Health insurance and consumption: Evidence from China s New Cooperative Medical Scheme Chong-En Bai Binzhen Wu * Tsinghua University Abstract We exploit a quasi-natural experiment arising from the introduction of a health insurance program in rural China to examine how the insurance coverage affects household consumption. The results show that on average, the health insurance coverage increases non-healthcare related consumption by more than 5 percent. This insurance effect exists even for households with no out-of-pocket medical spending. In addition, the insurance effect is stronger for poorer households and households with worse self-reported health status. These results are consistent with the precautionary saving argument. Moreover, the insurance effect varies by household experience with the program. Particularly, the effect is significant only in villages where at least some households have actually obtained reimbursement from the insurance program. Within these villages, the program stimulates less consumption among the new participants than among households that have participated in the program for more than one year. Keywords: Precautionary savings; Health insurance; Consumption; New Cooperative Medical Scheme; Chinese savings JEL Classification Nos.: D12; D91; E21; I18 * Corresponding author. Email: wubzh@sem.tsinghua.edu.cn; Tel: 86-10-62772371. For their valuable suggestions, we want to thank participants in the workshops for Saving and Investment in China at Tsinghua University. Chong-En Bai acknowledges supports from National Planning Office of Philosophy and Social Sciences Major Grant 10zd&007. Binzhen Wu acknowledges supports from the National Natural Science Foundation of China (Project ID: 70903042).
1 Introduction Over the past several years, considerable international attention has been directed at the high level of savings by Chinese households. The household saving rate rose by about 10 percentage points from 1995 to 2008, accounting for 28% of the disposable income in 2008. This increase is higher than that observed in most other countries including East Asian nations (Prasad, 2009). A popular explanation for China s high saving rate is that the dissolution of the traditional social safety net has prompted higher precautionary savings (Chamon and Prasad, 2010). Accordingly, the Chinese government has exerted extensive efforts to improve China s social safety net. The social insurance programs disbursed RMB 1.2 trillion in 2009, with an annual growth rate of 19.4% since 2000; the health insurance program disbursed RMB 0.28 trillion in 2009, with an annual growth rate of 47% since 2000 (Chinese Statistical Yearbook, 2010). Existing empirical literature provides mixed evidence on the importance of precautionary savings. The results range from minimal to substantially important. Recent studies have exploited the exogenous variations in insurance coverage caused by policy changes. These studies include Gruber and Yelowitz (1999), Engen and Gruber (2001), and Kantor and Fishback (1996) for the US; Atella, Rosati, and Rossi (2005) for Italy; Wagstaff and Pradhan (2005) for Vietnam; and Chou, Liu, and Hammitt (2003) for Taiwan. A few Chinese studies, including Ma, Zhang, and Gan (2010), Liu et al. (2010), and Brown, Brauw, and Du (2010), use the launch of public health insurance programs in China to investigate the size of precautionary savings in the country. However, they deliver different conclusions because of the differences in methodology and dataset. 1
This paper exploits one of the most important policy changes implemented in rural areas: the introduction of the New Cooperative Medical Scheme (NCMS) in July 2003. This public health insurance program was introduced sequentially in different counties, and household participation is voluntary. To estimate the insurance effect that excludes the effects of contemporary policy changes, we focus on the double-difference comparison of the insured and non-participants in the villages where the program has been launched. The difference-in-difference (DID) framework helps eliminate all time-invariant selection biases. Selection bias on observables is further reduced by enabling the temporal change in consumption to vary with observable characteristics such as income and health status, or by applying the DID matching method. Selection on unobservables is examined using the counterfactual test that applies the estimation model to the time periods at which households have not enrolled in the program. The data we use combine the longitudinal Rural Fixed-Point Survey (RFPS) from 2003 to 2006 and a household survey that we conducted on the NCMS for a subsample of the 2006 round of the RFPS. The results indicate that household consumption other than health expenditures has increased by about 5.6 percent or 149 yuan (18.64 US dollars at the exchange rate in June 2006) because of the health insurance coverage. The magnitude of the increase is much larger than the average premium of the insurance, which was typically 30 yuan in 2003 and 50 yuan in 2006. The program stimulates consumption more effectively than do the cash transfers from the government because the average propensity of rural households to consume is only about 0.44. The estimate is robust to different specifications and consistent with the macro time trend of rural household saving rate: the NCMS has been rapidly rolled out since 2005, and rural household saving rate has started to decline beginning that same year (Prasad, 2009). 2
