Financial Protection of a Rural Health Insurance Program in China

Transcription

1 Financial Protection of a Rural Health Insurance Program in China Julie Shi January 29, 2013 Abstract In 2003, the Chinese government launched a voluntary health insurance plan in rural area, called New Cooperative Medical Scheme (NCMS), aiming to prevent rural residents from becoming impoverished by health expenditure. This paper uses a national wide four-year panel data to address county selection problem, households participating behavior and the impact of the plan. The results show 1) Counties with greater fraction of rural households, higher rural income, higher government health expense and better facilities have higher probability of launching earlier than other counties. 2) Adverse selection was not severe on household level, although we found last year s out-of-pocket expense and head s age affect households participation. 3) The impact of the plan on reducing of out-of-pocket expense is greater for the rich than that for the poor, although on average was not significant. Key words: Health Insurance, NCMS, Out-of-pocket Expense I am grateful to Chong-en Bai for providing dataset, Randall Ellis, Wesley Yin and Daniele Paserman for helpful comments, and other seminar participants in Boston University. All errors are my own. Department of Economics, Boston University. 270 Bay State Rd, Boston, MA jushi@bu.edu 1

2 1 Introduction Does the introduction of health insurance to the uninsured reduce their out-of-pocket expenses? Are the effects the same between the poor and the rich? These are important questions for policy making especially in developing countries. The majority of the population receive low income, and some still live below the poverty line. Without health insurance, it is quite costly to receive health care. Thus people choose to either not seek treatment or borrow from others to pay for health care when they become ill or get injured. Both cases can decrease life quality. Health expense is one of the leading causes of poverty. China is one example of a country that faced these problems, and the rural residents suffer more because they have lower average income than urban residents. Gustafsson and Li [2003] showed that health expense caused 2.5% of households to fall below the poverty line in Liu et al. [2003] stated that medical expense had become an important source of transient poverty in According to the third National Health Service Survey in 2003, more than 79% of rural residents were not covered by any health insurance % of people who were sick did not seek health care, and 38.2% of this group claimed it was because they could not afford medical expenses. For inpatient care, the condition was worse. 70.0% of people who should have received inpatient treatment did not seek care because of the expense. To address these problems, several developing countries have established nation-wide health insurance during the last twenty years, with the aim of providing primary care and helping reduce out-of-pocket expense. However, it is unclear whether these insurance plans effectively achieve the goals that were originally set. Specifically, these 1 There was a health insurance in rural area prior to 1979, but it started to collapse when China began its economic reform. 2

3 insurance might do not improve utilization and/or reduce out-of-pocket expense. This paper sheds light on the reduction on expense by evaluating the impact of the New Cooperative Medical Scheme (NCMS), a rural health insurance established in China. NCMS was started in 2003 and heavily subsidized by the government. The primary goal was to prevent rural residents from becoming impoverished by health expense, and it aimed to increase accessibility of health care as well. The scheme started in 304 pilot counties (out of a total of 2860 counties) in 2003 and expanded rapidly. By 2010, almost all counties were covered by this program. NCMS was designed as a voluntary plan with the purpose of eliminating rural residents resistance toward paying money to the government. The reason was before the scheme there were several taxes and fees implemented by the government, and some of funds had been misused. In order to attract the healthiest individuals, government heavily subsidized the plan, so participants only needed to pay a small amount of money as premium. By using a nation-wide four year panel data, this paper attempts to answer three questions. First, some of counties were selected earlier than others to implement the scheme, and they were not randomly assigned, so it s natural to ask what kind of counties were selected first by the government. This question is interesting because we can detect government s preference between residents welfare and political reputation. From the one hand, it was easier to start the project in wealthy areas, where people have higher incentive to support it so government can collect the premium with less difficulty. In this way, government could gain a good reputation through issuing the scheme first in those places. From the other hand, people in poor areas were less protected than those in rich areas, so they typically had a higher need for this scheme. From the welfare perspective, poor areas should launch the plan before 3

4 wealthy areas. Second, which characteristics would affect households participation behavior? This is a voluntary scheme, under which households can choose whether or not to participate. A household level data is used to address the household adverse selection problem. In addition, we tested whether the selection was consistent over time. Third, how efficiently did the scheme protect households against medical care spending risk? Specifically, did the scheme significantly reduce households out-of-pocket expenses, and did the scheme impact the poor and the rich equally? Before NCMS, there were other attempts to provide health insurance, yet most of them failed because of limited financial resources and poor management. For NCMS, it is important to know whether people benefited from it or not, in the sense of an increase of utilization and/or a reduction of out-of-pocket expense, in order to at least predict whether the scheme could be sustained over a long period of time. (only the hypothesis of reducing out-of-pocket expense is tested in this paper, because the dataset does not include utilization information) Furthermore, the primary goal of NCMS was to prevent rural residents from becoming impoverished from health expenses, so it is important to test whether the impacts, if they were significant, were the same between the poor and the rich. However, the reimbursement was limited in the scheme with high co-payment rate, high deductible, low ceiling and narrow coverage. One concern is that the poor could not afford the co-payment even when they were insured, so the rich benefited more from the insurance, while they were not the group who were most in need of the plan. Contributions of this paper are: First, this research uses a nation-wide sample, compared to previous studies which used samples collected from only a few counties, so the result is more general and robust. Second, this is the first research to determine 4

