The Largest Insurance Program in History

The Largest Insurance Program in History: Saving One Million Lives Per Year in China Abstract: The developing world is adjusting to rising medical spending risk through expanded use of insurance. The largest single insurance expansion in the developed world was the National Cooperative Medical Scheme (NCMS) that was rolled out in China from 2003-2008, which provided comprehensive insurance to rural Chinese, and which at its peak covered 800 million people. Despite the size and importance of this program, there is little convincing evidence on its health impacts. We combine aggregate mortality data with individual survey data, and identify the impact of the NCMS from program rollout and heterogeneity across areas in their rural share. We find that there was a significant decline in aggregate mortality, with the program saving more than one million lives per year at its peak, and explaining 78% of the entire increase in life expectancy in China over this period. We confirm these mortality effects using micro-data, and expand our analysis to show correspondingly large effects on health care utilization and other health outcomes. We show that there were sizeable reductions in out-of-pocket spending at the top of the expense distribution. We also document a key potential mechanism through which the program improved health was encouraging increased private investment in public health care facilities. 1 Medical spending risk is already one of the greatest sources of financial risk in developed countries of the world. Fortunately, virtually every developed country has either government or privately-provided insurance programs that cover the vast majority of the population. In 2018, across all OECD nations, as a result, the share of health care costs borne out of pocket amounts to only 20.1% of health care costs, and 1.7% of incomes. 1 In developing nations, health insurance coverage of the population is more sporadic. Traditionally, this has not been a priority for development because both health care quality and spending were low. In 2018 for example, across all low-income countries (LIC), 44.1% of health care spending was out of pocket – and this amounted to 3.45% of incomes. 2 The need for health insurance has changed dramatically in the developing world in recent decades as higher quality medical techniques have been extended worldwide. The benefit of improved medical access in the developing world is one key source of dramatic improvements in life expectancy around the world. Over the past two centuries, every region of the world has seen its life expectancy double; the global average life expectancy today is higher than any country in 1950. 3 Even the least developed countries have seen their average life expectancy grow by almost 25 years since 1960. 4 But this improved 1 20.1%: https://stats.oecd.org/Index.aspx?ThemeTreeId=9. This is an unweighted average across all OECD countries. 1.7%: https://stats.oecd.org/Index.aspx?DataSetCode=AV_AN_WAGE. This is an unweighted average across all OECD countries (Australia, Colombia and Turkey are removed since they are missing data). We take per capita OOP costs and divide by average annual wage in each country and average these percentages. 2 44.1%: https://data.worldbank.org/indicator/SH.XPD.OOPC.CH.ZS?locations=XM. This is a simple average of OOP as a percentage of health care expenditure of low-income countries as designated by the world bank. The data for Syria is from 2012 and for Yemen is from 2015. 3.45%: https://data.worldbank.org/indicator/NY.ADJ.NNTY.PC.CD?end=2016&locations=XM&start=2016&view=bar. This is an unweighted average across all low-income countries as designated by the world bank (Eritrea, North Korea, Somalia, South Sudan, Syria, and Yemen are removed due to incomplete information). We take per capita OOP costs and divide by adjusted net national income per capita in each country and average these percentages. 3 https://ourworldindata.org/life-expectancyglobally#:~:text=Globally%20the%20life%20expectancy%20increased,more%20than%20twice%20as%20long. 4 https://data.worldbank.org/indicator/SP.DYN.LE00.IN?locations=XL 2 medical care comes at a cost. Total medical spending in low-income countries have risen from 4.2% to 5.3% of GDP over the past 20 years. 5 As medical costs rise, there has been a dramatic expansion of public insurance programs in developing countries. Table 1 lists the introduction of new insurance programs in some of the largest developing countries over the past 30 years, along with estimates of the number of individuals enrolled in these programs. The largest single expansion in Table 1 is the New Cooperative Medical Scheme (NCMS) program that was initiated in China in 2003. This program provided comprehensive coverage of hospital expenditures for rural households in China. Within five years of its introduction, it was covering over 800 million persons, making it the largest insurance program in the history of the modern world (Zhu et al., 2017). Expenditure on the program was 0.21% of GDP or 4.56% of total health care spending in China in 2008, growing to 0.43% of GDP, or 7.16% of health care spending, in 2015.6 Despite its rank as the largest insurance expansion in modern history, there is little evidence on the impacts of the NCMS on health outcomes. There are only 3 articles in the economics literature, as well as 6 more in the public health literature, assessing the impact of the NCMS on health outcomes. These papers generally find little benefit for health from this enormous program – consistent with limited positive findings for a number of massive insurance expansions in the developed world. But, as we discuss below, these papers suffer from a number of concerns over identification and precision. We provide a comprehensive evaluation of the impact of the NCMS on the health of rural residents in China. We combine multiple sources of data to paint a broad picture of 5 https://data.worldbank.org/indicator/SH.XPD.CHEX.GD.ZS?locations=XM Data on NCMS expenditure is from China Health Statistical Yearbook, 2008 & 2016; data on China's GDP is from China Statistical Yearbook, 2008 & 2016. 