THE IMPACT OF THE NEW COOPERATIVE MEDICAL SCHEME IN RURAL CHINA: DO THOSE WHO LIVE FAR FROM A MEDICAL FACILITY BENEFIT MORE FROM NCMS PARTICIPATION?

Transcription

1 THE IMPACT OF THE NEW COOPERATIVE MEDICAL SCHEME IN RURAL CHINA: DO THOSE WHO LIVE FAR FROM A MEDICAL FACILITY BENEFIT MORE FROM NCMS PARTICIPATION? A Thesis submitted to the Faculty of the Graduate School of Arts and Sciences of Georgetown University in partial fulfillment of the requirements for the degree of Master of Public Policy By Corey Lipow, B.A. Washington, DC April 15, 2010

2 Copyright 2010 by Corey Lipow All Rights Reserved ii

3 ACKNOWLEDGEMENTS Gratitude is due to my thesis advisor Yuriy Pylypchuk and to Eric Gardner for assisting with the data programming. iii

4 THE IMPACT OF THE NEW COOPERATIVE MEDICAL SCHEME IN RURAL CHINA: DO THOSE WHO LIVE FAR FROM A MEDICAL FACILITY BENEFIT MORE FROM NCMS PARTICIPATION? Corey Lipow, B.A. Thesis Advisor: Yuriy Pylypchuk, Ph.D. ABSTRACT Since the first pilot NCMS s in 2003, the New Cooperative Medical Scheme has offered catastrophic illness and other medical coverage to rural residents in China. Yet few studies have examined whether the scheme actually improves health status, utilization of health care services, and the burden of health care expenditures. This study employs data from three rounds of the China Health and Nutrition Survey to further examine this issue. For the analysis, I employ non-pooled and pooled Ordinary-Least-Squares regression and linear probability models. I also include an interaction term between the NCMS and distance from a medical facility to determine whether the effect of the NCMS differs by distance. I find that participating in the NCMS reduces the probability of being sick or injured in the past four weeks at most distances from a medical facility, and that as distance increases after approximately 8.6 minutes, so does the negative effect of NCMS participation on sick or injured in the past four weeks. The NCMS also appears to increase utilization of preventive care but does not increase utilization of other professional medical care. In two models, I find that the NCMS decreases the cost of preventive care at most distances from a medical facility, and that as distance from a medical facility increases after approximately 9.4 minutes, so does this reduction in costs for NCMS participants. However, in another model the NCMS appears to increase the cost of preventive care, counteracting the previous effect. NCMS participation reduces iv

5 treatment costs in one pooled model, but the estimate is not very robust. This study finds no other indication that the NCMS reduces the burden of health care expenditures for rural residents. v

6 TABLE OF CONTENTS Introduction...1 Literature Review...6 Data...10 Conceptual Model...12 Methodological Approach...14 Part I: Constructing the NCMS Indicator Variable...14 Part II: OLS and Linear Probability Model for 2004 Wave...19 Part III: Pooled OLS and LPM...22 Results...24 Part I: Descriptive Statistics and Differences in Means...24 Part II: Interpretation of Non-Pooled OLS and LPM Regressions...28 Part III: Interpretation of Pooled OLS and LPM...33 Conclusion...36 Appendix...41 Bibliography...57 vi

7 LIST OF TABLES Ia. Year and Number of Communities that First Implemented Cooperative Insurance in Jiangsu Province...15 Ib. Trend in Communities First Implementing Cooperative Insurance in Shandong and Hubei Provinces, and Whether Communities Were Determined As Having Implemented the NCMS...17 II. Descriptive Statistics 2000 Wave...41 IIIa. Descriptive Statistics 2004 Wave...43 IIIb. Descriptive Statistics by Distance 2004 Wave...45 IVa. Difference in Means between Rural Residents Enrolled in NCMS and Those without Insurance...47 IVb. Differences in Means between Rural Residents That Live 15 Minutes or More from a Medical Facility and Those That Live Within 15 Minutes from a Medical Facility (Far Near)...49 Va. Regression Estimates 2004 Only...50 Vb. Effect of NCMS on Probability of Sick or Injured in the Past 4 Weeks by Distance from a Medical Facility...30 VI. Pooled Regression Estimates 2000 and 2004 Waves...53 VII. Effect of Implementation of NCMS in 2005 or 2006 for Pooled OLS...56 vii

