with their health are more willing to purchase health insurance because they know that the amount they spend on healthcare will be larger than the premium they will pay (e.g., Rothschild and Stiglitz 1976; Akerlof 1970). Living near a high-quality facility that is covered by insurance may also increase the likelihood that a household will …nd insurance valuable. Traditional models also posit that the risk averse will be more likely to buy insurance because they su¤er greater loss of utility in the presence of health expenses. If risk aversion increases the probability of insurance purchase but decreases the amount of risk one takes with one’shealth, it may mitigate the impact of adverse selection (e.g., Chiappori, Jullien, Salanié, and Salanié 2006; Jullien, Salanié, and Salanié 2007). Those who anticipate liquidity constraints if they face large health expenses will have higher value for insurance. Thus, the ability to self-insure can reduce the demand for insurance (Morduch 1995; Townsend 1994; Udry 1994).

3.3.2 Recent Theory Recent theoretical work has focused on how the problem of adverse selection may also be mitigated by factors such as wealth (which could both increase the probability of insurance purchase and improve health outcomes) (e.g., Case, Lubotsky, and Paxson 2002; 113

Smith 2005; Currie and Stabile 2003), and optimism (where some people underestimate their accident probability, and thus don’t buy insurance, but are also less willing to take precautions, leading to a higher probability of a health shock) (e.g. Koufopoulos 2002). Cutler and Zeckhauser (2004) describe several other cases in which insurance markets act di¤erently from what standard theory would predict. For example, the authors give several examples of people over-insuring in a way that standard theory would deem irrational, and cite theories of why this may occur. People may buy more insurance than standard theory would predict if the risk is particularly salient in their mind, as may be the case when people buy additional insurance immediately before an airplane trip (Tversky and Kahneman 1974). Similarly, if a household knows someone who has been very ill or had high medical expenses in the past, they may increase their subjective probability of a costly shock. Present biased households or those with liquidity constraints may under-insure and also fail to invest in many other precautions. (This hypothesis is subtly di¤erent from the point that households not currently liquidity constrained will have a high value of insurance if they anticipate liquidity constraints in the case of a costly health shock.) More broadly, it is likely that demand for insurance will be lower from households that do not trust Western medicine or do not understand how health insurance works. If a household puts less weight on healthcare of certain household members, such as women, expected medical expenses will be lower for these household members, even when there is a high probability they are ill. Thus, illness among such groups may have a weaker e¤ect on insurance demand than illness among other household members.

3.3.3 Developing Country Context Most theories have been developed with developed countries in mind. On the one hand, potential customers in developing countries may be expected to behave in a similar way to those in more developed countries: those with higher expected health care expenses (or lower baseline health levels) are expected to purchase more insurance. On the other hand, there are several reasons to believe that clients in developing countries may behave di¤erently than what has been found in developing countries. For example, because insurance is a relatively new and unknown product, only those who are willing to take a risk on a new product may be willing to try it. Similarly, households often have not been exposed to insurance, and thus may not understand the concept of paying ahead for an uncertain risk (Giné, Townsend, and Vickery 2007) Credit constraints that bind households in developed countries may be an even more important factor in a context where many people often live on a dollar per day. Thus, wealthier households may be better able to a¤ord insurance. At the same time, wealthier households may be better able to self-insure and thus may be less likely to buy insurance (Giné, Townsend, and Vickery 2007). Developing countries also must contend with inconsistent quality at health facilities, and inability to travel due to poor quality roads or lack of transportation. Households may choose not to buy insurance if they perceive the quality of covered care as low, or if covered facilities are a long distance away (Fuller 1974; Mwabu, Ainsworth, and Nyamete 1993; Banerjee, Deaton, and Du‡o 2004). Finally, bias towards the care of males, or towards the care of productive household 114 members, may have more in‡uence on insurance purchase in a developing country setting (Gupta 1987; Pande 2003; Sauerborn, Berman, and Nougtara 1996).

3.3.4 Empirical Literature There is an extensive empirical literature on the extent of adverse selection in insurance markets in developed countries. This literature is discussed in Chapter 2. While many studies …nd evidence of adverse selection (Cutler and Zeckhauser 2000, give a good review of existing studies), others …nd a surprising lack of correlation between health utilization and purchase of health insurance (e.g., Finkelstein and McGarry 2006; Cardon and Hendel 2001). The lack of correlation may be explained by o¤setting e¤ects of “advantageous”or positive selection. Several studies …nd evidence of these non-traditional sources of selection into insurance. For example, Finkelstein and McGarry (2006) …nd that people with risk-avoiding behaviors are less likely to use a nursing home but more likely to buy long-term care insurance. Fang, et al., (2008) note that selection into insurance on any characteristic can be considered advantageous selection if that characteristic is both positively correlated with purchase and negatively correlated with health risk. They show that buyers of Medigap insurance in the U.S. tend to be healthier than non-buyers, and they show that the source of this advantageous selection includes income, education, longevity expectations, …nancial planning horizons, cognitive ability, and …nancial numeracy. They …nd that risk preferences are not in fact a major source of advantageous selection, as hypothesized in other papers. The authors hypothesize that the correlation of cognitive ability with insurance purchase coincides well with reports that seniors do not always understand the rules of Medigap. This …nding is relevant to the current research, in which we hypothesize that many potential customers do not understand fully the concept of insurance. Note that Medigap is similar to SKY insurance in that within the six month open enrollment period, private insurers cannot discriminate on previous health conditions, much like SKY does not. Despite the plethora of empirical work in developed nations, there have been far fewer studies of selection in developing countries, partly because there are far fewer insurance markets in developing countries. Non-experimental studies from developing countries have found enrollment to be more common in households with chronically sick members, which is evidence of adverse selection (e.g. Wagsta¤, Lindelow, Jun, Ling, and Juncheng 2009, in China), but commonly …nd higher enrollment rates in wealthier households, potentially leading to positive selection if wealthier people also tend to be healthier (e.g. Wagsta¤, Lindelow, Jun, Ling, and Juncheng 2009; Wagsta¤ and Pradhan 2005, in Vietnam; Jutting 2004, in Senegal; Lamiraud, Booysen, and Scheil-Adlung 2005, in South Africa). In contrast, Jalan and Ravallian (1999) found that wealthier households in rural China were better equipped to self-insure against income shocks. It is possible that households who are better able to self insure are less likely to buy insurance. Gine, Townsend and Vickery (2007) study the determinants of take-up of a rainfall insurance product in rural India. As predicted by traditional models of take-up behavior, those with fewer credit constraints, the wealthier, and those who plant more crops covered by the insurance were more likely to purchase insurance. Less in line with traditional models, farmers who were more risk averse were less likely to purchase insurance and those 115 more familiar with the insurer were more likely to buy insurance. We would expect these results to carry over to a health insurance product: a lack of knowledge regarding insurance, and an unwillingness to take a risk on a new product, may also be important determinants of take-up in the Cambodian health insurance context. Recall that the SKY insurance program covers care only at public health centers and hospitals with which they partner. The convenience of these health facilities (e.g., time to travel to the facility and operating hours) and perceived quality (e.g., availability of necessary equipment and cleanliness of the facility) may also in‡uence purchase. In Chile, Fuller (1974) found that distance to a fertility clinic was the single most powerful indicator of utilization of contraception and Mwabu, et al., (1993) found similar results in Kenya for use of health facilities. However, it has been more di¢ cult to …nd impacts of quality on utilization of care. In the Kenya study, Mwabu, et al., found no impact of drug availability on utilization (although results may have been biased downward by endogeneity issues). Similarly, Banerjee, et al., (2004) also found that low quality of care in rural Rajasthan, India, (measured by training of medical sta¤, unscheduled closing of facilities, infrequent testing accompanied high frequency administration of injections and drips) did not deter utilization of facilities, nor patient perception of care. It is possible that these measures of quality are not correlated with perceived measures of quality by patients. Several studies have shown gender discrimination amongst households, whereby male household members are favored in either nutrition or health care (e.g., in India, Gupta 1987 and Pande 2003). Other studies have shown no gender bias, but have shown that families spend more on health care for productive family members, and less on care of children or the elderly (Sauerborn, Berman, and Nougtara 1996). If families are less willing to pay for care of females, children, or the elderly, they may also be less willing to buy health insurance unless a male or productive household member is ill. Finally, because we induce random variation in the price of SKY, we can say something about how demand varies with price (price elasticity). There is a newly emerg- ing literature on demand for health and health care services. On the one hand, some studies have found that demand for coverage of acute illness (e.g., malaria), is relatively inelastic (Dupas 2011), possibly because households insure against health risk through social networks (Townsend 1994; Robinson and Yeh 2011, as referenced in Dupas 2011). On the other hand, demand for preventative services such as bednets, water treatment, and deworming products has been found to be very price elastic (Kremer, et al., 2011; Cohen and Dupas 2010; Kremer and Miguel 2007; Abdul Lateef Jameel Poverty Action Lab 2011). If we consider health insurance more akin to a “preventative”service, we may expect demand for insurance to be relatively elastic. The research presented here adds to the literature in several ways. First, empirical studies have taken place for the most part in developed countries. As discussed, selection among the poor in developing countries may be very di¤erent than that described in the existing literature. Second, the above empirical studies have taken place in more traditional competitive markets, whereas the SKY program in Cambodia is the only health insurance option in the rural markets targeted. Since there is no plan choice, adverse selection may show up di¤erently in this market, as individuals must choose SKY insurance or nothing at all. 116

Finally, because this study examines a population previously unexposed to insurance, di¤erences in characteristics of households who buy or decline insurance at the baseline have not been in‡uenced by prior insurance contracts. These in‡uences on take-up are an important aspect of both e¤ective targeting and …nancial viability of the insurance program. For the insurance program to have the greatest impact, it must be taken up by as many households as possible. Likewise, to maintain …nancially stability, marketing must target households in the most e¢ cient way possible.

3.4 Speci…cation

To investigate self-selection in take-up, we perform a probit estimation of the following equation, where i is a household-level observation and v is a village-level observation.

;H ;M ;M P ubl ;M P riv ;Z ;D ; ; fac ; qual ; finrisk ; SKY = F i i i i i i i i i v v i (3.1) i hlthrisk ; selfins ;W ; disc ; u ; sal ; trust ; pref ;" i i i i i i i i i

Here, the independent variable SKYi = 1 if the household accepts insurance. i is an indicator variable equal to 1 if a household received a large discount for SKY. Hi is subjective health, equal to 1 if at least one household member is in poor self-reported health; Mi is a measure of past health care shocks (presence of a health shock in the months prior to the SKY village meeting); P ubli and P rivi indicate a visit to a public facility or private facility, respectively, following a health shock; Zi is a measure of objective health characteristics of children under age …ve (an indicator variable equal to 1 if the household has a stunted or wasted child aged 5 or under); Di is a set of demographic characteristics of the household (number of household members, indicator variables for old or young members, education of the health care decision maker); facv are measures of distance and cost to travel to public facilities covered by SKY; qualv is a measure of the quality of public facilities; finriski is a measure of …nancial risk aversion; hlthriski is a measure of risks households take with their health; selfinsi is a measure of the ability to self-insure for health shocks; Wi is a measure of a household’s wealth, as observed by the enumerator; disci is a measure of a household’s discount rate; ui is a measure of a household’sunderstanding of insurance; sali is a measure of salience of health shocks; trusti are measures of trust of western medicine, including a variable equal to 1 if all children under 6 have received all recommended vaccinations and a variable equal to 1 for always covering water jugs; prefi are variables representing preference for care of male or working-aged ill household members; "i is an error term; and F ( ) is the probit function. Table 3.2 summarizes the hypothesized sign on each of the variables we analyze. Appendix 3.B describes a theoretical model of take-up behavior that informs our hypotheses. Supply and demand dictates that households facing the higher coupon and lower price will be more likely to buy SKY. Based on theory, we expect households with a member in self-reported poor health will be more likely to purchase SKY insurance, as will a household that had at least one member with a large health shock, measured as a health event that resulted in missing 7 or more days of normal household activities, resulted in an expense of over 100 USD, or resulted in a death. Of households that reported a shock, we 117 expect that households that used a public facility for care (health center or public hospital) will be more likely to purchase SKY, because SKY covers only public facilities. Similarly, we expect that households that used private facilities for care prior to the SKY village meeting will be less likely to purchase SKY. Households with stunted or wasted children should be more likely to buy SKY if stunting and wasting is an accurate proxy for poor health. Results of these adverse selection measures are presented in Chapter 2. We use household size as a control variable, but as SKY’s premium is based on the number of household members, and becomes slightly cheaper per person as household size increases, household size may be positively correlated with take-up. We predict that households with elderly or young members will be more likely to buy insurance if these groups have higher rates of illness, but will be less likely to take up if illnesses by these members are not frequently treated outside of the home. We predict that households that live far from a public facility, or who live near a public facility that is of poor quality, will be less likely to buy SKY, which only partners with public facilities. Households willing to take a …nancial risk may be less likely to buy SKY because they care less about ‡uctuations in income. However, because SKY is a new product, some households with low risk aversion for …nancial loss may be more likely to purchase SKY because they are less concerned that they will lose their money if SKY turns out to be a bad product. Households that take health risks may either be more or less likely to purchase SKY. On the one hand, these households may have higher expected health care costs, which would make them more likely to purchase SKY. On the other hand, these households may give their health needs less weight than other households, and may be less likely to seek preventative care, etc. If that is the case, these households may foresee lower health expenses than other households and thus be less likely to buy SKY. Some families may be able to pay for health care expenses without much sacri…ce, even in the absence of SKY. Households that, for example, can borrow from family, or can use savings, may feel less need for outside insurance, and thus we predict they will be less likely to purchase SKY. We expect that the coe¢ cients on wealth variables will be positive, because these households will be better able to a¤ord insurance. At the same time, if these households have a preference for private care, or are better able to self-insure in the absence of SKY insurance, they will be less likely to buy SKY. Insurance is a trade-o¤ of a small payment today to avoid a possible future loss. Households with a high discount rate may not be willing to sacri…ce consumption today for the possibility of increased consumption at a later date. Thus, we predict that households with a high discount rate will be less willing to buy SKY. Households that are better able to understand SKY will be more likely to buy, so we expect households that are better educated, are literate, and that were able to understand the survey question on risk aversion will be more likely to buy SKY. Whether or not their own future health care needs are actually high, a household may estimate higher expected medical costs and be more likely to purchase insurance if they know someone who has recently had a serious illness or injury with high costs. 118

Households that trust western medicine should be more likely to buy SKY, as SKY covers facilities that o¤er Western but not traditional medicine. One measure of trust we use is having all vaccines up to date for children under age 6. On the one hand, households with all vaccines ful…lled can be considered to trust western medicine, and may put a priority on health, and thus may be more likely to buy SKY. On the other hand, households with up to date vaccines may feel that they are less prone to a health shock and thus may be less likely to buy SKY. Households that use covers for their water jugs are also considered to trust western medicine (and its emphasis on preventing the spread of germs), and thus may be more likely to use western medicine and buy SKY. Finally, if Cambodian households favor health care for males or working-aged individuals, households with ill members with these characteristics are more likely to buy health insurance. At the same time, having a household member over the age of 64 or under the age of 6 may increase take-up, as stated above, if these members are more likely to be ill.

