Assessing adverse selection and health care demand in micro health insurance: Evidence from a community-based microfinance insurance model in India

Transcription

1 Assessing adverse selection and health care demand in micro health insurance: Evidence from a community-based microfinance insurance model in India Ketki Sheth* December 11, 2014 Abstract: Delivering health insurance to the poor through microfinance institutions (MFIs) has been gaining attention as a method to overcome informational asymmetries and increase demand for insurance. This study evaluates a micro health insurance contract by randomizing the insurance offer to women in microfinance Self Help Groups (SHGs) in rural Maharashtra, India. I assess the MFI s novel enrollment strategy to insure SHGs (rather than individuals) as a method to increase demand and reduce adverse selection. Comparing enrollment among SHG members, for whom the unit of enrollment was binding, with enrollment of additional family members suggests this method of bundling health insurance with MFIs successfully increased demand but did not protect against adverse selection. Both SHG members and their extended household members who chose to enroll in the health insurance were 9 percentage points more likely to report illness at baseline. Contrary to other studies on micro health insurance, I fail to find robust evidence of the health insurance increasing demand for health care, and even find limited suggestive evidence of a reduction in self-reported illness and health care utilization. This suggests scope for additional indirect benefits of increased health to insured members and improved financial sustainability of micro health insurance contracts. *Department of Economics at the University of California, Merced. ksheth@ucmerced.edu. I thank Craig McIntosh, Karthik Muralidharan, Gordon Dahl, Paul Niehaus, and seminar participants at UCSD for support and comments. I also thank the staff at Chaitanya, particularly Kalpana Pant, for overall support of the project, and the International Labour Organization s Microinsurance Innovation Facility and the European Development Research Network for financial assistance.

2 I. Introduction Though health care has been documented as a significant expenditure in poor households (Banerjee et al. 2009, Dupas and Robinson 2009), and informal risk pooling shown to be incomplete (Townsend 1994, Morduch 1999, Jalan and Ravallion 1998), until recently health insurance was virtually non-existent in most developing countries. 1 Missing insurance markets in poorer parts of the world has led to a growing effort among community based organizations, particularly micro finance institutions (MFIs), to use their organizations as platforms to provide and deliver health insurance, often referred to as micro health insurance (MHI). How best to leverage MFIs to deliver health insurance and whether MHI is effective at reducing vulnerability and increasing access to health care remain open questions. MFIs have been particularly promising at delivering insurance by providing an aggregated unit to insure though bundling insurance products with their loans. This arguably reduces concerns of both low demand leading to insufficient risk pools and adverse selection in which demand is limited to those with poor health. However, Banerjee et al. (2014) find that when MHI was made a mandatory product for receiving a loan (i.e., the aggregated unit being the entire MFI) clients preferred not to renew loans rather than be forced to purchase MHI and there was no evidence of adverse selection. The case highlights the difficulty in determining how to bundle MHI with MFI and suggests adverse selection may not be a concern in these settings. Using evidence based on a randomized introduction of MHI, this paper assesses an alternative method for bundling MFIs with MHI by leveraging smaller sub-units (SHGs 2 ) 1 According to WHO Core Health Indicators: 2 An SHG (Self-Help Group) is usually 15 to 20 women who come together to borrow and save. Many MFIs, including the government, have organized themselves around such SHGs in India.

3 common within many MFIs. I assess an MFI which offered MHI to their members at the SHG level. Conditional on their SHG enrolling in the insurance, SHG members were then free to enroll additional household members as they chose. Differences in the enrollment patterns by SHG members (where the aggregated level of SHG enrollment was binding) to enrollment patterns among their household members, provide suggestive evidence of the effectiveness of the bundling at increasing demand and preventing adverse selection. The significantly higher enrollment rates among SHG members (67%) relative to non-shg household members eligible for enrollment (17.5%) suggests providing insurance at an aggregated unit does significantly increase demand for the product. I also find no evidence of SHG members choosing to leave their SHG or the MFI when offered the health insurance. In this context of high demand for MHI, I find concerns for adverse selection persist despite providing insuring at the SHG level. Among both SHG members and non-shg household members, those who enroll in MHI are 9 percentage points more likely to be experiencing an illness at baseline than those who did not enroll. However, though households with greater health needs are more likely to enroll non-shg household members, I find no evidence that within households members with greater health needs are more likely to be enrolled. A primary purpose of MHI is to improve health and reduce vulnerability through mechanisms such as reducing health expenditure and increasing access and demand for health care. Though there was high demand for MHI, using the randomized introduction of the MHI, I find no evidence that the MHI offer affected the demand for health care or health expenditure. Observing future enrollment decisions, I estimate a difference-in-difference effect of the MHI and find limited suggestive evidence of a decrease in self-reported illness and seeking of health care for enrolled SHG members.