8 Introduction When China adopted the New Cooperative Medical Scheme (NCMS) in 2003, 80% of people in rural areas had no health insurance of any kind (Wagstaff et al 2009). By February of 2008, the NCMS had expanded to 730 million farmers and about 86% of all counties in China (Wen 2008). The NCMS is a voluntary health insurance program that is administered at the local level with funding from the central government, the local government, and individuals. The government not only frames the NCMS, but it also monitors the scheme and acts as the main provider of subsidies (Zhengzhong 2006). The NCMS was originally enacted to address the dearth in coverage for rural residents, which had declined from as many as 90% of rural residents in 1978 under the Cooperative Medical Scheme (CMS), the insurance scheme under Mao Zedong, to 20% in the early 2000s (Yip et al. 2008). This significant decline in insured was largely the result of the collapse of collectives in rural areas the collective economy in the early 1980s during which most villages lost their welfare funds and could no longer afford to provide the CMS. The government began implementing the NCMS in 2003 with the goal of providing health coverage to all of China s residents by 2010 (Lei and Lin 2009). The NCMS differs from the old CMS in several distinctive ways. The NCMS operates at the county level rather than the village level, the latter being the level at which the old system operated. The NCMS is voluntary, which raises the issue of self-selection bias in studies evaluating the scheme. Furthermore, the NCMS varies across counties in terms of design and implementation different counties have different copayments and deductibles, and counties can choose whether to emphasize inpatient or outpatient expenses over and above the catastrophic coverage requirement. Some counties offer 1

9 inpatient reimbursement and a household savings account which can be used for outpatient expenditures (World Health Organization 2004). The 2002 State Council Policy Document No. 13, Decisions of the State Council on Strengthening Rural Healthcare, gives a few specific guidelines for the NCMS that each county must follow, including that participation in the NCMS must be voluntary and that the administration must be conducted at the county level (Lei and Lin 2009). In addition, the NCMS should concentrate on catastrophic illness. This requirement has received a degree of criticism as concentrating on catastrophic illness has resulted in high deductibles and copayments, which could prevent poor rural residents from obtaining needed health services. The World Health Organization (WHO) has argued that because only a small number of farmers are likely to be hit by a catastrophic illness, the NCMS will not address the majority of health issues for the rural population. You and Kobayashi 2009 tabulated data from the Ministry of Health that put the percentage of farmers benefiting from the NCMS at about 3.0% of total participants in the east and 3.7% of total participants in central China. Sun et al 2008 estimated that 8.98% of rural households in Linyi County in 2004 fell into catastrophe determined as having medical expenditures at 40% or more of a household s capacity to pay (which itself was determined as total household disposal income minus total subsistence expenditure). Zhengzhong 2006 put hospitalization for severe diseases at 3.6%, with that percentage pertaining to all rural residents and not just those covered by the NCMS. Another provision of the NCMS is that an entire household must participate, not just one or two individuals in a household participating while other members opt out. This provision was intended to reduce adverse selection, or the tendency for sick 2

10 individuals to opt into the program and overwhelm the insurance pool, and healthier individuals to opt out due to rising costs. With an entire household participating, not just the sicker individuals would opt into the program the healthier members of the household would be required to opt in as well. Recognizing that adverse selection may still be an issue despite the household requirement, in some poorer provinces the central government has required a minimum 80% of eligible households to participate in the NCMS before matching the contributions of local governments (You and Kobayashi 2009). The financing of the NCMS involves a threefold effort in which the individual, the local government, and the central government all contribute. During the period from 2003 to 2005, the minimum requirement was a 10 RMB per beneficiary household contribution, a 20 RMB subsidy from the local government, and a 20 RMB matching subsidy from the central government for households in the poorer central and western provinces. According to Wagstaff et al 2009, the total NCMS budget averaged about 62.9 RMB per beneficiary in a sample of 189 counties and 17 provinces, indicating that local governments contributed more than the 20 RMB subsidy. Moreover, the NCMS budget varied considerably with respect to local income and program design. The 62.9 RMB per beneficiary average represents approximately twenty-five percent of average per capita health care spending in rural areas. This low reimbursement rate means that the burden of health care spending even for those enrolled in the NCMS is placed mostly on the individual seventy-five percent of health care spending in rural areas is done by households or individuals. The limited financing is reflected in high deductibles and services that are only partially covered or not covered at all (Wagstaff et al 2009). 3

11 This study examines the effect of the NCMS on several indicators of health status, utilization of health care services, and the burden of health care expenses. The indicators of health status include whether an individual was sick or injured in the past four weeks and whether an individual had difficulty carrying out daily activities due to illness in the past three months. The indicators of utilization of health care services include whether an individual who had identified as sick or injured in the past four weeks utilized health care services and whether an individual utilized preventive care services in the past four weeks. The out-of-pocket indicators are treatment cost so far, money spent on illness or injury, and preventive service cost in the past four weeks. All of the variables used in the analysis come from the China Health and Nutrition Survey (CHNS). When possible, this paper examines the effect of the NCMS and distance from a medical facility on health outcomes, utilization of health care services, and the burden of health care expenditures. This paper looks at whether the NCMS has a varied effect on health status, utilization, and the burden of health care expenditures based on distance from a medical facility. Unfortunately, the data being employed does not have a consistent measure of distance across survey rounds. For 2000 and 2004, distance is measured in minutes; in 2006, distance is measured in kilometers. Therefore, this study looks at the effect of distance only for 2000 and For 2006, this study looks at the effect of NCMS participation on the dependent variables without looking at the effect of distance or the effect of an interaction term between the NCMS and distance from a medical facility. I deduced two hypotheses for the varied effect of the NCMS based on distance. The first hypothesis is that the NCMS would have little effect for rural residents who live 4