3.5 Data

Our analyses use several sources of data: a household survey; SKY administrative data; a village chief interview; a health center survey; and a village meeting interview. In this section we describe these data sources. The sample is all households for which we have baseline data. Appendix 3.C gives descriptions of all variables used in our analyses.

3.5.1 Household Survey The principal component of data collection is a large-scale survey of over 5000 households . Most of the data for the selection study comes from the baseline survey, but we also use some data from the second round of the household survey which was administered one year after the baseline. For the baseline survey, we intended to visit households shortly after the village meeting, within two to seven weeks. However, logistical concerns meant that we could only interview households in two phases over the 13 months of meetings. The …rst phase of the baseline survey took place in July and August of 2008 and the second phase took place in December of 2008. Thus, households were interviewed anywhere from two to nine months after the SKY meeting in their village. The second round survey was administered in two phases in July and August 2009 and December 2009 to January 2010. The baseline survey collected data on household demographics, wealth indicators, self-perceived and objective health, health care utilization and spending, assets and asset sales, savings, debt, health risk behaviors, willingness to take …nancial risks, trust of institutions, means of paying for large health expenses, and willingness to trade current for future income. For most questions, the baseline survey interviewed a primary respondent, and requested that this respondent answer questions for other members of the family. 119

3.5.2 SKY Administrative Data For each household that joins SKY, SKY records registration date, date the household starts coverage, and date the household drops out of SKY. We use this SKY administrative data to determine if and when each household from the village meeting purchased SKY insurance. To match our baseline data to the SKY database, for each village, we matched the name of household member in our survey to the names listed in the SKY database.

3.5.3 Village Leader Survey In each village we interviewed the village chief or another village leader in order to collect general village-level information, including the distance to local public health centers.

3.5.4 Health Center Survey Households may be more likely to purchase SKY if the quality of the local public health clinic with which SKY partners is of good quality. To measure this, we administered a simple survey of health clinics in areas covered by our study. The survey consists of checklists of operating hours, drug supply, cleanliness, and equipment supply. To minimize data collection costs, the health center survey focuses on observations by SKY member facilitators. SKY hires member facilitators to be present at health facilities to facilitate treatment for SKY members and manage client complaints and questions as needed. Member facilitators typically work mornings at one particular Health Center. The survey consists of checklists of operating hours, drug supply, cleanliness, and equipment supply.

3.5.5 Village Meeting Survey At the end of each village meeting, our …eld team spoke to a village leader to collect village-level data relevant to our study. Data from this source includes, for example, distance to the nearest public referral hospital.

3.6 Background Results

3.6.1 Summary Statistics Table 3.3 and Table 3.4 present summary statistics for each variable used in the analysis of take-up. Means are presented for all households, and separately for buyers and decliners. From the summary statistics, we can see that households that buy SKY are di¤erent than those that decline SKY on a number of characteristics. Purchasers of SKY insurance are less likely to report the ability to mortgage land to pay for a large health expense (0.4% v. 0.8%, p < 0.05) and are less likely to borrow without interest to pay for such an expense (16.7% v. 21.1%, p < 0.001). 120

Buyers live near a health center that was open for more hours during the week of the clinic survey (97.3 versus 94.3 hours, p < 0.10), and received slightly higher scores on the facility quality index (79.5% versus 78.2%, p < 0.01). Buyers are also more likely to have a family member with a costly health shock in the past year (Table 3.4), but this includes members in the household, so it is essentially a measure of health status, and thus adverse selection into insurance. In our multivari- ate analysis we control for in-household health shocks so that this variable can be argued to represent only out-of-household health shocks, and thus can be counted as a salience measure. Buyers are somewhat richer, and this di¤erence is statistically signi…cant for enumerator-ranked subjective wealth (16.1% versus 12.8% are in the wealthiest group, p < 0.01, and 11.2% versus 13.0% are in the lowest ranked wealth group, p < 0.05). SKY buyers are signi…cantly more likely to have members in poor health for almost all ages and genders. However, SKY households are no more or less likely to have an ill female over the age of 64. Buyer/decliner di¤erences in other variables are not statistically signi…cant.

3.6.2 Characteristics of Ill Members In our analyses of take-up of insurance, we include characteristics of household members, and also of ill members, to gauge whether there is any discrimination by gender or age in insurance purchase. If households purchase less insurance for older household members, but older household members are more likely to have a health shock and pay for care, then households may not be acting rational in the neo-classical sense. Similarly, if households are more likely to purchase health insurance for a working-aged female than a working-aged male, then these households are acting rationally only if ill working-aged females are more likely to receive insurance-covered care. To ease interpretation of subsequent results, in this section we examine the likelihood that an individual of a given age and gender will be ill and seek care, and the likelihood that each of these individuals will seek care following a health shock. We look at only households that did not receive a large discount on insurance, because most of these households did not purchase SKY, and thus insurance is less likely to have in‡uenced health-seeking outcomes. Table 3.5 regresses poor health and health utilization on characteristics of individuals. Columns 1 through 6 present utilization results on all individuals; Columns 7 through 10 present these results using only individuals with major health shocks. In short, results show that the elderly are the most likely to be ill, but the least likely to receive care following an illness compared to other age groups. Females of working age are more likely to be ill, and about as likely to receive care for an illness than their male counterparts. Columns 1 and 2 use indicator variables for poor health and a recent health shock as dependent variables, respectively. The oldest household members are the most likely to be in poor health, followed by working-aged household members and those under age 6 (di¤erences between the working-aged and those under 6, and the working aged and those over 64, are signi…cant at p < 0.001). Those aged 6 to 15 (the excluded category) are least likely to be ill. 121

Females are more likely than males to be in poor health if they are working age (col. 1, p <0.001), and are more likely to have a health shock (col. 2, p < 0.001).2 For those 65 or older, females are more likely than males to be ill (col. 1, p < 0.01) but are not more likely to report a major health shock. Males under 6 are more likely than females under 6 to be ill or report a health shock (col. 1, p < 0.01, col. 2, p = 0.054). Looking at all individuals, whether or not they experienced a health shock, individuals over the age of 64 are the most likely to seek all types of care for a major health shock (col. 3 –6), followed by the working-aged, those under 6, and the excluded category of individuals aged 6 to 15. However, when we look at only those who have had a health shock (col. 7 –10), so that frequency of health shock does not come into play, we …nd that those over 64 are often the least likely to seek care. Note that almost all individuals (96%) receive some kind of care for a health shock. For both females and males, the youngest are the most likely to receive care, followed by the working aged, and those over age 64. For public care, the di¤erence between the working aged and those over 64 is signi…cant at p < 0.05; between the working aged and under 5, p < 0.01. Di¤erences for hospital care are not signi…cant. Comparing care by gender, females over the age of 64 are less likely to receive public or hospital care following a health shock than males (col. 9 - 10). Di¤erence for public care is signi…cant at p = 0.13, but other di¤erences are not. However, if we increase sample size to include households that received the large discount as well, the di¤erence becomes signi…cant for both public and hospital care. Di¤erences in private or any care are not signi…cant even when including households that received the large discount. In the same tables, we also investigate whether wealth in‡uences health and utilization of health services. Individuals in households that are in the highest enumerator-ranked wealth category are more around 2.0 percentage points less likely to be in self-reported poor health (col. 1, signi…cant only if we include sample who received large discounts, not shown) and 2.1 percentage points less likely to have a major health shock (col. 2, p < 0.01). Over all households (with and without a health shock), they are signi…cantly less likely to use care (col. 3 –6), but are no more or less likely to receive care at public or private facilities or hospitals following a health shock (col. 7 –10). Individuals in the poorest households are 14.0 and 3.2 percentage points more likely to be in poor health or have a major health shock, respectively, than indiviudals in other households, holding all else constant (p < 0.001). They are also more likely to use public or public hospital care following a major health shock (col. 9 and 10). While these di¤erences are not signi…cant, increased use of hospital care for the poor becomes signi…cant when we include households with a large discount (p = 0.051, not shown). Over all households (col. 3 – 6) poorer households are around 1 to 3 percentage points more likely to use care than the excluded category (mid-rated wealth households).

3.6.3 Qualitative Survey Responses To begin to understand why households buy SKY, we administered a survey of SKY Insurance Agents and Member Facilitators at the start of the study to ask these SKY

2Signi…cance of di¤erences between coe¢ cients for each age/gender group is not presented in the table. 122 sta¤ members why they thought households bought SKY. Insurance agents stated that households join SKY if they have a lot of illness in the family, have had positive experiences with public facilities, understand the bene…ts of SKY and are better educated. They also thought that some households buy SKY because SKY agents stationed at public health centers and hospitals can help deal with problems that arise. Insurance agents believed households dropped or did not buy SKY because they did not understand the bene…ts, did not foresee using SKY’s services, have other …nancial commitments, and because the health centers had a poor selection of medicines and health center sta¤ is low quality and rude (Domrei Research and Consulting and University of California, Berkeley 2010a; Domrei Research and Consulting and University of California, Berkeley 2010b). In addition, we administered a small household-level qualitative survey at the time of the baseline survey that asked SKY and once-SKY households why they bought SKY immediately, waited to buy, or bought and then dropped SKY. Some respondents stated that they did not join SKY at …rst because they did not trust SKY, did not understand the product, or they could not a¤ord the premium. One household waited to buy SKY so that they could …rst observe SKY activities. Households that dropped SKY did so because hospital sta¤ was rude, because of low quality care, lack of drugs at public facilities, or because it was too di¢ cult to travel to the hospital when sick. Many households seemed to not understand the concept of insurance. One household dropped SKY because they were told it did not make sense to have insurance because people are never sick every month of the year. One respondent stated that he understood SKY, but later dropped because nobody was sick. The respondent later re-joined SKY after an explanation from an insurance agent. The following quotations from these qualitative surveys illustrate some of the mo- tivations behind the decision to purchase SKY.

“My family didn’t join SKY immediately because I didn’t have enough money to pay the premium.” “I got some advice from my cousin and neighbor that because my family has a lot of members and because we have children with diseases (one has cancer of the nose and one more has typhoid with stomach ache and heart disease) we should become SKY members because SKY insures many diseases, especially serious diseases.” “I dropped out of SKY because I had a problem with blood pressure and I was treated at [the nearest] Hospital. At the hospital the sta¤ and the nurses were not friendly and were careless, and the place was dirty. I stayed there for three days and got only three tablets of medicine. It is the same as in the Pol Pot regime.”

3.7 Regression Results

Table 3.7 through Table 3.9 present the results of the selection probit analysis. I present results in a way that facilitates interpretation. The regressors could have been 123 included in many di¤erent combinations, and thus the particular regressions presented are somewhat arbitrary. However, unless noted, the signi…cant results presented are robust to inclusion of dependent variables in many di¤erent combinations.