4 This paper is one of the few studies to find that health care utilization does not increase upon enrollment in an MHI program (Jutting 2004, Chankova et al. 2008, Jakab and Krishnan 2004, Wagstaff and Lindelow 2008). Because previous studies evaluating the impact of MHI on health care demand have primarily relied on comparing the insured versus uninsured, it is unclear to what extent these results stem from the effect of being insured versus preexisting differences between those who choose to enroll in insurance and those who do not. This paper s results suggest that demand for health insurance is a function of baseline health and therefore selection bias is a significant concern in evaluating MHI. This paper adds to the literature by evaluating a promising method of bundling MHI with MFIs. It is also one of the few papers to provide suggestive evidence for adverse selection in MHI demand and a causal link between MHI and health care outcomes that does not rely on comparisons based on enrollment status. Rather, the paper s identification strategy exploits a randomized controlled trial design. As described above, many studies have been stymied by identification and compare users and non users of an insurance program; the randomized controlled trial methodology employed in this paper attempts to overcome this barrier and provide causal estimates without relying on differences between insurance status. Unlike the majority of the studies, I reject the null hypothesis of an increase in the use of health care and find limited financial protection against health expenditures. The remainder of this paper is organized as follows: Section 2 describes the MHI contract and context, Section 3 outlines the research design and data, Section 4 details the identification strategy and assumptions, Section 5 discusses results and robustness analysis, and Section 6 concludes. II. Overview of the MHI

5 Overview of the Micro Health Insurance Contract In January 2011, Chaitanya, a non-profit microfinance institution (MFI) working on women s empowerment and microfinance in Junnar sub-district of rural Maharashtra, expanded its community based MHI program, Dipthi Arrogya Nidhi (DAN). Though micro health insurance contracts differ in design, DAN shares many of the characteristics common to MHI. These include distributing through an existing MFI infrastructure (the most common provider of micro health insurance), reducing the cost of health care, implementing a coverage cap and copay, and charging a single premium price. The reduction in the cost of health care includes both price reductions and mechanisms such as improved signals of health care quality (e.g., empanelling facilities), easier access to health care, and increased saliency of health. The cost of membership to DAN is INR 200 (USD 4) per person per year if the household insures 1 or 2 persons, or INR 150 (USD 3) per person per year if the household insures 3 or more persons. The main provisions of the health insurance contract are discounted prices (5 to 20%) negotiated at private network medical facilities, which include hospitals, medical laboratories, and pharmacies. Additionally, for in-patient treatment, the member receives 60 percent reimbursement of their medical fees at network private hospitals, and 100 percent reimbursement at government medical facilities, up to a limit of INR 15,000 (USD 300) per event. 3 The product also includes a 24-7 medical help-line, health camps, and monthly village visits by a doctor to offer referrals and basic medicines. However, village visits by a doctor were intermittent and only one health camp was implemented during the timeframe of the research study. Bundling MHI with MFI: DAN capitalizes on Chaitanya s preexisting microfinance Self Help Groups (SHGs) structure comprised of 15 to 20 women members. The option to purchase the contract is limited to SHGs in which at least 80 percent of SHG members purchase the MHI. Conditional upon enrolling themselves, SHG members can decide the number of additional household members to 3 Specific illnesses may have lower coverage caps based on predefined categories of illness type. Relative to other micro health insurance plans, this limit is relatively generous. For example, VimoSEWA, a large micro insurer in India, has a limit of INR 2,000 6,000 (USD ) and RSBY (government insurance for BPL households) has a limit of INR 30,000 (USD 600) for the entire household (SEWA 2013, RSBY 2013a).

6 enroll. To the extent that SHGs are not sorted by baseline health status, this eligibility requirement reduces the likelihood of household characteristics being correlated with enrollment into the program. If all population heterogeneity was within SHGs, then the eligibility requirement would be the most effective in ensuring enrollment is uncorrelated with household characteristics. The eligibility requirement falls short of preventing household characteristics from being correlated with enrollment by three factors: heterogeneity across SHGs, members being free to choose additional household members to enroll, and requiring 80 percent compliance (as opposed to 100). III. Research Methodology and Data MHI Geographical Expansion and Randomized Offer The geographic expansion of DAN followed a randomized controlled trial methodology. In February 2011, half of the 43 villages in which Chaitanya was operational were randomly offered the MHI in the Junnar sub-district of Maharashtra. 4 The randomization was stratified upon three distinct areas, increasing rural, referred to as Block 1, 2, 3. In treatment villages, the initial enrollment into the MHI began in February 2011 in Block 1and the remaining two blocks in May In November 2012, the insurance product was made available to the remaining control villages. Data Sources The primary data source is an Endline Household Health Survey conducted in October 2012 on a randomly selected subsample of the population, approximately 18 to 21 months after the insurance was introduced in treatment villages. This survey was a detailed questionnaire on individual demographics and illnesses. Additionally, short health surveys were conducted during monthly SHG meetings from October 2011 to July These SHG Monthly Surveys asked 4 The randomization was originally done for 61 villages. However, in the early stages of the study it was realized that 18 of these villages were not operational and so were dropped from the study. These villages were equally assigned to treatment and control villages (see Appendix Table 1).