6 3 the effect of NCMS on mortality and other health outcomes, as well as on medical care utilization and spending. We rely on the fact that this program was rolled out at different times across areas, and that its impact across areas is proportional to the share of the population living in particular rural areas. The share of counties covered by the program rose from 11% in 2003 to 95% by 2008. Meanwhile, across counties in China, there is wide variation in the agricultural population eligible for this program; for example, in the sample of counties that we use for our mortality analysis, the agricultural share varies from 5% to 96%. As we show clearly throughout the paper, the areas that were differentially shocked by this new insurance program were on almost identical trends before the law's passage. We find that the passage of the law created enormous benefits for the health of the rural Chinese population. Within eight years of passage by 2010, the program had enrolled 88% of China's rural population and 62% of the country's total population. We estimate that this led directly to a reduction in the mortality rate among the enrolled population by 21%, and extended life expectancies by 4.3%. From 2003 to 2010, China saw a 2.5-year increase in life expectancies, and the introduction of the NCMS program explains over 78% (1.94 years) of that overall improvement. 7 Since 2008, we estimate this program saved more than one million lives per year were saved – a population larger than seven U.S. states. Moreover, this program was incredibly cost-efficient. We use data on program expenditures to illustrate that the cost per life saved was RMB 66.1 thousand ($9.8 thousand US) from 2004-2010. This is well below both estimates of the value of a life in China and other measures such as compensation for accidental death. 7 China Statistical Yearbook 2011. 4 We then explore the mechanisms through which this program achieved its success, using two micro-data sets on health care utilization and health outcomes among large samples of the Chinese population. This analysis allows us to estimate treatment-on-thetreated estimates of those who actually enrolled in the program, and to examine a rich array of health care utilization, health outcomes, and out-of-pocket spending protection. We find that the law dramatically improved health care utilization. Individual-level hazard models corroborate the aggregate reduction in mortality rates. In both data sets, the probability of utilizing medical care increased by between 56% and 77%; conditional on utilizing medical care, the healthcare expenditure increased by around 40%. Measures of illness such as limitations of activities of daily living, cognitive deficits, and self/interviewer assessed health improved dramatically as well, confirming that the life extension we find reflect broader improvements in health. We then document two key mechanisms through which these results operated: reductions in out-of-pocket costs and expansions in supply. We show that out-of-pocket costs fell substantially for those at the top of the spending distribution, providing protection against catastrophic expenditures. Moreover, starting from 2003, the Chinese government strongly encouraged private capital investment in public hospitals and clinics. Because the NCMS only reimbursed expenses incurred in public facilities, there was a massive capital injection into this sector. But the expansion in health care supply occurs at a much more rapid rate in areas that were more exposed to NCMS. This suggests that much of the improvement that arises from the NCMS may be due to supply responses (as was the case for Thailand in Gruber et al, 2014). To further explore this hypothesis, we note that the supply expansion was highly equalizing – there was a larger rise in supply in areas with the smallest stock of health care facilities. We therefore 5 test whether the health outcomes were equalizing – which allows us to contrast the hypothesis that the impact was due to more patient care at existing facilities versus creating new facilities. While imprecise, our evidence suggests that the supply mechanism is playing role, but not the predominant one, in explaining our findings. Our paper proceeds as follows. Section I provides background on health transformation in the developing world. Section II describes the NCMS program and its place in the trajectory of health care development in China. Section III describes our multiple sources of data. Section IV carries out our aggregate analysis of mortality rates, and computes cost effectiveness. Section V then turns to micro-data to explore impacts on health care utilization and self-reported health, confirming the health improvements behind our mortality results. Section VI then turns to the mechanisms behind our results, exploring impacts on both OOP costs and on interactions with expanded supply. Section VII concludes. Part I: Background on Health Transformation in Developing Nations China's health care system has gone through a radical transformation since the founding of the People's Republic of China in 1949. At that point, health care provision in China was dominated by traditional medicine, and life expectancy was low – particularly in rural areas. For example, in 1952, total health care spending was only 1.3% of GDP, compared to 5% in the US in 1960.8 Life expectancy in China was 35 in 1949, and it was even lower in rural areas. 