We also find that the insurance effect remains significant for households that do not spend on health care in a given year. This result cannot be explained by the crowd-in concept, which emphasizes that the insurance coverage reduces out-of-pocket health payments, thereby leaving insured households with more income for non-healthcare expenditure. Even the income effect of government subsidy cannot explain the magnitude of the increase in consumption. Moreover, the positive effect of insurance on consumption is much stronger for poorer households and households with worse self-reported health status. Given that these households tend to have a higher risk of incurring large health expenditure relative to income, these results are consistent with the explanation that points to precautionary savings. In addition, the positive insurance effect increases when the NCMS provides more generous compensation for household health expenditure at county-level hospitals, which is also consistent with the precautionary saving justification. Finally, the insurance effect varies depending on household experience with the program. Particularly, the level of trust accorded to the program can be crucial to the stimulation of consumption. The insurance effect on consumption is significant only in the villages where at least some households have obtained reimbursement from the program, through which trust in the program is established. Moreover, in such villages, the increase in consumption is much larger among the experienced members who have participated in the program for more than one year than that among the new members. By contrast, in the villages where no household has received any reimbursement, neither the average insurance effect nor the difference between the experienced and new participants is significant. To the best of our knowledge, this paper is one of few studies that exploit this policy change in systemically assessing the effect of the health insurance program on consumption in rural China. In addition, we have examined the importance of the precautionary saving 3
explanation and heterogeneity of the insurance effect, which are mostly unaddressed in related Chinese studies. The findings on the roles of trust and experience warrant more attention in terms of research and policymaking, particularly for public insurance programs in developing countries where transparency and trust are often lacking. The rest of the paper is organized as follows. Section 2 introduces the background of the NCMS and presents the literature review. Section 3 introduces the data and provides descriptive statistics. Section 4 discusses our econometric specifications. Section 5 shows the results for the baseline model, and Section 6 presents the robustness tests. Section 7 concludes. 2 Background 2.1 New Cooperative Medical Scheme Since the dissolution of the rural Cooperative Medical System in the early 1980s, illness has emerged as a leading cause of poverty in rural China, and the high cost of health care has deterred households from obtaining necessary health treatment (You and Kobayashi, 2009). In response to these problems, the Chinese government initiated a pilot program of the New Cooperative Medical Scheme (NCMS) in 2003. The pilot program was launched in 310 rural counties of China s more than 2800 rural counties in July 2003; it was expanded to 617 counties in 2005 and 1451 counties in 2006. By June 2007 when we collected the data for this study, the program had expanded to over 84.9% of all the rural counties and had covered 82.8% of all the rural residents. Several key features characterize the NCMS: 1) the program targets rural residents; 2) participation is voluntary but should be on a household basis; 1 3) participating households 1 The requirement of participation on household basis is imposed over 97% of the counties in our sample. 4
are required to pay a flat-rate premium, but the insurance is heavily subsidized by the governments; 4) the program reimburses participants mainly for in-patient expenses; 5) the program is operated at the county level rather than at the township or village level. The voluntary nature of the participation raises concerns on the adverse selection issue that can threaten the financial sustainability of the NCMS. However, the participation rates in pilot villages are generally very high, with an average of 86% from 2003 to 2006 in our sample. An important reason for the high participation rate is the generous government subsidies. The premium for each member in a subscribing household in 2003 was typically 10 yuan while the government paid 20 yuan a year. Since 2006, the government subsidy has increased to 40 yuan while household contribution has remained the same. 2 Along with the NCMS, the government also implemented some supporting policies, such as improving the quality and delivery of health care services and strengthening pharmaceutical governance. At the same time, some studies show that the average expenditure per visit increased after the introduction of the NCMS program (Yao and Kobayashi, 2009; Mao, 2005). These changes also affected households that chose not to participate in the NCMS. Finally, the government set up a parallel program, i.e., the medical assistance scheme, to help the poverty-stricken population. 3 Although the central government has issued broad guidelines for how the NCMS should be designed and implemented, provincial and county governments have retained considerable discretion over the details of the program, including the placement of the pilot Some studies suggest that local governments have made considerable efforts to attain high participation rates, including mandating households to participate. However, our survey shows that less than 1% of households report compulsory enrollment in 2007. 2 The poor and some other groups are exempted from contributions. In 2008, the government subsidies increased to 80 yuan per person a year. Household contribution was raised to 20 yuan per person a year. 3 In counties that have introduced the NCMS, this program helps poor families pay the premium of the NCMS and reimburses the health expenditure below the deductible or above the ceiling. For counties that have not launched the NCMS, this serves as a subsidized insurance for the poorest families. 5