3.7.1 Traditional In‡uences on Take-up Table 3.7 focuses on traditional in‡uences on take-up: distance to health facilities, and measures of risk aversion and health risk. The in‡uence of expected health care costs on take-up (adverse selection) is explored in Chapter 2. These covariates are included in the regressions presented below but not shown. Being o¤ered the steeply discounted price increases purchase of SKY by around 38 percentage points (col. 1-3). This coe¢ cient underestimates the impact of the premium because we include over-sampled low coupon buyers in the regression. If we do not include oversampled households that purchased SKY without a discount, the e¤ect of the 80% price discount is closer to 41 percentage points increase in purchase. Price elasticity for purchasing SKY within the …rst 6 months after the SKY village meeting is 7.7 (Table 3.6), meaning that demand for insurance is rather elastic in rural Cambodia. Chapter 2 discusses …nancial implications to SKY of the change in premium. As with other measures of adverse selection, we discuss take-up by households with older or younger household members in Chapter 2, but mention these results again in light of the more in-depth exploration of health and health utilization by age described above. Despite the results in Table 3.5 showing that members over the age of 64 are most likely to seek health care (because of the higher rate of illness, and despite the lower rate of health care use following an illness), households with a member over age 64 are no more likely to purchase SKY (Table 3.7, col. 1). In Table 3.7 we also explore other characteristics that may lead households to predict higher utilization of health facilities. One reason a household would not expect to use SKY-covered facilities is if they live far from public facilities, or if it is costly to get to these facilities. Our results (col. 2) show that cost of taking a moto (a small motorcycle) to the local public health center has a negative impact on purchase of SKY: An increase in cost of 1USD (around 2 standard deviations from the mean of 0.39USD) decreases take-up by around 3.7 percentage points. However, including walking time or moto time to the public health center in place of cost of moto (not shown) does not induce any signi…cant change in take-up of SKY. Distance from the village to the nearest referral hospital does not in‡uence take-up of SKY. A household that thinks of facilities as low quality may be less likely to utilize these facilities and thus less likely to purchase SKY. Our results show that households living near a facility of higher quality, according to our clinic survey quality scale, are more likely to purchase SKY. The average health center scored 0.786 on our quality scale, which consists of 25 quality checks, with a standard deviation of 0.091. Regression results (col. 2) show that an increase of 1 in the score leads to a 35.5 percentage point increase in probability of SKY purchase. A more realistic increase of 0.08 in the quality score (for example, going from a score of 19 out of 25 –the average health center score - to 21 out of 25), would lead to a 2.84 percentage point increase in take-up of SKY insurance (equal to 0.355 multiplied by 0.08). 124

A household may anticipate low use of public facilities if these facilities have few hours of operation. We …nd no signi…cant impact of facility hours on take-up of SKY (column 2). This coe¢ cient increases to 0.027 and gains some signi…cance (to P = 0.30) if we remove the quality score from the regression. A household that is risk averse may be more likely to buy SKY. We present results of two measures of risk aversion in Table 3.7. The …rst is a hypothetical question on a choice between several lottery gambles. We control for households that were either confused by the question or were hyper risk-avoiders, preferring a guaranteed gift of 500USD over a gamble of 500 or 1000USD. There is no signi…cant e¤ect of this hypothetical question on purchase of insurance (col. 3). Our second question on risk aversion, which asks about actual gambling behaviors, yields results that are contrary to traditional theory: households in which the respondent or spouse “plays games of chance for money” (gambles) are 5.75 percentage points (p < 0.05, col. 3) more likely to buy SKY insurance than decliners. We also look at health risk behavior, which we consider separate from …nancial risk aversion. A household that takes risks with their health may be considered less risk averse, but may also be more willing to buy SKY to cover potential health care costs. However, there is no evidence that a household that would choose a riskier job (in terms of health) is more likely to buy SKY. There is also no signi…cant evidence that a household that exhibits behaviors that are risky to health, measured by a respondent or spouse having had an accidental injury, is more likely to buy SKY. (Similarly, households that never cover water jugs are no more likely to buy SKY. We include this as a measure of trust of western practices, but it can also be considered a measure of health risk behavior.) Table 3.8 examines how self-insurance in‡uences take-up of insurance. We expect households that have easy ways to pay for health shocks in the absence of SKY will be less likely to purchase SKY. We have already shown that wealthier households are more likely to purchase SKY, which does not support this hypothesis. Holding wealth and other variables constant, we examine whether the ways in which a household could pay for a hypothetical health shock increase or decrease the likelihood of SKY purchase. It is a bit di¢ cult to interpret individual responses to this question because even a di¢ cult way to pay for care is better than no way at all, and because households were asked to list any ways they could pay for care, and can list more than one way to pay for care. In addition, as the survey took place after households had had the opportunity to buy SKY, "SKY would pay" is included as an option, and although households were prompted to list alternatives to SKY paying for care, a few households chose only this option. Thus we run three separate regressions to check robustness of results, and include "SKY would pay" as a regressor in all regressions. Coinciding with theoretical predictions, households that could pay for a large health expense with a no-interest loan (presumably from family or a friend), or have a family, friend, or association (e.g., a rotating savings group) that could help pay for health care have a lower probability of SKY purchase, although these results are not signi…cant (Table 3.8, col. 1). Similarly, the ability to cover expenses with cash, doctor credit, or health equity funds reduces the likelihood of insurance purchase (results not signi…cant). Households that could pay for health care expenses with savings or a loan with interest have a higher likelihood of insurance purchase, although not signi…cantly. 125

We also run the regression looking at households that list only what we consider expensive ways to pay for care (col. 2). A household that would only be able to pay for care with the sale of assets or borrowing with interest are more likely to purchase SKY although not signi…cantly. A household that would have to seek extra work would be less likely to purchase SKY (not signi…cantly). Households that list no inexpensive ways to pay for care (could not pay with cash or savings, could not get help from family, friends, or a savings association, could not borrow without interest, and are not Health Equity Fund members) are 2.5 percentage points more likely to purchase SKY (col. 3, P = 13.8). The hypothetical question asks for any ways to pay for care, so many households chose more than one option. Column 3 looks at the number of options a household lists to pay for care, holding constant whether the household stated that SKY would pay for care. Households that list more options are less likely to buy SKY, but not signi…cantly.

3.7.2 Other In‡uences on Take-up Table 3.9 includes other in‡uences on insurance purchase that are less in line with traditional theory. All previous regressors, except for self-insurance measures, are included as controls but are not shown. The exception is the …nal column, which does not control for health utilization measures for ease of interpretation of coe¢ cients. Households that are in the wealthiest subjective wealth category are 6.7pp more likely to buy SKY (col. 1, P < 0.01). The poorest are not signi…cantly more or less likely to buy SKY. Discount rate has no signi…cant impact on purchase of health insurance (col. 1). Measures of understanding of insurance did not in‡uence SKY purchase as we predicted it would (col. 2). In relation to the excluded category of 5 or more years of education (4.7 years is average) having a respondent with 1 to 4 years of education increased take-up of SKY by 4.2 percentage points (p < 0.05), as did having 0 years of education (not signi…cant). We included measures of education in several ways (not shown) and years of education always had a negative impact on purchase of SKY. Illiterate households were less likely to purchase SKY, but this result is not statistically signi…cant. Households in which respondents did not understand the hypothetical risk aversion question (choosing the option with a certainly lower payout) were similarly less likely to purchase SKY, but not signi…cantly so (Table 3.7, col. 3). Tests of joint signi…cance of these three variables also yielded no signi…cant results. Knowing someone who received a health shock increases purchase of SKY, even if that person is not in your household. Having a neighbor with a large health expense increases SKY purchase by 5.0 percentage points (col. 3, p < 0.05), even though this neighbor would not be covered by a household’spurchase of SKY. The point estimate of this variable is somewhat sensitive to the inclusion of other regressors, and sometimes becomes only marginally signi…cant. Knowing a family member who spent more than 100USD on a health shock in the last year increases SKY purchase by 4.2 percentage points (P < 0.05). We include controls for spending more than 100USD on a member living in the household (an in-family individual with a 100USD health expense), so the assumption is that these results hold for even a family member living outside of the household, who would not be covered by SKY purchase by the household. 126

There is also the possibility that households are buying SKY not because they know someone who is ill, but because they know someone else who bought SKY, which is more likely if they know someone that is ill. Future analyses will examine this possibility. Families with children under age 6 that have all WHO-recommended vaccines …lled are 7.1 percentage points (P < 0.01, col. 4) more likely to buy SKY. We interpret vaccinations to be a signal of trust of western medicine, which will increase expected utilization of public health centers as compared to traditional healers or drug sellers. At the same time, covering water jugs, which can be interpreted as willingness to take a health risk or as an adherence to western medicine, has no signi…cant impact on insurance purchase (col. 4). The coe¢ cient of this measure does not change when we eliminate potentially collinear regressors such as other risk measures (results not shown). Above we describe the likelihood that members of di¤erent ages and genders will be in poor health, and likelihood of these members using health facilities. We found that the elderly are the most likely to be ill (Table 3.5, col. 1 and 2), but the least likely to receive care following an illness (col. 7 –10). Females of working age are more likely than males to be ill (col. 1 and 2), and about as likely to receive care following an illness(col. 7 - 10). Based on health levels, adverse selection theory would predict that household members with an elderly member should be most likely to purchase SKY. However, we found above (Table 3.7) that households are not more likely to purchase SKY when they have an older member. Now we test whether households are any more or less likely to purchase SKY depending on the characteristics of the ill members of the household. If past health care utilization predicts future use, based on observed patterns of health care utilization, households would be more likely to purchase for ill males over age 64 than ill females over age 64 (because males at this age use more public care following a shock), and less likely to purchase for the ill elderly than for the ill in other age groups, whether male or female. Indeed, we …nd that households with an older ill member are signi…cantly less likely to purchase SKY than households with a working-aged member in poor health (Table 3.9, col. 5, p < 0.01, signi…cance for this di¤erence not shown) and less likely to purchase SKY than households with a young member in poor health (p = 0.17, signi…cance not shown in table). Households with older ill males are more likely to be in households with insurance than are households with older ill females, but this di¤erence is not statistically signi…cant. Households are more likely to purchase for an ill male under 6 than an ill female under 6 (by 3.9 percentage points, di¤erence not statistically signi…cant). Households with an ill female member of working age (16 to 64) are around 3.7 percentage points more likely to purchase SKY than households with an ill working-aged male (p = 0.15, signi…cance not shown in table).

3.8 Robustness Tests

The regressions above were run many times using di¤erent combinations of independent variables. Signi…cant results presented in the tables above do not change mean- ingfully depending on which other independent variables are included in the regressions. In 127 addition, the following robustness checks were run. Appendix Table 3.11 presents results of testing changes in sample, and results of adding village-level indicator variables. The …rst column of this table presents results from the full sample (as presented in our main results) for comparison purposes. Self-insurance measures are not included in regressions for this table, but separate regressions show that signi…cant results for self-insurance measures do not change with change in sample (not shown). For all other regressors, only signi…cant co- e¢ cients are shown in the table, although all variables from the above tables were included in the regressions. Appendix Table 3.12 interacts wealth and health. Keep in mind that column 1 of Appendix Table 3.11 includes all regressors at once, whereas the main results added a few regressors at a time. Thus, the coe¢ cients for some variables change meaning somewhat. For example, because we are including a covariate for “All vaccines ful…lled for members under 6”, the coe¢ cient on “At least one member age 5 or under”is now interpreted as having a member 5 or under that did not receive a vaccine. The coe¢ cients on other variables also may have also changed slightly, but overall results and signi…cance remains the same.

3.8.1 Interview Lag and Delayed SKY Purchase Due to delays in survey implementation, some households were not interviewed for up to 274 days after the village meeting. The data on pre-meeting health shocks may be less accurate for these households due to poor recall. This also means that some questions, in particular, self-reported poor health, are reported several months after the start of SKY. This should not lead to problematic bias, because if anything, SKY members should have increased health over time, leading any late responses to bias downwards the illness of the insured. As a robustness check, we include only households that were interviewed within 3 months (93 days) of the SKY village meeting (Appendix Table 3.11, col. 2). We also look at only households that purchased SKY within 2 months (63 days) of the village meeting (Appendix Table 3.11, col. 3). For these households, health at baseline may be more likely to in‡uence take-up of SKY. Households who bought after this date are left out of the regression. In a separate regression we restrict the sample to include only early buyers and early surveys (and decliners) (Appendix Table 3.11, col. 4). With these added …lters we have less than a third of the original sample, so signi…cance drops for many variables, although most point estimates do not change sign for signi…cant coe¢ cients. Coe¢ cients that do change sign are for variables that are marginally statistically signi…cant at best. Quality of health facilities is a positive in‡uence on take-up for early interviews (col. 2) and early buyers (col. 3), but becomes negative when we look at early buyers and early interviews (col. 4). However, as this is a village-level measure, this may be due to the small number of villages in these regressions: signi…cance for column 4, for example, drops to p = 0.52.

3.8.2 Village Controls If we include a variable for each village in the sample (Table A 3.11, column 5), there are some small changes in the results, but point estimates do not change sign. For 128 example, the coe¢ cients on salience variables (knowing family or a neighbor ill) are reduced or are no longer signi…cant. This makes sense: Households may buy SKY if they have a member that is ill, or if they know someone that is ill. In villages with many ill members, there will be high take-up of SKY. In these same villages, households are more likely to know someone that is ill. Thus, living in certain villages is collinear with knowing someone that is ill. A regression of the percentage of households with poor health in a village on the percentage of households who know someone that is ill shows that average number of households with poor health is positively correlated with the average number of households knowing someone who is ill.

3.8.3 Wealth Interactions In the main results section we found that wealthier households were more likely to buy SKY. It is possible that wealthy households may buy SKY even when a member is not currently ill (i.e., for pure insurance/consumption smoothing reasons in the case of a health shock), but that poorer households buy SKY only when they have a very ill household member because of budget constraints. Alternatively, the poorest households, even with the sickest members, may forgo all care, and not purchase SKY at all. To test this, we interact wealth with subjective health variables. Table 3.10 gives summary statistics by wealth and health. Holding nothing else constant, for all wealth levels, households with a member in poor health are more likely to buy SKY, but wealthier households are even more likely to purchase SKY when they have a member in poor health. The wealthier are also more likely to purchase SKY if no member is in poor health. Looking at this from a di¤erent perspective, SKY households of all wealth levels are more likely to have a member in poor health than households without SKY at that wealth level, but there are fewer members in poor health in wealthier households for both buyers and non-buyers. When we interact wealth and poor health in regression form (Appendix Table 3.12), we …nd that while the rich are more likely to buy as a whole (the coe¢ cients on “Highest ranked wealth” and “Poor health X Highest ranked wealth” are jointly positive and statistically signi…cant) the rich that are healthy are not more likely to purchase SKY (the coe¢ cient on “Highest ranked wealth” is not statistically signi…cant). We interpret this as meaning that the rich are not signi…cantly more likely to buy for pure insurance reasons (“just in case”), but the poor are more likely to abstain from buying SKY because of budget constraints.