7 basic questions on household s rate of illness and health care utilization since the previous SHG meeting (i.e., a one month recall period). Though this provides panel data on health care use, it is limited to the unit of the household (rather than individual) and is dependent on whether the SHG meeting was held in the given month. Additionally, two pilot SHG Surveys were conducted in February and July The Appendix describes the data sources in more detail, including concerns and robustness to non-response. Enrollment, claims, and the insurance s doctor village visits are accessed from Chaitanya s internal records. Enrollment data is observed from January 2011 to August The data collected are midline and endline data for treatment villages, but are baseline information for control villages which were not yet offered DAN at the time of data collection. For all estimations, analysis is limited to SHG members at the start of the insurance offer, preventing estimates from being driven by the entry and exit of members. Enrollment and Demographic Table 1 describes the demographics of the SHG members and their household members in the research study 5. Typical of a non-profit MFI, a significant number of individuals belong to households below the poverty line, castes recognized as disadvantaged by the government, and participate as agricultural laborers for employment. Column 4 suggests there may be some preexisting differences between treatment and control villages. Generally, treatment households have higher socioeconomic status, though only disadvantaged caste is statistically different at the 5 percent level. Table 2 describes the health characteristics prior to the introduction of MHI. The observations are therefore limited to control villages, for which the data reflects characteristics prior to the offer of MHI. In the previous week, 12% of individuals experienced being ill, 3% 5 The demographic data was collected in the Endline Survey, but can be considered baseline characteristics for both treatment and control villages as these variables are unlikely to have been affected in the course of 20 months.

8 were admitted to health facilities, and average health expenditure was Rs. 95 (approximately USD 2). Table 3 documents the enrollment into the MHI for both treatment and control villages. In treatment villages, 68.34% of SHG members enrolled in the health insurance within the first 20 months of the program. At the time of the Endline Survey in October 2012, 51% were still enrolled in the program. In compliance with the randomized controlled trial design, no members in control villages had enrolled prior to the October After 20 months of being introduced to the program in control villages, 61% of SHG members had enrolled in the program for some time. Enrollment for additional household members is significantly lower. Conditional upon the SHG member enrolling in the MHI, 19% of non-shg household members enrolled sometime during the first 20 months, with 18% still being enrolled in October In control villages, the enrollment rate is similar at 16%. Figure 1 illustrates the number of additional household members enrolled in the program. In both treatment and control villages, the majority of SHG members (65%) chose to not enroll additional household members. Due to the pricing schedule of the MHI, the next most common option was to enroll two additional household members. In both treatment and control villages, conditional on SHG member enrollment, 91% of MFI members enrolled 2 or less additional household members. Renewal rates for MFI members are also similar in both treatment and control villages, 64% and 59% respectively. However, the renewal rates for additional household members are significantly different by treatment status, a mere 14% in control villages compared to 52% in treatment villages. The relatively high reenrollment rate among SHG members suggests there is robust demand for the product.

9 Compliance with the enrollment eligibility requirement, 80% of SHG members must enroll in the product, is generally followed. 71% of SHGs in control villages complied by having either no SHG members enroll or a minimum of 80% of SHG members enroll. The compliance was even higher in treatment villages (85%). Figure 2 displays the distribution of SHG member enrollment by SHGs. IV. Estimating Equations: Demand for Insurance: To estimate whether the demand for MHI is a function of baseline health status I observe the enrollment decisions of control villages that were offered the insurance after the completion of the randomized phase-in (i.e., after November 2012). For these control villages, I observe health data prior to the insurance offer and their enrollment decisions for 18 months after the MHI is made available. I compare individuals who will choose to enroll in the future relative to those who choose not to enroll: (1) Enrolled ihgvt = δ 1 HealthStatus ihgvt 1 + θ ihgvt where Enrolled is an indicator for whether the individual enrolled in DAN; HealthStatus is a proxy for the individual s health in the time prior to the introduction of MHI, and subscript i indicates the individual, subscript h indicates the household, subscript g indicates the SHG to which the household belongs, subscript v indicates the village, and subscript t indicates the time period where the MHI was offered to the household. I estimate Eqn (1) separately for SHG members and additional household members. To estimate how households selectively choose additional family members to enroll, I expand Eqn (1) to include household fixed effects for the additional household members.