program and insurance package. This kind of local authority has led to considerable heterogeneity in the benefit, coverage, and management packages across counties. Table A.1 shows the main parameters of the insurance packages for the 54 counties, about which we have detailed information. We first observe that the insurance typically does not offer generous coverage. In particular, deductibles are high, ceilings are low, coinsurance rates are high, and outpatient expenditure is usually not fully covered. At the township clinics, for example, the deductible was about 125 yuan on average, the ceiling was 14838 yuan, and the coinsurance rate was 49.1% in 2006. However, the insurance program can still substantially reduce the out-of-pocket health care payments of the insured. For households whose health expenditure is more than 14838 yuan, they can save 7489 yuan at most. Furthermore, the insurance plans have become more generous over time for all levels of health care centers, particularly for the township clinics. Moreover, the insurance covers most kinds of disease, including childbirth, as long as the health expenditure is related to in-patient service. Table A.1 also shows that although all counties cover in-patient care, the coinsurance rate varies substantially, ranging from 20% to 75%. Moreover, most counties offer better packages to the lower level health care centers. This feature has magnified over time. By contrast, the difference in the coinsurance rate for different amounts of health expenditure is small and declines over time. Finally, the procedure for claiming reimbursement has become simpler over time. In 2003, about 86% of counties asked households to pay the providers upfront for all their costs and go to insurance organizations to claim the reimbursement. The rate declined to 52% in 2006. However, the number indicates that in 6
most counties, the households bear the risk if the government does not pay. 4 2.2 Related literature After the seminal theoretical papers by Zeldes (1989), Deaton (1991), and Carroll (1992), many studies have used micro data to examine the strength of the precautionary saving motive. Simulations or structural estimations mostly find that precautionary savings can explain a sizeable portion as much as 50% of US savings (Gourinchas and Parker, 2002; Hubbard, Skinner, and Zeldes, 1994; etc.). By contrast, other empirical studies have drawn mixed conclusions: Dynan (1993), Guiso et al. (1992), and Starr-McCluer (1996) find little or no precautionary saving, whereas Banks et al. (2001) (for the UK), Carroll and Samwick (1998) (for the US), and Fuchs-SchÜndeln and SchÜndeln (2005) (for Germany) find economically important precautionary motives. Early studies typically examine the issue by relating wealth accumulation to some measures of income risks that households encounter. The mixed results stem partly from the variation in the measure of income uncertainty (Engen and Gruber, 2001). Various measures have been tested, including the variability of income (Carroll and Samwick, 1998; etc.), variability of consumption (Dynan, 1993), expectations of future job loss (Guiso et al., 1992; Lusardi, 1998), actual job loss (Carroll, Dynan, and Krane, 2003), a proxy based on job characteristics or education (Skinner, 1988), and household insurance coverage (Starr-McCluer, 1996; Guariglia and Rossi, 2004). In the Chinese context, literature along this line generally finds strong evidence for the importance of precautionary savings (Meng, 2003; Jalan and Ravallion, 2001; Kraay, 2000). However, these studies all suffer from the potential bias caused by the likely correlation between income risks and underlying 4 In addition, around 48% of the counties provided insurance for migrants, although the reimbursement is usually much less generous for health care expenditure in hospitals outside the county. 7
preferences for savings. Recent studies have exploited the exogenous variations in insurance coverage caused by policy changes. These studies include Engen and Gruber (2001), Gruber and Yelowitz (1999), and Kantor and Fishback (1996) for the US; Atella and Rosati (2005) for Italy; and Wagstaff and Pradhan (2005) for Vietnam. Although these studies focus on different programs, including unemployment insurance, worker compensation, and health insurance, most studies confirm that social insurance programs reduce asset accumulation by reducing income or expenditure risk. Related research in developing countries is still in its early stages. Wagstaff and Pradhan (2005) study a case in Vietnam and find that the introduction of a health insurance program increased nonmedical household consumption. Chou, Liu, and Hammitt (2003) find that the universalization of health insurance in Taiwan reduced the household saving rate by about 2.5 percentage points. For mainland China, Ma, Zang, and Gan (2010) examine the effect of the NCMS specifically on food consumption among rural households. However, food consumption can be much less elastic than other kinds of consumptions. Liu et al. (2010) investigate the effect of introducing the public health insurance program on the consumption of urban households. Nevertheless, rural residents respond differently to income uncertainty compared with urban residents (Zhang and Pei, 2007). Both studies apply the Difference-in-Difference method and find a significant positive insurance effect on consumption. The paper that is most similar to our study is Brown, Brauw, and Du (2010). They apply the matching propensity score method to household survey data on two provinces, and find that the NCMS reduces food consumption but does not significantly affect non-healthcare and total consumption. However, their study does not thoroughly 8
address the selection problem. These studies do not examine whether the increase in consumption results from a reduction in precautionary savings. In addition, the literature on developed countries indicates that the strength of precautionary saving can vary among different income groups (Carroll, Dynan, and Krane, 2003), but the analysis of the heterogonous effects are lacking in these Chinese studies. 3 Data and descriptive statistics Our data come from the longitudinal Rural Fixed-Point Survey (RFPS) from 2003 to 2006, and a supplementary household survey that aims to evaluate the NCMS. The sample in the RFPS is selected on the basis of a multi-stage stratified random sampling strategy. The 2006 round includes 19,488 households in 313 villages drawn from 26 Chinese provinces. The survey uses the weekly book accounting information maintained by the households as the primary information source. It provides detailed information on income and expenditure. 5 The supplementary survey was conducted by Tsinghua University in May 2007. It surveyed a subsample of the 2006 round of RFPS and covered 23 provinces, 143 villages, and 5728 households. It collected detailed information about the time at which a household enrolled in the NCMS, and retrospective information on each member s health care utilization and expenditure in each year from 2003 to 2006. The survey oversampled households with economically meaningful health care expenditure. 6 5 The identification code for tracking individuals and households is ridden with mistakes. We use conservative rules based on individual age, gender, and education to match individuals and households over years. If more than half of the household members cannot be matched across two years, we exclude the household from our sample. Altogether, we exclude around 8% of the households in our sample because of the inconsistency in the identification code. 6 More specifically, the survey first ranks all the households in the 2006 round of the RFPS on the basis of their average health care expenditure from 2003 to 2006. Then, it randomly draws 80% of the observations in the top one-third of the sample, and 50% of the observations in the remaining two-thirds of the sample. 9