3.8.4 Coupon Status A companion paper (Chapter 2) compares selection results for high coupon versus low coupon households. Comparisons on baseline characteristics are limited due to the low number of households that purchased SKY with a low coupon (the higher price). Results for health status show that low and high coupon households are equally likely to buy SKY for members in poor health as measured by the baseline survey. 129

3.9 Conclusion

We collected baseline data from over 5000 households in rural Cambodia as the voluntary health insurance program SKY was introduced to the region. We use a simple logistical model of take-up on baseline characteristics to examine how baseline characteristics of households in‡uence the decision to purchase insurance. Households had not previously been exposed to insurance, so baseline estimates of health and health utilization were not in‡uenced by previous insurance status. We …nd evidence of both traditional and less traditional incentives to purchase SKY. In results not shown, we …nd signi…cant evidence of adverse selection: households that have a member in poor self-reported health, or have a member that has used a public health facility for care of a major health shock in the months preceding the SKY village meeting, are more likely to purchase SKY (results presented in Chapter 2). Interestingly, we …nd that simply knowing someone in poor health, even if they would not be covered by a household’s purchase of SKY (e.g., a neighbor), induces an increase in purchase of SKY. We interpret this as a salience e¤ect: knowing someone ill and witnessing high health expenses increases the perceived likelihood of illness in the mind of the potential insurance purchaser. However, we cannot rule out the possibility that knowing someone ill is correlated with knowing someone who has purchase SKY (since increased illness increases SKY purchase), and thus that it is not knowing someone ill but knowing someone who has purchased SKY that is what is increasing the purchase of insurance. It is also possible that, especially if an illness is contagious or due to a common external factor, knowing someone ill is in fact correlated with a household’s own probability of getting ill, and that purchase is not due to salience but instead to the real increased risk of illness. Contrary to adverse selection theory, although the elderly are more likely to utilize health care (because of a higher rate of illness), households with an elderly member are not more likely to buy insurance. Households that take risks with their health are also no more likely to purchase SKY. Also contrary to traditional theory, like some recent studies in developing countries (Giné, Townsend, and Vickery 2007), we …nd no evidence that households that are more risk averse are more likely to buy SKY. In fact, the limited evidence we have indicates that it is the less risk averse households are those that are purchasing SKY insurance. We interpret this as a “…rst mover” e¤ect, whereby households that are willing to take risks with their money are also more willing to spend money on an untested new product. In a related explanation, Bryan (2010) proposes that ambiguity aversion, whereby households fear they will not get paid when they most need it, decreases the demand for some types of insurance, and that without holding ambiguity aversion constant, it will appear that risk averse households are less likely to purchase insurance. Our result means that, unlike some recent evidence that …nds that risk aversion can counteract adverse selection (positive selection), we do not …nd that this is the case for SKY. This is an important result: that the risk averse purchase more insurance is a long-accepted theory in the insurance literature. Recent studies theorize that purchase due to risk aversion may even o¤set some adverse selection. Our results adds one more piece of evidence that this theoretical hypothesis is not always true empirically. Households were less likely to buy the greater the cost to arrive at the local public 130 health facility. They were more likely to buy if the nearest SKY-partnered health facility was measured to be higher quality, although this result were not robust to some changes in sample. This result is in accordance with traditional models of adverse selection, as households near higher quality facilities may be those that are most likely to utilize these SKY-covered services. It is interesting to note that unlike some previous studies, that found little relationship between measured quality and utilization, our quality measures are all attributes of the health centers that are relatively observable to households, such as availability of equipment and cleanliness of the facility. These attributes, along with other factors such as politeness of sta¤ and waiting time, may be more important to households than sta¤ training and diagnostic skills. We found some evidence that households with limited ways to cheaply self insure are more likely to purchase insurance, but results were not statistically signi…cant. However, the wording of our question makes it di¢ cult to interpret responses and thus this result must be treated with caution. Individuals in wealthier households are less likely to be in poor health and have lower utilization than poorer households, but these households are more likely to buy SKY than poorer households. This can be interpreted as evidence of positive selection, if we consider that wealth is negatively correlated with probability of health shock but positively correlated with purchase of insurance (Fang, Keane, and Silverman 2008). Poorer households with an ill member are less likely than richer households with ill members to purchase SKY, presumably due to budget constraints. Education, cognitive ability, and discount rate, which had been shown in other studies to o¤set adverse selection (Fang, Keane, and Silverman 2008), do not have any statistically signi…cant e¤ects on take-up of SKY insurance. It is surprising that the ability to understand SKY, which we measure by education of the health decision maker, ability to understand a risk aversion question, and a literacy and numeracy test, does not have a positive impact on take-up of SKY. It was clear from questions at the SKY village meeting and from our in-depth qualitative interviews that some households did not understand the concept of insurance. In addition, 82% of our sample did not seem to understand the hypothetical risk aversion question, and answered that they would prefer a guarantee of $500 over a 50/50 gamble of $500 or $1000. Thus, although our results show do not show a positive impact of education or understanding on SKY take-up, we believe this may in fact be an important reason that households remain uninsured. It is possible that our measures of cognitive ability are not accurately capturing a household’sunderstanding of SKY insurance, or that greater understanding of insurance is correlated with some unmeasured factor that is negatively in‡uencing SKY purchase. As in other studies of developing countries, we …nd some evidence of age and gender inequality in SKY purchase. Households are less likely to purchase SKY for ill household members aged 65 or older than for younger ill members. This is rational when we consider that while older individuals are more likely to be ill, they are less likely to receive treatment for their illnesses. However, irrationally, although overall health care utilization is highest amongst older members (due to high rates of illness), households are not more likely to purchase SKY when an elderly member is in the household, as mentioned above. In sum, while we …nd some support for traditional models of insurance take-up 131

(adverse selection), our evidence also gives support for less traditional in‡uences on insurance purchase (budget constraints, salience of illness, age preference), and provides evidence that counters the long-accepted theory that the risk averse will be more likely to purchase insurance. Our results cover a single insurer in a few regions of one nation. It is important to see how results would vary with di¤erent products, di¤erent health care systems, and so forth. In addition, we relied largely on survey measures of risk aversion and other behavioral factors. It would be useful to measure these factors more objectively with behavioral measures or experimental games. It is important to understand how baseline characteristics of households, which we interpret as expected utilization of health care, translates into di¤erences in utilization once SKY is purchased. In a related paper (Chapter 2), we …nd evidence of adverse selection in utilization above and beyond self-selection based on factors consumers can observe but which we do not measure in our survey. Our results suggest that insurers in developing countries must contend with the same adverse selection issues as those in developed countries if they are to become …nancially sustainable without donor support. In addition, they must contend with barriers to take- up that are less traditional, and may be unique to a developing country context. Insurers must take these characteristics into account when determining how to market their product to consumers. Finally, if insurance is to be used as a policy tool, policy makers must understand how to cover the targeted population. 132

3.10 Tables

2000 2005 First Treat 2nd Treat 3rd Treat First Treat 2nd Treat 3rd Treat % of People Seeking Treatment (by provider type) (a) Public 18.8% 4.0% 1.2% 22.0% 6.0% 2.1% (b) Private medical practitioners 33.0% 6.7% 2.1% 46.7% 13.1% 5.0% (c) Private non•medical providers 34.0% 10.0% 3.4% 21.6% 7.7% 3.2% Total % Seeking Treatment 85.8% 20.7% 6.7% 90.3% 26.8% 10.3% Source: Cambodia DHS 2005

Table 3.1: Treatment Behavior of Ill Households 133

Measure Variable Description Predicted Sign Traditional Premium Household received a large coupon for SKY + Expected Health Costs Subjective health Household has at least one member in poor health, as reported by + respondent Recent health shock Household member has had a major health shock (death, 7 days unable + to work, or cost of over 100USD) in the three months prior to the SKY village meeting. Past use of public care Household member has used public care for a major health shock in the + three months prior to the SKY village meeting Past use of private care Household member has used private care for a major health shock in the • three months prior to the SKY village meeting Stunting and Wasting Household has a stunted or wasted child under age 6 + Demographic Number of household members + Characteristics Indicator variable for under 6 +/• Indicator variable for over 64 +/• Distance to public care Cost of moto to local health center • Kilometers to nearest public referral hospital (square root) • Facility Quality Local health center weekly number of hours open (divided by 100) + Average Clinic Survey score for availability of drugs, equipment, and + cleanliness Risk aversion Financial: Risk aversion ranking of 1•4 based on hypothetical choice of +/• certain versus riskier monetary pay•offs Financial: Respondent or spouse likes to gamble +/• In health: Respondent or spouse has ever received care for an accidental +/• Risks with health injury In health: Respondent chooses the highest pay but riskiest job over safer +/• options in a hypothetical question. Self insurance Various ways to pay for a hypothetical 100USD health care bill varies Could only pay by selling an asset + Could only pay with extra work + Could only pay by borrowing with interest + Household could pay for hypothetical 100USD health care bill with • cash, savings, family help, or borrowing w/ or w/o interest Number of ways to pay for care of a hypothetical 100USD health care • bill Other Budget Constraints Poorer household, by subjective wealth ranking +/• Wealthier household, by subjective wealth ranking +/• Discount rate Household has the highest discount rate as measured by a hypothetical • question asking the respondent to choose between a smaller payoff in two weeks time or a larger payoff in one year. Understanding of Insurance Education of health•decision maker: 0 years • 1•4 years • Omitted: > 4 years N/A Answered all literacy/numeracy test questions correctly + Confused by risk question: Chose guaranteed $500 over a 50/50 chance • of $500 or $1000 in hypothetical risk question Salience Respondent has a neighbor with a large health shock in the past 12 + months Respondent has (out of household) family with a large health shock in + the past 12 months (we include a control for in•household shocks in the past 12 months) Trust of Western Medicine All children under age 6 are up to date on vaccinations + Respondent always covers water jugs +/• Gender and Age preference Male household member in poor health, as reported by respondent + Working•aged, younger, or older household member (age 15•65, under +/ • / •, 6, over 64) in poor health, as reported by respondent respectively

Table 3.2: Summary of Hypotheses 134

Pooled Standard Buyer Decliner Clustered Mean Deviation Mean Mean ttest Traditional Measures Premium Offered a Deep Discount 0.481 0.500 0.761 0.340 •20.192 *** Risk Aversion/Health Risks Respondent or spouse ever needed care for accidental injury 0.134 1.870 0.132 0.135 0.045 Would accept 25% salary increase for riskier job 0.068 0.252 0.064 0.070 0.823 Hypoth. Local Risk Aversion: 1•4, least to most risk averse. Confused also = 4. 1.213 0.710 1.222 1.208 •0.744

Plays games of chance for money (gambles) 0.101 0.301 0.113 0.095 •1.704 + Self Insurance Could sell asset to pay for a large health expense 0.464 0.499 0.455 0.469 0.523 Family, friend, or association would pay for a large health expense 0.159 0.366 0.153 0.163 0.942 Could mortgage land to pay for a large health expense 0.007 0.083 0.004 0.008 2.291 * Health Equity Fund would pay for a large health expense 0.001 0.039 0.002 0.001 •0.237 Could use cash to pay for a large health expense 0.300 0.458 0.300 0.299 0.100 Would pay for a large health expense with savings 0.091 0.287 0.099 0.086 •1.296 Could borrow with no interest to pay for a large health expense 0.197 0.398 0.167 0.211 3.948 *** Could borrow with interest to pay for a large health expense 0.413 0.492 0.397 0.421 1.612 Doctor would give credit for a large health expense 0.001 0.033 0.001 0.001 1.088 Would get extra work to pay for a large health expense 0.046 0.210 0.039 0.050 1.593

SKY would pay for a large health expense 0.059 0.236 0.173 0.003 •49.734 *** Health Facility Quality Health Center total open hours / 100, actual (survey week) 0.953 0.500 0.973 0.943 •1.897 + Average score for inventory, hygiene, and equipment (positive/total outcomes) 0.786 0.091 0.795 0.782 •3.245 ** Health Facility Distance Cost of moto from village to health center (USD) 0.392 0.505 0.374 0.401 1.738 + Sqrt distance from village to referral h o sp i ta l (km ) 3.122 1.467 3.122 3.122 0.150 Observ ations 5229 1754 3475 Notes: Ttests clustered at the village level. + p<0.10, * p<0.05, ** p<0.01, *** p<0.001. Health facility quality Notes: measures are from the Clinic Survey. Distance to health centers is from interviews with village leaders. Distance to Ttestregional clustered hospital at the village is from level. village meeting data. Hypothetical risk aversion and accidental injury are from the second round survey. All other data is from the baseline survey. Sample is all SKY decliners and all SKY buyers who first Buyerspurchased are all households SKY after that the first Villagebought SKY Meeting. following the village meeting.

Decliners are all households that never bought SKY.

Table 3.3: Summary Statistics (Traditional In‡uences on Take-up) 135

Pooled Standard Buyer Decliner Clustered Mean Deviation Mean Mean ttest Other Measures Salience Family member had a > 100 USD health shock in the last year 0.194 0.396 0.211 0.186 •2.343 * Knows a neighbor with a > 100 USD health shock in the last year 0.138 0.345 0.148 0.133 •1.277 Trust of Western Medicine All vaccines fulfilled for members under 6, 0 if no under 6, pre•Meeting 0.259 0.438 0.273 0.251 •1.890 + Household never uses covers for water jugs 0.276 0.447 0.282 0.274 •0.580 Understanding of Insurance Confused by risk aversion question: Chose $500 over a 50/50 chance of $500/$1000 0.823 0.382 0.816 0.826 0.807 Education of health decision•maker (years) 4.651 3.419 4.674 4.640 •0.535 Answered all literacy/numeracy questions correctly 0.151 0.358 0.156 0.148 •0.754 Budget Constraints Highest ranked wealth by enumerator 0.139 0.346 0.161 0.128 •3.047 ** Lowest ranked wealth by enumerator 0.124 0.329 0.112 0.130 1.983 * Discount Rate Prefers 20USD now over 60 or 120USD in 12 months 0.654 0.476 0.655 0.654 0.421 Gender and Age Preference Male household member, under 6, in poor self• reported health 0.074 0.261 0.084 0.068 •2.064 * Female household member, under 6, in poor self•reported health 0.061 0.240 0.072 0.056 •2.117 * Male household member, age 6 to 15, in poor self•reported health 0.091 0.288 0.100 0.086 •1.509 Female household member, age 6 to 15, in poor self•reported health 0.085 0.279 0.102 0.077 •3.081 ** Male household member, age 16 to 64, in poor self•reported health 0.303 0.460 0.352 0.279 •5.773 *** Female household member, age 16 to 64, in poor self•reported health 0.480 0.500 0.557 0.442 •7.309 *** Male household member, over 64, in poor self• reported health 0.083 0.276 0.097 0.076 •2.372 * Female household member, over 64, in poor self•reported health 0.138 0.345 0.139 0.138 •0.061 Observations 5229 1754 3475 Notes: Ttests clustered at the village level. + p<0.10, * p<0.05, ** p<0.01, *** p<0.001. Use of water jug covers, Ttesthypothetical clustered atrisk the aversion village level. question, and literacy test are from the second round survey. All other data is from the Buyersbaseline are survey.all households Sample that first is allbought SKY SKY decliners following and all SKY buyers who first purchased SKY after the Village Meeting. the village meeting.