10 The identifying assumption for consistently estimating δ 1 is that individuals are not altering their health care behaviors in anticipation of being offered and enrolling in MHI. This is unlikely as the expansion of the MHI required time and control villages would not have not known how long they would have to wait before enrolling in MHI. Specifically, the median coverage date started four months after the completion of the survey. MHI Effects The randomization of the insurance offer assists in estimating the casual effect of the health insurance offer in the community. Using the following equation, I estimate the effect of the insurance offer on illness, health care utilization, and health expenditures in the past week. (2) y ihgv = βtreamentvillage v + BlockFixedEffects v + ε ihgv where y is the outcome of interest, TreatmentVillage is an indicator of whether the household lives in a village that was offered DAN, BlockFixedEffects are indicators for whether the household lives in the area upon which the randomization was stratified. The identifying assumption for consistent estimates of β is that TreatmentVillage is not correlated with the error term, ε igv. The randomization of the MHI offer makes this a likely assumption, though Table 1 does raise some concerns of the validity of the assumption. For this reason, I also include baseline characteristics in the estimating equation and perform a series of robustness checks on the effect of the MHI offer using techniques that do not rely on this identification assumption. Difference-in-Difference Observing Future Enrollment By observing the future enrollment decisions for those in control villages, one can estimate the Treatment Effect on the Treated (TET) using a difference-in-difference technique with the enrolled households in the control and treatment villages.

11 (3) y ihgv = θ 1 TreamentVillage v + θ 2 Enrolled igv + θ 3 TreatmentVillage Enrolled igv + BlockFixedEffects v + ε igv where Enrolled is now an indicator of the individuals who became in the treatment and control villages. θ 1 is the average difference between unenrolled members in treatment and control villages (which may be due to pre-existing differences, externalities from the insurance program, and/or differences in the selection into the insurance by treatment status), θ 2 is a measure of the type of individuals who choose to enroll in the program, and θ 3 is the parameter of interest the effect of the insurance on those who enrolled (as opposed to the insurance offer). The identifying assumption for consistently estimating θ 3 as the TET is that demand for MHI is the same for both treatment and control villages; i.e., θ 2 accurately describes the selection process for both treatment and control villages. Unlike Eqn (2), estimating θ 3 does not require preexisting balance between treatment and control villages, though it does require preexisting differences to be the same across enrollment status. Time Variation and Additional Robustness Tests The varied timing of enrollments provides an additional opportunity for robustness analysis by including household fixed effects and estimating the effect of the treatment by comparing households before and after enrollment into MHI. (4) y hgvt = θ 3 Enrolled hgvt + α h + ε hgvt where α h are household fixed effects, and Enrolled is an indicator for whether any member in the household is enrolled in the given month. This is estimated using the monthly health surveys collected at the household level. Therefore, the analysis cannot be done at the individual level and ignores differences based on the number of household members enrolled. Additionally, the SHG monthly data is collected at a time when most treatment households have

12 already been enrolled into MHI, and thus the estimates are based on a subset of SHG members which enrolled late into the expansion of MHI. The identifying assumption is parallel time trends between those who insure relative to those who do not. One concern of violating this assumption is the timing of the enrollment may not be exogenous. For example, we may expect that households choose to become enrolled into the health insurance contract when they foresee health consumption in the near future, biasing θ 3 upwards. Table 4 shows the average duration between enrollment and the submission of the first claim for treatment households until October On average, households submitted a claim 7 months after being enrolled in DAN, suggesting that such endogenous timing of enrollment is not a concern. A series of additional robustness tests are conducted to estimate whether the above MHI effects are consistent. This includes trimming and controlling for potential pre-existing differences (both directly controlling for observable differences and using propensity score matching) between the treatment and control arms for the estimated β from Eqn (2). V. Results MHI Demand MHI demand is assessed by observing future enrollment decisions of control villages. Table 5, Panel A, reports differences among enrolled and non-enrolled SHG members in control villages on baseline demographic characteristics. There appears to be no statistical relationship between such characteristics as education, poverty status, or age in an SHG member s decision to enroll. Panel B assesses how results may differ if SHGs perfectly complied with eligibility requirements for MHI by making the following assumption all SHG members in an SHG with an observed enrollment of less than 50% are assumed to have not enrolled in MHI under perfect

13 compliance and all SHG members in an SHG with an observed enrollment over 50% are assumed to be enrolled in MHI under perfect compliance. As Panel B illustrates, even if insuring the SHG unit had been more enforced, it is unlikely that SHGs which chose to enroll would differ along these dimensions from non-enrolled SHGs. Unlike Table 5, Table 6 finds that demand for MHI is a function of baseline health individuals that enroll in MHI are more likely to experience illness at baseline than individuals that choose not to enroll. Though health expenditures and admit rates are similar across enrollment status, Panel A finds that SHG members that choose to enroll in the program are 9 percentage points more likely to have been ill and visit a doctor in the previous week. Interestingly, these magnitudes decrease to almost half when greater compliance with SHG level requirements are assumed (see Table 6, Panel B). This is suggestive that insuring at the SHG level reduces the relationship between baseline health and demand for MHI, though it still falls short of eliminating it. Panel C shows a similar relationship among non-shg household members, with enrolled individuals being approximately 9 percentage points more likely to have been ill in the preceding week but no statistically significant difference in admit rates and health expenditures. Surprisingly, Panel D finds that within households, it is not household members that are more ill that are more likely to be enrolled in the program. In other words, households are not choosing to enroll unhealthier family members, but those households that enroll multiple family members are unhealthier than those who do not. Conditional upon having enrolled, there is no pattern of renewal being a function of greater health needs at baseline. This may be due to the health variables not being a strong indicator of health at the time of renewal, or that the decision to drop out of MHI is not determined by health status as much as initial enrollment. Table 7 finds that the decision to