5 which counties were selected earlier than others to launch the scheme. Third, this is the first research to compare nation-wide impacts on different income groups with the plan. The paper is organized as follows. Section 2 is a literature review. Section 3 is the introduction of NCMS. Section 4 describes methods. Section 5 presents the data. Section 6 shows the results and section 7 contains a conclusion and discussion. 2 Literature Universal health insurance was issued as early as a half century ago in developed countries and in recent two decades in developing countries, in order to protect people against health spending risk and improve health status. A substantial amount of research has been conducted on U.S. cases. Currie and Gruber [1996b] and Currie and Gruber [1996a] studied the effect of Medicaid. The former focused on pregnant women and found that the expansion of Medicaid improved health status, by reducing the incidence of infant mortality and low birth weight. The latter concentrated on low-income children and showed that eligibility for Medicaid significantly increased the utilization of medical care. Card et al. [2008] found a similar result for Medicare and showed that eligibility for Medicare caused a sharp increase in the use of health care services. In Canada, Hanratty [1996] found that the implementation of Canadian national health insurance was associated with an improvement in infant health outcomes. For developing countries, besides utilization and health status, research has paid attention to out-of-pocket expense as well. Health insurance played an important 5

6 role in reducing financial burden and preventing people from becoming impoverished. Both Jowett et al. [2003] and Sepehri et al. [2006] found that public health insurance in Vietnam significantly reduced out-of-pocket expenditure. Chen et al. [2007] showed that Taiwan s National Health Insurance has increased utilization among elderly while there was no significant effect on health status. Sosa-Rubí et al. [2009] showed that Seguro Popular (a program to finance health care for the poor) in Mexico had a significant impact on access to obstetrical services. Since 2003, when China issued NCMS, several researches have been conducted. These studies can be sorted into three categories: enrollment and expenditure and utilization. Wang et al. [2008] found that adverse selection existed in the scheme, and gender, socioeconomic status and other factors were significantly associated with enrollment. Zhang and Wang [2008] found a similar result and showed that health status and income significantly affect enrollment choice.the second category is expenditure and utilization. Wagstaff et al. [2009] concluded that when outpatient and inpatient utilization were increased, out-of-pocket expense was not reduced. Lei and Lin [2009] also found that the scheme had no impact on reducing out-of-pocket expenditure, increasing use of formal medicine or improving health status. Sun et al. [2009a] concentrated on catastrophic medical payment and claimed that outof-pocket medical payments remained a burden for rural households and financial protection was modest. In addition to the two categories above, Sun et al. [2009b] also studied prescribing behavior of village doctors on use of drugs, antibiotics and injections, and concluded that over-prescribing was common in villages and worse with NCMS health insurance. Besides NCMS, another government subsidized health scheme was launched in a town located in western China, called Rural Mutual Health Care (RMHC). Yip and 6

7 Hsiao [2009] determined that the scheme had a similar premium as the NCMS but provided first dollar coverage for primary care and a lower ceiling for hospital services. Similar researches had been done on it. Wang et al. [2006b] found that individuals with worse health had higher incentive to participate. Zhang et al. [2006] showed that community-level social capital (measured by reciprocity index) and individual-level social capital (measured by trust index) were significantly and positively associated with willingness-to-join. Wang et al. [2006a] found that net benefits (benefits net of payment) are not only between the rich and the poor, but also between the sicker and the healthier. Yip and Hsiao [2009] compared RMHC and NCMS, and concluded that the former was more effective at reducing medical impoverishment than the latter. Some research studies focused on different impacts of health insurance plans among different income groups is not a new finding. Currie and Gruber [1996b] found that earlier changes in Medicare eligibility, which were restricted to specific low-income groups, had much larger effects on birth outcomes than broader expansions to women with higher income levels. Both Jowett et al. [2003] and Sepehri et al. [2006] found that public health insurance in Vietnam reduced out-of-pocket expenditure more for the poor than for the rich. Wang et al. [2006a] is the most pertinent research for this paper. They found that richer/sicker participants obtained greater net benefits (benefits net of payment) than poorer/ healthier did and wealthy farmers benefited the most from the scheme at every level of health status. 7