Table 3.4: Summary Statistics (Other Measures) 136 (10) 1359 0.113 0.007 0.120* 0.103* 0.148* 0.0475 0.0806 0.0232 0.0258 Hospital (9) 1359 0.014 0.170* Public 0.0697 0.0394 0.0269 •0.0383 [0.0732] [0.0825] 0.248*** (8) 1310 0.009 For Each Incident 0.0162 0.0392 0.0279 Public/ Private 0.00085 •0.0732 [0.0282] [0.0677] [0.0739] [0.0233] [0.0456] [0.0402] (7) 1081 0.025 Any Care Any (6) 0.026 13607 Hospital (5) 0.023 13607 Public 0.0267* 0.00914 (4) 0.028 13607 0.0122 Public/ Private (3) All individuals 0.031 13607 Any Care Any (2) Pre• 0.031 13607 Shock [0.0304] [0.0303] [0.0296] [0.0255][0.0251] [0.0237] [0.0251] [0.0283] [0.0247] [0.0465] [0.0188][0.0147] [0.0718] [0.0180] [0.0148] [0.0720] [0.0123] [0.0145] [0.0334] [0.0127][0.0142] [0.0609] [0.0114] [0.0143] [0.0602] [0.0142] [0.0124] [0.00926] 0.220*** 0.212*** 0.194*** 0.114***0.218*** 0.0934*** 0.217*** •0.0214 0.201*** •0.0355 0.0791*** 0.0718*** 0.0116 Meeting 0.0390** 0.0415** 0.0405** 0.0400** 0.0296** [0.00932] [0.00935] [0.00904] [0.00709] [0.00617] [0.0107][0.00949] [0.00949] [0.0246] [0.00920] [0.00726] [0.0481] [0.00621] [0.0481] [0.0120] [0.0240] [0.0453] [0.0435] •0.0207** •0.0180*[0.00755] [0.00757] •0.0169* [0.00737] •0.0128** [0.00466] •0.00365 [0.00364] [0.00900] [0.00901] [0.00865] [0.00672] [0.00487] [0.0107] [0.0246] [0.0411] [0.0352] (1) Poor 0.085 13602 health •0.0199 [0.0128] [0.0122] [0.0212] [0.0152] [0.0220] [0.0228] [0.0157] [0.0186] 0.148*** 0.0421*** 0.0429*** 0.0403*** 0.0278***0.274*** 0.0249*** 0.00822 0.0794*** 0.0805*** 0.0728*** 0.0478***0.548*** 0.0345*** 0.0184 •0.00036 0.0873+ 0.609*** 0.184*** 0.140*** 0.0324*** 0.0325*** 0.0261** 0.0180** 0.0118* 0.00537 •0.0313 0.0944*** 0.00408 0.00637 Table 3.5: Health and Utilization Regressed on Characteristics of Members Male age 16 to 64 (d) Male Female age 16 to 64 (d) 64 (d) over Male Female over 64 (d) under 6 (d) Male Female under 6 (d) subjective in wealthiest Household ranking (d) in poorest subjectiveHousehold ranking (d) Observations Pseudo R•squared Col. (1): Indicator for poor self•reported health; Col. (2): Indicator LHS variables: for a major observations. individual•level Notes: Regression uses effects;Marginal Standard errors in brackets health shock pre•meeting; a Cols. (3 • following 6): Indicators any care, public or private hospital respectively, for receiving Robust standard errors clustered at level. the village forhealth shock pre•meeting a major no major (includes zeros members health shock pre meeting, with with health shock); Col. (7•10): For individuals Data is from survey. the baseline for Major health shock is definedindicators use ofcare, public or as a shock causing hospital care, respectively. any care, public or private did not receive a large discount for who insurance. all individuals Sample includes of or more. + 100USD expense p<0.10, * p<0.05, ** or that effects; results in a health p<0.01, *** p<0.001. of Marginal death, 7 days disability, expense). a major(death, 100USD with health incident 7 days of disability,or Where households only indicated, sample includes that did Standard errors in brackets. Sample is all households All data is from Robust standard errors the baseline survey. clustered at the village level. (d) for discrete change of from 0 to dummy variable 1 not receive a large discount for insurance. (d) for discrete change of from indicator variable 0 to 1. + p<0.10, * p<0.05, ** p<0.01, *** p<0.001 137

Regular Price Large Discount Price, in Months, for 6 months insurance 5 1 Purchase within 6 months of meeting 172 1233 Number of households receiving price offer 2536 2539 % SKY 6.8% 48.6% Price Elasticity of Demand •7.7 SampleNotes: includes Sample only includes randomized randomized households, sample, not over•sampled not over•sampled buyers. buyers. Take•up is the number of households purchasing within 6 months of the Village Meeting, 41.8% even if a household drops within this period. Price elasticity of demand equals (%Change in Take•Up)/(%Change in Price).

Table 3.6: Price Elasticity of Demand 138

(1) (2) (3) Characteristics of Households: Demographic Clinic Risk

Offered a Deep Discount (d) 0.382*** 0.383*** 0.385*** [0.0191] [0.0189] [0.0189]

At least one member over 64 (d) •0.0193 •0.02 •0.0164 [0.0176] [0.0174] [0.0177]

At least one member age 5 or under (d) •0.000861 •0.00107 •0.00351 [0.0179] [0.0179] [0.0180]

Cost of moto from village to health •0.0368* •0.0361+ center (USD) [0.0185] [0.0189]

Average score for inventory, hygiene, 0.355* 0.369* and equipment (positive/total outcomes) [0.147] [0.151]

Health Center total open hours / 100, 0.00233 •0.00583 actual (survey week) [0.0295] [0.0303]

Sqrt distance from village to referral 0.00183 0.00514 hospital (km) [0.0105] [0.0111]

Hypoth. Local Risk Aversion: 1•4, least 0.00912 to most risk averse. Confused also = 4. [0.0139]

Confused by risk aversion question: •0.015 Chose $500 over a 50/50 chance of $500/$1000 (d) [0.0249]

Plays games of chance for money 0.0575* (gambles) (d) [0.0263]

Would accept 25% salary increase for •0.00667 riskier job (d) [0.0276]

Respondent or spouse ever needed care •0.0166 for accidental injury [0.0250]

Observations 4898 4871 4740 Pseudo R•squared 0.144 0.147 0.149 Notes: LHS variable: 1 if bought SKY, 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** p<0.01, *** Marginalp<0.001. effects; Marginal Standard effects; errorsStandard in brackets errors in brackets. Robust standard errors clustered at the village level. Clinic LHShours Variable: and hygiene 1 if bought and inventory SKY, 0 andif declined equipment (SKY score Admin are data) from the Clinic Survey. Distance to health facilities is from interviews with village leaders. Hypothetical risk aversion and accidental injury are from the second round (*) Major health shock means shock that caused 7 days of missed work, a 100USD expense, or a death. Utilization is following a major shock. survey. All other data is from the baseline survey. Subjective health and health care utilization measures are includedRobust standard in the regression errors clustered but not atpresented. the village Sample level. is all SKY decliners and all SKY buyers who first purchased SKYPre•meeting after the indicates Village Meeting. prior to the (d) SKY for discrete village meeting.change of indicator variable from 0 to 1. Clinic actual hours and hygiene and inventory and equipment score are from the Clinic Survey. Distance to health facilities is from interviews with village leaders.

Table 3.7: In‡uence of Traditional Selection Measures (Household Demographics, Clinic Characteristics, and Risk Characteristics) on SKY Purchase 139

(1) (2) (3) (4) Ways to Self• Any difficult Number of Characteristics of Households: Insure Difficult ways way Options

Could sell asset to pay for a large 0.00594 health expense (d) [0.0208] Family, friend, or association would pay •0.0125 for a large health expense (d) [0.0242] Could mortgage land to pay for a large •0.0957 health expense (d) [0.0857] Health Equity Fund would pay for a large •0.111 health expense (d) [0.185] Could use cash to pay for a large health •0.0158 expense (d) [0.0178] Would pay for a large health expense 0.0331 with savings (d) [0.0271] Could borrow with no interest to pay for •0.0289 a large health expense (d) [0.0218] Could borrow with interest to pay for a 0.0106 large health expense (d) [0.0185] Doctor would give credit for a large •0.166 health expense (d) [0.165] Would get extra work to pay for a large •0.0373 health expense (d) [0.0355] SKY would pay for a large health expense 0.709*** 0.703*** 0.702*** 0.709*** (d) [0.0152] [0.0147] [0.0147] [0.0150] Selling asset only option to pay for 0.0238 health expense (d) [0.0238] Extra work only option to pay for health •0.0687 expense (d) [0.105] Borrowing with interest only option to 0.0144 pay for health expense (d) [0.0230] Limited self•insurance options (no 0.0253 family, borrow w/o interest, etc.) (d) [0.0170] Number of ways to pay for a hypothetical •0.00716 health shock [0.0107] Observations 4739 4778 4778 4739 Pseudo R•squared 0.231 0.225 0.225 0.229 Notes: LHS variable: 1 if bought SKY, 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** Marginalp<0.01, effects; *** p<0.001. Standard Marginalerrors in brackets effects; Standard errors in brackets. Robust standard errors clustered at LHSthe villageVariable: level. 1 if bought All data SKY, is 0 fromif declined the (SKYbaseline Admin survey. data) Regression includes all regressors in the previous Regressiontable, plus includes controls all regressorsfor rich and in previous poor households, table, not shown not here. presented here. Sample is all SKY decliners and all ControlsSKY buyers included who for firstrich and purchased poor households SKY after(not shown). the Village Meeting. (d) for discrete change of indicator Datavariable shown from is from 0 to the 1. baseline survey.

Table 3.8: In‡uence of Self-Insurance Measures on SKY Purchase 140

(1) (2) (3) (4) (5) Trust of Liquidity/ Understands Western Characteristics of Households: Patience Insurance Salience Medicine Gender/ Age Highest ranked wealth by enumerator (d) 0.0674** 0.0674** 0.0676** 0.0683** 0.0656** [0.0229] [0.0228] [0.0228] [0.0231] [0.0231] Lowest ranked wealth by enumerator (d) 0.00515 0.00368 0.00644 0.00848 0.00422 [0.0230] [0.0230] [0.0232] [0.0237] [0.0241] Prefers 20USD now over 60 or 120USD in 0.00805 0.00783 0.00905 0.00964 0.00789 12 months (d) [0.0170] [0.0170] [0.0168] [0.0167] [0.0169] Health decision•maker has 0 years of 0.0177 0.0204 0.0197 0.0176 education (d) [0.0214] [0.0215] [0.0217] [0.0216] Health decision•maker has 1 to 4 years 0.0424* 0.0424* 0.0411* 0.0389+ of education (d) [0.0203] [0.0203] [0.0202] [0.0201] Answered all literacy/numeracy questions •0.0233 •0.0218 •0.0215 •0.0198 incorrectly (d) [0.0163] [0.0164] [0.0165] [0.0163] Knows a neighbor with a > 100 USD health 0.0498* 0.0514* 0.0498* shock in the last year (d) [0.0252] [0.0255] [0.0253] Family member had a > 100 USD health 0.0415* 0.0416* 0.0411* shock in the last year (d) [0.0206] [0.0206] [0.0206] Control: Household member spent over •0.0161 •0.0104 •0.00787 100USD on a given individual, past 12 m (d) [0.0257] [0.0260] [0.0258] Household always uses covers for water •0.0231 •0.0239 jugs (d) [0.0159] [0.0160] All vaccines fulfilled for members under 0.0707** 0.0696** 6, 0 if no under 6, pre•SKY (d) [0.0245] [0.0244] Male household member, under 6, in poor 0.0548+ self•reported health (d) [0.0296] Female household member, under 6, in 0.0163 poor self•reported health (d) [0.0338] Male household member, age 16 to 64, in 0.0267 poor self•reported health (d) [0.0189] Female household member, age 16 to 64, 0.0632*** in poor self•reported health (d) [0.0181] Male household member, over 64, in poor 0.0113 self•reported health (d) [0.0347] Female household member, over 64, in •0.034 poor self•reported health (d) [0.0322] Observations 4739 4739 4739 4695 4695 Pseudo R•squared 0.15 0.151 0.153 0.153 0.155 Notes: LHS variable: 1 if bought SKY, 0 if declined (SKY Administrative data). + p<0.10, * p<0.05, ** p<0.01, *** Marginalp<0.001. effects; Marginal Standard effects; errors inStandard brackets errors in brackets. Robust standard errors clustered at the village level. Use of LHSwater Variable: jug covers 1 if bought and SKY, literacy 0 if declined tests are (SKY from Admin the data) second round survey. All other data is from the baseline survey. RegressionsRegression include includes all previous all regressors regressors in except the previousself•insurance table, measures not presented (not shown). here. Gender/age regression does not Gender/agecontrol for regression household•level does not controlself•reported for household•level health or self•reported health shocks health (allor health other shocks. regressions do, not shown). Sample is Robustall SKY standard decliners errors and clustered all SKY at the buyers village who level. first purchased SKY after the Village Meeting. (d) for discrete change of Allindicator other data variable is from thefrom baseline 0 to 1.survey.