12 far from a medical facility, because they would be disinclined to make the trip and utilize health care services regardless of enrollment status. Rural residents who live far from a medical facility might be disinclined to utilize health services because of the long distance they would have to travel, their lack of knowledge about health insurance, and/or the high deductibles and copayment rates of the NCMS. The second hypothesis is that the NCMS would have more of an effect for those who live far from a medical facility, as uninsured individuals who live far from a medical facility would be much less likely to utilize health services. 5

13 Literature Review The following analysis builds on two recent studies in the field: the first is Wagstaff et al 2009 and the second is Lei and Lin These studies find mixed evidence that the NCMS improves health outcomes, utilization of care, or the burden of health care expenditures for rural residents. The survey that each study employs in its analysis is different; Wagstaff et al 2009 uses data from the National Health Service Survey (NHSS), administered by the Center for Health Statistics and Information (CHSI) of the Ministry of Health, while Lei and Lin 2009 utilizes the CHNS (the same survey employed in this analysis). The NHSS collects data from approximately 54,000 households in 900 villages with counties, townships, villages, and households chosen using multi-stage stratified random sampling. Wagstaff et al 2009 uses the 2003 round of the NHSS and a follow-up survey implemented by the CHSI in 2005 that explicitly evaluated the effect of the NCMS program. The follow-up survey used the same households from the 2003 NHSS for counties in which the NCMS was implemented (10 counties in total), as well as five additional counties that had not implemented the NCMS between 2003 and 2005 as a control group. The CHNS, in contrast, was conducted in several rounds, including 2000, 2004, and The CHNS has data from before any county implemented the NCMS (2000), when the NCMS had just begun being implemented (2004), and several years after the first pilot NCMS s had been implemented (2006). The additional years involved in the CHNS makes it more appealing for the current study. Wagstaff et al 2009 found that the NCMS had an appreciable positive impact on the utilization of both outpatient and inpatient services, with broadly similar percentage 6

14 impacts for both. The study discerned between the poorest 20% of the sample and the rest of the sample and noted that the poor saw a larger increase in outpatient care after enrolling in the NCMS while the rich saw a larger increase in inpatient care, but only for the increase in inpatient care was the difference statistically significant. The NCMS did not appear to have an impact on out-of-pocket expenditures, however. The authors suggest that the increase in utilization could explain the lack of a reduction in household out-of-pocket expenditures in counties that implemented the NCMS. Additionally, the cost of a typical outpatient visit was not reduced by the NCMS, which could indicate that services became more expensive as a result of the scheme. Lei and Lin 2009 found that enrollment in the NCMS improves preventive-care utilization but does not improve anything else, such as access to inpatient care or the health statuses of individuals. They also found no evidence that NCMS participation leads to relieved financial burden, as measured by patients out-of-pocket expenditures in the last four weeks. They attribute these results to any one of the following: (1) deductibles tended to be high, which might have prevented utilization of more formal medical care; (2) the household medical savings accounts in many counties could have been sufficient for preventive care expenses, but not for other medical expenses; and (3) some counties allowed participants a free preventive service once per year, provided that the participants had not used any medical services paid for by the NCMS in that year. While this paper builds off the results of these two studies, other studies have incorporated the NCMS in some way. Sun et al 2008 centered on the effect of the NCMS on Catastrophic Medical Payments (CMPs) by households in Linyi County in Shandong Province. In their study, catastrophic medical payments by households were 7

15 8.06 times households capacity to pay (again, determined as total household disposal income minus total subsistence expenditure) before NCMS reimbursement, and 6.34 times households capacity to pay after NCMS reimbursement. Sun et al 2008 concluded that although the NCMS reduced CMPs, out-of-pocket expenditures were still too harsh for the rural households and a disaster in the context of their meagre lives. A study by Yip et al (2008) examined the effect of Rural Mutual Health Care (RMHC) on utilization of care. The RMHC is an insurance intervention program that provides first-dollar coverage for outpatient and inpatient services. The authors found that the RMHC increased utilization of outpatient services and reduced the probability of self-medication. They compared the RMHC to another government program with medical savings accounts and high-deductible catastrophic coverage that they refer to as the NCMS, and found that the alternative program had little effect on either outpatient services or self-medication. 1 An analysis of the pilot program of the NCMS by Zhengzhong (2006) outlined equity issues and compared expenditures and reimbursement rates across regions in China. There is thus little evidence that the NCMS improves utilization of more formal medical care or out-of-pocket expenses for poor rural residents. While Wagstaff et al 2009 found that the increase in outpatient services for the poor was not statistically significant, Lei and Lin 2009 observed no increase in either outpatient or inpatient services. Both papers found little if any evidence of reduced out-of-pocket expenditures for rural residents. Lei and Lin 2009 found no evidence of improved health statuses of individuals due to the NCMS. 1 I am grateful to Lei and Lin 2009 for directing me to this working paper. 8