14 reenroll in MHI is not a function of baseline health for both SHG members (Panel A) and their additional household members (Panel B). The Effect of MHI on Health Care Table 7, Panel A, estimates the effect of the MHI offer on SHG members and finds no change in self-reported health, visiting health facilities, or health expenditure. Table 8, Panel A, estimates the effect for non-shg household members and does find an increase of 4 percentage points for self-reporting illness and 3 percentage point increase in visiting a doctor. The observation of enrollment choices after the Endline Survey provides the opportunity to compare the difference between enrolled and non-enrolled in treatment villages relative to control villages. Table 10 estimates a difference-in-difference on demographic characteristics among SHG members (Panel A) and non-shg household members (Panel B) to confirm selection is similar across treatment status. Most parameters are statistically insignificant, though it does suggest SHG members below the poverty line were less likely to enroll in the MHI in treatment villages. For non-shg household members, individuals with greater poverty, proxied by house infrastructure, and larger household sizes were more likely to enroll. The difference in enrollment rates and renewal rates for non-shg household members by treatment status suggests the identifying assumptions for Eqn (3) may not be valid among non-shg household members. For SHG members, though, the enrollment process seems relatively similar suggesting the difference-in-difference strategy outlined in Eqn (3) will provide consistent estimates of TET despite pre-existing differences between treatment and control villages. Table 8, Panel B, estimates that MHI reduced self-reported illness and visits to a doctor by 13 percentage points for SHG members. There continues to remain no impact on being admitted or health expenditure. Table 9, Panel B, estimates the same for non-shg household

15 members and does not find the dramatic reduction in weekly illness or doctor visits, but does find a reduction in reporting poor health that day. The coefficient on treatment village remain large and statistically significant, suggesting that if selection into MHI is the same by treatment status, this reflects either very large externalities to non-enrolled individuals or pre-existing differences by treatment status. Estimating TET Using Time Variation Table 11 uses the timing of the household enrollment, along with household fixed effects, to compare households before and after enrolling in MHI. The estimates suggest that even when comparing within households, households are less likely to be ill and have lower health expenditures after being enrolled in the program. This result is supportive of the primary findings in Table 8 and do not rely on identification assumptions requiring balance across treatment status. Preexisting Differences by Treatment Status One concern is that the estimated effect of the MHI offer is due to preexisting differences by treatment status and not the MHI offer. In addition to the slight imbalance by treatment status shown in Table 1, pilot SHG surveys conducted in February and July 2011 were also explored the concern of preexisting differences (see Appendix). These surveys recorded the proportion of the SHG that had experienced household illness in the past month, and prolonged bed rest or high health expenditure in the past three months. Unfortunately, this data has relatively low response-rate, a slight imbalance in the response rate by treatment status, and identification only at the SHG level (not at the household level). Also, these surveys were technically conducted after the start of the intervention though insurance coverage only began in February 2011, and enrollments had only minimally begun in Block 2 and 3 by July Nonetheless, the results of

16 these initial surveys are disconcerting as they report that SHGs in treatment villages had potentially lower levels of illness even prior to the insurance program. Due to these concerns of imbalance between treatment and control arms, the main tables in the remainder of the paper are expanded to control for baseline characteristics in the Appendix. Inclusion of these characteristics does not change the estimates presented in Tables 8 to 10 in the primary text. Survey Non-Response and Robustness to Lee Bounds: Survey non-response is also a primary concern for the consistent estimates of the insurance program. The response rate of the October 2012 Endline Survey was 80% in both treatment villages and control villages. Very few households refused consent and the majority of households not surveyed were due to relocation, which seems unlikely to be a result of the insurance offer. There is some additional non-response for the individual questions on the survey. One primary reason for the low-response stemmed from three villages which were experiencing difficulties with the MFI due to high defaults. This made it difficult for surveyors to contact households in these villages, and thus accounts for over 50% of the unknown nonresponse. In a small number of households, a shorter survey was implemented which asked basic health and expenditure questions. Unlike the Endline survey, the SHG Monthly Surveys suffered from even higher non-response rates and differential response rates by treatment status. However, all results are robust to the inclusion of Lee (2009) bounds (see Appendix) 6. VI. Conclusion The success and effectiveness of MHI depends critically on MHI demand and changes in health care demand conditional upon being insured. I find that bundling MHI with MFIs by 6 Bounding by the other method was also done at the quintile levels, but created bounds too large to have any meaningful contribution (not shown).