Table 1 shows the enrollment rate of the villages and households from 2003 to 2006. The enrollment of our sample villages spread over different years: 16.4% of the villages enrolled in the program in 2003 and the rate increased to 77.1% in 2006. Similarly, the enrollment rate of households gradually increased from 9.5% in 2003 to 72.3% in 2006. These figures are consistent with national data (Chinese Statistical Yearbook 2010). In the villages that have launched the NCMS (referred to as NCMS-villages hereafter), the participation rate of households increased from 63.7% in 2003 to 94.6% in 2006. Moreover, most households participated in the year of program launch; the first-year participation rate was 63.7% in 2003 and 96.2% in 2006. These numbers also indicate that quite a few households (14.4% on average) chose not to participate in the program in the first year the program was introduced. Over the four years from 2003 to 2006, about 12.8% of households in the NCMS-villages did not participate. To relate consumption to household enrollment status in the NCMS, we exclude some outliers, such as households that terminated membership in the NCMS or participated in some cooperative insurance programs from 1993 to 2002. Also excluded are households that purchased commercial insurance or did not participate in the NCMS but enrolled in some government insurance programs in 2007. 7 Finally, given that the NCMS was first piloted in July 2003, we exclude all the observations in 2003 for villages that launched the program in 2003 (but keep the observations in other years). In so doing, we avoid the potential complication that arises from the effect of the NCMS actually beginning in the middle of that year. 8 As a result, year 2003 is regarded as the year during which no counties 7 Only 107 households ever terminated membership in the program; 4.6% of the households enrolled in some cooperative insurance programs from 1993 to 2002, and most of them participated after 1997. About 6% of the households have some commercial insurance, and less than 1% enrolled in some government insurance programs but not in the NCMS in 2007. 8 We include these observations as a robust test, and as expected, find a slightly smaller insurance effect of the 10
introduced the NCMS. The final sample includes 520 villages and 17,715 households over the 4-year sample period. Table 2 shows the descriptive statistics for three groups: the insured households and two kinds of uninsured households: the non-participants who live in the NCMS-villages but chose not to participate in the program, and the non-exposed households located in the non-ncms villages. Because more villages and households join the program over time, the household compositions of these three groups vary over time. Therefore, we use the 2003 values of the variables that may change over time as a proxy for the underlying household characteristics at the time before the implementation of the program. The table illustrates that compared with the non-participants, the households that chose to participate generally had higher incomes, total consumption, and non-healthcare consumption in 2003. The evidence for adverse selection is mixed. There are five categories of self-reported health status: excellent, good, fair, bad, and no working capacity, the last two of which are collectively called poor. The participants had more members reporting fair or worse health status and spent more on in-patient health care than did the non-participants in 2003. However, the participants had fewer members with poor health status (including bad and no working capacity ) and had less total health-care expenditure in 2003. Column 3 of Table A.2 shows that even the positive evidence for the adverse selection disappears when we focus on within-village comparison by controlling for the village fixed effect. For the demographics, the heads of the participating household are slightly older, more educated, and less likely to be single or be a non-agricultural worker. These households are also more likely to have communist members, and less likely to be a minority or a household in poverty ( Wubao ). These differences are confirmed by NCMS on consumption. 11
the regression results in Table A.2 in the Appendix. 9 Additionally, the table indicates that although the non-participants differ from the insured in observable characteristics, they are more similar to the insured than to the non-exposed households in terms of income and consumption. This observation is not surprising because households located in the same village are more likely to be similar to one another than to those located in a different village. The comparison of village characteristics indicates that the placement of the pilot programs may not be random: the NCMS-villages are richer and have fewer clinics but more children receiving vaccinations than the non-ncms villages. They have fewer migrants and more laborers, and the residents have higher educational levels. They are also less likely to be in mountainous, western, or central areas. Column 4 of Table A.2 confirms that these differences are significant. 4 Baseline empirical model Our empirical analyses exploit the quasi-natural experiment arising from the NCMS to examine the effect of the insurance coverage on household non-healthcare consumption. We exclude health expenditure because we want to focus on precautionary savings, and health expenditure is affected by the insurance through other channels. To simplify exposition, consumption refers to all consumption expenses net of health expenditures throughout the paper, unless otherwise specified. We begin by applying the Difference-in-Difference (DID) framework to the four-year panel. More specifically, the effects of the NCMS are identified by differences in dynamic changes in consumption between the insured and uninsured households in the time periods 9 Table 2 also indicates that the subscribing households have, on average, fewer members older than 65 or younger than 10, more migrants, and are less likely to have a female head. However, the regression in Table A.2 shows that these differences are not significant or that the differences take the opposite direction. 12