9 In the 70 years since 1949, the Chinese health care system has undergone a fundamental shift towards Western medicine. As a result, health care spending has risen to 8 9 Research Report on China Healthcare Expenditure, Ministry of Health 2008. (In Chinese) China Health Statistical Yearbook 2019 6 6.6% of GDP in 2018. 10 And China has seen one of the most amazing improvements in life expectancy in the modern world, with around 120% increase to 76 in 2015. 11 But as China and other countries have modernized their health care systems, they have faced a fundamental question of how to share this burden within society. Some countries, such as India, Pakistan, and Singapore, have decided to place the burden mostly on households, with the share of spending coming out of pocket above 50% in each case. Other countries, such as Thailand and South Africa, have moved to a model comparable to many European nations, with only about 10% of costs borne out of pocket. Nations such as Vietnam, China, Peru and Brazil have followed a middle ground, with roughly 25-33% of health care spending being out of pocket. Developing nations have undertaken a variety of approaches to covering the burden of health care spending. One approach is to reduce the user fees often attached to use of public medical services. Gertler, Locay and Sanderson (1987), Litvack and Bodart (1993), and Souteyrand et al. (2008) explore user fees via financing proposals in Peru, a field experiment in Cameroon and HIV/AIDS treatment in developing countries respectively; all find equity, quality and access concerns are reduced by the elimination of fees. Gertler, Locay and Sanderson (1987) in particular, finds that the loss in consumer welfare falls on the poor. Zombré, De Allegri and Ridde (2017) find that removing user fees for young children in Burkina Faso increased usage dramatically. Most comparably for our study, there is a very large literature evaluating large scale insurance expansions in developing nations (see Appendix 1). A structured literature review finds 15 articles reviewing the impacts of the 1990 program in Brazil that covered roughly 200 million residents, finding significant improvements in health outcomes and reductions 10 11 China Health Statistical Yearbook 2019. China Statistical Yearbook 2016. 7 in health inequities. Seventeen articles study the impacts of the 1992 Vietnamese program that covered 80 million residents, suggesting the program offered some financial protection and positive spillover effects for children, but was limited effects due to non-zero premiums and misplaced incentives which lead to high utilization but little health improvement. We found 19 articles focused on the 1993 health care expansion in Colombia to 34 million people, with strong evidence of higher health care utilization but more mixed evidence on health outcomes. We found 10 studies of the 1995 Philippine program that covered as many as one hundred million people, but once again there is little evidence of improved outcomes. We found 30 studies exploring 1999/2005 expansions in Rwanda/Ghana to 11/12 million people, with dramatic effects on utilization, positive spillover effects for children, reduced financial exposure and once again little evidence of improved health outcomes. And we found 8 articles covering the 2002 expansion to more than 50 million people in Thailand, with strong reductions in OOP and medical utilization, and some evidence of improved outcomes. Taken together, this large literature suggests that enormous insurance expansions in developing countries can play a major role in reducing OOP and in increasing utilization. But evidence on health impacts is surprisingly limited. More generally, there are remarkably few comprehensive evaluations of the impact of health insurance reform on OOP spending, utilization and health outcomes. Given its size, there is a surprisingly modest literature evaluating the NCMS on health. We are aware of only three economics papers that focus primarily on the health impacts of the program. Chen and Jin (2012) use a cross-sectional approach using aggregate data from the 2006 Census to examine the impact of the NCMS on mortality, and find little effect, perhaps unsurprisingly given differences between early versus late adopters of the 8 program. Lei and Lin (2009) use micro-data covering the early years of the program (20002006) and find little impacts on self-reported health status using an individual fixed effects regression (which doesn't control for endogeneity of NCMS enrollment) or cross-sectional instrumental variables using county dummies (problematic as noted above). Cheng et al. (2015) use propensity score methods to attempt to control for endogenous enrollment in their micro-data study; they find NCMS improves activities of daily living, but not mortality outcomes. There are six papers in the public health literature evaluating the health effects of NCMS. Perhaps the closest paper to ours is Zhou et al. (2017), who carry out a differencesin-differences study on aggregate mortality data using the timing of NCMS expansion across rural counties. They find no effect on mortality, although their confidence intervals are fairly large. As we explain below, the substantial variation in exposure of