Electronic Thesis and Dissertations UC Berkeley

(1)

eScholarship provides open access, scholarly publishing services to the University of California and delivers a dynamic research platform to scholars worldwide.

Peer Reviewed Title:

Health Insurance in Rural Cambodia: Impacts and Selection Author:

Polimeni, Rachel Acceptance Date: 2011

Series:

UC Berkeley Electronic Theses and Dissertations Degree:

Ph.D., EconomicsUC Berkeley

Advisor(s): Miguel, Edward Committee:

Levine, David I, Lee, Ronald

Permalink:

http://escholarship.org/uc/item/5jh52309 Abstract:

Copyright Information:

(2)

Rachel A. Polimeni

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

Doctor of Philosophy in Economics in the GRADUATE DIVISION of the

UNIVERSITY OF CALIFORNIA, BERKELEY

Committee in charge: Professor Edward Miguel, Chair

Professor David Levine Professor Ronald Lee

(3)

(4)

Health Insurance in Rural Cambodia: Impacts and Selection by

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 con-sequences. 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 unin-sured 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

(5)

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 insur-ance purchase. However, households that paid more for health insurinsur-ance 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 in-surance may decrease the likelihood of buying inin-surance 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 house-holds 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.

(6)

(7)

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

in SKY . . . 86

2.3 Proportion of Households using SKY-covered Health Facilities for Care, by

(11)

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

(12)

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 7 days . . . 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

(13)

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

(14)

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.

(15)

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’s household 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

1

David 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.

(16)

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 ran-domizing 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 cat-egories 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 insur-ance 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

(17)

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

(18)

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 insur-ance should not increase utilization of care. On the other hand, because the SKY insurinsur-ance 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

(19)

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 med-ical providers, private drug sellers (with and without pharmaceutmed-ical training), and tradi-tional 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

(20)

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’s principal 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.

(21)

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 house-hold 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

(22)

lowing 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 in-surance is most important for larger shocks, we also estimate whether inin-surance decreases the occurrence of costs exceeding 250 USD following a single incident (the top 10th per-centile), 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 im-pact 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

(23)

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 ran-domized controlled trial that allows us to identify the impact of health insurance indepen-dently 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.

3_{The 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 ap-plications 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.

(24)

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’s name 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.

(25)

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 householdi at timetwith ordinary least squares:

Yit= SKYit+"i (1.1)

the estimated coe¢ cient _OLS can have very large bias because SKY membership is

en-dogenous. 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 (M onthsit):

SKYit = 1 Ti+ 2 M onthsit+ 3 M onthsit Ti+uit (1.2) 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 montht,t

is de…ned as the date of the incident,monthsit is de…ned as the number of months between the village meeting and timet, and the instrumentM onthsit Ti isM onthsitmultiplied by

1 if householdireceived a high coupon and 0 if the household received a low coupon. SKY

status in montht,SKYit, is de…ned as a three-month average membership rate centered in montht, to account for imperfect recall of the timing of health incidents. Thus,SKYitcan take on the values 0, 1₃;2₃or 1. For example, for a health incident occurringt months after the village meeting,SKYit equals 1 if householdiwas insured in monthst 1,t, andt+ 1, but equals 1₃ if the household was insured in only time t 1.

Similarly, for birth outcomes,tis de…ned as the month of the birth, andmonthsit

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…neM onthsit 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…neSKYit 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.

(26)

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 mea-sures 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 sam-ple, 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

(27)

The surveys cover demographics, wealth, objective health measures, health care utilization and spending, assets and asset sales, savings, debt, trust of health care insti-tutions, 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.

(28)

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

(29)

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 montht, and thatmonthsis de…ned as the number of months between the village meeting and montht. 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

(30)

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 pur-chased 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’s theory 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 mea-sures) 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 immu-nizations 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.

4_{The 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.

(31)

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.

5

Results hold if we include households that did not have a death or missed 7 days, but spent over $100USD on care.

6

We 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.

(32)

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 out-comes 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 accu-mulation 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).

7

We test some outcomes holding baseline constant in Appendix Table 1.14.

8_{To 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.

(33)

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 indi-viduals 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