before and after the launch of the NCMS in the villages. The framework can eliminate all time-invariant selection biases. This is crucial to our context because participation in the program is voluntary, implying that the households that chose not to participate can differ from the participants in both observable and unobservable characteristics. Moreover, program placement over villages can be non-random. As a result, insured households and uninsured households may have consumed differently in the absence of the NCMS. The double-difference method can still deliver unbiased and consistent estimates as long as the temporal changes in household consumption would have been parallel were there no NCMS. As mentioned previously, there are two types of uninsured households in each period. One is composed of the non-participants in the NCMS-villages and the other comprises the non-exposed households in the non-ncms villages. To examine the precautionary saving motive, we focus on the double difference between the insured and non-participants in the NCMS-villages. This focus is driven by two reasons. The first is that other changes occurred along with the introduction of the NCMS. In particular, the governments implemented supporting policies to improve the quality and delivery of health care services. In addition, anecdotal evidence indicates that the price of health care services increased after the introduction of the NCMS (Mao, 2005). For the precautionary saving explanation, we need to identify the insurance effect of the program that occurs only through the insurance coverage and excludes the effects of the aforementioned contemporary policies or changes. This insurance effect can be estimated by the double difference between the insured and non-participants within the NCMS-counties because both groups were affected 13
by these changes. 10 By contrast, the double difference between the insured and non-exposed determines the gross effect of the NCMS on the insured, which includes both the insurance effect of the NCMS and effects of other associated changes. 11 The second reason is related to the identification assumption for the DID model: the consumption dynamics of the insured and that of the control group should be parallel even in the absence of the NCMS. We argue that the assumption is more problematic for the comparison between the insured and non-exposed than that between the insured and non-participants. First, households in the same village are more likely to be similar to one another than to households located in a different village. This argument is partly justified by Tables 2 and A.2, where we see significant differences between the NCMS-villages and non-ncms villages. The argument is further confirmed in the matching procedure, in which balancing the observable village and household characteristics between the insured and non-exposed groups is much more difficult than balancing the characteristics between the insured and non-participant groups. 12 Second, consumption can grow more similarly among people living in the same geographic areas than among those living in different areas, particularly when the different areas have various incomes, and hence, consumption. To implement the DID framework for the panel data, the baseline model applies the fixed-effect regression that controls for both household and year fixed effects. All the 10 The effects of these policies can differ for the insured and non-participants. Thus, the estimate of the insurance effect has incorporated this difference. 11 Wagstaff et al. (2009) focus on the double difference between the insured and non-exposed to evaluate the effect of the NCMS on health care expenditure. Their choice is reasonable when the focus is a program evaluation, so that the gross effect may be more important. Furthermore, when the outcome is health care expenditure, avoiding the selection bias caused by the voluntary participation is crucial because the insured and non-participants have different expectations on future health expenditure. However, they also point out that the NCMS-villages and non-ncms villages are different, and they actually focus on the comparison between the insured and non-participants in the early versions of their paper. 12 In terms of reducing bias, matching the insured with non-participants in the NCMS-villages is considerably more successful than matching the insured with the non-exposed. 14
time-invariant effects of household characteristics are controlled by the household fixed effects, and the yearly time trend of consumption that is common to all households is controlled by the year fixed effects. Refinements such as matching DID and tests of the identification assumptions are discussed in Section 6. More specifically, the regression model for the double-difference comparison between the insured and non-participants is as follows: Y ijt Family _ insured it T D X t t i i ijt ijt, (1) where Y ijt represents the log value of household non-healthcare consumption for household i located in village j in period t. Family_insured it is the binary variable that indicates whether household i subscribes to the NCMS in year t. T t includes three year dummies. D i includes all the household indicators. X ijt includes the observable household and village variables that vary over time and may affect consumption and participation decision. Such variables include log(household income), household size, share of members over age 65, share of members under age 10, whether there are communist party members, whether households are officially categorized as poor ( Wubao households), and log(average income per person in the village). 13 In Eq. (1), γ measures the effect of the insurance coverage on consumption. The precautionary saving explanation indicates that >0. Given that we control for log(income), also represents the effect of the NCMS on the average propensity to consume non-healthcare expenditure because log(average propensity to consume) is equal to the difference between log(consumption) and log(income). A primary concern in this model is that the identification assumption may not hold even 13 Some of the family characteristics do not vary much over time. However, the results are highly robust to whether these variables are included. 15
