94958 The Impact of RSBY on Hospital Utilization and Out-of-Pocket Health Expenditure Douglas Johnson and Karuna Krishnaswamy December 2012 Abstract In 2008, India launched Rashtriya Swasthya Bima Yojana (RSBY), a government subsidized health insurance scheme for the poor that covers secondary hospital care for Below Poverty Line (BPL) households. Designed to improve access to healthcare and reduce the financial burden of healthcare expenses, RSBY currently covers 32 million households. We exploit the phased rollout of RSBY to estimate the impact of the scheme in its early days on hospitalization rates and out-of-pocket health expenditure using a difference-in-differences approach with matching. We use secondary data from the National Sample Survey Organization conducted in 2009-10. We find that the scheme has led to a small decrease in out-of-pocket household outpatient expenditure and consequently total medical expenditure. We also find limited evidence of increase in the number of households that have had a hospitalization case though there are regional variations in the nature of impact. It is unclear whether the improved access to inpatient care is reducing outpatient expenses through decreased need for outpatient care or because some people or hospitals are colluding to convert outpatient treatment to inpatient to avail of the scheme. We suggest that additional evidence with a more recent data set allowing for more time for RSBY to display effects is needed to strengthen these early findings. Contacting emails: karuna.krishnaswamy@gmail.com and doug892@gmail.com. This study was made possible by funding and support from the World Bank. The authors thank Changqing Sun, Robert Palacios, Maria Jos and Ajay Mahal for helpful suggestions. All errors are our own. I - Introduction India’s health market is characterized by poor access and outcomes, low levels public health spending, and high levels of out-of-pocket (OOP) private spending. While the public health system in India is, in theory, accessible to all, use of the private sector for healthcare, even among the poor is high. India has one of the highest rates of out-of-pocket health spending in the world at 78% of total health spending, and 94% of all private health spending (Rao et al, 2005). High OOP spending on health implies that access is limited to those who can afford it. As well, low income households which experience ill health face considerable expenses. Hospitalization expenses, in particular, cause considerable financial burden. Twenty four percent of all Indians who are hospitalized each year fall below the poverty line due to the hospitalization (Peters et al, 2002). In 2008, the government of India launched Rashtriya Swasthya Bima Yojana (RSBY), the country’s first centrally funded, publicly subsidized health insurance program. The objective of RSBY is to increase access to medical services among the poor and shield them from large hospitalization bills. As of February 2012, 27 million households in 24 states (396 districts) were enrolled in RSBY and 3.1 million hospitalization cases had been covered under the scheme1. In this paper, we employ difference in differences to estimate the impact of RSBY on hospitalization and OOP health spending using data from the National Sample Survey Organization from 2004-05 and 2009-10. We find that the scheme has led to a small decrease in outpatient and total medical expenditure of target households and find some limited evidence of increased hospital utilization rates. This paper is organized as follows. Section 2 describes the scheme in brief. Section 3 reviews the literature on the impact of health insurance schemes in India. Section 4 describes the data source. Section 5 presents the empirical strategy used. Section 6 presents and discusses the results. Section 7 concludes. 1 www.rsby.gov.in II - RSBY RSBY provides cashless coverage of up to Rs. 30,000 ($600 2 ) each year to each enrolled household for hospitalization procedures at any hospital, public or private, empanelled under the scheme. While almost all procedures (both surgical and non-surgical) are covered, the scheme does not cover all diagnostic tests or all pre and post operative care. Rs. 100 ($2) is provided per visit to a hospital as transportation allowance. RSBY does not cover any outpatient procedures. All households officially designated as “Below Poverty Line” (BPL) by their respective state government are eligible for the scheme but must pay Rs. 30 for each year of coverage3. Enrolled households may nominate up to five members to be covered under RSBY. Implementation of RSBY is contracted out to insurance companies at the district level in a public-private partnership model. In each district where RSBY is implemented, each year, a single insurance company is responsible for enrolling households and recruiting hospitals to participate in the scheme. Enrolments of identified BPL households are conducted in enrolment camps. Households are given invitation slips and awareness campaigns are conducted inter alia through the Gram Panchayat. In the first year 43% of BPL households in those districts where the scheme was implemented enrolled in the scheme (Krishnaswamy and Ruchismita, 2011). Households which choose to enroll in the scheme receive a smart card which they may use to receive care at any participating hospital across the country. Insurance companies are paid a fixed price per household enrolled and must settle all claims with the hospitals directly based on rates fixed by the central government. For further details of the scheme, see Hou and Palacios (2010) and Krishnaswamy and Ruchismita (2011). III - Existing Evidence of the Impact of Health Insurance While there is a body of literature on the impact of health insurance globally, overall the findings show mixed and limited evidence of impact of health insurance. We summarize below three 2At $1 = Rs. 50 3This Rs. 30 represents only a small fraction of the total cost of the program. The remainder of the premium, typically between Rs. 400 and 650 is paid in part (75%) by the central government and in part (25%) by the respective state government. papers that study three schemes in India. Aggarwal (2010) evaluates India’s Yeshasvini community-based health insurance programme which covers highly catastrophic and less discretionary in-patient surgical procedures and offers free outpatient diagnostics and lab tests at discounted rates. The study uses a one-time survey of 4109 households and employed propensity score matching to identify suitable control households. They find that Yeshasvini increased the use of health services, reduced OOP spending and improved health outcomes. Fan, Karan and Mahal (2012), evaluate the impact of Arogyashri, the state of Andhra Pradesh’s health insurance scheme using a difference in differences strategy. They find that Arogyashri significantly reduced OOP inpatient expenditures and, to a lesser extent, outpatient expenditures, while they find no impact on catastrophic health expenses. Karan and Selvaraj (2012) use NSSO data to evaluate the impact of RSBY. However their study suffers from several methodological shortcomings. First, they aggregate data at the district level which significantly reduces the power of their analysis to detect true impact and complicates interpretation of their effects. Second, they fail to provide standard errors of their estimates preventing the reader from knowing whether the results are statistically significant. Third, and most importantly, they appear to misinterpret their own findings – despite showing large positive impacts of RSBY they claim that their analysis proves RSBY to be ineffective. Devadasan et al. (2010) evaluate a Community Based Health Insurance scheme, the ACCORD-AMSASHWINI scheme. 