16 In terms of distance, few studies have assessed the sensitivity of utilization to distance to a medical provider in rural areas. Erlyana 2008 concluded that uninsured people in Indonesia who live in rural areas are more sensitive to distance from a medical facility, while uninsured people who live in urban areas are more sensitive to price. For those with a higher income, the effect of distance from a medical facility lessens; and for the insured, there does not appear to be an effect of distance on demand for ambulatory care services. Müller et al 1998 assessed distance decay effects on attendance rates at a health center in Wosera, Papua New Guinea. They found that the effect of distance was highly statistically significant, and that the effect of distance was differently influenced by gender in different age groups. Gender differences in distance effects were most pronounced in adolescents. Müller et al also examined whether distance effects were influenced by major diagnosis groups, and found that for the two most prevalent diagnosis groups malaria and acute respiratory infection (ARI) distance effects did differ. This paper does not employ a complicated interaction term between the NCMS, distance, and gender or income level. The only interaction term employed in this paper is between the NCMS and distance from a medical facility. 9

17 Data The data set used for the analysis is the China Health and Nutrition Survey (CHNS). The CHNS is divided up into seven rounds including the years 2000, 2004, and 2006, which are of primary interest to the study because the NCMS began implementation in The CHNS is a joint project by the Carolina Population Center at the University of North Carolina at Chapel Hill and the National Institute of Nutrition and Food Safety at the Chinese Center for Disease Control and Prevention. Each round of the survey took place over a 3-day period using multistage, random cluster sampling. The sample included over 4000 households with a total of between 15,000 and 19,000 individuals for each round. The sample covers nine provinces that vary substantially in geography, economic development, public resources, and health indicators, according to the CHNS website. 2 The nine provinces are Heilongjiang, Liaoning, Shandong, Jiangsu, Henan, Hubei, Hunan, Guizhou, and Guangxi. They are all in the eastern part of China. The data set includes community-level data as well as individual-level and household-level data. For the community-level data, a community questionnaire was collected from a knowledgeable respondent on community infrastructure, services, population, prevailing wages, and other variables. The community-level data is essential to the study because although the individual-level data records whether an individual had Cooperative Insurance, it does not differentiate between the old CMS and the current NCMS. Although the community-level data does not explicitly mention the NCMS, the data provides information on when Cooperative Insurance was first made available for a community. If the community first implemented Cooperative Insurance before 2003,

18 we can deduce that Cooperative Insurance refers to the old CMS. If the community first implemented Cooperative Insurance in or after 2003, we can for the most part expect that Cooperative Insurance refers to the NCMS. 3 If an individual lived in a community that, from the community-level data, was deduced as having implemented the NCMS, and the individual reported that s/he was enrolled in cooperative medical insurance, then that individual was defined as having been covered by the NCMS. Another issue with the data is that the unit of analysis alternates between the individual and household for survey data that was not gathered at the community level. For some of the variables used in this study, data was recorded for individuals, such as whether an individual had health insurance. For other variables, data was recorded for households. An example of a variable that was recorded at the household level is distance from a medical facility. Each household gave the various distances from medical facilities where individuals in that household sought health care services. Thus, I had to coordinate between the household-level variables and the individual-level variables. 3 Part I of the Methodological Approach section provides more detail about how I determined whether a community had implemented the NCMS. 11

19 Conceptual Model There are a few conceptual issues involved in addressing the impact of the NCMS. One is moral hazard. If the NCMS changes people s behavior so that they are more accidentprone, the impact of the NCMS could be exaggerated as a result. For example, a construction worker who knows that his or her health care is covered (at least to some degree) by the NCMS may be more liable to have accidents at the work site. Thus the increase in utilization of health care services is not necessarily due to increased access to care; it may instead be due to peoples changing responses to risk. They may take more risks as a result of being enrolled in the NCMS. I do not consider moral hazard as a significant problem for the current study because of the generally risk-free lifestyle of rural residents, most of whom make their living by agricultural means. Many of these residents are also from low-income households and may not be able to afford health care services even with NCMS enrollment. The deductibles and copayments of the NCMS may simply not be enough for some low-income households, so that enrollment in the NCMS is not likely to substantially, if at all, change the behavior of individuals in such households. A provision of the NCMS is that enrollment is voluntary. People who decide to enroll might therefore share some characteristic such as an aversion to risk or disbelief in the rural health care system at large that differentiates them from others who do not enroll in the NCMS. This is a self-selection issue: the risk-aversion factor or a disbelief in the rural health care system would influence both enrollment in the NCMS and the measures of health status, utilization rates, and health care expenses. Although self-selection biases could contaminate the results of this study, most rural residents living in a community 12