after conditioning on the observable characteristics. The most likely situation is related to the adverse selection problem: households that expect substantial health expenditure in the next year are more likely to participate; hence, their consumption dynamics would have differed from that of households that have no such expectation were there no NCMS. By excluding health expenditures from consumption, we partially avoid the complication arising from the possibility that participants would spend more on health care than the non-participants would in the absence of the program. This selection bias tends to underestimate the precautionary saving motive because families who expect to incur huge health expenditure are more likely to be frugal in other consumption. To address the potential selection bias, we first use the self-reported health status to proxy the unobservable expectation on future health expenditure. Then, by adding the interaction term between year and measures of household health status, we allow households with different health statuses to have varied time trends in consumption. Health status can be affected by the insurance coverage. For the estimations in our sample, therefore, we use the self-reported health status in 2003 at which time none of the villages introduced the program. Similarly, we add the interaction between year and income to allow the linear time trends in consumption to vary with the income. This approach address the concern over the phenomenon that the insured are generally richer than the non-participants, and that different income groups may have varied income growth rates. 5 Results for the baseline model 5.1 Average treatment effect on the treated groups Table 3 reports the results for the baseline model that focuses on the double-difference comparison between the insured and non-participants. The first column assumes that all 16
households have the same counterfactuals of the time trends in consumption. It shows that the insurance coverage has stimulated non-health care consumption by 5.5 percent for the insured, an effect that is not negligible. Given that the average non-healthcare consumption per person for the participating households was about 2660.7 yuan in 2003, an increase of 5.5 percent implies an increase of 146.3 yuan, which is much higher than the total premium of the insurance that was typically 30 yuan in 2003 and 50 yuan in 2006. Moreover, the program more effectively stimulates consumption than do the cash transfers from the government because the average propensity of rural households to consume is only about 0.437. The result is quite robust when we relax the assumption by allowing the linear time trend in consumption to vary with the observable characteristics. In particular, column 2 controls for the interaction term between year and household income and that between year and village average income. 14 Column 3 additionally controls for the interaction term between year and initial household health status, which is measured by the share of members reporting fair or worse health status and share of members reporting poor health status in 2003. 15 Both columns show an insurance effect similar to that in the first column, including both the magnitude and significance level. Particularly, column 3 shows that after being covered by the NCMS, the consumption of the insured households increased by 5.6 percent or 149 yuan. The results are also similar when we allow the difference in trends to vary year by year through the control of the interaction terms between the year dummies and household income, village average income, and household health status in 2003 (column 4). Aside 14 When we use income in 2003 instead of income in the current year, the results are almost unchanged. 15 Results are similar when we additionally control for the share of members with good health statuses, or instead control the mean value of the health status in the household. Finally, we also consider the health status in the current year instead of the health status in 2003 to increase the number of observations. 17
from income and health status, other differences are observed between the insured and non-participants. As in the propensity score matching method, we can summarize the differences using a one-dimensional variable, a household s propensity score of joining the program. The estimation of the propensity score is discussed in detail in Section 6. Here, we control for the interaction between year and propensity score to allow the trend in consumption to vary with the propensity score. The estimate is shown in column 5, which shows a slightly stronger insurance effect on consumption. Controlling for the interaction between the year dummies and propensity score yields a similar estimate (column 6). 16 Given that an increasing number of counties and households enroll in the program, we have an unbalanced panel in Table 3. Table A.3 reports the estimates of the balanced panel, in which we have much fewer observations. The insurance effect is stronger. Particularly, the insurance effect on non-healthcare consumption is, on average, 9.6 percent after we allow the trend to vary with the income and health status in 2003 (column 3). The difference in the magnitude of the insurance effect between the balance and unbalanced panel can be attributed to the fact that the balanced panel has a higher proportion of experienced NCMS members who have participated in the program for more than one year, and at the same time, the insurance effect on consumption for the experienced members is stronger than that for the new members (shown later in the paper). In summary, the estimates of the positive insurance effects on non-healthcare consumption are robust to the specifications that enable the linear time trends in consumption to vary with the observable variables that are the important determinants of participation decisions. The results also withstand the other robustness tests shown in 16 The consistency of the estimate in the last two columns requires an additional assumption: the conditional expectation of the outcome given that the propensity score is linear. In addition, the standard errors here are not adjusted for the first-stage estimation of the propensity score. 18