counties to the program, as well as underlying trends in non-rural counties, explain the much stronger effects we find. There is a larger literature on the other effects of the program, but it is once again quite mixed. Several studies find increases in health care utilization: Long et al. (2012) find a higher use of Cesarean-section delivery, and Wagstaff et al. (2009) find higher inpatient and outpatient utilization. But Lei and Lin (2009) and Babiarz et al. (2010) find no change in overall medical use. Lei and Lin (2009) and Wagstaff et al. (2009) find no effect on OOP expenditure, while Sun et al. (2009) find a sizeable reduction in catastrophic medical spending; Liu (2016) suggests that families are sufficiently self-insured against health spending by non-insurance mechanisms. Part II: The New Cooperative Medical Scheme 9 NCMS is the successor of the Cooperative Medical Scheme (CMS). The CMS was established in the 1950s alongside the establishment of the People's Republic of China. The CMS covered up to more than 90% of rural residents from early 1950s to the economic reform from 1979 (Liu and Cao, 1992). It was funded by contributions from enrollees and subsidies from collective welfare funds at the commune level. 12 The primary healthcare providers in rural areas were the so-called barefoot doctors. Although only receiving preliminary healthcare training, these doctors played a critical role in providing primary healthcare and public health services under government subsidies. However, the CMS began to collapse after 1979 when the rural communes were dissolved under the marketoriented economic reform. As a result, the CMS coverage of rural residents declined sharply from nearly universal to almost zero in the 1990s (Blumenthal and Hsiao, 2015). Meanwhile, the vast majority of barefoot doctors left the profession. Only a few remained for private practice, leaving the rural areas again with undersupply of qualified medical providers. In the absence of any health insurance and qualified medical providers in rural areas, individuals or families became the only payer for all medical costs in the 1990s. The share of out-of-pocket expenses paid by individuals and families out of the total healthcare expenses at the national level tripled, from 20% in 1979 to nearly 60% in 2000.13 During this period, the rise in out-of-pocket healthcare expenses dramatically outpaced income growth. In real terms, total out of pocket spending rose 30-fold from 1970 to 2000, while disposable income rose only 5-fold (4.5-fold) for urban (rural) residents.14 With the rising out-of-pocket 12 The rural commune was the administrative organization that owned the land, organized, and distributed the production, and provided social services such as healthcare before 1978. It was later transformed into the village. 13 China Health Statistical Yearbook 2001. 14 China Health Statistical Yearbook 2001 and China Statistical Yearbook 2001.Disposable income indicates the after-tax income that can be saved or used for consumption, consisting of wage, business, property, and government transfer. It is reported by the National Bureau of Statistics at the end of every year. 10 expenses, the rural families were fully exposed to health risks and felt great financial burden, which led the Chinese government to implement a new public health insurance program for rural families at the onset of the new millennium. In 2003, China initiated the NCMS program. This new program was administered separately in each of the 2,800 county-level administrative units in China since most rural residents are scattered across the townships and villages governed by the county government.15 Under broad guidelines issued by the central government, at least two to three pilot counties were selected in each province in the first year, and more counties were gradually included, aiming at achieving full coverage by 2010 (State Council, 2002). The rollout of the program across counties is shown in Figure 1. In 2003, the first year when the program was available, only 11% of counties implemented it. There followed a rapid expansion so that by 2008, coverage was virtually universal in all counties. Its coverage fell back after 2008 because some counties attempted to replace the NCMS with a new scheme covering both rural and urban residents after a parallel program started to roll out in urban areas. The NCMS targeted rural residents, and agricultural hukou serves as the eligibility requirement. The Hukou system is a household registration system implemented after the establishment of the People's Republic of China. This system categorizes each Chinese citizen into two types: an agricultural hukou holder or a non-agricultural hukou holder. The hukou is used to determine one's eligibility for social service and welfare to his registered place of residence. One's Hukou type is inherited from parents, and its change is highly restricted by the government. According to the China Population Census, the share of agricultural hukou remained relatively flat during the first ten years of the 15 On average, one county has 14 townships and each township contains 16 villages. 