8 3 NCMS NCMS was launched in 304 selected counties as pilots in (National Statistical Year Book on the NCMS, 2004). After acquiring experience from those areas, the scheme expanded rapidly. Figure 1 shows how fast the expansion was. There are 143 counties included in my sample. The fraction of newly issued counties was the same during the first three years. In 2006, more counties were added to the scheme. By the time the data was collected, NCMS was launched in 76% of all counties, while by the end of 2008, almost all counties were covered by the insurance scheme (The Fourth National Health Service Survey, 2008). NCMS operated at the county level. Each county constructed a risk pool, which was much larger than the village level pool in the old Cooperative Medical Scheme. Since counties varied significantly from one to another in China, it was not appropriate to set a single policy for all counties. Local governments were more informed about the features of their own counties than the central government, so local counties were assigned the role of policy maker in the scheme. For this reason, local governments had great freedom to design the policy, which included: household premium, coverage on different health care types (outpatient or inpatient), reimbursement rules and so on. The scheme was financed by contributions from the central government, local (provincial, county and township) governments and individuals. In 2003, the central government subsidized 10 RMB ($1.4) per individual in poorer central and western areas, and total amount of local governments subsidies were required to be no less than 10 RMB. Each individual had to pay 10 RMB as annual premium. In wealthy coun- 2 By the end of 2002, there were total 2860 counties in China 8

9 ties, local governments could choose to subsidize more than the minimal level, and the premiums could also be set higher than 10 RMB. Generally, the NCMS budget per participant was higher in rich areas than that in poor areas. The budget per person increased over time. In 2006, subsidy from both central and local governments increased to 20 RMB per individual while personal premium remained the same. In 2008, both central and local government subsidies increased to 40 RMB and individuals were required to pay 20 RMB per year. At the time the scheme was initiated, there were already several kinds of taxes and fees which created a heavy burden for rural residents, so they would have resisted the scheme if it had been compulsory. For this reason NCMS was designed as a voluntary scheme that allowed households to decide individually whether to participate. However, as mentioned previously governments heavily subsidized the scheme in order to make it affordable and attract low-risk individuals. Figure 2 represents households participation rate over time (the participation unit is household instead of individual), and it can be concluded that adverse selection was not a severe problem in NCMS. As residents became more familiar with the scheme through time, more and more chose to participate, which is indicated by two dimensions. One is that within the same county, participation rates increased. For the counties which lunched NCMS in 2003, the average participation rates were 62.76%, 81.03%, 93.18% and 95.21% in 2003, 2004, 2005 and 2006, respectively. The other is in different counties, the participation rates in the counties where the NCMS was first implemented increased. The first year participation rates were 62.76%, 71.06%, 80.14% and 91.92%, respectively, for counties that launched the scheme in 2003, 2004, 2005 and Reimbursement inculed three specific features. First, since the budget was small, the protection of the scheme was limited, performed as high deductible, low ceiling, high 9

10 copayment rate and narrow coverage. Second, benefit packages in rich counties were generally better than those in poor counties, because governments in rich areas were able to collect more money per participant. Third, benefit packages in lower level hospitals were better than those in higher levels (basically there are three levels of hospital in China: township health center, county hospital and higher level hospital), in order to encourage people to receive health care in lower level facilities. The national average copayment rates in township health center, county hospital and higher level hospital ranged between 45-60%, 50-70% and 55-80%, respectively (Evaluation Reports on China s NCMS, 2006). Figure 3 shows average deductibles in different level of hospitals in different regions. In county hospitals, the average deductibles were RMB 500, 300 and 200 in eastern, central and western areas respectively (Evaluation Reports on China s NCMS, 2006). For the coverage, the majority of counties only cover inpatient care in early years, for the purpose of protecting people against catastrophic illnesses. Other than that, some counties cover a few kinds of chronic illnesses in outpatient services, and some counties used household saving account to encourage people to enroll 3. For provider payment methods, most counties applied fee-for-service, only a few tried other alternatives. For example, some counties paid capitation (10 RMB per participant per year) to village doctors and township health centers for primary health services. Case-based payment systems were used in some counties. care 3 Household saving account usually contains 8-10 RMB which can be directly paid for outpatient 10

11 4 Methods 4.1 County selection Although there was no formal guideline for government to decide which county should be selected first, county characteristics were necessary to be considered before decision. Local governments had their own utility, which contained lots of features like government s expense, quality of facilities and the increase of residents welfare after being insured. If the expected utility of launching NCMS was greater than not launching it, government would opt to issue the scheme. However, the decision process was not clear from government documents. It could be central or provincial governments chose counties directly, while considering local governments interests at the same time. Or local governments showed their willingness to launch the scheme to provincial government first, and then provincial governments approved these cases. Here I assume it was the first case that counties were selected by higher level governments. A logit model is used to capture the behavior, since the dependent variable is a binary choice variable. The model is as follows: P (y c = 1 x c ) = G(x c β) (1) where y c is a dummy variable indicating whether county c is selected to launch NCMS (y c = 1 means the county is launched NCMS) in a given year, x c are a group of county level characteristics, which can represent capacity of local government, economic development level and the quality of local facilities. P (y c = 1 x c ) is the 11