297 insured with matched 248 uninsured were observed for a year and interviewed when faced with an ailment. The authors find that insured patients had a hospital admission rate 2.2 times higher than uninsured patients. Potential Impact of RSBY RSBY is somewhat unique in that only hospitalization expenses are covered. While conceding that the scheme was in its early years at time of data collection and hence is not likely to reveal its long term impact, we would expect the following. We expect that RSBY should increase utilization of hospitals by BPL households that may have otherwise postponed non-critical procedures because they could not afford it. However, as Finkelstein et al (2012)4 and others have pointed out, the potential impact of health insurance on 4 http://economics.mit.edu/files/6796 OOP spending is not always intuitively obvious. Given that there may still be OOP payments needed to be made by the patient on drugs and tests and post procedure care, it is possible that increased hospitalization rates may increase OOP spending on total inpatient expenses overall, although the premium is 95% subsidized. Hence the sign of the impact of inpatient expenses is unclear. Given the coverage of Rs. 30,000 per year, while there may be some increase in low levels of inpatient expenses, the number of households with high amounts of OOP inpatient spending should fall. IV - Data Our data on consumer health expenditure comes from the National Sample Statistics Organization’s (NSSO) survey data. We use NSSO round 61 (conducted in 2004 -05) and round 66 (conducted in 2009-10), as the pre and post surveys for measuring the potential impact of RSBY. The household is the unit of observation. These two quinquennial surveys are used since they are twice the sample size of the annual surveys, thereby making district level analysis more meaningful by reducing sampling errors. The sampling design used in these surveys used stratified multi-stage random sampling with villages in rural and Urban Frame Survey Blocks in urban areas as first stage units and households as the second stage units. We also use round 55 (1999-2000) to compare historical trends between treatment and control groups. The survey instruments of NSSO rounds 61 and 66 contain detailed data on consumer expenditure, including spending on healthcare broken up by whether the spending was on outpatient or inpatient care. Inpatient healthcare expenditure data uses a one year recall period, while outpatient expenses are for a 30 day recall period. We divide healthcare expenditure by household size and adjust for inflation using the consumer price index5 to generate variables for inflation adjusted per capita OOP spending on inpatient and outpatient healthcare. We generate a dummy variable for whether any member of a household was hospitalized by looking at whether the household incurred any inpatient healthcare expenditure at all. We note here that even if many members of a household were to be hospitalized, the indicator would still be one, thereby understating the impact of the scheme on hospital utilization. 5 We use the rural labour CPI for rural areas and the non-manual CPI for urban areas. Our base year is 1999-2000. Although inflation rates vary by state, we only adjust for inflation nationally. Except in Kerala and Himachal Pradesh, only households officially designated as BPL were eligible for RSBY in the early years of the scheme. BPL status is included in NSSO round 61 data (our pre-round) but not in round 66 data (our post round). We use round 61 data to fit a logit model estimating the probability a household has BPL status based on total household expenditure, state of residence, whether the household lived in a rural or urban area, number of kilograms of rice and atta (wheat flour) purchased from the ration shop, religion, caste, household size, highest educational attainment of any member of the household, amount of money spent on food per capita, number of meals eaten in school, and cooking and lighting facilities used6. We then use coefficients from this fitted model to estimate the probability a household included in round 66 data has BPL status. We are not aware of any changes to the government’s BPL identification strategy between rounds 61 and 66. Hence, we believe, the forecast model for round 61 is applicable in round 66. For both rounds, we tag households as BPL if the predicted probability of BPL status from the model is above 50%. This threshold is selected based on the simulation results from in-sample tests to minimize the incidence of incorrectly marking a non BPL holder as BPL and maximize the rate of identifying correctly BPL card holders. For the round 61 data, our model has a high rate of identification of true positives (83% out of 18,000 BPL households were correctly identified) but also a high rate of false positives (50% of non-BPL households were incorrectly identified as BPL). Since this would only understate the impact results since non-BPL households are not eligible to receive RSBY in all but two states, it does not take away from the results. We use administrative data from RSBY on policy coverage dates by district in order to identify the treatment and control districts as described in detail in the next section. Finally, the District Level Household Survey (DLHS, 2005-06) and Census 2001 data are used for generating district level control and matching variables. Determining treatment status Determining treatment status of a household is tricky since in some cases, the NSSO survey was being carried out at around the same time that RSBY was being rolled out. Round 66 was conducted between July 2009 and June 2010 while RSBY was launched in different districts at different times. Typically the insurance enrolments took about three to four months to complete. 