11 2000s, at 72.31 % in 2000 and 70.86 % in 2010; due to rapid industrialization of China over the past decade, it has fallen to 55.62% today.16 The central government issued broad guidelines for the design and implementation of the NCMS program; meanwhile, local governments in each county had considerable discretion over details such as premiums and coverage of spending. As a result, the program's benefits packages vary geographically. Generally, the program focuses on inpatient care, the largest source of health care spending risk, and provides a "catastrophic" insurance coverage design. During the initial years , individuals who were hospitalized were responsible for a deductible of around 200-500 RMB (about $25-$62), which was about 615% of net annual income per capita in 2005 (You and Kobayashi, 2009). Beyond this deductible, there was a coinsurance rate of 30-50%. But overall exposure is capped at 15,000 RMB (about $1,875), roughly five times net annual income per capita of rural residents in 2005 (Bai and Wu, 2014). As time went by, the NCMS benefits became more and more generous. The coinsurance rate decreased to around 30% in 2012, and the reimbursement limit reached over 60,000 RMB (about $8,862) (Liang and Langenbrunner, 2013) . More recently, the coinsurance rate continued to decrease to 10%-30%, and the limit is adjusted annually to stay above eight times the net annual income per capita of rural residents (Burns and Liu, 2017). In addition, the program provided more limited coverage of outpatient care. As noted in Du and Zhang (2007), only 18% of counties provided payment for outpatient care in 2006; of that group, three-fifths reimbursed all expenses with a coinsurance rate beyond a deductible, and the other two-fifths merely covered certain critical diseases. 17 In another 16 http://www.stats.gov.cn/english/PressRelease/202002/t20200228_1728917.html. Critical diseases refer to those incur catastrophic outpatient cost, such as nephrosis and hepatopathy. The list of critical diseases covered was designed by local government, and therefore differed across counties. 17 12 65% of counties, the program set up mandatory household medical savings accounts that would be used to pay for outpatient care.18 These accounts are funded by the individual contribution to plan premiums, which started as roughly 10 RMB and rose to roughly 20 RMB by 2012 (Liang and Langenbrunner, 2013). Over time, this savings mechanism has been gradually abolished and replaced by a more typical outpatient reimbursement program, with a reimbursement rate of 50% to 80% (Milcent, 2018). The NCMS was paid for by an insurance premium. The central government set a minimum level, and the local government adjusted it according to local income and benefits packages. At the start of the program, the minimum premium was 30 RMB ($3.63) per year, of which the government (the individual) paid 2/3 (1/3). By 2008, the minimum premium had risen to 100 RMB ($14.50), of which the government paid 80%. In counties with savings accounts for outpatient expenditures, Individual premiums were the funding source for the savings accounts. Participation in this program was voluntary. However, due to the modest premiums, intense government subsidies, and strong government mobilization ability, participation rates were as high as 75% in the three years of initiation and increased to 96% in 2010. Individuals who paid the premium and enrolled in the program received a certificate card that could be used when visiting clinics and hospitals. There was a significant debate over the potential efficacy of the NCMS. Concerns were raised along three dimensions. The first was that the benefits were not generous enough – that the incomplete coverage of outpatient care and the relatively high cost sharing would mitigate any health benefits (You and Kobayashi, 2009). The second was that 18 Establishing and contributing to the household savings account was mandatory in counties that applied this scheme. The amount of contribution, which was by reallocation of the total premiums collected from both individuals and the government, was determined by local governments, and varied across counties (You and Kobayashi, 2009). 13 participation was voluntary, leading to low enrollment and potential adverse selection (Milcent, 2018). The third was that the decentralized administration of the NCMS at the county level would be problematic because risk pools would not be large enough and local governments would not have the capacity to support the program (Burns and Liu, 2017). For this reason, the health benefits of NCMS were an open question. Notably, there was a corresponding change in the role of the private sector in financing health care supply around the same time. Before 2000, the Chinese healthcare market was almost exclusively run by public sector. Although private clinics were allowed to operate, they were extremely limited through unfavorable tax and reimbursement policies. In addition, private investment in public medical facilities was illegal. In 2003, the government began allowing private investment in public health facilities in order to address the coming increase in the demand for medical care. Private capital was encouraged to invest in the public healthcare sector via forms such as entrusted operation and joint venture. Private investors were allowed to send delegates to the administrative boards of public facilities (although board leadership was still public); these delegates would take part in operations of public facilities and may propose to hire professional managers to improve the management efficiency. The fixed-asset investment in China's healthcare industry by the sources of funds is shown in Figure A1 in Appendix 2. We see that the drastic increase in the investment has been mainly driven by equity investment, starting in 2003. In addition, health insurance coverage was extended through a new program introduced for urban areas in 2008, Urban Resident Basic Medical Insurance (URBMI). This program was much smaller than the NCMS, targeting 210 million unemployed urban residents