12 probability of county c to launch NCMS conditional on county s characteristics x c. Assuming it has a logistic distribution, so logit model is applied. 4.2 Households participation Households participation decision is similar to governments decision. Households enroll in the scheme if utility of participation is greater than that of staying uninsured. A logit model is used: P (y i = 1 x i ) = G(x i β) (2) where y i is a dummy variable indicating whether a household enrolled in NCMS (y i = 1 means the household enrolls NCMS) in a given year, x i are a group of households characteristics. P (y i = 1 x i ) is the probability of household i to participate conditional on household s characteristics x i. Assuming it has a logistic distribution, so logit model is applied. This approach is used to detect households decision making during the first year when NCMS was launched. The reason is households had to decide separately each year whether to participate or not, so factors affected participation decisions could be changed through time. In the first year when the scheme was newly introduced, social economic status, like income and education, and other factors, like health status, may play important roles. However, in following years, people could be better informed through their own experience or hearing from others, so this information may be more important in participating decision. Regressions are run separately for the four years, and can identify year-specific and 12

13 common factors that influence the participation. Though, the reason of factor change is not clear. There are two possibilities that could cause this phenomenon. First, residents in later launching counties could learn information about the scheme through media or directly from people in already issued counties, causing their consideration to change. Second, the county selection mechanism distinguished characteristics of counties, making different groups diverse in their decision making. 4.3 Scheme impact on out-of-pocket expense A panel data regression is used to test whether health insurance significantly reduced participatants out-of-pocket expenses. The model is as follows: y it = θ t + parti it β + x it γ + c i + ε it (3) where y it is household i s out-of-pocket expense in year t, parti it is a dummy variable indicating whether household i participated NCMS in year t or not (parti it = 1 if household i enrolled in NCMS in year t), x it is a group of household level characteristics and c i is household fixed effect. However, since counties are not randomly selected and this is a voluntary program, there may be some unobserved characteristics in the error term affecting both participation decision and outcome. Hence, an instrumental variable is used for parti it to solve the endogeneity problem. The IV is a dummy variable that indicates the issuance status of the county household i belong to in year t. It equals to 1 if household belongs to a county that has issued NCMS in the current year. It is correlated with households participation choice. In this way, this should be an efficient IV becaue only households in issued counties can 13

14 choose to participate. Although county level characteristics do affect which county was selected first to issue NCMS, each household s characteristics in the error term are uncorrelated to county level selection. In this sense, it s a valid IV. In order to test whether the effects are different among different income groups, another regression can be applied: y it = θ t + income it α + parti it β + income it parti it γ + x it δ + c i + ε it (4) where y it, parti it, x it and c i have the same meaning as equation (3), income it is a group of dummy variables indicating which income group households i in year t belongs to. Four dummies are generated: income1 it, income2 it, income3 it, income4 it. Households were sorted by per capita income each year, and were distributed to different income groups. income1 it indicates the lowest income group. Specifically, income1 it = 1 indicates household i in year t belong to the lowest income group. As the endogeneity still presents, the same IV will be applied to this regression. 5 Data Three datasets are used for this analysis. Two of them came from a survey conducted in China in 2007 through the investigation system of the Ministry of Agriculture villages (one out of each county) were randomly selected and two questionnaires 4 The system was founded in About 300 villages were randomly chosen all over the country to join the system. Households in these villages were paid to record their daily behavior. Each village was assigned an investigator to collect information from households. In this way, panel data can be collected year by year. 14

15 were distributed. One was designed for households to collect information on participation, health care and social economic status. For this part, 5756 households (approximately individuals) were included. The other was for the investigators in villages. They were required to collect correspond information about the county the village belonged to. The third dataset was from the system s annual survey. Since they all came from the same investigating system, they ccould be easily merged together. This is a retrospective survey, which requires households to recall their previous 4 years behavior. Since households in these villages kept their daily record, it was not difficult for them to cover the questionnaire accurately. For the county level questionnaire, the investigator is required to collect data from county s yearbook. In this way, a 4-year panel data was collected from 2003 to is the first year to launch NCMS, and 2006 is the latest year we can survey at the time the data was collected. Since the rate of hospital visit for rural residents was very low, households who spent more on health care were over-sampled in order to get more non-zero observations for outcome. Households were ranked in their villages by average health expenditure through 2004 to 2006, and from the first one-third part a proportion of fout-fifth were randomly selected, while from the last two-third a proportion of one-second were selected. In the following analysis, observations are weighted in regressions when necessary to accommodate. Although 5756 questionnaires were distributed, 5728 questionnaires (22125 individuals) in 143 villages were returned, so the attrition rate is about 0.5% for household survey. For county survey, 100 questionnaires were returned, so the attrition rate is 30% for county survey. 15