17 insuring an aggregated sub-unit, the SHG, may be a promising method to increase demand. Unlike Banerjee et al. (2014), I do not find that the members were so compelled to not purchase the insurance that there was no relationship between health needs and enrollment. In contrast, I find that those individuals with poor health are more likely to demand insurance, both among SHG members and non-shg household members. This suggests that SHGs are sorted along health characteristics and insuring at this aggregated unit falls short of eliminating concerns of adverse selection. However, it also provides evidence that those who are most likely to benefit from the MHI are being included in the program. DAN has been operational for over 4 years suggesting that the relationship between health needs and enrollment may not result in an unraveling of the MHI, though it does suggest that healthier individuals who may still benefit from insurance are opting out. Though adverse selection continues to be a threat, I fail to find convincing evidence of increases in health care utilization among enrolled members. Instead, I find limited suggestive evidence of the MHI reducing reported illness and demand for health care (though no effect on health expenditure). In general, the potential of MHI to increase health status and lower the amount of health care warrants further research. Numerous factors in the design of the MHI may be responsible for decreasing the barriers of access to health care and potentially reduced health shocks: direct price reductions, network facilities with quality checks, and local doctors being monitored. Further research is required to decipher which of these factors led to a decrease in reported illness and health care utilization and how these can be promoted and integrated into the designs of MHI programs.

18 References Arnott, Richard J., and Joseph E. Stiglitz The basic analytics of moral hazard. The Scandinavian Journal of Economics 90, (3, Information and Incentives. Vol. 1: Organizations and Markets) (Sep.): Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan The Miracle of Microfinance? Evidence from a Randomized Evaluation. Banerjee, Abhijit, Esther Duflo, and Richard Hornbeck Bundling Health Insurance and Microfinance in India. American Economic Review: Papers and Proceeding. 104(5):1 7. Chankova, S., Sulzbach, S., and Diop F Impact of mutual health organizations: evidence from West Africa. Health Policy and Planning. 23(4): Das, Jishnu, Jeffrey Hammer and Carolina Sanchez Paramo The Impact of Recall Periods on Reported Morbidity and Health Seeking Behavior. Policy Research Working Paper World Bank. Dupas, Pascaline Health Behaviors in Developing Countries. Prepared for the Annual Review of Economics, Vol 3. Dupas, Pascaline, and Jonathan Robinson Savings Constraints and Microenterprise Development: Evidence from a Field Experiment in Kenya. National Bureau of Economic Research Working Paper No Ekman, Bjorn Community-based health insurance in low-income countries: a systematic review of the evidence. Health Policy and Planning. 19(5): Jakab, Melitta and Chitra Krishnan Review of the Strengths and Weaknesses of Community Financing. In Health Financing for Poor People: Resource Mobilization and Risk Sharing, edited by A. S. Preker and G. Carrin. Washington D.C.: The World Bank. Jalan, Jyotsna and Martin Ravallion Are the poor less well insured? Evidence on vulnerability to income risk in rural China? Journal of Development Economics, 58: Jutting, J Do community-based health insurance schemes improve poor people s access to health care? Evidence from rural Senegal. World Development 32: Morduch, Jonathan Between the State and the Market: Can Informal Insurance Patch the Safety Net? World Bank Research Observer, 14(2): Morduch, Jonathan Micro-insurance: The Next Revolution? in Understanding Poverty, edited by Abhijit Banerjee, Roland Benabou, and Dilip Mookherjee. Oxford University Press. Townsend, Robert M Risk and insurance in village India. Econometrica 62, (3) (May): Wagstaff, Adam and Magnus Lindelow Can insurance increase financial risk? The curious case of health insurance in China. Journal of Health Economics, 27:

19 Figure 1: MHI Demand Among Non-SHG Household Members HH Members Demand Fraction of SHG Members Number of Additional HH Members Enrolled Figure 2: SHG Compliance with MHI Eligibility Requirement by Treatment Status (Left: Control Villages, Right: Treatment Villages) SHG Enrollment Compliance SHG Enrollment Compliance 0 1 Fraction of SHGs Proportion of SHG Members Enrolled

20 Table 1: Demographic Statistics and Balance All Villages Control Villages Treatment Villages Treatment - Control Female (0.500) (0.500) (0.500) (0.0227) Age (20.29) (20.52) (20.05) (1.129) Education (4.546) (4.447) (4.641) (0.347) Below the Poverty Line (0.491) (0.498) (0.478) (0.0448) Poor House (0.644) (0.648) (0.630) (0.0594) Disadvantaged Caste ** (0.499) (0.449) (0.478) (0.0876) Agricultural Cultivator (0.407) (0.412) (0.403) (0.0732) Agricultural Laborer (0.472) (0.456) (0.484) (0.0623) Household Size (3.250) (3.268) (3.230) (0.196) Notes: All variables are indicators, except Age (ranging from 0 to 101), Education (ranging from 0 to 15), Poor House (ranging from 1 to 3), and Household Size (ranging from 1 to 22). All ranges are identical across treatment status, except the maximum household size is 20 in control villages and 22 in treatment villages. Treatment - Control includes block area fixed effects with standard errors clustered at the village level. Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population. Means are reported with standard deviations and errors underneath in parentheses. Statistical significance levels are as follows: *10%, **5%, ***1%.