Table 3.9: In‡uence of Other Selection Measures on SKY Purchase 141

Poorest Medium Wealthiest

Percent Purchasing SKY No Poor Health 24% 24% 27% Poor health 31% 37% 43%

Percent with a Member in Poor Health No SKY 78% 66% 63% SKY 83% 78% 78%

Notes: Sample includes decliners or buyers that first purchased SKY after the Village Meeting. "Poor health" is defined as having at least one household member in self• reported poor health.

Table 3.10: SKY Purchase, by Wealth and Poor Health

3.A Supplementary Tables 142

(1) (2) (3) (4) (5) Early Early Early surveys and All HHs surveys buyers buyers Village FE Offered a Deep Discount (d) 0.386*** 0.369*** 0.366*** 0.325*** 0.439*** [0.0192] [0.0278] [0.0178] [0.0297] [0.0133] At least one member over 64 (d) •0.0115 •0.0206 •0.0202 •0.0144 •0.0116 [0.0177] [0.0260] [0.0152] [0.0227] [0.0176] At least one member age 5 or under (d) •0.0498* •0.0323 •0.0302 •0.0157 •0.0374 [0.0247] [0.0437] [0.0224] [0.0363] [0.0255] Cost of moto from village to health •0.0342+ •0.0453 •0.0392 •0.0262 center (USD) [0.0179] [0.0543] [0.0261] [0.0451] Average score for inventory, hygiene, 0.382* 0.0125 0.135 •0.135 and equipment (positive/total outcomes) [0.151] [0.226] [0.122] [0.210] Health Center total open hours / 100, •0.00741 0.042 0.034 0.0535 actual (survey week) [0.0304] [0.0344] [0.0259] [0.0375] Sqrt distance from village to referral 0.00513 •0.0165 •0.00998 •0.0308+ hospital (km) [0.0110] [0.0199] [0.0107] [0.0180] Plays games of chance for money 0.0498+ 0.0227 0.0377 0.0195 0.0702* (gambles) (d) [0.0264] [0.0346] [0.0250] [0.0340] [0.0280] Highest ranked wealth by enumerator (d) 0.0683** 0.134*** 0.0700** 0.133*** 0.0840*** [0.0231] [0.0388] [0.0233] [0.0399] [0.0235] Lowest ranked wealth by enumerator (d) 0.00848 •0.0123 •0.0387+ 0.00384 •0.0167 [0.0237] [0.0480] [0.0203] [0.0462] [0.0247] Health decision•maker has 0 years of 0.0197 0.0483 •0.00623 0.0199 0.00787 education (d) [0.0217] [0.0396] [0.0207] [0.0360] [0.0220] Health decision•maker has 1 to 4 years 0.0411* 0.0364 0.0327+ 0.0431 0.029 of education (d) [0.0202] [0.0351] [0.0193] [0.0311] [0.0182] Answered all literacy/numeracy questions •0.0215 •0.0136 •0.0214 •0.00458 •0.0249 incorrectly (d) [0.0165] [0.0259] [0.0148] [0.0219] [0.0165] Knows a neighbor with a > 100 USD health 0.0514* 0.0853* 0.0253 0.0472 0.0246 shock in the last year (d) [0.0255] [0.0388] [0.0221] [0.0371] [0.0233] Family member had a > 100 USD health 0.0416* 0.038 0.0374+ 0.0247 0.0328+ shock in the last year (d) [0.0206] [0.0313] [0.0203] [0.0279] [0.0199] Household always uses covers for water •0.0231 •0.00687 •0.0322* •0.00364 •0.0321* jugs (d) [0.0159] [0.0281] [0.0137] [0.0245] [0.0162] All vaccines fulfilled for members under 0.0707** 0.0902+ 0.0404+ 0.0569 0.0628* 6, 0 if no under 6, pre•SKY (d) [0.0245] [0.0465] [0.0222] [0.0395] [0.0269] Observations 4695 1611 4222 1485 4684 Pseudo R•squared 0.153 0.177 0.183 0.185 0.254 Notes: LHS variable: 1 if bought SKY, 0 if declined (SKY Administrativ data). + p<0.10, * p<0.05, ** p<0.01, *** Marginalp<0.001. effects; Marginal Standard effects; errors inStandard brackets errors in brackets. Robust standard errors clustered at the village level, LHSexcept Variable: in Col. 1 if (5),bought where SKY, village•level0 if declined (SKY fixed Admin effects data) are included (not shown). Clinic hours and hygiene and (*)inventory Major health and shock equipment means shock score that are caused from 7 thedays Clinic of missed Survey. work, a Distance 100USD expense, to health or afacilities death. Utilization is from is interviews following a majorwith shock. Robustvillage standard leaders. errors Hypothetical clustered at therisk village aversion, level. accidental injury, use of water jug covers, and literacy tests are from Pre•meetingthe second indicates round survey.prior to the All SKY other village data meeting. is from the baseline survey. All previous regressors, except for self• Clinicinsurance actual hoursmeasures, and hygiene are includedand inventory in regressions,and equipment butscore only are fromsignificant the Clinic regressors Survey. are shown. Col. (1) and (5) use Distancethe full sampleto health facilitiesof households is from interviews that declined with village insurance leaders. or bought for the first time after the Village Meeting. Col. (2) Hypotheticalincludes only risk householdsaversion, accidental that wereinjury, surveyedwater jug covers, within and 93 literacy days testsof the are Village from the Meeting. round 2 survey. Col. (3) includes only Allhouseholds other data is that from purchased the baseline survey.SKY within 63 days of the Village Meeting. Col. (4) includes only households that were Samplesurveyed includes within all 93SKY days decliners and and bought all buyers SKY who within first purchased63 days ofSKY the after Village the village Meeting. meeting (d) unless for otherwisediscrete indicated.change of indicator variable(d) for discrete from change 0 to 1. of dummy variable from 0 to 1

Table 3.11: Robustness Checks 143

Wealth Interaction

Offered a Deep Discount (d) 0.379*** [0.0185]

Household member in poor health (d) 0.132*** [0.0153]

Highest ranked wealth by enumerator (d) 0.0296 [0.0375]

Poor health X Highest ranked wealth (d) 0.0522 [0.0429]

Observations 5228 Pseudo R•squared 0.143 Notes: LHS variable: 1 if bought SKY, 0 if declined (SKY Admin data). + Marginalp<0.10, effects; * p<0.05, Standard ** p<0.01, errors in *** brackets p<0.001. Marginal effects; Standard errors in LHSbrackets. Variable: Robust 1 if bought standard SKY, 0 errorsif declined clustered (SKY Admin at thedata) village level. All data is from (*)the Major baseline health survey.shock means Sample shock thatis all caused SKY 7decliners days of missed and work,all SKY a 100USD buyers expense, who first or a death. Utilization is following a major shock. Robustpurchased standard SKY errors after clustered the Village at the villageMeeting. level. (d) for discrete change of indicator Pre•SKYvariable indicates from 0 priorto 1. to the SKY village meeting.

Table 3.12: Wealth/Health Interaction 144

3.B Theoretical Model

Previous research has theorized that several factors can be in‡uencing insurance purchase. This section describes a simple model of take-up behavior that highlights the key factors that will a¤ect the take-up decision and generates predictions for the empirical analysis. We focus on three sets of factors: health parameters, preference parameters, and access and supply-side factors. The design of this model borrows heavily from the model produced for rainfall insurance take-up in Giné, et al., (2007). The model was modi…ed to …t the health sector. In the following model, household utility depends on both consumption and health, so that the household seeks care for health shocks and must pay for care via either out-of-pocket expenses or insurance. Several other features are added to the model, such as parameters representing a personal discount rate, the ability to understand insurance, and the ability to self-insure, To start, assume that each household has utility dependent on consumption c and health H: U(c; H). Health depends on health status before care, d, and medical care, M: Health status d equals 1 if sick and 0 if healthy. Thus, when ill, H = H [1;M], and when healthy, H = H [0; 0]. Assume that illness is 100% cured with care, and households always get care for illness. Thus, H [1;M] = H [0; 0]. Because care cures illness, health does not a¤ect utility, and thus let U (c; H) = U (c) for simplicity of notation. A household must decide whether to buy insurance for a price . Assume quadratic expected utility (see Giné, et al., 2007), so E [U (c)] = E (c) var(c). is the weight that the household puts on the variance of consumption. The higher is , the more risk averse the household is to loss of consumption income. Let Y equal income. Assume Y = y, and that the only possible shock to consumption is medical costs (m), so that c = y m. Medical costs m depends on whether the household is insured, and whether or not the household becomes ill. m equals mNI (NI = no insurance) in the absence of insurance. mNI equals 0 if healthy, which happens with probability (1 p) or msick if unhealthy 2 (probability p). Let E [mNI ] = mNI = p msick, and variance of mNI = . Assume no m saving or borrowing, so c = y mNI in the absence if insurance. With insurance, m = , and thus c = y . Let t be the fraction of medical costs not covered by SKY insurance, or costs a household perceives are not covered. SKY’s o¢ cial policy is to cover all medical costs at public health centers, all medical costs at public hospitals with a referral, and all drugs prescribed and purchased at public facilities. Transportation costs to hospitals are covered in case of emergency. However, there are several reasons a household may perceive that SKY will not cover all medical costs. t can be larger for a given household if a household must pay high costs for transportation to public facilities or if public facilities are far away, if a household uses mostly private facilities or drugsellers (not covered by insurance), if a household does not trust that SKY will pay for treatment, or if a household anticipates that they will need to make "thank you" payments to doctors. t may also be higher if the household does not understand insurance, and thus does not understand that treatment costs will be covered by SKY. Taking into account t, consumption with insurance becomes c = y t mNI . 145

Some households may …nd it more di¢ cult to pay for medical expenses in the absence of insurance. If a household must take out high interest loans to pay for care, or if the household must sell productive assets, the costs of medical care are e¤ectively higher. In contrast, if a household can cheaply self-insure, e.g., family can help pay for care, or if the household has savings, they do not have these added costs of a medical shock. To model this, I in‡ateuncovered medical costs by a factor q (q > 1), where q is higher for households with stricter liquidity constraints. To take into account that the insurance premium is paid today but payout from insurance is in the future, we can in‡ate by a personal discount rate ( < 1). With this discount rate, c = y t q mNI (1+) for the insured. For the uninsured, c = y q mNI . Finally, assume that there are some unobserved qualities of households that in‡u- ence their decision to purchase insurance. For simpli…cation assume that these unobserved factors in‡uence consumption, so that c = y t q mNI (1 + ) + "I for the insured and c = y q mNI + "NI for the uninsured. A household buys insurance if their expected utility with insurance exceeds that without insurance: E [UI ] > E [U]. Substituting in utility = U(c), we can simplify to the decision rule to buy insurance if:

2 2 2 1 < (1 t) q mNI + q 1 t (1 ) + " (3.2) m 1 where " = ("I "NI ) (1 ) . Thus, a household will buy insurance if the premium, , is less than the expected future health care payments from the insurance company (the …rst term) plus the utility gain from reducing uncertainty about consumption (the second term), plus any unobserved in‡uences on insurance purchase ("). Both terms are reduced by the discount rate () and by consumers’belief that SKY may renege on promised insurance or may not pay all costs (t), and increased by the value consumers put on avoiding liquidity constraints (q). So far, the model includes two factors related to the ability or willingness to pay for either health care premium or medical expenses: measures the extent to which a household discounts future income and q represents liquidity constraints on future possible health expenses (i.e., the ability to self-insure). A third reason a household may not buy insurance is because of present budget constraints, or the inability to a¤ord the insurance premium. Even if a household would not be able to self-insure in the future, they cannot purchase insurance if they cannot a¤ord it. For simplicity of notation, assume q = t = 1, and t = 0, and ignore unobserved characteristics, ". As in Giné, et al., (2007) assume households have existing wealth W , and they must have that wealth before they can spend it either on insurance at price or on investment that increases income y. Thus, W + I . Let y be an increasing function of I with decreasing marginal returns to I, so f is concave: y = f(I). If wealth is su¢ ciently high, a household can choose whether to buy insurance and put their preferred amount in I. If is not su¢ ciently high, a household that chooses to buy insurance must decrease investment: I = W if they buy insurance, and I W if they don’tbuy. Thus, consumption becomes c = f(W ) for the insured, and c = f(W ) mNI for the uninsured. 146

A budget-constrained household will buy insurance if E [UI ] > E [UNI ], which is equivalent to E [f (W )] var [f (W )) ] > E [f(W ) mNI ] var [f(W ) mNI ]. 2 Simplifying, we get f(W ) > f(W ) mNI m. Thus, a household buys 2 insurance if f(W ) f(W ) < mNI + m. We assumed decreasing returns to investment. Thus, for a given premium , the foregone return from lower investment (that is, the di¤erence f(W ) f(W )) declines as wealth increases. Intuitively, willingness to pay for insurance increases because returns to investment are decreasing. Thus, the model predicts that when credit constraints are binding, households will be more likely to buy insurance the higher is their income. To summarize, the comparative statics predictions derived from this model are that a household will be more likely to buy insurance if expected medical expenses (mNI ) are higher; insurance is expected to cover a higher percentage of medical expenses (t is 2 lower); risk aversion for income loss ( ) is higher; variance of expected loss (m) is higher (not measured); the household cannot cheaply self-insure (q is higher); the household is not present biased ( is lower); and if a household can a¤ord the premium (W is not too low). Further, recall that expected medical expenses mNI = p msick (probability of becoming ill times expected medical expenses when ill). Anything that increases either p or msick will also increase the probability of buying insurance. We can break down p and msick in the following way. Probability of a household member becoming ill, p, will be higher if a household member is currently in poor health; a household member has had a recent health shock; a household member is stunted or wasted; a household member is under age 6 or over age 64; a household has many household members; a household member is accident-prone; or if household members take risks with their health. Subjective probability of a health shock may also increase if an individual knows someone with a recent health shock (Tversky and Kahneman, 1974). We include this as a factor in‡uencing p: A household that knows someone who has been very ill or had high medical expenses in the past will adjust the probability of shock or expectation of costs upwards. We assume that expected medical expenses when ill (msick) are in‡uenced by past health care utilization. Thus, a household that has had high medical expenses in the past will have higher expected future medical expenses. If a household puts less weight on healthcare of certain household members, such as the elderly, expected medical expenses will be lower for these household members, even when there is a high probability they are ill. A household may believe that SKY will cover only a small fraction of medical costs (t is low) if this household: uses non-public (non-covered) health facilities for treatment (past use of public health facilities will raise t); prefers private facilities; believes public facilities are low quality; must pay high costs for transportation to public care (includes costs of lost time); does not trust Western medicine; does not understand how SKY works; does not trust SKY to pay for treatment (not measured); must make thank you payments (not measured). In our analyses, we break down these factors into traditional measures and more 147 recent extensions to these measures. Table 3.2 organizes theoretical results into traditional and other measures.