6Despite the stated objective of the government to grant BPL status only to poor households, household consumption expenditure along is a poor predictor of BPL status. We hence define treatment status as follows. A household is deemed treated if the policy start date in that district was two month prior to the date of the interview in order to give the household sufficient time to undergo a procedure. The cost of a longer treatment period is that fewer districts are available to be included in the treatment group reducing the size of the treated group. We drop three states, Andhra Pradesh, Karnataka and Tamil Nadu since these states have their own state health financing schemes similar to RSBY (Arogyashri in Andhra Pradesh, Yeshasvini in Karnataka and the Chief Minister’s Health Insurance scheme in Tamil Nadu). Further, two of these schemes, Arogyashri and the Chief Minister’s Health Insurance scheme were launched between our pre and post survey rounds. Identifying control districts is tricky since RSBY treatment was not assigned randomly. We conduct falsifications tests on the non-treatment districts and find that the outcome variables exhibit time trends in the two surveys prior to the introduction of RSBY (explained further in the Robustness Checks section). We divide the non-treatment districts into two categories. Our first category (controls 1) consists of those districts where RSBY was planned (and an insurer identified), but not launched at the time of the survey. These districts also represent states where some districts had been treated at the time of the survey while the remaining had not. Our second category (control 2) consists of districts where RSBY was not planned at the time. The Control 1 districts are similar to the treatment districts in terms of district level covariates that are likely to influence selection (and order of selection) of RSBY such as the medical infrastructure in the district such as availability of hospitals with sufficient number of beds and good quality, and administrative capacity at the state and Panchayat levels, crime, overall health insurance penetration and distance to roads, while the Control 2 districts are substantially different. Table 1 presents district covariates calculated from DLHS and Census 2001 datasets. Table 1: Means of district covariates by group Treatment Control 1 Control 2 Percentage with Health Insurance 1.156 1.116 7.237 Percentage with Government hospital 42.51 54.77 77.21 No. villages with med facility 338.8 308.4 250.0 No. times GP met per year 3.319 3.100 3.202 No. beds per district 50.38 40.88 39.87 No. Serious Crimes 931.3 725.6 683.4 Distance to nearest town 12.97 16.70 17.16 Sec. care in village 6.456 4.987 7.208 Tert. care in village 13.53 7.176 10.01 Allopathy hospitals in village 24.79 17.11 8.288 Hospital quality index 259.3 221.3 213.0 The above district level characteristics that reflect the possible reasons why the districts were selected in the order that they were, such as, need for the scheme, ease of implementation, etc., may also be correlated to outcomes. Hence, we construct a set of treatment and control groups (from among all non-treatment districts) by matching on district level covariates that are likely to be correlated with the probability of being selected an RSBY district. Finally, we have the difficulty that 90 districts in round 55 (20 in round 61) split into two or more smaller districts in the post round 61 (in round 66). These split districts end up having different national identification codes – one of them takes the parent district’s code while the remainder were given new codes. In order to retain comparability, we assigned all the post district codes into the parent code. Our sample We have 297 control and 204 treatment districts with a total of 186,065 households. Out of these, 102,810 are from the PRE intervention round and 83,255 from the POST round (Table 2a). Table 2a: RSBY Rollout PRE POST RSBY implemented at the time of survey (Treatment) 57,813 46,446 RSBY planned and not planned during survey (Control) 44,997 36,809 102,810 83,255 The RSBY policy date starts on the same day for all enrollees in a district. The earliest RSBY rollout in a treatment district was May 2008 and the latest was March 2010. Hence, there is a wide range of duration of treatment, the maximum being 25 months and the average being about 6 months. Out of the 83,255 households in the POST round (66) observations, 25,548 households were surveyed two months after RSBY was introduced and hence treated. Out of these, 12,995 were predicted to be a BPL card holder and hence in effect the treated sub-sample (Table 2b). Table 3b: No. of households in the post round Non-BPL BPL Treated (2 months of RSBY prior to survey) 12,553 12,995 Control 29,051 28,656 41,604 41,651 V - Empirical Strategy The household level difference-in-differences (DiD) estimator is our canonical analysis. This analysis is restricted to the BPL sub-sample of our data. The identifying assumption of this model is that, in the absence of RSBY, the change in mean outcomes for the treatment group would have been the same as the change in mean outcomes for the control group. This is commonly interpreted as implying that outcomes are additive in a time component, a group component, and a constant treatment effect though difference in differences does allow for heterogeneous treatment effects as long as individual time specific shocks are not correlated with treatment status. We estimate an equation of the form: Yit = Di + β1 ∗ POSTt + β2 ∗ POSTt ∗ Tit + ∗ Xit + τst + εit (1) Yit is an outcome measure that RSBY is likely to influence such as total OOP spending on healthcare, OOP spending on outpatient and in-patient services by type – fees, drugs and tests and total, share of population with “catastrophic” levels of spending on inpatient services (defined as greater than 10%, 20% and 40% of total spending or greater than Rs. 5,000 and Rs. 10,000), and household hospital utilization rate which is a dummy equal to one if there was a positive inpatient spending by the household in a one year recall period, and finally the share of OOP spending on in-patient services and total health care out of total household expenditure. Di is a district fixed effect since RSBY rollout is at the district level. POSTt is a time dummy equal to 1 (post RSBY) or zero (pre RSBY) Tit is a treatment dummy equal to one if the observation came from a household in an RSBY district three months after RSBY was implemented Xit is a vector of household covariates used for improving precision, such as season of survey date, monthly per capita income, cooking and lighting type, caste, religion, maximum household educational attainment, urban or rural, average household size, age, landownership. τst is a state and time level error term εit is a household level error term Standard errors are clustered by state-time cell (Mullainathan, Bertrand and Duflo, 2004). Survey weights are used. Our estimator of interest is β2 . Triple Differences We may improve our estimates by performing triple differences using non-BPL as a second control. Triple differences allows for endogeneity in the selection of districts as long as selection is not correlated with change in outcomes over time between BPL and non-BPL in a district. In other words, triple differences allows for changes in outcomes between pre and post rounds to differ across RSBY and non-RSBY districts as long as non-BPL observations experience similar trends to BPL observations. We believe this is more credible since district characteristics have played a part in the order in which RSBY districts were selected and RSBY’s key performance indicators – borne out by interviews with RSBY stakeholders and