21 Table 2: Health Summary Statistic (Control Villages) Mean Obs (Ind) Poor Health (Today) Experienced Illness Visited Doctor Admited Health Expenditure Notes: All variables reflect a weekly recall period, except Health Today. All variables are indicators, except Poor Health (increasing in poor health from 1 to 4, with a standard deviation of.4762) and Health Expenditure (ranging from 0 to 10,150, with a standard deviation of 588.) Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population.

22 Table 3: MHI Enrollment Rates Control Villages Treatment Villages Panel A: SHG Members Enrolled 61% 74% Enrolled Before November % 68% Enrolled in October % 51% Renewal Rate 58.56% 64% Obs (SHG Members) Panel B: Additional Household Members (Conditional upon SHG Member Enrollment) Enrolled 15.64% 19.42% Enrolled Before October % 17.93% Enrolled in October % 13.73% Renewal Rate 13.51% 52.44% Obs (Non-SHG Members) Panel C: Household Enrollment Proportion of Additional Household Members Enrolled (0.3148) (.3277) Obs (Enrolled SHG Members) Notes: Enrolled is an indicator of having ever been enrolled in the MHI. Renewal rate is conditional on enrolled individuals with a coverage start date of June 1, 2013 or earlier (435 and 558 SHG Members and 293 and 380 Additional Household Members for control and treatment villages respectively.) Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population.

23 Table 4: Duration Between Enrollment and First Claim (Conditional upon Claim Submission) (1) (2) (3) (4) (5) Panel A: Summary Statistics Obs (HH) Mean SD Min Max Duration Between Enrollment and Panel B: Duration by Poverty Status Dependent Variable: Below the Poverty Line (BPL) (2.047) Constant 8.472*** (0.961) Obs (HH) 31 Notes: Observations are limited to treatment households which submitted a claim as of October Observations in Panel B are limited to treatment households selected and surveyed in the Household Health Survey and are weighted to be representative of the target population. Standard errors are in parentheses and are robust standard errors. Statistical significance levels are as follows: *10%, **5%, ***1%. Duration Between Enrollment and First Claim (Months)

24 Table 5: SHG Member Enrollment by Demographic Characteristics (1) (2) (3) (4) (5) (6) (7) (8) Age Education Below the Disadvantaged Agricutural Agricultural Household Poor House Poverty Line Caste Cultivator Laborer Size Panel A: Observed Compliance Enrolled (0.987) (0.358) (0.0592) (0.0659) (0.0656) (0.0543) (0.0455) (0.273) Constant 40.83*** 5.327*** 0.472*** 1.741*** 0.736*** 0.791*** 0.693*** 5.995*** (0.679) (0.254) (0.0446) (0.0475) (0.0559) (0.0460) (0.0363) (0.213) Panel B: Perfect Compliance Enrolled (1.032) (0.383) (0.0698) (0.0747) (0.0840) (0.0673) (0.0535) (0.313) Constant 40.51*** 5.278*** 0.501*** 1.726*** 0.787*** 0.793*** 0.719*** 6.229*** (0.784) (0.291) (0.0530) (0.0595) (0.0656) (0.0562) (0.0419) (0.248) Obs (SHG Members) Notes: Perfect compliance assumes all SHG members enrolled in the insurance if there is an observed enrollment rate of above 50% and no SHG members Observations are limited to households surveyed in the Household Health Survey and are weighted to be representative of the target population. Standard errors are in parentheses and are clustered at the SHG level. Statistical significance levels are as follows: *10%, **5%, ***1%.

25 Table 6: MHI Demand by Baseline Health (1) (2) (3) (4) (5) Poor Health Experienced Visited Health Admited (Today) Illness Doctor Expenditure Panel A: SHG Members (Observed Compliance) Enrolled * *** *** (0.0392) (0.0265) (0.0273) (0.0173) (45.34) Constant 1.245*** 0.165*** 0.141*** *** 122.7*** (0.0298) (0.0193) (0.0196) (0.0103) (34.60) Obs (SHG Members) Panel B: SHG Members (Perfect Compliance) Enrolled ** ** (0.0457) (0.0252) (0.0263) (0.0166) (49.44) Constant 1.290*** 0.181*** 0.154*** *** 153.0*** (0.0388) (0.0198) (0.0211) (0.0106) (43.01) Obs (SHG Members) Panel C: Additional Household Members: Enrolled 0.142*** *** *** (0.0317) (0.0192) (0.0164) ( ) (37.02) Constant 1.152*** *** *** *** 78.38*** (0.0103) ( ) ( ) ( ) (9.831) Obs (Additional HH Members) Panel D: Additional HH Members (HH Fixed Effects) Enrolled ** (0.0480) (0.0299) (0.0277) (0.0132) (60.46) Obs (Additional HH Members) Notes: Observations are limited to households surveyed in the Household Health Survey and are weighted to be representative of the target population. Standard errors are in parentheses and are clustered at the SHG level. Statistical significance levels are as follows: *10%, **5%, ***1%.