16 Table 1 reports the descriptive statistics of county level variables in Only statistics in 2003 are shown in this table, because statistics in 2004 are similar. Due to limited sample size, this analysis only considers about the cases in 2003 and There are 78 observations in 2003, and 68 in The number of observations in 2004 is less than 2003 s, because counties introduced NCMS in 2003 were dropped out from the sample in The first row in the table shows dependent variable select, and it equals to 1 if the county was launched NCMS in The second part are counties basic conditions, including population, GDP and government income and expense. The third part are characteristics of county level hospitals which capture the quality of facilities in those counties. [Table 1 should be here] Table 2 reports the descriptive statistics of household level variables. Although sample size is 5728 for the household survey, observation of balanced panel data is only for 4 years. There are two reasons: The first is that original matched dataset was not perfectly balanced. The investigation system changes a small fraction of villages each year, either including new villages or dropping out old villages. In micro-level, the system kept records for all households living in these villages, so residents were added in when they moved in and dropped when they moved out. The second reason is that I dropped observations whose key variables, such as household s participation status, were omitted. [Table 2 should be here] The dependent variable of household adverse selection is parti, which equals to 1 if household participates the scheme in that year. The average participation rate in each year is already shown in the introduction of NCMS part, so I skip it here. The first row shows the dependent variable l_expense, which is natural log of 16

17 one plus household out-of-pocket health expense. The index contains expenses from both outpatient and inpatient care (including doctor visiting fee, surgery fee and payment for beds, while drug fee is not included). The following part are a group of independent variables, including: (1)demographic conditions: family size, head s age and gender (male=1), (2)social economic status: per capita income, head s education and occupation, assets, (3)household s members health status: head s self-report health status (1 indicates the best and 5 the worst), the fraction of people within household whose age is less than 12 or greater than 60, the fraction of people whose can t work, (4)other factors which may affect the demand of health insurance: the fraction of people who stayed at home less than 180 days in a year. 6 Results 6.1 County selection Table 3 shows the results of county selection. Instead of coefficients, odds ratios (OR) are presented for interpretation. Each column represents ORs in a single model, and model 1 through model 6 are similar ones with slightly change in control variables. The first three models are for 2003 and the last three are for [Table 3 should be here] In 2003, the ORs for fraction of rural households are , which indicates that there is a 16% to 25% increase in the odds of being selected to launch NCMS if the fraction increases 1%, after controlling for all other variables. This result is 17

18 significant at p=0.05 level. The OR for population is slightly less than 1, while it s not significant. We can see counties with more rural households launched the scheme first when other conditions are the same. The reasons are: first, more rural households lead to higher demand, since most of them hadn t been covered by any health insurance before; second, more households means a larger risk pool, so it would be safer on finance than small counties. The OR for rural income is in model 3 and is significant, which means when rural income increases by 100 RMB, the odds of launching NCMS would increase by 9%. As mentioned before, the government faces a trade off between issuing the scheme in high or low income counties, where the former means safer and the latter can generate more social welfare. From the result we can see the government was conservative and it focused more on the success of the scheme in the first year. Decision makers prefer to select high income areas, where the incentives to participate the scheme are higher. Government income of the counties doesn t significantly affect the selection while government expense on health does. The ORs of government expense on health in model 2 and model 3 are 1.23 and 1.19, which means when health expense increase 0.1 million RMB, odds of being introduced the scheme would increase 19-23%. The implementation of the scheme is quite complex. Government needs to hire and train new workers and buy new equipment during the process. For local government, a high level of health expense means sufficient financial sources to implement the scheme and also subsidy participants. However, this result can either be due to a high historical health expense to guarantee government can afford NCMS, or only comes from sudden increase of health expense because of issuing the scheme. The two effects are entangled in model 1 through model 3 since the data on previous year s 18