by Krishnaswamy and Ruchismita (2011). We estimate an equation of the form: Yit = Di + POSTt ∗ (β1 + β2 ∗ Tit ) + β3 ISBPL + ISBPL ∗ POSTt ∗ (β4 + β5 ∗ Tit ) + Xit + τst + εit (2) Yit is now the household level outcome in each period for both BPL or non-BPL. ISBPL is a dummy which equals 1 for a BPL observation and 0 for non-BPL. The other terms are as before. Our estimator of interest is β5 . Treatment Intensity In addition to estimating the overall impact of RSBY on target households, we also estimate how the length of time a household has been covered by RSBY prior to the date of survey affects outcomes. The number of days elapsed between the start of the policy and the date of the NSSO survey in a district varies substantially across households since each survey round took a year to complete. The longer the length of time between the start of the RSBY policy and the survey date, the higher the likelihood that someone may have had need for hospitalization and utilized RSBY for the same. There are several other reasons why the impact of RSBY may vary with the time since implementation. First, in many districts, awareness of RSBY is extremely low in the first few months after initial rollout despite households being given smartcards. (Johnson and Kumar, 2011 find that a large proportion of enrolled households are only vaguely aware of the purpose of the smartcard.) Second, hospitals make take time to understand and become familiar with the scheme. We estimate the impact of the duration of RSBY implementation on outcomes using a modified difference in difference approach similar to Duflo (2001). We first define “treatment intensity” as the number of months elapsed between the launch of RSBY in a district and the date of the survey and hence the amount of time available to use the scheme before the survey. The treatment intensity varies between the households within a district and between different districts. We then estimate the impact of treatment intensity on outcomes using the following regression borrowed from Duflo (2001): Yit = Di + POSTt ∗ β1 + MTHS_RSBYt ∗ β2 ∗ Post it + Xit + τst + εit (3) Here MTHS_RSBY is the number of months elapsed between the launch of RSBY in the district and the date of the survey and β2 is the coefficient of interest indicating the impact of each additional month of treatment on outcome measures. Matching Basic difference-in-differences, without the inclusion of any covariates, requires a strong identifying assumption in order for results to represent true causal impact. We may weaken the identifying assumption required for difference-in-differences by allowing potential outcomes to vary by observable characteristics of households. This is the motivation for including covariates in our difference-in-difference specification above. This specification allows potential outcomes to vary linearly with observable characteristics of households. Matching may allow us to further weaken our identifying assumption by allowing potential outcomes to vary in an unspecified way with household characteristics. Matching attempts to select a subset of households from the treatment and control groups with similar characteristics in each time period. The use of matching combined with difference-in-differences for repeated cross section data was first proposed by Heckman et al (1997). Ho et al (2007) show that matching may be viewed as a type of non-parametric pre-processing of data and thus may be combined with a variety of parametric models for estimating causal impact. Matching has been used in many health insurance impact studies such as Wagstaff et al (2009). There is now a body of literature on matching and a variety of approaches to use in performing matching. We use the coarsened exact matching approach developed by Iacus, King, and Porro (2011). In this method, covariates on which treatment and control units are to be matched are first “coarsened” by grouping values into distinct categories and treatment and control are then matched exactly based on the coarsened values of these covariates. The purpose of the first step is to allow for exact matching in cases of where the dimensionality of the covariate space may be extremely high such as if some of the covariates are continuous. Despite its simplicity, Iacus, King, and Porro (2011) provide evidence that coarsened exact matching outperforms propensity score matching and other common techniques in many applications. As mentioned earlier, we also use matching to select a set of control districts from the non- treatment districts. Table 3a presents the means of the covariates before and after matching. Table 3a: Means of district covariates before and after matching Unmatched Matched Control Treatment Control Treatment Percentage with Health 4.348 1.149 1.695 0.926 Insurance Percentage with 67.55 42.30 40.79 40.87 Government hospital No. villages with med 275.9 338.8 277.5 326.4 facility No. times GP met per 3.086 3.303 2.994 3.326 year No. beds per district 39.39 50.19 26.83 33.72 No. Serious Crimes 698.7 931.3 667.9 796.7 Distance to nearest town 16.94 12.92 14.56 13.80 Allopathy hospitals in 12.29 24.79 15.43 18.31 village Hospital quality index 213.4 257.4 179.1 219.1 Sec. care in village 6.208 6.392 3.808 4.559 Tert. care in village 8.695 13.44 5.771 6.715 This matching process dropped 79 control districts and 46 treatment districts leaving us with 376 districts. Table 3b shows the number of household observations left after district level matching. Table 3b: No. of observations before and after district matching PRE POST Matched Unmatched Matched Unmatched Non-BPL 17,375 33,821 14,085 27,519 BPL 16,335 35,278 13,148 28,503 Next we selected a subset of households matching by households covariates that were chosen based on our subjective judgment of their relevance. Tables 4 display average values of household covariates for treatment and control groups in our pre and post surveys before and after matching. Although only 320 observations were dropped unmatched, since the weights change after matching, the means of the household covariates are better lined up after matching. Table 4: Household covariates by group Baseline Endline Matched Unmatched Matched Unmatched Monthly Per Capita Expenditure 805.63 768.25 811.93 708.46 1372.32 1291.37 1406.13 1182.82 [859.97] [759.14] [959.79] [1232.03] [1635.25] [1425.32] [2856.39] [1301.79] Religion 1.3452 1.2806 1.3557 1.2785 1.3483 1.2829 1.3653 1.2594 [.9105] [.7648] [.9277] [.782] [.9299] [.7675] [.958] [.736] Caste 4.6837 4.6487 4.6861 4.8088 4.584 4.6973 4.5758 4.8635 [3.1158] [3.1144] [3.1169] [3.2488] [3.0947] [3.1352] [3.0923] [3.2673] Education 5.3495 5.2469 5.3578 4.8853 7.7622 7.6242 7.7589 7.2554 [2.6028] [2.5623] [2.6067] [2.5164] [3.0576] [3.0196] [3.066] [2.9202] PDS item purchased 1 1.0699 0.9143 1.1273 0.965 2.8465 2.0838 2.983 2.0863 [4.469] [4.0122] [4.728] [4.8355] [6.2663] [5.4901] [6.7409] [5.3477] PDS item purchased 2 2.1286 2.6084 2.1412 2.6384 1.8967 2.1013 1.932 2.3503 [2.4822] [2.7678] [2.7031] [2.6487] [1.8859] [2.059] [2.0503] [2.2363] PDS item purchased 3 1.6621 1.0379 1.7937 1.1305 4.1791 