20 that has implemented the NCMS decide to enroll. The percentage of rural residents living in communities that implemented the NCMS in 2003 or 2004 who decided to enroll was 77.94% (see Table IIIa). The percentage of rural residents living in communities that implemented the NCMS between 2004 and 2006 who decided to enroll was just under 95%. Thus it would seem the majority of rural residents decide to enroll when their community implements the program. Yet it may be that the small percentage of those residents who decide not to enroll share an aversion to risk or disbelief in the rural health care system at large that could compromise the results. The high percentage of rural residents participating in the NCMS suggests that risk aversion or a disbelief in the rural health care system are not widespread in the rural population and thus probably not a significant factor in peoples decision not to enroll in the NCMS. If communities that implemented the NCMS in 2003 or 2004 shared a characteristic such as an older population who tended toward utilizing health services more than the general rural population, endogeneity issues may be present in this study. The increased utilization rates by people enrolled in the NCMS versus those not enrolled, if determined, would then be attributed not to the NCMS but to the older population of those enrolled in the NCMS. It seems unlikely that the older age of NCMS participants represents an endogeneity issue. For the 2004 round of the CHNS, the average age for those enrolled in the NCMS was 47.1 years, versus 46.3 years of age for all communities surveyed (see Table IIIa). This difference is not statistically significant. Furthermore, I controlled for age in both the non-pooled and pooled models. Nevertheless, communities that implemented the NCMS may share other characteristics that also affect the dependent variables. 13

21 Methodological Approach For identifying the effect of the NCMS on several indicators of health status, utilization of health care services, and the burden of health care expenditures, I employ non-pooled and pooled Ordinary Least-Squares (OLS) regression and linear probability models. Part I: Constructing the NCMS Indicator Variable. To construct a variable indicating enrollment in the NCMS, it was first necessary to coordinate between the community-level data and the individual-level data, as has been outlined in the Data section. The individual-level data does not distinguish between the old CMS and current NCMS; it merely shows whether individuals had cooperative medical insurance. Fortunately, the community-level data records the first year that Cooperative Insurance became available. Since the NCMS was first implemented in 2003, if a community recorded first implementing Cooperative Insurance in or after 2003, one might reasonably assume that Cooperative Insurance referred to the NCMS and not the old CMS. In several provinces, however, the years immediately preceding 2003 saw similar numbers of communities first implementing Cooperative Insurance as 2003 or afterward. For these provinces, I could not determine that Cooperative Insurance first made available in 2003 or afterward referred to the NCMS and not the old CMS or some other insurance scheme. For example, in Jiangsu province five communities reported first implementing Cooperative Insurance in 2001, four communities reported first implementing Cooperative Insurance in 2002, and six communities reported first 14

22 implementing Cooperative Insurance in See Table Ia for an extended comparison. Table Ia. Year and Number of Communities that First Implemented Cooperative Insurance in Jiangsu Province Year No. of Communities The above table shows when either a rural or urban community in Jiangsu province first implemented Cooperative Insurance. Since the years immediately preceding 2003 saw similar numbers of communities first implementing Cooperative Insurance as in 2003 or 2004, I could not deduce with any degree of certainty that Cooperative Insurance 15

23 first implemented in 2003 or 2004 referred to the NCMS and not the old CMS model or some other insurance scheme. 4 Alternatively, if the years immediately preceding 2003 saw lower numbers of communities first implementing Cooperative Insurance than in 2003 or afterward, I defined the communities that first implemented Cooperative Insurance in or after 2003 as having implemented the NCMS. Shandong and Hubei provinces serve as two examples. According to the community-level data, Shandong province saw one community first implementing Cooperative Insurance in 2001, one in 2002, two in 2003, four in 2004, six in 2005, and one in The uptick in Cooperative Medical Insurance right around the time the first pilot NCMS s were being implemented indicates that the Cooperative Insurance first implemented in Shandong province in 2003, 2004, and 2005 were the NCMS and not the old CMS. Similarly, Hubei province saw one community first implement Cooperative Insurance in 1968, ten in 2005, and one in I deduced from the uptick in Cooperative Insurance in 2005 that those ten communities had implemented the NCMS. The following guideline summarizes how I defined communities as having implemented the NCMS versus as having implemented the old CMS model or some other insurance scheme: if there was a recognizable uptick in communities that first implemented Cooperative Insurance in 2003 or afterward, all communities included in the uptick were defined as having implemented the NCMS. If only one community in the province first made available Cooperative Insurance in 2003 or afterward, regardless of whether the years preceding 2003 saw any communities first implementing Cooperative Insurance, then that one community was not defined as 4 So far I have not distinguished between rural and urban communities. Later, when dealing with the individual-level and household-level data, I distinguish between rural and urban communities and exclude the latter communities from the analysis. 16

24 having implemented the NCMS. Table Ib shows the trend in communities first implementing Cooperative Insurance in Shandong and Hubei provinces, as well as whether I defined those communities as having implemented the NCMS or as having implemented some other insurance plan. Table Ib. Trend in Communities First Implementing Cooperative Insurance in Shandong and Hubei Provinces, and Whether Communities Were Determined As Having Implemented the NCMS Shandong Province Year Frequency Defined as Having Implemented the NCMS? No No Yes; both communities defined as having implemented NCMS Yes; all 4 communities defined as having implemented NCMS Yes; all 6 communities defined as having implemented NCMS No Hubei Province Year Frequency Defined as Having Implemented the NCMS? No Yes; all 10 communities defined as having implemented NCMS No Because there was no recognizable uptick in communities implementing Cooperative Insurance in Jiangsu province following 2002 (see Table Ia), none of those communities were defined as having implemented the NCMS. In contrast, there were recognizable upticks in number of communities first making available Cooperative Insurance in Shandong and Hubei provinces; all communities included in the uptick in either province 17