in 2010, which is only one-fourth of those covered by the NCMS. In 2008, the government started combining the URBMI with the NCMS to form a new public health 14 insurance program, Urban and Rural Residents Basic Medical Insurance (URRBMI) (Milcent, 2018). Part III: Data We use a variety of data sets for our analysis. To define our treatment variable, we focus on China's 2000 Population Census. This provides the share of the population with agricultural Hukou in each county. The distribution of this share across counties in China is shown in Figure 2. For aggregate mortality, we turn to the China Death Surveillance Point Dataset (DSP). This is a nationally representative data set which records all deaths and population counts by sex and age. The data includes information from a sample of counties throughout China. Round 1, from 1991-2000, included 145 counties, while Round 2, from 2004-2012, included 161 counties. We hand collected data on the dates of implementation of NCMS, and were able to gather that data for 121 of the 161 counties included in Round 2.19 This forms our analysis sample. Our key dependent variables are age-adjusted mortality rates and life expectancy at birth at the county/year level.20 We focus our analysis on the years 2004 (the first year available in the second panel) to 2010. A potential confounder of our analysis is the new insurance program introduced for urban areas in 2008 which, while much smaller, weakens 19 We actively searched through official announcements of NCMS implementation from local governments' official websites and news reports. We compared the characteristics of counties with missing information on the starting year of NCMS and those of the remaining counties using the DSP data and 2000 China Population Census. We did not detect any significant difference between the two groups. See Table A1 in Appendix 3. 20 We obtain the county-level mortality rates by age from recorded deaths and population counts. These mortality rates are used to calculate the life expectancy at birth and age-adjusted mortality rates. The age adjustment is based on direct standardization (Curtin and Klein, 1995), where the standardized population by age is from China's 2000 Population Census. 15 the comparison to some extent with non-rural counties. As we show below, excluding the later years strengthens our results. The means from these data are shown in the first panel of Table 2. For comparability, we also show means for these variables for the nation as a whole as well. We find that, along every dimension, there are no significant differences between our sample and the entire nation. We then turn to micro-data analyses of individual impacts. For this, we use two data sets. The Chinese Longitudinal Healthy Longevity Survey (CLHLS) provides data from individual face-to-face interviews among the elderly, in seven waves during 1998-2014 from 22 provinces in China. 21 Survivors in each wave are re-interviewed, while the deceased are replaced by new participants to maintain sample balance; however, the deceased family members are interviewed for information about the deceased, such as the date of death. The sample began with those elderly above age 80, then added those aged 65-79 since 2002, and included the elders' adult children aged 35-64 in the 2002 and 2005 waves. For our analysis, we include individuals aged 65 - 110 for all waves from 1998-2014, except for new participants in 2014.22 Our final sample for the mortality analysis is 78,446 observations of 33,307 unique individuals, among which 19,857 died as of 2014; more than 80% of the sample is above age 80. We restrict our sample to the survivors in the analysis of other health outcomes and healthcare utilization. We use several key variables from these data, as summarized in the first panel of Table 3. The survey asks whether the individual suffered from a serious illness in the last 21 The information was collected in 1998, 2000, 2002, 2005, 2008, 2011, and 2014, respectively. The official website of the CLHLS: https://sites.duke.edu/centerforaging/programs/chinese-longitudinal-healthy-longevitysurvey-clhls/. 22 Sample age is restricted because data on mortality and health status before dying between waves were only collected for the elders aged 65-110. 16 two years, whether they have limitations in their activities of daily living23, their mental health status24, their self-reported health status and corresponding interviewer-reported health status25. Questions regarding healthcare utilization include whether the individuals consider themselves as getting adequate medical services, whether they refuse medical services due to lack of money, and their total medical spending during the previous year. The survey also asks about alcohol consumption and smoking. The second is the China Health and Nutrition Survey (CHNS). This is a longitudinal survey of about 7200 households, including over 30,000 individuals in 15 provinces that vary substantially in geography, economic development, public resources, and health indicators.26 The Survey uses a multistage, random cluster process to draw the sample of households, which ensures its national representativeness.27 There are ten waves from 1989 to 2015. We use waves 2000, 2004, 2006, 2009 and 2011 in our analysis, and exclude the six provinces that are not covered by the survey until 2011. Our final sample consists of 19,037 individuals residing in 225 communities of 54 counties and 9 provinces. 28 23 Limited ADL is an indicator that equals one if the individual reports a need of help with eating, dressing, moving, toileting, bathing, and continence. 