2.6172 4.3075 2.8921 [6.164] [4.64] [7.0009] [5.1284] [8.6189] [6.9015] [9.1634] [7.1249] Household size 4.9303 4.9186 4.9284 4.8771 4.6938 4.6841 4.7315 4.6882 [2.4906] [2.4962] [2.5078] [2.4387] [2.3202] [2.2766] [2.3969] [2.288] Is rural 0.7224 0.7237 0.7214 0.7735 0.7039 0.7126 0.7033 0.7667 [.4478] [.4472] [.4483] [.4186] [.4565] [.4525] [.4568] [.4229] Observations 44503 27588 44997 28113 35459 21890 36274 22326 Standard deviations in brackets VII - Results Our main results are based on the matched household data. Table 5 shows the base regression output for the unmatched household observations, while Table 6 shows the duration of treatment model output/ Tables 7 and 8 show the same output with matched households. Tables 9 and 10 show the matched output excluding Uttar Pradesh and Haryana. Our first observation is that both the double and triple differences show a consistent significant negative impact on per capita outpatient and total medical expenses in the basic data, and in the matched household data. In the duration of treatment model, we see a consistent significant negative impact on per capita outpatient and total medical expenses in the double differences model, and partly supported by the triple differences model. As regards hospital utilization, we see a consistent positive impact in all the triple differences models, although it is insignificant in the double differences models. Hence, we re-ran regressions with regional sub-samples dividing the country into four zones. The North zone appears to run contrary to the rest of the country in this outcome variable with a negative impact. We find a positive significant impact in the double differences if we drop Uttar Pradesh and Haryana. However, in this sub-sample analysis, outpatient expenses lose some significance. We do not find any impact of inpatient expenses or components of outpatient and inpatient such as fees, drugs and tests. Appendix C shows the regression outputs without matching on district covariates. We find significant effect on outpatient expenses in the double differences and on hospital utilization on the triple difference estimates. In other regression specifications, we experimented with labeling a household as treated if it had even fewer than two months of treatment. The main results hold up. However, we find a significant positive impact on the number of households with inpatient expenses of Rs. 5,000 and more and for Rs. 10,000 and more with 15 days of treatment. However, we do not believe that a gap of 15 days between the start of the policy period and the date of the survey is sufficient time for utilization and since these results are not consistent as we increase treatment intensity, we do not report these results (available upon request). Table 4: Impact of RSBY (without household matching) (1) (2) (3) (4) (5) (6) (7) (8) (9) (11) (10) (12) (13) OP Exp IP Exp. Total IP drug IP fees. IP hosp. Was Has OP IP > Rs. Ratio IP/ IP > Rs. Ratio Ratio (Rs.) (Rs.) Medical + tests (Rs.) fees. hospitalized visit 5000 HHD10,000 IP/ IP/ Exp. (Rs.) (Rs.) (Rs.) Exp > (Rs.) HHD HHD (Rs.) 10% Exp > Exp > 20% 40% Triple - -8.938 -13.42** -4.697 -1.774 -1.585 0.0249** 0.0147 - - 0.00537 - - differences 4.478** 0.00475 0.00598 0.00182 0.00472** (0.049) (0.104) (0.046) (0.116) (0.343) (0.166) (0.018) (0.523) (0.363) (0.142) (0.334) (0.611) (0.011) Observations 124957 124957 124957 124957 124957 124957 124957 124957 124957 124957 124957 124957 124957 Adjusted R2 0.071 0.017 0.050 0.020 0.006 0.008 0.040 0.085 0.023 0.017 0.017 0.015 0.012 Difference - 1.106 -3.610** 0.384 -0.0769 0.387 0.0157 -0.0105 0.00436 0.00407* 0.00640 0.00252 0.00135 in 4.716*** differences (0.001) (0.461) (0.025) (0.538) (0.866) (0.341) (0.473) (0.517) (0.260) (0.099) (0.341) (0.531) (0.478) Observations 63676 63676 63676 63676 63676 63676 63676 63676 63676 63676 63676 63676 63676 Adjusted R2 0.065 0.015 0.046 0.013 0.019 0.004 0.050 0.093 0.025 0.019 0.026 0.024 0.019 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month Table 5: Duration of treatment model (without household matching) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) OP Exp IP Exp. Total IP drug IP IP hosp. Was Has OP IP > Rs. IP > Rs. Ratio IP/ Ratio Rati (Rs.) (Rs.) Medical + tests fees. fees. (Rs.) hospitaliz visit 5000 10,000 HHD IP/ o Exp. (Rs.) (Rs.) (Rs.) ed (Rs.) (Rs.) Exp > HHD IP/ 10% Exp > HH 20% D Exp > 40 % Triple -0.230 -0.811* - -0.452* -0.107 -0.174** 0.00299** 0.00229 - - 0.000193 - - * Differences 1.04 0.00057 0.00080 0.00025 0.000564*** 1* 1 9* 0 (0.357) (0.066) (0.07 (0.080) (0.487 (0.042) (0.006) (0.261) (0.235) (0.051) (0.726) (0.528) (0.001) 5) ) Observation 124957 124957 1249 124957 12495 124957 124957 124957 124957 124957 124957 124957 124957 s 57 7 Adjusted R2 0.071 0.017 0.05 0.020 0.006 0.008 0.040 0.084 0.023 0.017 0.017 0.015 0.012 0 DiD -0.280** - -0.282* -0.0114 -0.000743 0.022 0.000672 0.00023 0.00019 0.00016 0.000018 0.0000352 - 0.002 1 1 9 4 6 0.0001 77 06 (0.033) (0.98 (0.076) (0.845) (0.984) (0.51 (0.720) (0.891) (0.571) (0.513) (0.975) (0.932) (0.566 4) 3) ) Observati 63676 63676 63676 63676 63676 6367 63676 63676 63676 63676 63676 63676 63676 ons 6 Adjusted 0.065 0.015 0.046 0.013 0.019 0.004 0.050 0.093 0.025 0.019 0.026 0.024 0.019 R2 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month Table 6: Impact of RSBY (Matched districts and households) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) OP Exp IP Exp. Total IP drug IP fees. IP hosp. Was Has OP IP > Rs. Ratio IP/ IP > Rs. Ratio Ratio (Rs.) (Rs.) Medical + tests (Rs.) fees. hospitalized visit 5000 10,000 HHD IP/ IP/ Exp. (Rs.) (Rs.) (Rs.) (Rs.) Exp > HHD HHD (Rs.) 10% Exp > Exp > 20% 40% Triple -3.767* -7.683 -11.45* -4.381 -1.281 -1.347 0.0259** 0.0156 - - 0.00563 - - differences 0.00352 0.00450 0.00120 0.00421** (0.071) (0.143) (0.053) (0.122) (0.526) (0.185) (0.019) (0.497) (0.473) (0.223) (0.294) (0.730) (0.025) Observations 124637 124637 124637 124637 124637 124637 124637 124637 124637 124637 124637 124637 124637 Adjusted R2 0.081 0.019 0.057 0.021 0.007 0.009 0.039 0.085 0.023 0.018 0.017 0.014 0.012 Difference - 1.183 -3.751** 0.353 -0.0205 0.436 0.0171 - 0.00514 0.00455* 0.00737 0.00329 0.00145 in 4.934*** 0.00957 differences (0.001) (0.413) (0.015) (0.561) (0.962) (0.251) (0.437) (0.557) (0.174) (0.057) (0.269) (0.408) (0.433) Observations 63537 63537 63537 63537 63537 63537 63537 63537 63537 63537 63537 63537 63537 Adjusted R2 0.069 0.014 0.049 0.014 0.018 0.002 0.050 0.096 0.024 0.018 0.027 0.024 0.018 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month Table 7: Duration of treatment model (Matched districts and households) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) OP Exp IP Exp. Total IP drug IP fees. IP hosp. Was Has OP IP > Rs. IP > Rs. Ratio IP/ Ratio Ratio (Rs.) (Rs.) Medical + tests (Rs.) fees. hospitalize visit 5000 10,000 HHD IP/ IP/ Exp. (Rs.) (Rs.) d (Rs.) (Rs.) Exp > 10% HHD HHD (Rs.) Exp > Exp > 20% 40% Triple -0.136 -0.677 -0.813 -0.417* -0.0591 - 0.00311 0.00237 - - 0.00022 - - differen 0.148* *** 0.00044 0.000669 3 0.0002 0.000532** * * ces 9 09 (0.511) (0.117) (0.109) (0.098) (0.723) (0.058 (0.005) (0.244) (0.323) (0.083) (0.682) (0.597) (0.002) ) Observ 124637 124637 124637 124637 124637 12463 124637 124637 124637 124637 124637 124637 124637 ations 7 Adjuste 0.080 0.019 0.057 0.021 0.007 0.009 0.039 0.085 0.023 0.018 0.017 0.014 0.012 d R2 Difference in - - - - 0.00349 0.0260 0.0007 0.000260 0.00025 0.00020 0.000077 0.0000982 - differences 0.312* 0.00457 0.31 0.0210 15 1 5 6 0.000 * 6** 101 (0.025 (0.972) (0.04 (0.720) (0.917) (0.400) (0.706) (0.882) (0.453) (0.387) (0.893) (0.807) (0.58 ) 1) 3) Observations 63537 63537 6353 63537 63537 63537 63537 63537 63537 63537 63537 63537 6353 7 7 Adjusted R2 0.068 0.014 0.04 0.014 0.018 0.002 0.050 0.096 0.024 0.018 0.027 0.024 0.018 9 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month Table 8: Impact of RSBY (matched districts and households) – No Uttar Pradesh and Haryana (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) OP IP Exp. Total IP drug IP fees. IP hosp. Was Has OP IP > Rs. IP > Rs. Ratio Ratio Ratio Exp (Rs.) Medical + tests (Rs.) fees. hospitalized visit 5000 10,000 IP/ IP/ IP/ (Rs.) Exp. (Rs.) (Rs.) (Rs.) (Rs.) HHD HHD HHD (Rs.) Exp > Exp > Exp > 10% 20% 40% Triple -3.650 -10.52 -14.17* -5.875 -1.715 -2.349** 0.0269** 0.0195 - - 0.00999 - - differences 0.00169 0.00450 0.000988 0.00459** (0.204) (0.153) (0.096) (0.105) (0.590) (0.025) (0.042) (0.486) (0.786) (0.327) (0.101) (0.807) (0.041) Observations 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 Adjusted R2 0.083 0.020 0.056 0.021 0.007 0.012 0.048 0.074 0.026 0.020 0.019 0.017 0.013 Difference - 1.734 -1.144 0.646 - 0.670 0.0543*** 0.00124 0.00825** 0.00464* 0.0150** 0.00487 0.000442 in 2.878*** 0.00261 differences (0.010) (0.346) (0.403) (0.326) (0.997) (0.126) (0.005) (0.943) (0.029) (0.090) (0.021) (0.352) (0.801) Observations 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 Adjusted R2 0.066 0.016 0.047 0.013 0.018 0.007 0.061 0.082 0.024 0.015 0.030 0.026 0.018 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month Table 9: Duration of treatment model (Matched districts and households) (No Uttar Pradesh and Haryana) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) OP IP Exp. Total IP drug IP fees. IP hosp. Was Has OP IP > Rs. IP > Rs. Ratio Ratio Ratio Exp (Rs.) Medical + tests (Rs.) fees. hospitalized visit 5000 10,000 IP/ IP/ IP/ (Rs.) Exp. (Rs.) (Rs.) (Rs.) (Rs.) HHD HHD HHD (Rs.) Exp > Exp > Exp > 10% 20% 40% Triple -0.186 -0.679 -0.865 -0.316 -0.121 -0.165 0.00419*** - -0.000665 0.000688 - - differences 0.000363 0.000135 0.000390* (0.496) (0.292) (0.241) (0.349) (0.636) (0.120) (0.000) (0.551) (0.173) (0.242) (0.695) (0.060) Observations 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 104482 Adjusted R2 0.083 0.020 0.056 0.021 0.007 0.012 0.047 0.026 0.020 0.018 0.017 0.013 Difference -0.122 0.0322 -0.0895 0.00209 0.00788 0.0437 0.00349* 0.000401 0.0000750 0.000612 0.000219 -0.000112 in differences (0.314) (0.834) (0.560) (0.973) (0.839) (0.217) (0.076) (0.271) (0.757) (0.276) (0.590) (0.472) Observations 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 53603 Adjusted R2 0.066 0.016 0.047 0.013 0.018 0.007 0.060 0.024 0.014 0.030 0.025 0.018 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month To place our coefficients of impact in perspective, Table 11 lists the average values of outcome variables for the BPL sub-sample in the POST round pooling control and treatment districts. The average outpatient expenses per capita per month was Rs. 31.3, more than double that of inpatient (Rs. 14). Drugs are the biggest source of medical expenditure. We see that 12.9% of the BPL households have had at least one hospitalization case in a one year recall period. In terms of catastrophic expenses, 1.2% (2.59%) of the households has had inpatient expenses of more than Rs. 10,000 (Rs. 5,000) per annum. On the other hand 2.77% (7.74%) of the households had total OOP greater than Rs. 10,000 (Rs. 5,000) per annum. This suggests that outpatient care appears to cause a bigger dent on expenses. We would expect that outpatient expenses are spread across household members and over time more than inpatient, but that information is not available in this data. 13.4% of the households have had a ratio of OOP to total household expenses greater than 10%. Table 10: Average values of outcome variables of interest (2009-10) Variable Value OP Total in current prices per month (Rs.) 31.3 OP Family Planning in current prices per month (Rs.) 0.1 OP Other in current prices per month (Rs.) 0.5 OP Tests in current prices per month (Rs.) 0.9 OP Fees in current prices per month (Rs.) 3.3 OP Drugs in current prices per month (Rs.) 26.5 IP Total in current prices per month (Rs.) 14.0 IP Other in current prices per month (Rs.) 1.0 IP Tests in current prices per month (Rs.) 1.4 IP Fees in current prices per month (Rs.) 2.1 IP Hospital in current prices per month (Rs.) 2.8 IP Drugs in current prices per month (Rs.) 6.7 % of HHD with at least one hospitalization last year 12.9 % of HHDs with IP/ Total HHD. exp > 10% 3.32 % of HHDs with IP/ Total HHD. exp > 20% 1.28 % of HHDs with IP/ Total HHD. exp > 40% 0.3 % of HHDs with IP > Rs. 5,000 2.59 % of HHDs with IP > Rs. 10,000 1.2 % of HHDs with OOP to total exp. 4.79 % of HHDs with OOP/Total HHD. exp > 10% 13.4 % of HHDs with OOP/Total HHD. exp > 20% 4.49 % of HHDs with OOP/Total HHD. exp > 40% 0.79 % of HHDs with OOP > Rs. 5,000 7.74 % of HHDs with OOP > Rs. 10,000 2.77 Source: Authors’ calculations from NSSO unit data We have evidence from triple differences point estimates that, roughly speaking, RSBY has increased hospital utilization rates by about 20%, and decreased outpatient expenses by about 15%, while total medical expenses have dropped by about 8%. The significant negative impact on outpatient expenses is desirable considering the large OOP expenditure on outpatient care. However it is not clear whether patients are now seeking the “correct” inpatient procedure for a condition that they could not afford formerly and instead were containing with lower cost (per visit or episode) outpatient care or whether the inpatient care is reducing need for outpatient visits or if patients and hospitals are colluding to convert outpatient to inpatient to take advantage of the insurance. It is expected that RSBY would enable people to undergo procedures that they would have not done without health financing and to that extent, the rise in hospital utilization indicates that the scheme is achieving its intended goals. We detect no impact on the mean expenditure on inpatient services although the premium is 95% subsidized. While this might simply be because the scheme has not been in existence long enough at the time of the survey, it may be the case that increased utilization can increase OOP expenses on hospitalization since RSBY does not cover all expenses. Robustness Checks The results pass a number of robustness checks and tests: Test 1: The difference in differences calculated between control and treatment districts should be driven by the intervention and not by pre-intervention trends in values of outcome