25 were defined as having implemented the NCMS. A key drawback to this study is the lack of concrete information about whether communities implemented the NCMS, the old CMS, or some other insurance program. After I defined which communities implemented the NCMS, I determined individuals who resided in those communities and reported having cooperative medical insurance as having enrolled in the NCMS. I then excluded urban residents from the analysis. For the 2004 wave, 660 rural residents were determined as having enrolled in the NCMS, from three provinces: Heilongjiang, Shandong, and Hunan. These three provinces vary considerably with respect to region: Heilongjiang is in the northernmost part of China, Shandong is in central China, and Hunan is in the south. 18

26 Part II: OLS and Linear Probability Model for 2004 Wave A non-pooled OLS model and linear probability model are employed for examining the effect of the NCMS and distance from a medical facility on the dependent variables. In both models, an interaction term between the NCMS and distance from a medical facility shows whether the NCMS varies by distance. The equation for this model is: Y ic = ß 0 + ß 1 NCMS ic + ß 2 TRAVELMIN ic + ß 3 NCMS TRAVELMIN ic + ß 4 X ic + u ic The dependent variable is the measures of health status, utilization of health care services, and out-of-pocket expenditures for individual i in community c; NCMS is a dummy variable for if the individual enrolled in the NCMS; TRAVELMIN indicates distance from a medical facility in minutes; NCMS TRAVELMIN ic is an interaction term between enrollment in the NCMS and distance from a medical facility; X ic represents observable characteristics including household income, age, gender, marital status, education, and ethnicity; and u ic is the error term. The interpretation of ß 1 is a bit tricky. For the dependent variable, ß 1 shows the effect of participating in the NCMS on the dependent variable when TRAVELMIN=0. In other words, ß 1 shows the effect of participating in the NCMS when the distance from a medical facility is 0 minutes. For a person to live 0 minutes from a medical facility, s/he would have to live in the medical facility or so close to it that it takes nearly 0 minutes to travel there. For people who live at any distance from a medical facility other than 0, the interaction term NCMS TRAVELMIN is employed in the analysis. At 15 minutes from a medical facility, the effect of participating in the NCMS on the dependent variable would be estimated as ß 1 + ß 2 (15). 19

27 An issue with the OLS model is that it does not control for unobservable characteristics such as risk aversion or a more active lifestyle. These unobservable characteristics are included in the error term u ic but could be correlated with the enrollment in the NCMS and the dependent variable. More risk-averse people may be both more inclined to enroll in the NCMS and more inclined to utilize health services. Therefore, the effect of the NCMS on utilization would be overestimated; the estimate on NCMS would be higher than if risk-aversion had been incorporated into the model and thus controlled for. Those who live a more active lifestyle may be both more inclined to enroll in the NCMS and healthier, so that the estimate on NCMS would be more negative than it would otherwise be for the dependent variable sick or injured in the past four weeks. Also, the OLS model does not resolve endogeneity issues. If a community s implementation of the NCMS is based on a sicker overall population in that community rather than the health status of the population being influenced by the NCMS, a multiple regression model such as the one described above would not provide accurate estimates on the independent variables. Whether or not a county implements the NCMS could also be correlated with the percentage of people living in poverty. As people in poverty tend to be sicker or utilize heath services less, such a correlation would again bias the estimates on the independent variables. However, a difference in means test for monthly wage/salary between rural residents enrolled in the NCMS and those without insurance is not statistically significant at any conventional alpha (see Table IVb). As well, a difference in means test for age between rural residents enrolled in the NCMS and those without insurance is not statistically significant at conventional alphas. Neither monthly 20

28 wage/salary nor age appears to have influenced implementation of the NCMS or enrollment in the NCMS. Some of the dependent variables are binary variables, such as the health variables indicating whether an individual was sick or injured in the past four weeks or an individual had difficulty carrying out daily activities due to illness in the past three months, and the utilization variable, What did you do when you felt ill? (=1 if sought care from a health worker or doctor). For these variables, I employ a linear probability model, where the coefficient estimates show a change in the probability that the dependent variable equals one. The model is an adaptation of the OLS model given earlier: P(y=1/x) = ß 0 + ß 1 NCMS + ß 2 TRAVELMIN + ß 3 NCMS TRAVELMIN + ß 4 X + u Several of the dependent variables are binary variables and so the linear probability model will be used frequently for the regressions. The same issues regarding unobservable characteristics and endogeneity that could bias the coefficients in the OLS model are extant in the linear probability model. 21