26 Table 7: MHI Renewal by Baseline Health (1) (2) (3) (4) (5) Poor Health Experienced Health Visited Doctor Admited (Today) Illness Expenditure Panel A: SHG Members Renew (0.0645) (0.0454) (0.0457) (0.0237) (60.99) Constant 1.316*** 0.220*** 0.202*** *** 132.7*** (0.0488) (0.0363) (0.0374) (0.0135) (33.15) Obs (SHG Members) Panel B: Non-SHG Household Members Renew (0.135) (0.0888) (0.0601) (0.0452) (278.6) Constant 1.309*** 0.197*** 0.171*** *** 127.3*** (0.0388) (0.0230) (0.0200) ( ) (32.70) Obs (Non-SHG HH Members) Notes: Observations are limited to households surveyed in the Household Health Survey and are weighted to be representative of the target population. Observations are conditional upon enrollment prior to June 1, Standard errors are in parentheses and are clustered at the SHG level. Statistical significance levels are as follows: *10%, **5%, ***1%.

27 Table 8: The Effect of MHI on Health Care for SHG Members (1) (2) (3) (4) (5) Poor Health Experienced Health Visited Doctor Admited (Today) Illness Expenditure Panel A: MHI Offer (ITT) Treatment Villages (0.0363) (0.0244) (0.0244) (0.0131) (0.0319) Obs (SHG Members) Panel B: MHI Effect (TET) Treatment Villages * (0.0583) (0.0342) (0.0362) (0.0191) (0.0515) Enrolled ** *** *** (0.0350) (0.0275) (0.0252) (0.0159) (0.0451) Treatment * Enrolled *** *** (0.0635) (0.0411) (0.0399) (0.0248) (0.0669) Obs (SHG Members) Notes: Regressions include block fixed effects and standard errors are clustered at the village level. Enrolled is an indicator of having ever been enrolled in the MHI. Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population. Statistical significance levels are as follows: *10%, **5%, ***1%.

28 Table 9: The Effect of MHI on Health Care for non-shg Household Members (1) (2) (3) (4) (5) Poor Health Experienced Health Visited Doctor Admited (Today) Illness Expenditure Panel A: MHI Offer (ITT) Treatment Villages *** ** (0.0172) (0.0128) ( ) ( ) (0.0145) Obs (Non-SHG HH Members) Panel B: MHI Effect (TET) Treatment Villages ** ** (0.0185) (0.0156) (0.0118) (0.0109) (0.0172) Enrolled 0.137*** *** *** (0.0309) (0.0234) (0.0181) ( ) (0.0415) Treatment * Enrolled *** (0.0433) (0.0432) (0.0354) (0.0197) (0.0557) Obs (Non-SHG HH Members) Notes: Regressions include block fixed effects and standard errors are clustered at the village level. Enrolled is an indicator of having ever been enrolled in the MHI. Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population. Statistical significance levels are as follows: *10%, **5%, ***1%.

29 Table 10: Enrollment Demand Across Treatment Status (1) (2) (3) (4) (5) (6) (7) (8) Age Education Below the Poverty Poor House Disadvantaged Agricutural Agricultural Household Caste Cultivator Laborer Size Line Panel A: SHG Members Treatment Villages ** (1.480) (0.529) (0.0592) (0.0944) (0.0860) (0.0797) (0.0781) (0.316) Enrolled (0.951) (0.448) (0.0367) (0.0574) (0.0855) (0.0707) (0.0390) (0.259) Treatment * Enrolled * (1.775) (0.612) (0.0538) (0.0921) (0.106) (0.0871) (0.0604) (0.401) Obs (SHG Members) Panel B: Non-SHG Members Treatment Villages *** (0.920) (0.261) (0.0462) (0.0594) (0.0843) (0.0759) (0.0624) (0.355) Enrolled ** *** (2.357) (0.385) (0.0445) (0.0550) (0.0936) (0.0568) (0.0679) (0.410) Treatment * Enrolled * * (2.625) (0.471) (0.0598) (0.0797) (0.105) (0.0949) (0.0870) (0.538) Obs (Non-SHG HH Members) Notes: Regressions include block fixed effects and standard errors are clustered at the village level. Enrolled is an indicator of having ever been enrolled in the MHI. Observations are limited to households surveyed in the Endline Health Survey and are weighted to be representative of the target population. Statistical significance levels are as follows: *10%, **5%, ***1%.

30 Table 11: Enrollment Time Variation (1) (2) (3) Dependent Variable: Illness Health Shock Health Expenditure Enrolled HH ** *** *** (0.0329) (0.0170) (111.3) Obs (HH Month) Notes: Regressions include household fixed effects with standard errors clustered at the household level. Dependent variables are self-reported monthly recall for whether any member in the household experienced the event (illness, health shock), and the total health expenditure for the household in the past month. Health shock is an indicator for being admited to a health facility or being on prolonged bedrest. Statistical significance levels are as follows: *10%, **5%, ***1%.