indicators. We use rounds 55 and 61 to verify the assumption that the control and treatment groups' time trends prior to the intervention are similar and hence are likely to have been similar in the absence of the intervention. A visual inspection of Figure 1 below shows the trends of the outcomes for the BPL sub-sample between the control and treatment groups. Figure 1: Trends between control and treatment – BPL sub-sample Was Hospitalized 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 55 61 66 Treatment Control OP Expenses 30 25 20 15 10 5 0 55 61 66 Treatment Control However, since we use triple differences, we also attempt to see the difference between the first differences between BPL and non-BPL in each survey round. Figure 2 below shows trends in outcome variables, which are the first difference between the average values for BPL and non- BPL households. However, while visual inspection is helpful, it is more meaningful to conduct statistical tests as below. Figure 2: Trends between control and treatment – BPL-Non-BPL difference Was Hospitalized 0 55 61 66 -0.01 -0.02 -0.03 -0.04 -0.05 -0.06 -0.07 Treatment Control OP Expenses 0 -5 55 61 66 -10 -15 -20 -25 -30 -35 -40 -45 Treatment Control Test 2: There should be no significant “impact” of being in a future RSBY district on outcome variables between rounds 55 and 61. We ran our model assuming that round 61 is the fake “post RSBY” round and round 55 is the fake “pre RSBY” round to see if in addition to the trends test, whether the regressions throw up any significant trends in outcome variables when comparing treatment with control districts. This is similar to the falsification tests used by Fan, Karan, and Mahal (2012) and Duflo (2001). We note two results. There are no effects on the matched control group or when using only districts where RSBY was planned but not launched as controls. However, a number of outcome variables show up as significant in this “fake” impact evaluation, when we include in the control group, those districts where RSBY was not planned at the time of the survey (called Control 2 in Section 3). (See Appendix B). Our interpretation is that RSBY district selection is endogenous. Test 3: Sub-sample analysis. We ran our suite of impact regressions on non-BPL households and find no effects similar to that of our main results on any of the outcome variables. We also ran the usual sub-sample robustness checks selecting 80% of the observations randomly and the results hold. Conclusion This paper provides evidence on the impact of RSBY, one of the largest government health schemes in the developing world, in its early days (the number of months of availability of RSBY for the average household was six months). RSBY has decreased outpatient and consequently decreased medical expenditure among the overall BPL population. There is some limited evidence that RSBY has been helpful in increasing access to hospitalization through increased hospitalization rates. It is not clear whether the increased access to inpatient care is now reducing expenses on outpatient treatment or whether patients and hospitals are colluding to convert outpatient into inpatient cases to avail of the coverage. In any case, RSBY has, in fact, piloted an outpatient insurance component in two districts and is planning to launch it at scale. It is expected that offering primary care will lead to reduction of hospitalization claims if preventable illnesses are contained before becoming worse. There are regional variations in the impact of RSBY, with the North zone, in particular UP, Haryana and Punjab appearing to be impacted in a different way compared to the rest of India. Allowing for longer duration of treatment, and using more recent NSSO data when available or household surveys in a controlled environment will shed more light on the impact of RSBY and strengthen our findings. We hope this paper will inform future research on RSBY and other government health schemes. Of principal interest to schemes that cover only inpatient care is to know the exact reason for the drop in outpatient expenses. Broader research should attempt to find out how exactly the scheme helps recipients, particularly in terms of avoiding catastrophic expenses and entailing impoverishment both from outpatient and inpatient expenses, to further inform the design of the scheme. References Abadie, Alberto. 2005. Semi-parametric Difference-in-Differences Estimators. Review of Economic Studies. 72(1), 1-19. Attanasio O, Augsburg, B, De Haas, R, Fitzsimons, E, Harmgart, H. December 2011. Group lending or individual lending? Evidence from a randomised field experiment in Mongolia. European Bank for Reconstruction and Development, Working Paper No. 136 Axelson H, Bales S, Minh P, Ekman B, Gerdtham U. 2009. Health financing for the poor produces promising short-term effects on utilization and out-of-pocket expenditure: evidence from Vietnam. International Journal for Equity in Health 8(1): 20. Aggarwal A. 2010. Impact evaluation of India’s ‘Yeshasvini’ community based health insurance programme. Health Economics 19(S1): 5-35. Bauhoff S, Hotchkiss, Smith O. 2011. The impact of medical insurance for the poor in Georgia: A regression discontinuity approach. Health Economics doi: 10.1002/hec.1673. Berman P.A. 1997. Rethinking Health Care Systems: Private Health Care Provision in India. Harvard School of Public Health. July Bloom, Howard S. 1995. Minimum Detectable Effects: A Simple Way to Report the Statistical Power of Experimental Designs. Evaluation Review 19(5): 547-5 Devadasan, N., B. Criel, W. Van Damme, S. Manoharan, P.S. Sarma, P. Van der Stuyft. 2010. Community health insurance in Gudalur, India, increases access to hospital care. Health Policy & Planning 25(2), 145–54. Dow W, Schmeer K. 2003. Health insurance and child mortality in Costa Rica. Social Science and Medicine 94(2): 975-978. Duflo, E. 2001. Schooling and labor market consequences of school construction in Indonesia: evidence from an unusual policy experiment. American Economic Review. 91(4), pp. 795-813 Duggal R and Amin S. 1989. Cost of Health Care: A Household Survey in an Indian District. The Foundation for Research in Community Health, Bombay, India. Heckman, James, Hidehiko Ichimura, and Petra Todd. (1997) Matching as an Econometric Estimator. Review of Economic Studies. 64: 605-654. Ho, Daniel, Kosuke Imai, Gary King, and Elizabeth Stuart. 2007. Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference. Political Analysis. 15: 199-236. Iacus, Stefano, Gary King, and Giuseppe Porro. 2011. Multivariate Matching Methods That are Monotonic Imbalance Bounding. Journal of the American Statistical Association. 493: 345-361. Johnson, Doug and Sushil Kumar. 