19 government expense was not obtained in Number of physicians and beds for patients represent provider s size and quality, and both of them have positive effects. This is natural because the better the facilities are, the better they can handle increased demand, so the higher probability these counties are selected to issue the scheme. Situations in 2004 are different than those in Only the size of population and fraction of rural households affect selection of second-wave counties. The OR for population is larger than 1, which means second-wave counties have more population than those uninsured counties in The OR for fraction of rural households is less than 1, which means this variable has negative effect on selection. Other ORs are not significant, including last year s government expense on health, and pseudo R 2 s in model 4 to model 6 are much lower than those in model 1 to 3. The reason is after observing the scheme performed well in first-wave counties, the government gained confidence and reputation on NCMS. The selection may be not as prudent as it in 2003, so it became more random. In this sense, all models in 2004 have less predictive power than models in Households participation Table 4 shows results on households participation decision. Odds ratios and standard errors are presented in the table. Robust standard errors are presented in the parentheses. Column1 to column 3 display ORs in regressions in year 2003, 2004 and 2005 respectively. The three models are the same except last year s health expense is added as a control variable in 2004 and in I skip analysis in 2006 for two 19

20 reasons. The one is participation rate was more than 90% in 2006, and even 100% in some villages. Only a few exceptional households did not enroll in the scheme, so features that affected the decision may not be covered by social economic status or health conditions. Hence including the year introduces a lot of noise, making it more difficult to identify. The other reason is for regression technique, some variables predict dependent variable perfectly, so these observations were automatically dropped by software. This is also the reason why in 2004 and 2005 the numbers of newly-issued counties are not less than the number in 2003, but observations in the two years are less than that in [Table 4 should be here] Generally speaking, the model doesn t have much prediction power. Probably because premium is low and participation rate is high, adverse selection is not severe in the scheme. Income level did not have significant effect on participation. In these years, households only need to pay premium at 10 RMB per person, so it s not a heavy burden for them. Last year s health expense has positive and significant effect both in 2004 and If last year s health expense increases 1%, the odds ratio of participation will increase 0.19% and 0.17% in 2005 and in 2006 respectively, after controlling other variables. Head s age has positive effect in On the one hand old people care more about family members health than the young does, so we suppose to observe a positive effect. On the other hand young heads accepted a new program more easily. Because of the two effects, coefficients are not significant in early years, but is positive in 2005 when NCMS in not that new. I still control for other variables in the model, such as family size, fraction of people whose health is not good or who can t work are also included in the model 5, but none of them are 5 This variable is not significant maybe because those family already bought some commercial health insurance before NCMS, so they don t need to participate this one 20

21 significant. 6.3 IV estimation Table 5 shows how the scheme impacted on out-of-pocket health expense. Four model are shown with coefficients and standard errors. Model 1 and model 2 display results in equation (3), that aims to detect the average impact on out-of-pocket expense. Model 1 runs the regression without IV, and model 2 with IV. Model 3 and model 4 display results in equantion (4), that add interaction of participation and income group. As we mentioned above, these models can test whether the scheme had different impacts on expense between the poor and the rich. Model 3 runs regression without IV, and model 4 with IV. Fixed effect is assumed in all models, and time trend, head s health condition and head s occupation are also controlled. Since adverse selection existed in the scheme, results in Model 2 and Model 4 are more reliable, so interpretaitons are based on the two models. [Table 5 should be here] In model 2, the coefficient on participation is and is not significant, which means the scheme had no impact on households out-of-pocket health expense, when controlling for other variables. One thing needs to point out is the result is the average treatment effect in 4 years through all counties. In fact, different counties can have different effects since they obtain different policies. The budgets in later years are larger than those in earlier years, so generally treatment effects are larger in later years. I also noticed that coefficient on participation in model is negative and significant. However, I don t have a good example in reality to explain why 21

22 the coefficient is underestimated when endogeneity exists. One possibility is that people s preference toward health care changed through time. They might get some information thourgh media or some other places that both caused them spend less on health (for example, using generic drug instead of imported drug) and also encouraged them to enroll in health insurance. Other than participation, the effects of per capita income, fraction of people in household whose health is not good or who can t work and asset are all positive and significant. These indicate that household with more per capita income, more people with bad health or more assets spend more on out-of-pocket health expense. Model 3 and model 4 distinguish the impact on the poor from that on the rich. pincome1 is omitted in the regression, so coefficients of product parti*pinocme2, parti*pincome3 and parti*pincome4 displays the difference of impacts on expense between the poorest and other groups. Coefficient of parti*pincome4 is in model 4, which demonstrates that the percentage reduction of out-of-pocket expense is 56% more for the rich than those for the poor on average. The effects of per capita income, fraction of people in household whose health is not good or who can t work and asset are all positive and significant, that are consistent with results in previous two models. 7 Conclusion and discussion This paper first detects government s preference in launching the scheme. The results show that counties with greater fraction of rural households, higher rural income, higher government health expense and better facilities had higher probability 22