29 Part III: Pooled OLS and LPM To incorporate two years into the analysis, I employ pooled OLS. The main years of interest are 2000 and Again, pooled OLS does not address the unobserved effects or endogeneity issues. The pooled OLS model with the interaction term is the following: Y ict = ß 0 + ß 1 D04 t + ß 2 NCMS ict + ß 3 TRAVELMIN ict + ß 4 NCMS TRAVELMIN ict + ß 5 X ict + u ict where D04 is a dummy variable. D04 equals zero if the time period is 2000 and D04 equals one if the time period is D04 controls for overall changes in the dependent variable across survey years. NCMS ict indicates whether individual i in community c and time period t participated in the NCMS. While I also briefly examined the 2006 CHNS data, the variable for distance from a medical facility was measured differently in 2006 than it was in previous rounds and so I did not include it in the 2004/06 pooled OLS model. In 2006, distance from a medical facility is given in kilometers, while in 2000 and 2004 it is given in minutes. It would be difficult to equate distance from a medical facility in kilometers with distance from a medical facility in minutes. The pooled OLS model for the years 2004 and 2006 is: Y ict = ß 0 + ß 1 d06 t + ß 2 NCMS06 ict + ß 3 NCMS04 ict + ß 4 X ict + a i + u ict The variable TRAVELMIN is not included in this model, nor is the interaction term between enrollment in the NCMS and distance from a medical facility. The dummy variable d06 t equals zero if the time period is 2004 and one if the time period is The dummy variable NCMS06 indicates whether an individual in a community that first implemented the NCMS in 2005 or 2006 participated in the NCMS (or equivalently whether an individual first enrolled in the NCMS in 2005 or 2006). The coefficient on NCMS06 is a measure of the difference in the dependent variable between individuals 22

30 who first enrolled in the NCMS in 2005 or 2006 and those that were uninsured or had other insurance. The dummy variable NCMS04 indicates whether an individual first enrolled in the NCMS in 2003 or Not including NCMS04 in the model would mean that all individuals who first enrolled in the NCMS in 2003 and 2004 would be grouped together with the uninsured and those covered by other insurance plans. In other words, without NCMS04 in the model, the coefficient on NCMS06 would show the difference in the dependent variable between individuals who first enrolled in the NCMS in 2005 or 2006 and the following: the uninsured, those who had other insurance and individuals who first enrolled in the NCMS in 2003 and Controlling for NCMS04 allows us to compare the effect of NCMS participation for individuals that first enrolled in the NCMS in 2005 and 2006 to individuals who were not enrolled in the NCMS during that period. 23

31 Results Part I: Descriptive Statistics and Differences in Means The descriptive data for 2000 and 2004 are given in Tables II and IIIa. From Table IIIa, it appears that rural residents participating in the NCMS are less likely to be sick or injured in the past four weeks than the overall rural population as well as the uninsured. Approximately 12.08% of rural residents enrolled in the NCMS reported that they had been sick or injured in the past four weeks, while 14.98% of all rural residents surveyed for the 2004 wave and 14.42% of uninsured rural residents reported that they had been sick or injured in the past four weeks. Table IVa shows whether the difference between those enrolled in the NCMS and those without insurance is statistically significant at the 10%, 5%, or 1% levels. It turns out that the difference in sick or injured the past four weeks is not statistically significant at any of those levels, but it may still be notable. Similarly, those enrolled in the NCMS saw a decline in difficulty carrying out daily activities during the past three months. 5.28% of enrollees reported that they had experienced difficulty carrying out daily activities during the past three months, versus 7.14% of the rural population and 7.07% of the uninsured. Again, the difference is not statistically significant at the 10%, 5%, or 1% levels. For the utilization variable, What did you do when you felt ill?, rural residents enrolled in the NCMS appear more likely to see a local health worker or doctor than the general rural population or the uninsured. However, the difference is again not statistically significant at the conventional levels. The difference in means for only one dependent variable was statistically significant at conventional levels: that was preventive care utilization in the past four 24

32 weeks. Utilization of preventive care services increased from 1.76% of uninsured rural residents to 4.15% of those enrolled in the NCMS in The difference is statistically significant at the 10% level. The descriptive statistics in Table IIIa also indicate that those enrolled in the NCMS pay much more for treatment of their illness or injury. The average amount spent on treatment so far for NCMS participants was yuan, versus yuan for the overall rural population and yuan for the uninsured. Higher costs of treatment for NCMS participants may reflect a higher income among participants, increased uptake or an increase in the cost of services for those who are insured. The difference is not statistically significant. Even if it was, a higher income among NCMS participants does not appear to affect the data; Table IIIa indicates that NCMS participants might even make less than the uninsured (although the difference in monthly wage is not statistically significant). The higher costs of treatment could have to do with the type of treatment offered or a higher utilization of treatment among NCMS participants. Additionally, the sample size is somewhat small, which could compromise the estimates; only 36 NCMS participants reported on treatment costs. Many of the independent variables are statistically significant at conventional alphas for the differences in means. In particular, differences in gender, marital status, ethnicity, and education level are all statistically significant. A higher proportion of women, people with spouses, those of the Han ethnicity, and those without a vocational or university degree enroll in the NCMS compared to the uninsured. The uninsured therefore exhibit a higher proportion of men, single individuals (divorced, widowed, 25