3.C Description of Variables

Variables from the Baseline survey, the Village Leader Interview, the Health Center Survey, and Village Meeting interviews used in Chapter 3 and some Robustness Checks in Chapter 2.

Variable Questionnaire Question Description Name Subjective How healthy is each household 1 if respondent describes poor health member? (Excellent health, health of any household mem- good health, poor health). ber as “poor”, 0 otherwise Primary respondent to questionnaire gives subjective response for all household members. Major health Three questions: In the last 1 if respondent answers “yes” shock, 2-4 year, were there any health to any of these three health months pre- problems in your household questions, AND the month meeting that made someone unable to of the health shock was 2-4 work or go to school for one months prior to the date of the week or more? In the last year SKY meeting. did anyone in your household pass away? In the last year did anyone in your household spend more than 400,000 riel ($100 USD) on a single health problem? Visit pub- [If household member expe- 1 if, following a major health lic facility rienced major shock in 2- shock in the 2-4 months pre- for a major 4 months pre-meeting:] Did meeting, a household member health shock, [sick member] seek treatment visited a public health center 2-4 months for this health problem? If yes, or hospital for …rst or subse- pre-meeting where? [Respondent chose quent treatment, 0 otherwise “Health center” or “public hospital”] 148

Variable Questionnaire Question Description Name Visits a pri- [If household member expe- 1 if, following a major health vate facility rienced major shock in 2- shock in the 2-4 months pre- for a major 4 months pre-meeting:] Did meeting, a household member health shock, [sick member] seek treatment visited a private doctor for 2-4 months for this health problem? If …rst or subsequent treatment, pre-meeting yes, where? [Respondent 0 otherwise chose “private doctor (village or town)”] Household has Height, age, and weight mea- 1 if household has a child that a stunted or sured for all children age 5 and is stunted or wasted (zscore wasted child under for height-for-age or weight- for-height is less than -2) according to WHO growth standards, 0 otherwise (including if household has no child age 5 or under) Household size Household roster: Name of Number of household mem- (used as a con- people who usually sleep here bers listed in the household trol only, not (slept in the house 5 out of the roster presented) 7 nights immediately preceding the interview) Household has Date of birth of each house- 1 if any household member is a member age hold member age 65 or older, 0 otherwise 65 or older Household Date of birth of each house- 1 if any household member is member has a hold member age 5 or under, 0 otherwise member age 5 or under Risk averse Round 2 survey: Which would Hypoth. Local Risk Aversion: you prefer: 1) a gift of $500 1-4, least to most risk averse, or the chance of either $500 confused also = 4. 1= chose or 2) a game which gives 250/850 over 500. 2 = chose you 50% chance to win $250 250/1000 ver 500, but not and 50% chance to win $1000 250/850. 3 = chose 250/2000 [Also asked to choose between over 500, but not 250/1000. 4 $500 and 250/850, 500 and = chose 500 over all gambles, 250/2000.] or was confused by question (see explanation below). Gambles Do you or your spouse play 1 if plays games of chance for games of chance for entertain- money, 0 otherwise ment with money? 149

Variable Questionnaire Question Description Name Household Round 2 survey: Have you 1 if respondent or spouse re- member has or your spouse ever needed ceived care for an accidental received care health care because of acciden- injury in the past year, 0 if not for accidental tal injury? injury Risky health Hypothetical question: I’mgo- 1 if family chose B, would ac- behavior ing to ask you to make some cept salary increase for the choices about taking a job at riskier job, 0 otherwise two di¤erent factories with dif- ferent salaries. Suppose that 100 people work at each fac- tory. Please tell me which job you prefer to take. A) Daily wage of 4000 riel (1 USD), no injuries B) Daily wage of 5000 riel, 3 people injured in the past year, spent two days in hospital Household How could you pay for a Several measures: Indicator does not have 400,000 riel (100 USD) health variables for each option; an ways to self- expense? (Multiple answers indicator variable for only insure against possible out of cash on hand, costly options (no family gift, health care savings, family gift this loan (with or without inter- costs province, family gift other est), help from association, province, borrow no interest, cash on hand, savings); a borrow with interest, …nd count of the number of options extra work, SKY would pay, listed sell asset, other) Poor house- Enumerator subjective 1 if enumerator rates household wealth ranking: poor- hold as poor, 0 otherwise est/medium/better o¤ Better-o¤ Enumerator subjective 1 if enumerator rates house- household wealth ranking: poor- hold as better-o¤, 0 otherwise est/medium/better o¤ High discount If a trusted relative wanted 1 if prefers 20USD now over rate to give you a gift, would you 120 USD 12 months from now, choose 20 USD now or 120 0 otherwise USD in 12 months? 150

Variable Questionnaire Question Description Name Education Who makes the decisions Education from 1 to 13 (13 of health about healthcare in your = tertiary education). If re- decision-maker family? What is the highest spondent decides with another (years) grade this person completed? household member, use max- What is the highest grade you imum education of the two completed? members. Indicator variables for 0 years or 1 to 4 years used in regressions. Respondent is Four literacy/numeracy ques- 1 if respondent answers all lit- illiterate and tions: Draw a line from each eracy and numeracy questions innumerate picture to the correct word; incorrectly, 0 otherwise Write the name of the village, commune and district where you live; Write the correct number of objects in the pictures, and what it the object is; Tell me what time it is (picture of a clock shown) Confused by Round 2 survey: Which would 1 if confused by risk aversion risk aversion you prefer: 1) a gift of $500 question: Chose $500 over a question or the chance of either $500 or 50/50 chance of $500/$1000, 2) if you are lucky $1000? If even after options were ex- respondent chooses $500: Are plained further you sure? In the second option you will get at least $500 and you may get $1000. In option 1 you will always get $500. Knows family Do you know anyone who has 1 if family member had a with shock spent 400,000 riel (100 USD) 100USD health shock in the on health care in the last year? last 12 months (pre-survey), 0 If yes, who? (Family, neigh- otherwise bor, friend, other) Knows neigh- Do you know anyone who has 1 if a neighbor had a 100USD bor with shock spent 400,000 riel (100 USD) health shock in the last 12 on health care in the last year? months (pre-survey), 0 other- If yes, who? (Family, neigh- wise bor, friend, other) 151

Variable Questionnaire Question Description Name Control: Indi- For each household member(s) 1 if a single household member vidual expense treated for a major shock in 12 spent over 100USD on health of over 400k months pre-survey, what was care in the past 12 months the total cost of treating ALL for a major health shock (used health problems (at any facil- to eliminate household mem- ity) bers from “Knows family with health shock”question), 0 otherwise All vaccines For each child age 5 and under, 1 if all children in the house- are up to date, enumerator recorded dates of hold are 100% up to date on pre-Meeting vaccines from yellow immu- vaccines prior to the Village nization card Meeting, according to WHO standards, 0 otherwise (including if household has no child age 5 or under) Always covers Round 2 survey: Do you use 1 if household always uses cov- water jars covers for your water jars? ers for water jugs, 0 if they (no, has some covered, has all sometimes or never do. (Wa- covered) ter jars collect rain for household consumption and use in cooking. Covering water jars keeps waer clean and prevents spread of disease, e.g., dengue fever, by mosquitos.) Male or female How healthy is each household 1 if a male (female) household household member? (Excellent health, member is reported in poor member in good health, poor health). health, 0 otherwise poor health, Primary respondent to ques- as reported by tionnaire gives subjective re- repondent sponse for all household members. Working-aged, How healthy is each household 1 if a household member of a younger, or member? (Excellent health, given age group is reported in older house- good health, poor health). poor health, 0 otherwise hold member Primary respondent to ques- (age 15-64, un- tionnaire gives subjective re- der 6, over 64) sponse for all household mem- in poor health, bers. as reported by respondent Table 3.13: Baseline Survey Variables 152

Variable Questionnaire Question Description Name Cost of moto How much does it cost to go Moto cost to local health cen- to local health [to the nearest health center] ter, in USD center by moto?

Table 3.14: Village Leader Survey variables 153

Variable Questionnaire Question Description Name Total hours On Monday [Tuesday, Sum of total hours the health health center is Wednesday, etc.], what center was open during the open per week time did the health center week of the survey, divided by open? What time did the 100 (for ease of presentation) health center close? Health center Drug inventory checklist, Average of: a) Number of quality score Equipment checklist, observ- drugs in stock divided by to- able hygiene and cleanliness tal in list b) Number of equip- questions ment available divided by total in list, and c) Number of negative hygienic practices divided by total number of hygiene questions

Table 3.15: Health Center Survey variables

Variable Questionnaire Question Description Name Km from vil- Village Meeting: How many Square root of number of kilo- lage to regional kilometers from village to clos- meters from village to public hospital est (public) regional hospital? regional hospital

Table 3.16: Village Meeting Variables 154

Bibliography

(2005): “DHS Demographic and Health Survey, Cambodia,”http://www.measuredhs.com.

(2009): “SKY Health Insurance Schemes,”www.sky-cambodia.org/e…rstrestuls.html.

Abdul Lateef Jameel Poverty Action Lab (2011): “The Price is Wrong,” http://www.povertyactionlab.org/the-price-is-wrong.

Akerlof, G. A. (1970): “The Market for "Lemons": Quality Uncertainty and the Market Mechanism,”The Quarterly Journal of Economics, 84(3), pp. 488–500.

Annear, P. (2006): “Study of Financial Access to Health Services for the Poor in Cam- bodia. Phnom Penh,” Research report, Cambodia Ministry of Health, WHO, AusAID, RMIT University.

Antos, J. R. (2007): “Health Care Financing in Thailand: Modeling and Sustainability,” Mission report, World Bank.

Asfaw, A. (2003): “How Poverty A¤ects the Health Status and the Health Care De- mand Behaviour of Households: The Case of Rural Ethiopia,”Conference paper, Chronic Poverty Research Centre (CPRC).

Ausubel, L. (1999): “Adverse Selection in the Credit Card Market,”University of Mary- land.

Banerjee, A., A. Deaton, and E. Duflo (2004): “Wealth, Health, and Health Services in Rural Rajasthan,”The American Economic Review, 94(2), pp. 326–330.

Bitran, R., V. Turbat, B. Meesen, and W. Van Damme (2011): “Preserving Equity in Health in Cambodia: Health Equity Funds and Prospects for Replication,”Discussion paper.

Brook, R. H., J. E. Ware, W. H. Rogers, E. B. Keeler, A. R. Davies, C. A. Donald, G. A. Goldberg, K. N. Lohr, P. C. Masthay, and J. P. Newhouse (1983): “Does free care improve adults’health? Results from a randomized controlled trial.,”The New England Journal of Medicine, 309(23), 1426–1434.

Bryan, G. (2010): “Ambiguity and Insurance,” Job market paper, University California of Los Angeles. 155

Cambodian Government (2004): “Administrative and Health Facility Mapping Health Coverage Plan 2004-2005,”Cambodian Government Publication.

Cameron, A. C., and P. K. Trivedi (1991): “The role of income and health risk in the choice of health insurance : Evidence from Australia,”Journal of Public Economics, 45(1), 1 –28.

Card, D., C. Dobkin, and N. Maestas (2007): “The Impact of Health Insurance Status on Treatment Intensity and Health Outcomes,”Discussion paper.

Cardon, J. H., and I. Hendel (2001): “Asymmetric Information in Health Insurance: Evidence from the National Medical Expenditure Survey,” The RAND Journal of Eco- nomics, 32(3), pp. 408–427.

Case, A., D. Lubotsky, and C. Paxson (2002): “Economic Status and Health in Child- hood: The Origins of the Gradient,” The American Economic Review, 92(5), pp. 1308– 1334.

Cawley, J., and T. Philipson (1999): “An Empirical Examination of Information Bar- riers to Trade in Insurance,”The American Economic Review, 89(4), pp. 827–846.

Central Intelligence Agency (2010): “The CIA World Factbook,” https://www.cia.gov/library/publications/the-world-factbook/.

Chiappori, P.-A., B. Jullien, B. Salanié, and F. Salanié (2006): “Asymmetric Infor- mation in Insurance: General Testable Implications,”The RAND Journal of Economics, 37(4), pp. 783–798.