Section 6. These tests increase our confidence that the baseline model provides a reliable estimate of the insurance effect. Therefore, we first examine the economic explanations for the estimates and heterogeneity of the effects, and discuss robustness thereafter. 5.2 Precautionary saving explanation This section examines whether the increase in consumption net of health expenditures represents the reduction in precautionary savings. All of the estimates in this section control for the household and year fixed effects, time-variant household and village characteristics, as well as the interaction term between year and household income, between year and average income in the village, and between year and household health status in 2003. Aside from the precautionary saving perspective, a potential explanation for the positive effect of the NCMS on non-healthcare consumption is that the insurance reduces out-of-pocket health payments, thereby leaving the insured households with more income for other consumption expenditure. This is a simple ex post crowd-in effect, a concept that is applicable only to the households that incurred health expenditure in the current year. However, column 1 of Table 4 shows that the insurance effect remains significant for households with no health expenditure in the current year. The magnitude is even stronger than that when we pool all households together, although the significance level declines. Therefore, the ex post crowd-in story cannot explain the positive insurance effect. 17 If the ex post crowd-in effect is the only explanation, then the higher non-healthcare consumption by the participants is caused only by their lower out-of-pocket healthcare expenditure, which implies that participation in the NCMS may have little effect on total 17 We have also estimated the effect of the NCMS on the out-of-pocket health expenditure. The result shows no significant effect, which contradicts the crowd-in perspective. However, we find that the insurance coverage stimulates more visits to health care facilities among the insured households. These findings are consistent with the results of the studies that evaluate the NCMS (Wagsaff et al., 2009; Lei and Lin, 2009; and Mao, 2005). 19
consumption. However, the second column of Table 4 shows that the insurance coverage stimulates total consumption by 6 percent, which is even higher than the effect of the NCMS on non-health care consumption. Another potential explanation is related to the income effect of government subsidies. Participants in the NCMS receive a government subsidy of 20 or 40 yuan for the premium payment. However, the effect of an income increase of 40 yuan on consumption is only about 17 yuan, implied by an estimated propensity to consume of 0.437. Even if households treat the subsidy as being permanent and the propensity to consume is around 1, the increase in consumption is no more than 40 yuan. The amount is much smaller than 149 yuan, our estimates for the insurance effect of the NCMS. This result strongly suggests that the income effect of subsidy is not the primary explanation. The precautionary saving explanation indicates that the insurance effect will strengthen when the insurance program becomes more generous, reducing the expenditure risk faced by the consumer. Columns 3 and 4 of Table 4 test this hypothesis by exploiting the detailed information on the NCMS program for 54 counties. The result shows that households respond to the generosity of the insurance scheme for the health expenditure at county facilities: the lower the deductible or the coinsurance rate, the more consumption the insurance program can stimulate. However, the generosity of the insurance scheme for the health expenditure at the village clinics does not exhibit such significant effects. This insignificant effect is somewhat reasonable given that most households seek health care services in county hospitals when they encounter serious health problems that most strongly demand the insurance. Finally, the precautionary saving explanation implies that the insurance effect is stronger for those who have a higher risk of incurring health care costs that are expensive 20
relative to income. Table 5 examines how the insurance effect varies with uncertainty about future health expenditure. We first look at the difference between income groups. Poor households are more likely to be unable to afford large health care expenditures than rich households; thus, their precautionary saving motive would have been stronger without the insurance program. As a result, the effect of the insurance on consumption should be stronger among the poor than among the rich. Column 1 of Table 5 confirms this conjecture: the positive effect of the NCMS on consumption decreases with income. The second and third columns run the regression separately for the bottom half (the poor) and top half (the rich) of the income distribution, and the insurance effect is significant only for the poor. The second part of Table 5 focuses on the risk related to household health status. On the basis of the self-reported health status of each member in a household, we construct two measures of household health status. Columns 4 to 6 consider the first measure of the health risk: whether at least one household member report fair or worse health status in 2003. In our sample, about 89% of individuals report good or excellent health. Therefore, reporting fair or worse health indicates serious health problems that may demand substantial health expenditure in the future. The results confirm that after being covered by the insurance, households that have members with fair or worse health status consume much more than do households with no such members. The second measure first assigns an ordinal value to each category of the self-reported health status: 5 for excellent, 4 for good, 3 for fair, 2 for bad, and 1 for no working capacity. It then calculates the average value of the health status of all the members in a household in 2003. Columns 7 to 9 yield similar results: the positive insurance effect on consumption decreases as the household health status improves, and the effect is significant only for the bottom half of the household health distribution (designated as poor health ). 21