23 to launch the scheme earlier than other counties. Higher fraction of rural households represents higher need of health insurance and larger pool for finance. Higher income usually means higher incentive to participate the scheme, since it s voluntary. Meanwhile, the scheme was heavily subsidized by central and local government, and counties with higher government health expense reflected their higher attention on residents health status, so these counties may have higher incentive to insure their residents. The result is not a surprise. At the time when NCMS was initiated, government didn t have much reputation among rural residents, not only because they had not been covered by any national health insurance for a long time, but also because taxes and fees had made a heavy burden for them. Government tried several ways in order to gain success on the scheme. One of them is to issue the scheme in rich counties first, where the resistance was less than that in poor areas. Another is to make the scheme voluntary instead of compulsory. The second finding is that even though the scheme is voluntary, adverse selection is not severe. Average participation rate in first-wave counties is more than 63%, it keeps increasing and as time goes. In 2006, the rate is more than 90%. The reason is that premium was not high and did not make a financial burden for households, so they were willing to participate the scheme. Other than that, I still find households with higher last year s health expense and older head s age had more incentive to enroll. Economists always worry about adverse selection when a schmem is voluntary, however, the effect can be negligible when the subsidy is large enough, or when almost everyone needs it. Third, the reduction of out-of-pocket expense is more for the rich than that for the poor because of the scheme, even though the average impact is not significant. This 23

24 finding is consistant with previous research. Wagstaff et al. [2009] found that outof-pocket expense had not been reduced on average, and Wang et al. [2006a] stated that richer/sicker participants obtained greater net benefits (benefits net of payment) than poor/ healthier did. It is conclude that the protection of the scheme against health spending risk was limited. That means the primary goal of NCMS protect people against impoverishment was not achieved, because financial burden was not reduced especially for the poor. Combined with previous research, a further conclusion can be made. Wagstaff et al. [2009] claims that the poor and the other group 6 have disparities on utilization. Inpatient visits for the poor increased less than those for the other group at township health center and county hospital. It can be inferred that the poor are less protected and may be less benefited than the rich from the scheme, when noting that NCMS mainly cover inpatient care. The limitation of this inference is they used a sample from only 10 counties. Since each county can make its own policy, the result may be not suitable to be applied in my paper. Other developing countries can draw lessons from China s experience. First, compulsory is not a necessary feature for health insurance. Combined with other policies, adverse selection can be a minor problem. Second, sometimes policy effect may not meet the purpose as it originally stated. The primary goal of NCMS is preventing people from impoverishment by health expense from catastrophic illness. Since poor people face more risk in impoverishment than the rich when they spend the same expense, the scheme should protect more for the poor than the rich. From the result I find the effect is the opposite: the rich are protected more than the poor. One possible reason is that budget is too small so the poor can t hold the expense even 6 In their paper, they compared the poorest 20% and the other, not the same subgroups 24

25 after reimbursing from the scheme. In this way, health care remained unaffrodable for the poor, and only the rich benefited from the scheme. For other countries, policies need to be considered carefully in order to guarantee that scheme impact meets the desiged purpose. 25

26 References D. Card, C. Dobkin, and N. Maestas. The impact of nearly universal insurance coverage on health care utilization and health: Evidence from medicare. American Economic Review, 98(5): , L. Chen, W. Yip, M. cheng Chang, H. sheng Lin, S. dye Lee, Y. ling Chiu, and Y. hsuan Lin. The effects of taiwan s national health insurance on access and health status of the elderly. Health Economics, 16: , August J. Currie and J. Gruber. Health insurance eligibility, utilization of medical care, and child health. Quarterly Journal of Economics, 111:431 66, May 1996a. J. Currie and J. Gruber. Saving babies: The efficacy of cost of recent expansions of medicaid eligibility for pregnant woman. Journal of Political Economy, (104): , 1996b. B. Gustafsson and S. Li. Expenditures on education and health care and poverty in rural china. China Economic Review, 15: , M. J. Hanratty. Canadian national health insurance and infant health. American Economic Review, 86(1): , March M. Jowett, P. Contoyannis, and N. Vinh. The impact of public voluntary health insurance on private health expenditures in vietnam. Social Science and Medicine, 56(2): , January X. Lei and W. Lin. The new cooperative medical scheme in rural chian: Does more coverage mean more service and better health? Health Economics, 18:S25 S46, Y. Liu, K. Rao, and W. C. Hsiao. Medical expenditure and rural impoverishment in china. Journal of Health Population and Nutrition, 21(3): , September M. of Health (PRC). National statistical yearbook on the new cooperative medical scheme. Beijing, M. of Health (PRC). Research on national health services: an analysis report of the fourth national health services survey in Beijing, A. Sepehri, S. Sarma, and W. Simpson. Does non-profit health insurance reduce fiancial burden? evidence from the vietnam living standards survey panel. Health Economics, 15(6): , June