24 The corresponding questions in the survey are about the frequency of the following five emotions: (1) looking on the bright side of things; (2) being happy as younger; (3) feeling fearful and anxious; (4) feeling lonely or isolated; and (5) feeling useless with age. Options include "Always", "Usually", "Sometimes", "Seldom", and "Never". Mental health is a dummy that equals one if "Always" or "Usually" is reported for the first two positive emotions, and "Seldom" or "Never" is reported for the last three negative emotions. 25 Both self-reported and interviewer-reported health status are categorical variables but on different scales. Self-reported health status is rated on a 1-5 scale: 1 for "Very good", 2 for "Good", 3 for "So-so", 4 for "Bad", and 5 for "Very bad". Interviewer-reported health status is on a 1-4 scale: 1 for "Surprisingly healthy", 2 for "Relatively healthy", 3 for "Moderately ill", and 4 for "Very ill". We define self-reported health as an indicator for "Very good" or "Good" health, and interviewer-reported health as an indicator for "Surprisingly healthy" or "Relatively healthy". The correlation between these two measures is 0.378. 26 The provinces covered in the survey changed across waves. There were nine provinces from 2000 to 2009. Three mega cities (at the same administrative level as province) were included in 2011, and another three provinces in 2015. 27 The official website of the CHNS: https://www.cpc.unc.edu/projects/china. 28 In China, there are five administrative levels: central government, province, prefecture, county, and community. Community, which refers to villages, townships, or neighborhoods in counties, is the lowest-level administrative unit. 17 Each wave of the CNHS includes three modules for individual, household, and community. The individual and household modules include a variety of measures summarized in the second panel of Table 3. The survey provides information about limitations of daily activities29 and cognitive ability30 for those above 55 in waves 2000 to 2006. The survey asks whether the individual is diagnosed with hypertension and diabetes for all aged above 12. In addition, the survey records information on healthcare utilization during the past four weeks, including any use of medical service when feeling sick or getting injured, any inpatient or outpatient care, and any use of preventive care. The survey also asks about the total medical spending and the proportion paid by the health insurance. This allows us to compute the out-of-pocket (OOP) spending and the ratio of OOP spending to total spending. Finally, the CHNS has questions on whether the individuals consume alcohol or cigarettes, and the amount of consumption per day. For the community module, government officials report data describing the health care infrastructure and services in their community. These data are collected at the level of the community (of which 4-5 are typically surveyed, for a total of 225 communities in 54 counties). For each facility, the module records the type of facility, the location, the number of healthcare workers, the number of beds, and more. We focus on facilities that provide health care to the general public, which includes clinics and hospitals.31 All hospitals are public, while clinics can be divided into private and public. We show descriptive statistics for these data in the final panel of Table 3. 29 Limited ADL is an indicator that equals one if the individual has any limitation in bathing, dressing, eating, toileting, combing hair, shopping, cooking, using public transportation, managing money, and using telephone. 30 Questions on cognitive ability differ across waves. We choose six questions that are constantly asked in all three waves, including one counting exercise and five subtraction exercises. We define an indicator of Cognitive deficit that equals one if the individual did not answer all the questions correctly. 31 We exclude clinics and hospitals which exclusively serve a specific group such as all employees in a stateowned enterprise, town family planning stations, and drug stores. 18 Part IV: Aggregate Impacts Aggregate Analysis using DSP We begin with county/year level analysis using the DSP. Using these data, we estimate equations of the form: log⁡(Ycpt ) = 𝛼0 + 𝛼1 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒𝑐,2000⁡ × 𝑃𝑜𝑠𝑡𝑐𝑡 + 𝑿′𝑐𝑡 𝛼𝑘 + 𝛿𝑐 + 𝛿𝑝𝑡 + 𝛿𝑡𝑐0 ,𝑡 + 𝛿𝑐 × 𝑡 + 𝜖𝑐𝑝𝑡  where 𝑌𝑐𝑝𝑡 denotes age-adjusted mortality rate (the number of deaths per 100,000) or the life expectancy at birth for county 𝑐 in province 𝑝 in year 𝑡; 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒𝑐,2000⁡ is the share of the population with an agricultural Hukou in county 𝑐 in the year 2000; 𝑃𝑜𝑠𝑡𝑐𝑡 indicates that county 𝑐 has adopted NCMS in year 𝑡; and 𝑋𝑐𝑡 includes GDP per capita and the average wage of urban employees.32 We also include a rich set of fixed effects: county fixed effects (𝛿𝑐 ), province-by-year fixed effects (𝛿𝑝𝑡 ), a separate set of year fixed effects for each NCMS timing group (i.e., NCMS-timing-by-year FE, 𝛿𝑡𝑐0,𝑡 ), and a county-specific linear time trend (𝛿𝑐 × 𝑡).33 The variable 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒𝑐,2000⁡ is absorbed by 𝛿𝑐 . With this rich specification, our key coefficient 𝛼1 is identified only by changes in outcomes in a county relative to that county's trend, differentially between counties in the same province and year, in counties with higher and lower shares of the agricultural population. Standard errors are clustered at the county level, and regressions are weighted by county population. Our results are presented graphically in