Cohen, J., and P. Dupas (2010): “Free Distribution or Cost-Sharing? Evidence from a Randomized Malaria Prevention Experiment,” The Quarterly Journal of Economics, 125(1), 1–45.

Cohen, J., P. Dupas, and S. Schaner (2011): “Prices, Diagnostic Tests and the Demand for Malaria Treatment: Evidence from a Randomized Trial,” Unpublished Manuscript cited in Dupas, 2011.

Collins, W. (2000): “Medical Practitioners and Traditional Healers: A Study of Health Seeking Behavior in Kampong Chhnang, Cambodia,”Discussion paper.

Currie, J., and J. Gruber (1996): “Health Insurance Eligibility, Utilization of Medical Care, and Child Health,”The Quarterly Journal of Economics, 111(2), 431–66.

(1997): “The Technology of Birth: Health Insurance, Medical Interventions, and Infant Health,” NBER Working Papers 5985, National Bureau of Economic Research, Inc.

Currie, J., and M. Stabile (2003): “Socioeconomic Status and Child Health: Why Is the Relationship Stronger for Older Children?,” American Economic Review, 93(5), 1813–1823. 156

Cutler, D. M., and S. J. Reber (1998): “Paying for Health Insurance: The Trade- O¤ between Competition and Adverse Selection,”The Quarterly Journal of Economics, 113(2), pp. 433–466.

Cutler, D. M., and R. J. Zeckhauser (1998): “Adverse Selection in Health Insurance,” Forum for Health Economics & Policy, 1, 2.

(2000): “Chapter 11 The anatomy of health insurance,”vol. 1, Part 1 of Handbook of Health Economics, pp. 563 –643. Elsevier.

(2004): “Extending the Theory to Meet the Practice of Insurance,”Working paper, Harvard University and NBER.

Das, J., J. Hammer, and K. Leonard (2008): “The Quality of Medical Advice in Low- Income Countries,”Journal of Economic Perspectives, 22(2), 93–114.

Domrei Research and Consulting, and University of California, Berkeley (2010a): “Insurance Agent/Member Facilitator Interview transcripts,” Phnom Penh, Cambodia.

(2010b): “Who Buys SKY Health Insurance,”Phnom Penh, Cambodia.

Dow, W., P. Gertly, R.-F. Schoeni, J. Strauss, and D. Thomas (1997): “Health Care Prices, Health and Labor Outcomes : Experimental Evidence,”Discussion paper.

Dupas, P. (2011): “Health Behavior in Developing Countries,” Prepared for the Annual Review of Economics, Vol. 3 (forthcoming, September 2011).

Ellis, R. P. (1989): “Employee Choice of Health Insurance,” The Review of Economics and Statistics, 71(2), pp. 215–223.

Fang, H., M. P. Keane, and D. Silverman (2008): “Sources of Advantageous Selection: Evidence from the Medigap Insurance Market,”Journal of Political Economy, 116(2), pp. 303–350.

Fihn, S., and J. Wicher (1988): “Withdrawing routine outpatient medical services,” Journal of General Internal Medicine, 3, 356–362, 10.1007/BF02595794.

Finkelstein, A. (2005): “The Aggregate E¤ects of Health Insurance: Evidence from the Introduction of Medicare,”Working Paper 11619, National Bureau of Economic Research.

Finkelstein, A., and K. McGarry (2006): “Multiple Dimensions of Private Information: Evidence from the Long-Term Care Insurance Market,”The American Economic Review, 96(4), pp. 938–958.

Finkelstein, A., and R. McKnight (2008): “What did Medicare do? The initial impact of Medicare on mortality and out of pocket medical spending,” Journal of Public Economics, 92(7), 1644–1668. 157

Finkelstein, A., and J. Poterba (2004): “Adverse Selection in Insurance Markets: Policyholder Evidence from the U.K. Annuity Market,” Journal of Political Economy, 112(1), pp. 183–208.

Fort, C., B. B. Ravenholt, and H. Stanley (1998): “Cambodia Private Sector As- sessment Report,” Discussion paper, Reproductive and Child Health Alliance, Engen- derHealth, BASICS, and the Family Planning Service Expansion and Technical Support (SEATS) Project, Phnom Penh, Cambodia.

Fuller, G. (1974): “On the Spatial Di¤usion of Fertility Decline: The Distance-to-Clinic Variable in a Chilean Community,”Economic Geography, 50(4), pp. 324–332.

Gertler, P. (2002): “Insuring Consumption Against Illness,”American Economic Review, 92(1), 51–70.

Gertler, P., D. I. Levine, and E. Moretti (2003): “Do Micro…nance Programs Help Families Insure Consumption Against Illness?,” Development and Comp Systems 0303004, EconWPA.

Giné, X., R. Townsend, and J. Vickery (2007): “Patterns of rainfall insurance participation in rural India,”Discussion paper.

Gupta, M. D. (1987): “Selective Discrimination against Female Children in Rural Punjab, India,”Population and Development Review, 13(1), 77–100.

Hanratty, M. J. (1996): “Canadian National Health Insurance and Infant Health,”The American Economic Review, 86(1), pp. 276–284.

Jacoby, H. G., and E. Skoufias (1997): “Risk, Financial Markets, and Human Capital in a Developing Country,”Review of Economic Studies, 64(3), 311–35.

Jalan, J., and M. Ravallion (1999): “Are the poor less well insured? Evidence on vulnerability to income risk in rural China,” Journal of Development Economics, 58(1), 61–81.

Jowett, M., P. Contoyannis, and N. D. Vinh (2003): “The impact of public voluntary health insurance on private health expenditures in Vietnam,”Social Science & Medicine, 56(2), 333–342.

Jullien, B., B. Salanié, and F. Salanié (2007): “Screening risk-averse agents under moral hazard: single-crossing and the CARA case,”Economic Theory, 30(1), 151–169.

Jutting, J. P. (2004): “Do Community-based Health Insurance Schemes Improve Poor People’s Access to Health Care? Evidence From Rural Senegal,” World Development, 32(2), 273–288.

Karlan, D., and J. Zinman (2009): “Observing Unobservables: Identifying Information Asymmetries With a Consumer Credit Field Experiment,” Econometrica, 77(6), 1993– 2008. 158

Keeler, E. B. (1992): “E¤ects of Cost Sharing on Use of Medical Services and Health,” Journal of Medical Practice Management, 8, 317–321.

Kenjiro, Y. (2005): “Why Illness Causes More Serious Economic Damage than Crop Failure in Rural Cambodia,”Development and Change, 36(4), 759–783.

Koufopoulos, K. (2002): “Asymmetric Information, Heterogeneity in Risk Perceptions and Insurance: An Explanation to a Puzzle,”FMG discussion papers, Financial Markets Group.

Kremer, M., J. Leino, E. Miguel, and A. P. Zwane (2011): “Spring Cleaning: Rural Water Impacts, Valuation, and Property Rights Institutions,”The Quarterly Journal of Economics, 126(1), 145–205.

Kremer, M., and E. Miguel (2007): “The Illusion of Sustainability.,”Quarterly Journal of Economics, 122(3), 1007 –1065.

Lamiraud, K., F. Booysen, and X. Scheil-Adlung (2005): “The Impact of Social Health Protection on Access to Health Care, Health Expenditure and Impoverishment: A Case Study of South Africa,”International Labour O¢ ce, Department of Social Security, Extension of Social Security No. 23, 2005.

Levine, D. I., and R. Gardner (2008): “Brie…ng Paper: Health Care in Cambodia,” Discussion paper.

Levine, D. I., R. Gardner, and R. Polimeni (2009): “Brie…ng Paper: A Literature Review on E¤ects of Health Insurance and Selection into Health Insurance,” Discussion paper.

Lichtenberg, F. R. (2002): “The E¤ects of Medicare on Health Care Utilization and Outcomes,”in Frontiers in Health Policy Research, Volume 5, NBER Chapters, pp. 27– 52. National Bureau of Economic Research, Inc.

Lurie, N., N. Ward, M. Shapiro, C. Gallego, R. Vaghaiwalla, and R. Brook (1986): “Termination of Medi-Cal Bene…ts,” New England Journal of Medicine, 314, 1266–1268.

Manning, Willard G, e. a. (1987): “Health Insurance and the Demand for Medical Care: Evidence from a Randomized Experiment,” American Economic Review, 77(3), 251–77.

Morduch, J. (1995): “Income Smoothing and Consumption Smoothing,”The Journal of Economic Perspectives, 9(3), pp. 103–114.

Mwabu, G., M. Ainsworth, and A. Nyamete (1993): “Quality of Medical Care and Choice of Medical Treatment in Kenya: An Empirical Analysis,”The Journal of Human Resources, 28(4), pp. 838–862.

Pande, R. P. (2003): “Selective gender di¤erences in childhood nutrition and immunization in rural India: the role of siblings,”Demography, 40(3), 395–418. 159

Pauly, M. V., P. Zweifel, R. M. Scheffler, A. S. Preker, and M. Bassett (2006): “Adverse Selection and Impacts of Health Insurance in a Developing Country: Evidence from a Randomized Experiment in Cambodia,”Health A¤airs, 25(2), 369–379.

People’s Daily Online (2008): “China’s New Rural Cooperative Medical Plan,” http://english.peopledaily.com.cn/90002/95607/6524401.html.

Polimeni, R. (2006): “Adverse Selection and Impacts of Health Insurance in a Developing Country: Evidence from a Randomized Experiment in Cambodia,”Research prospectus, University of California, Berkeley.

Robinson, J., and E. Yeh (2011): “Risk-coping through Sexual Networks: Evidence from Client Transfers in Kenya,”Discussion paper.

Rosenzweig, M. R., and K. I. Wolpin (1993): “Credit Market Constraints, Consump- tion Smoothing, and the Accumulation of Durable Production Assets in Low-Income Countries: Investment in Bullocks in India,”Journal of Political Economy, 101(2), 223– 44.

Rothschild, M., and J. Stiglitz (1976): “Equilibrium in Competitive Insurance Mar- kets: An Essay on the Economics of Imperfect Information,” The Quarterly Journal of Economics, 90(4), pp. 629–649.

Sauerborn, R., P. Berman, and A. Nougtara (1996): “Age bias, but no gender bias, in the intra-household resource allocation for health care in rural Burkina Faso.,”Health transition review the cultural social and behavioural determinants of health, 6(2), 131–145.

Sekhri, N., and W. Savedoff (2005): “Private health insurance: implications for developing countries,”Bulletin of the World Health Organization 2005, 83, 127–134.

Siegelman, P. (2004): “Adverse Selection in Insurance Markets: An Exaggerated Threat,” The Yale Law Journal, 113(6), pp. 1223–1281.

Smith, J. P. (2005): “Unravelling the SES health connection,”Working papers, RAND.

Tarozzi, A., A. Mahajan, B. Blackburn, D. Kopf, L. Krishnan, and J. Yoong (2011): “Micro-Loans, Insecticide-Treated Bednets and Malaria: Evidence from a Ran- domized Controlled Trial in Orissa (India),” Economic Research Initiatives at Duke (ERID) Working Paper No. 104.

Thomas, D., E. Frankenberg, J. Friedman, J.-P. Habicht, N. Jones, C. McK- elvey, and et al. (2004): “Causal E¤ect of Health on Labor Market Outcomes: Evi- dence from a Random Assignment Iron Supplementation Intervention,”Working papers, UC Los Angeles: California Center for Population Research.

Townsend, R. M. (1994): “Risk and Insurance in Village India,” Econometrica, 62(3), pp. 539–591.

Tversky, A., and D. Kahneman (1974): “Judgment under Uncertainty: Heuristics and Biases,”Science, 185(4157), 1124–1131. 160

Udry, C. (1994): “Risk and Insurance in a Rural Credit Market: An Empirical Investiga- tion in Northern Nigeria,”The Review of Economic Studies, 61(3), pp. 495–526.

Van Damme, W., L. Van Leemput, I. Por, W. Hardeman, and B. Meessen (2004): “Out-of-pocket Health Expenditure and Debt in Poor Households: Evidence from Cam- bodia,”Tropical Medicine and International Health, 9(2), 273–280.

Vietnam Social Security (2010): “History Development and Expanding Program of Voluntary Health Insurance in Vietnam,”http://www.asean-ssa.org.

Vong, S., J. F. Perz, S. Sok, S. Som, S. Goldstein, Y. Hutin, and J. Tulloch (2005): “Rapid assessment of injection practices in Cambodia, 2002,”BMC Public Health, 5(1), 56.

Wagstaff, A., M. Lindelow, G. Jun, X. Ling, and Q. Juncheng (2009): “Extending health insurance to the rural population: An impact evaluation of China’snew cooperative medical scheme,”Journal of Health Economics, 28(1), 1–19.

Wagstaff, A., and M. Pradhan (2005): “Health insurance impacts on health and non- medical consumption in a developing country,” Policy Research Working Paper Series 3563, The World Bank.

Wagstaff, A., and E. Van Doorslaer (2003): “Catastrophe and impoverishment in paying for health care: with applications to Vietnam 1993-1998.,” Health Economics, 12(11), 921–934.

Wang, Y. (2007): “Development of the New Rural Cooperative Medical System in China,” China & World Economy, 15(4), 66–77.

Wolfe, J. R., and J. H. Goddeeris (1991): “Adverse selection, moral hazard, and wealth e¤ects in the medigap insurance market,” Journal of Health Economics, 10(4), 433–459.

World Bank (2006): “Cambodia: Halving Poverty by 2015? Poverty Assessment 2006, East Asia and the Paci…c Region,”Economic Report 35213-KH, The World Bank.

World Health Organization (2007): “Social Health Protection. Factsheet No. 320,” Factsheet 320, World Health Organiztion.

Yip, W., and P. Berman (2001): “Targeted health insurance in a low income country and its impact on access and equity in access: Egypt’s school health insurance,” Health Economics, 10(3), 207–220.