5.3 Dynamics of the insurance effect and trust After the dissolution of the old CMS in the 1980s, most households in rural areas were not covered by any health insurance for a long period. Moreover, the NCMS differs from the CMS in many aspects. As a result, households need time to understand and establish trust in the new program. Column 1 of Table 6 controls for a dummy for the experienced participants who have participated for more than a year. This enables the insurance effect on the experienced members to be different from the effect on the new members who have participated in the NCMS for less than a year. The result shows that the insurance effect on consumption is significant among the new members (about 4.5 percent). Moreover, the effect among the experienced participants is much higher, and the difference is about 6.7 percentage points, which indicates that compared with the consumption of the non-participants, that of the experienced members increases by 11.2 percent because of the insurance coverage. 18 The dynamics of the insurance effect can result from the fact that the experienced members learned more about the benefits of the insurance program. However, an alternative explanation is that the NCMS coverage becomes more generous over time and people reduce precautionary saving in response to the rising generosity. Column 2 examines the issue by controlling for an interaction term between household insurance status and year. The result yields a negative answer to the alternative explanation because no significant increase in the insurance effect occurs over time. The third column further confirms that the difference between the new and experienced members remains significant after the time trend of the insurance effect on consumption is controlled for. 18 When we further allow the insurance effect to differ in the second, third, and fourth years of household subscription in the program, we find that the increase in the insurance effect occurs primarily in the second year of subscription. 22
Another explanation for the dynamics of the insurance effect is related to household trust in the program. The effect of knowledge on the program is double edged. Particularly, if households find that the alleged benefits of the program do not take effect, more knowledge about the insurance cannot reduce precautionary savings. Therefore, household trust in the insurance program may be the factor that matters most. To identify the trust effect, in column 3 we control for the interaction between household insurance status and the indicator of whether some households in the village have received some reimbursement from the NCMS ( village reimbursement =1 if yes; 0 otherwise). The result shows that the insurance effect on consumption becomes significantly stronger by 17 percentage points when the benefits of the insurance are experienced by the residents. For the villages that have not received any reimbursement, the insurance effect is even negative. 19 These results are confirmed by the succeeding estimates on the subsamples. Columns 4 and 5 consider only the villages that have not experienced any reimbursement. Herein, no significant increase in consumption is observed among the participants (column 4), regardless of whether the participant is experienced (column 5). By contrast, when we consider only the villages where some reimbursements have been received, the insurance program stimulates consumption by 6.3% on average for the insured households (column 6), which is higher than the average insurance effect of 5.6% reported in the baseline model. In addition, the experienced participants in these villages exhibit significantly more consumption than do the new members (column 7), which may be because they acquired more information about the insurance or because they accord more trust to the program. These results emphasize that only when households trust the program do they begin to 19 We need to be cautious in explaining the negative insurance effect here because the result is sensitive to specifications, and there are only a few villages around 18.6% from 2004 to 2006 that have launched the program but no household has received reimbursement. 23
reduce precautionary savings and consume more. The responses related to trust are also consistent with the pattern of participation decision. We find that for those who did not participate immediately after their villages introduced the program, the participation rate is 67% in the village that have received some reimbursement, but only 36% in the villages that have not received reimbursement. 6 Robustness, refinements, and gross effects 6.1 Counterfactual tests and other robustness tests To test whether the identification assumption of the baseline model holds, we apply the same model to the periods at which households were not covered by the insurance. In columns 1 to 5 of Table 7, we consider only the households that have not enrolled in the program in each year. In addition, household status of insurance coverage in period t is defined as the status in the succeeding period (t+1). That is, the treatment group in this model represents the households who are not insured in period t but are insured in period t+1. By construction, the NCMS should not affect household consumption in period t. The first column shows no significant insurance effect, as it should be. This is also true when we consider only households who have no health expenditure in the current year (column 2). These results enhance our confidence that the estimate of the insurance effect in the baseline model represents the causal effect of the NCMS. The next three columns confirm that the heterogeneity in the insurance effect also disappears. The primary concern of the baseline model is that households that enrolled in the NCMS are not comparable to households that did not. Particularly, we are alerted that households that have not participated by 2007 may have been covered by other insurance policies or have very special concerns regarding the participation. Therefore, we exclude 24