33 separated, or never married), national minorities, and those that have received a vocational or university degree. The higher proportion of people with spouses among NCMS participants indicates that people in more stable households opt into the NCMS program. Here we can observe a possible self-selection issue. If people in more stable households decide to enroll in the NCMS and are more inclined to utilize health care services, then a perceived increase in utilization attributed to the NCMS may actually be due to living in a more stable household. That the uninsured exhibit a higher proportion of those with higher levels of education largely reflects that way that I constructed the education variable. It does not necessarily indicate that the uninsured are more educated. In fact, if I had reconstructed the variable such that it equaled one if the highest education level achieved was either graduation from primary school or lower middle school degree, and zero otherwise, the difference in means between the uninsured and NCMS participants would have been = Constructing the education variable in this way, the uninsured would have a lower proportion of those with higher levels of education than NCMS participants. Table IIIa shows that approximately 2.59% of rural residents were enrolled in the NCMS in It also shows that approximately 77.94% of rural residents living in communities that implemented the NCMS in 2003 or 2004 enrolled in the NCMS. Another notable statistic in Table III is the difference in minutes to a medical facility. The mean number of minutes to a medical facility for NCMS participants is 12.72, while for the uninsured it is The difference is statistically significant at the 10% level, 26

34 but not at the 5% level. On average, the uninsured appear to live farther from medical facilities than NCMS participants. This raises issues of whether the NCMS reaches rural residents who live far from medical facilities, or rural residents who live in areas that are more isolated. Table IVb provides the differences in means between rural residents that live 15 minutes or more from a medical facility and those that live within 15 minutes from a medical facility (Far Near). The table has a few significant results. One is the increase in utilization of health care services for people that live 15 minutes or more from a medical facility. These results are peculiar because I would expect distance from a medical facility to be negatively correlated with utilization of health care services. The results signify that those who live farther away from a medical facility utilize services more than those who live within 15 minutes. Müller et al 1998 suggest that distance effects can be reversed by social factors, such as stigmatizing diseases, so that people may prefer to utilize more distant facilities. Table IVb indicates that preventive service costs are reduced for NCMS participants that live more than 15 minutes away from a medical facility versus NCMS participants that live within 15 minutes from a medical facility. The reduction in costs could reflect reduced utilization of health care services for individuals who live more than 15 minutes away from a medical facility. 27

35 Part II: Interpretation of Non-Pooled OLS and LPM Regressions The results of the non-pooled OLS and linear probability models are provided in Table Va. For the dependent variable, sick or injured in the past four weeks, the coefficients on the NCMS, minutes from a medical facility, and the interaction term are all statistically significant at the 5% level. The coefficients on minutes from a medical facility and the interaction term are statistically significant at the 1% level. At first glance it might appear that the coefficient on the NCMS dummy variable implies that participation in the NCMS increases the probability of being sick or injured in the past four weeks by percentage points. However, the estimate on the NCMS dummy variable indicates the effect of participating in the NCMS on the probability of being sick or injured in the last four weeks when minutes from a medical facility equal zero. There were only 12 observations out of 515 total (2.3%) for which TRAVELMIN=0. The negative coefficient on the interaction term implies that as distance from a medical facility increases, the effect of NCMS participation on the probability of being sick or injured in the past four weeks first decreases in magnitude, reaches zero, and then becomes more negative. Taking the mean distance from a medical facility among NCMS participants from Table IIIa and plugging it into the equation, we find the effect of NCMS participation to be: (12.72) = The model says that, holding the other variables in the model fixed, such as age and gender, for an individual who lives minutes from a medical facility, participating in the NCMS is expected to decrease the probability of being sick or injured in the past four weeks by 9.76 percentage points. This gap in the probability of sick or injured in the past 28

36 four weeks between those enrolled in the NCMS and those not enrolled in the NCMS becomes larger as distance from a medical facility increases. NCMS participation decreases the probability of being sick or injured in the past four weeks more at larger distances from a medical facility than at shorter ones. The linear probability model has a clear limitation. At large distances from a medical facility, such as 100 minutes away, the effect of NCMS participation on the probability of being sick or injured in the last four weeks becomes illogical. Plugging 100 into the model for TRAVELMIN yields: (100) = = The interpretation is this: holding the other variables in the model fixed, for a person who lives 100 minutes from a medical facility, participation in the NCMS reduces the probability of being sick or injured in the past four weeks by 239 percentage points, on average. The model is clearly flawed, most likely because it assumes the relationship between the interaction term and the probability of being sick or injured in the past four weeks as linear. In contrast, this relationship is likely to be non-linear; possibly, the effect of participating in the NCMS on the probability of being sick or injured in the past four weeks peaks at some distance from a medical facility and afterward declines. This model assumes that the effect of the NCMS keeps on growing as distance from a medical facility increases. The following table shows the effect of NCMS participation on sick or injured in the past four weeks at various distances from a medical facility. 29