2011. Do Health Camps Make People Healthier? Evidence from an RCT of Health Camps on Usage of RSBY. Available at http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1873269 Karan, AK and Selvaraj, S. 2012 Why Publicly-Financed Health Insurance Schemes Are Ineffective in Providing Financial Risk Protection. Economics and Polictical Weekly. Vol - XLVII No. 11 Krishnaswamy K, and Ruchismita R. 2011 Rashtriya Swasthya Bima Yojana - Performance Trends and Policy Recommendations. An evaluation of a Government Health Insurance Scheme. RSBY Working Paper. CIRM Miller G, Pinto DM, Vera-Hernandez M. 2009. High-powered incentives in developing country health insurance: evidence from Colombia’s regimen sobsidiado: NBER Working Paper, 15456. Peters, D., Yazbeck, A., Sharma, R., Ramana, G., Pritchett, L., & Wagstaff, A. (2002). Better Health Systems for India's Poor, World Bank. Stuart, E.A. (2010). Matching Methods for Causal Inference: A review and a look forward. Statistical Science 25(1): 1-21. Thornton, Rebecca, Laurel Hatt, Erica Field, Mursaleena Islam, Freddy Solís, Martha Azuzena González Moncada. 2010. Social Security Health Insurance for the Informal Sector in Nicaragua: A Randomized Evaluation. Health Economics, Volume 19, Issue S1, September. Pages: 181–206. Victoria Fan, Anup Karan, and Ajay Mahal. 2012. State Health Insurance and Outof-Pocket Health Expenditures in Andhra Pradesh, India. CGD Working Paper 298. Washington, D.C.: Center for Global Development. World Bank. 1995. India Policy and Financing Strategies for Strengthening Primary Health care Services. Report no. 13042-IN, May 15 Wagstaff A. 2009. Social health insurance reexamined. Health Economics 19(5): 503-517. Wagstaff A. 2007. Health insurance for the poor: initial impacts of Vietnam‘s health care fund for the poor. The World Bank, Policy Research Working Paper Series: 4134 2007. Wagstaff A. 2010. Estimating health insurance under unobserved heterogeneity: The case of Vietnams Health Care Fund for the poor. Health Economics 19(2): 189-208. Wagstaff A, van Doorslaer E. 2003. Catastrophe and Impoverishment in Paying for Health Care: With Applications to Vietnam 1993-98. Health Economics 12(11): 921-933. Wagstaff A, Lindelow M. 2008. Can insurance increase financial risk? The curious case of health insurance in China. Journal of Health Economics 27(4): 990-1005 Wagstaff A, Lindelow M, Jun G, Ling X, Juncheng Q. 2009. Extending health insurance to the rural population: an impact evaluation of China's new cooperative medical scheme. Journal of Health Economics 28(1): 1-19. Wang H, Yip W, Zhang L, Hsiao WC. 2009. The impact of rural mutual health care on health status: evaluation of a social experiment in rural China. Health Economics 18: S65–S82. APPENDIX A: Number of districts by group RSBY RSBY planned but RSBY not State implemented not implemented planned as of Total during survey during survey 2011 Jammu & Kashmir 0 0 10 10 Himachal Pradesh 4 8 0 12 Punjab 15 2 0 17 Chandigarh 1 0 0 1 Uttaranchal 2 11 0 13 Haryana 19 0 0 19 Delhi 1 0 0 1 Rajasthan 0 4 28 32 Uttar Pradesh 61 9 0 70 Bihar 11 26 0 37 Sikkim 0 0 4 4 Arunachal Pradesh 0 4 9 13 Nagaland 3 3 2 8 Manipur 0 2 7 9 Mizoram 0 5 3 8 Tripura 2 2 0 4 Meghalaya 1 4 2 7 Assam 1 5 17 23 West Bengal 4 14 0 18 Jharkhand 6 9 3 18 Orissa 6 7 17 30 Chattisgarh 15 1 0 16 Madhya Pradesh 0 0 45 45 Gujarat 9 16 0 25 Daman & Diu 0 0 2 2 D & N Haveli 0 0 1 1 Maharastra 27 2 5 34 Goa 2 0 0 2 Lakshadweep 0 0 1 1 Kerala 14 0 0 14 Pondicherry 0 0 4 4 A & N Islands 0 0 1 1 Total 204 134 162 500 APPENDIX B1: Falsification tests using planned RSBY districts only as controls (Using Round 55 and Round 61 data) Triple Differences – RSBY planned districts only (1) (2) (3) (4) (5) (6) (7) (8) OP Exp per mth IP Exp. pm Total Medical Was IP > Rs. 5000 IP > Rs. 10,000 OOP > Rs. OOP > Rs. (Rs.) (Rs.) Exp. (Rs.) hospitalized 5,000 10,000 Triple diffs impact 0.191 -3.194 -3.002 0.00140 0.00336 -0.00282 0.00949 0.00250 of RSBY (0.938) (0.237) (0.549) (0.863) (0.334) (0.296) (0.433) (0.677) Observations 132495 132495 132495 132495 132495 132495 132495 132495 Adjusted R2 0.048 0.034 0.065 0.141 0.033 0.016 0.074 0.047 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month deflated Difference in differences - BPL Sub-sample - RSBY districts only (1) (2) (3) (4) (5) (6) (7) (8) OP Exp per mth IP Exp. pm Total Medical Was IP > Rs. 5000 IP > Rs. 10,000 OOP > Rs. OOP > Rs. (Rs.) (Rs.) Exp. (Rs.) hospitalized 5,000 10,000 DiD impact of RSBY 3.605 -0.718 2.887 0.0389 0.00220 0.000667 0.0192* 0.0140** (0.188) (0.791) (0.364) (0.377) (0.758) (0.868) (0.063) (0.033) Observations 66138 66138 66138 66138 66138 66138 66138 66138 Adjusted R2 0.104 0.058 0.118 0.159 0.019 0.010 0.044 0.017 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month deflated APPENDIX B2: Falsification tests including non RSBY planned districts as controls Triple Differences – using non-RSBY planned districts as control (1) (2) (3) (4) (5) (6) (7) (8) OP Exp per mth IP Exp. pm Total Medical Was IP > Rs. 5000 IP > Rs. 10,000 OOP > Rs. OOP > Rs. (Rs.) (Rs.) Exp. (Rs.) hospitalized 5,000 10,000 Triple diffs impact of -5.496*** -3.960*** -9.456*** -0.00610 0.00389** 0.00190 -0.00702 -0.000574 RSBY (0.000) (0.005) (0.000) (0.759) (0.046) (0.178) (0.307) (0.851) Observations 141586 141586 141586 141586 141586 141586 141586 141586 Adjusted R2 0.046 0.032 0.063 0.133 0.031 0.017 0.072 0.046 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month deflated Difference in differences - BPL Sub-sample - using non-RSBY planned districts as control (1) (2) (3) (4) (5) (6) (7) (8) OP Exp per mth IP Exp. pm Total Medical Was IP > Rs. 5000 IP > Rs. 10,000 OOP > Rs. OOP > Rs. (Rs.) (Rs.) Exp. (Rs.) hospitalized 5,000 10,000 DiD impact of RSBY 3.242* 3.422** 6.664*** 0.0271 0.0128** 0.00611** 0.0253*** 0.0182*** (0.086) (0.032) (0.002) (0.727) (0.022) (0.016) (0.004) (0.001) Observations 70672 70672 70672 70672 70672 70672 70672 70672 Adjusted R2 0.099 0.047 0.109 0.154 0.017 0.009 0.040 0.016 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month deflated Appendix B3: Falsification tests using matched districts (Using Round 55 and Round 61 data) (1) (2) (3) (4) (5) (6) (7) (8) (4) (5) (6) (7) (4) (5) OP IP Exp. Total IP IP IP hosp. Was Has OP IP > IP > Rs. Ratio IP/ Ratio Rati Exp (Rs.) Medical drug + fees. fees. hospitalized visit Rs. 10,000 HHD IP/ o IP/ (Rs.) Exp. (Rs.) tests (Rs.) (Rs.) 5000 (Rs.) Exp > HHD HH (Rs.) (Rs.) 10% Exp > D 20% Exp > 40% Triple -0.719 -1.666 -2.385 0.687 0.0243 -1.021 0.335 -0.430 -0.0116 0.00504 0.000058 -0.00505* 0.00037 - differences 9 8 0.0020 8 (0.722 (0.465 (0.540) (0.229) (0.936 (0.293 (0.553 (0.703 (0.261) (0.387) (0.986) (0.066) (0.901) (0.276) ) ) ) ) ) ) Observatio 13586 13586 135862 135862 13586 13586 13586 13586 135862 135862 135862 135862 135862 13586 ns 2 2 2 2 2 2 2 Adjusted 0.047 0.035 0.062 0.060 0.025 0.018 0.003 0.004 0.184 0.026 0.014 0.007 0.030 0.014 R2 Difference in 3.858 0.287 4.145* - 0.232 0.478 0.565 0.0206 0.024 0.0056 0.00273 - 0.0014 -0.000772 * differences 4.374* 0 8 0.0037 5 * 9 (0.125) (0.900) (0.017 (0.027 (0.507 (0.586) (0.368) (0.976) (0.522 (0.553) (0.667) (0.273 (0.841) (0.848) ) ) ) ) ) Observations 67816 67816 67816 67816 67816 67816 67816 67816 67816 67816 67816 67816 67816 67816 Adjusted R2 0.117 0.054 0.123 0.128 0.111 0.011 -0.003 -0.002 0.203 0.021 0.009 0.009 0.016 0.007 p-values in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Note columns 1 through 6 are per capita expenses per month deflated