27 S. G. Sosa-Rubí, O. Galárraga, and J. E. Harris. heterogeneous impact of the "seguro popular" program on the utilization of obstetrical services in mexico, : A multinomial probit model with a discrete endogenous variable. Journal of Health Economics, 28:20 34, X. Sun, S. Jackson, G. Carmichael, and A. C. Sleigh. Catastrophic medical payment and financial protection in rural china: Evidence from the new cooperative medical scheme in shandong province. Health Economics, 18: , 2009a. X. Sun, S. Jackson, G. A. Carmichael, and A. C. Sleigh. Prescribing behaviour of village doctors under china s new cooperative medical scheme. Social Science and Medicine, 68(10): , May 2009b. N. P. E. Team. Evaluation reports on China s new cooperative medical scheme. Beijing: People s Medical Publishing House, A. Wagstaff, M. Lindelow, G. Jun, X. Ling, and Q. Juncheng. 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, January H. Wang, W. Yip, L. Zhang, L. Wang, and W. C. Hsiao. Community-based health insurance in poor rural china: The distribution of net benefit. Health Policy, 76 (2): , April 2006a. H. Wang, L. Zhang, W. Yip, and W. C. Hsiao. Adverse selection in a voluntary rural mutual health care health insurance scheme in china. Social Science and Medicine, 63(5): , September 2006b. H. Wang, D. Gu, and M. E. Dupre. Factors associated with enrollment, satisfaction, and sustainability of the new cooperative medical scheme program in six study areas in rural beijing. Health Policy, 85:32 44, W. Yip and W. C. Hsiao. Non-evidence-based policy: How effective is china s new cooperative medical scheme in reducing medical impoverishment? Social Science and Medicine, 68(2): , January L. Zhang and H. Wang. Dynamic process of adverse selection: Evidence from a subsidized community-based health insurance in rural china. Social Science and Medicine, 67(7): , October L. Zhang, H. Wang, L. Wang, and W. C. Hsiao. Social capital and farmer s willingness-to-join a newly established community-based health insurance in rural china. Health Policy, 76(2): , April

28 Figure 1: Accumulated number of counties that were issued NCMS in each year Figure 2: Average participation rate in each year in different counties 28

29 Figure 3: Average deductibles in three level hospitals in estern, central and western regions in

30 Table 1: Description of county level variables in 2003 Variables label Unit Obs Mean Std. Dev. Min Max select =1 if issued NCMS in popu population 10, hh_rural Number of rural households 10, frac_hh_rural fraction of rural households income_rural rural per capita income 1 RMB 77 2, , ,156 6,915 gov_income government income 100 Million RMB pc_g_income per capita government income 100 RMB g_h_exp government expense on health 100 Thousand RMB pc_h_exp per capita government expense on health 100 RMB beds* # of beds for patients ,386 beds_occup beds occupation rate phys # of physicians *: variables beds, beds_occup and phys are from county level hospitals 30

31 Table 2: Description of household level variables from 2003 to 2006 Variables description Obs Mean Std.Dev. Min Max log_health natural log of one plus household 18, out-of-pocket health expenses log_pincome natural log of one plus per capita income 18, famsize family size 18, headage head s age 18, headsex head s gender 18, headedu head s years of education 17, age12 fraction of people aged less than 12 18, age60 fraction of people aged more than 60 18, health4* fraction of people whose health is not good or who can t work 18, health5 fraction of people who can t work 18, , homeday180 fraction of people who stayed at home less than 180 days in a year asset** asset 18, head_health head s self-report health status headoccup*** head s occupation 18, *: five categories of self-report health status: 1 is the best and 5 is the worst **: assets=1 if household obtains any of color TV, fridge, motor car, mobile phone, computer or car ***: 8 categories of occupation: 1: household labor for agriculture, 2: household labor for non-agriculture, 3: labor force employed by others, 4: cooperated entrepreneur, 5: private company owner, 6: government officer, 7: workers for education/science/medical/arts department, 8: misc. 31

32 Table 3: Odds ratio on logit model in county selection Variables select OR (Model 1) OR (Model 2) OR (Model 3) OR (Model 4) OR (Model 5) OR (Model 6) Popu * 1.02 * 1.03 ** (0.02) (0.03) (0.03) (1.01) (0.01) (0.01) frac_hh_rural 1.16 ** 1.25 ** 1.19 ** * 0.95 * (0.08) (0.12) (0.10) (0.03) (0.03) (0.03) income_rural * (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) gov_income (0.32) (0.41) (0.49) (0.07) (0.11) (0.10) g_h_exp 1.23 ** 1.19 ** 0.99 (0.10) (0.09) (0.03) g_h_exp(n-1) 1.02 (0.04) phys 1.01 ** 1.01 ** (0.00) (0.00) beds 1.01 * (0.00) (0.00) (0.00) (0.00) Pseudo R Obs Standard errors in parentheses. * Significant at 10%; ** significant at 5%; *** significant at 1%. 32