Figure 3, which shows log mortality rate and log life expectancy over time. These graphs are in event study format, where year zero is the year that a county adopted NCMS, and the sample is divided into counties with above or 32 Data sources: China Statistical Yearbooks 2005-2011. NCMS-timing fixed effects, in other words, indicates that we have separate year effects for all counties that implemented in 2004, 2005, etc. 33 19 below agricultural shares at the median. We first regress log mortality rate and log life expectancy on NCMS-timing-by-year fixed effects and a county-specific linear trend. We then plot the residuals from this regression relative to the year of adoption; observations more than 3 (5) years before (after) the NCMS adoption are binned into groups. The results are striking. Both groups of counties have similar (relatively flat) trends before the program, with the more urban counties having a significantly lower mortality rate (by about 8%) and a longer life expectancy (by about 1.5%). But these differences quickly shrink as soon as the NCMS is put into place, and then reverse. While mortality is slightly rising and life expectancy is slightly falling in more urban counties after the year of NCMS adoption, mortality is substantially reduced, and life expectancy is dramatically extended in more rural counties. By the time a new steady state is achieved 2-3 years after passage, mortality rates are roughly 4% lower in more rural counties, and life expectancy is roughly 0.7% higher in more rural areas. The regression estimates corresponding to equation (1) are shown in Table 4. We see a substantial drop in mortality associated with counties with a high agricultural share post the NCMS, relative to before. We estimate that living in a county with the median agricultural share (78.5%), relative to a county with no agricultural population, implied a decline in mortality of 15.3% at the time of the NCMS. Similarly, we show a strong and significant correlation between agricultural share and the change in life expectancy after the NCMS. Living in a county with the median agricultural share resulted in a rise in life expectancy of 3.1% relative to a county with a zero agricultural share. We also estimate the dynamic version of this specification: ′ 𝛽 +𝛿 +𝛿 + log(𝑌𝑐𝑝𝑡 ) = 𝛽0 + ∑𝑗=−3,𝑗≠−1 𝛽𝑗 𝟏{𝑡 − 𝑡𝑐0 = 𝑗} × 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒𝑐,2000 + 𝑋𝑐𝑡 𝑘 𝑐 𝑝𝑡 𝑗=5 𝛿𝑡𝑐0 ,𝑡 + 𝛿𝑐 × 𝑡 + 𝜖𝑐𝑝𝑡      20      where 𝑡𝑐0 is the first year when county 𝑐 adopted the NCMS. The dummy variable 𝟏{𝑡 − 𝑡𝑐0 = 𝑗} indicates that 𝑡 is 𝑗 years relative to 𝑡𝑐0 . The year before NCMS adoption is omitted, so the estimates are normalized to zero in that year. Other variables are defined in equation (1). The results are shown in Figure 4. We see that there is no effect before the passage of the NCMS, and a gradually growing effect thereafter, consistent with our causal interpretation. The remaining columns of Table 4 show robustness checks on our results. In the second set of columns, we present the results from specifications removing the set of covariates (GDP per capita and average wage) described earlier; the results are nearly identical, confirming the independence of our identifying variation from other factors. In the third set of columns, we shorten the sample period to 2004-2007 in order to remove any influence of the 2008 introduction of an urban insurance program; our results get stronger but less precise when we do so, and we cannot reject that they are the same as for the full period. Finally, we use the earlier mortality panel that runs from 1994-2000 to run a placebo test on our results. We take each implementation data for NCMS and shift it back by 10 years, so that the program is (counterfactually) phased in from 1994-1997. We then redo our difference-in-differences analysis on this earlier data. As shown in the last columns of Table 4, we find no effect on either of our dependent variables. Interpretation To estimate the total impact of the NCMS, we use the specification shown in the first two columns of Table 4. For each county, we then estimate the result in each year based on their agricultural share and their year of adoption of NCMS. Doing so, we estimate that over 21 the 2004-2007 period, the NCMS saved 550,000 lives per year, and from 2008 onwards was saving more than one million lives per year. 34 Similarly, based on our estimates the NCMS raised the life expectancy of the entire nation of China by 1.94 years from 2003 to 2010. 35 Over this period, the life expectancy of China roses by 2.5 years.36 This suggests that the implementation of the NCMS can explain 78% of the entire rise in life expectancy over this period. We can also assess the cost effectiveness of NCMS based on our data. Expenditures on NCMS in 2008 were 63.7 billion (in 2010 RMB, hereafter the same), and we estimate that 1,012 thousand lives were saved that year for the whole nation. That implies a cost of 63.0 thousand RMB ($7900 U.S.) per life saved – completely ignoring any of the additional medical benefits that we show below. How does this compare to typical estimates of the value of a life? Prior studies show that the statistical value of a life in rural China varies from 148 – 296 thousand RMB (Hammit and Zhou, 2006) to 600 thousand – 1.0 million RMB (Wang and He, 2014). The legal compensation for death from injury in China is set to 20 times the local dispensable income per capita, which amounts to 122 thousand RMB in rural China in 2008.37 The limit on compensation for death by a traffic accident in liability insurance is 110 thousand RMB. 38 By any of these metrics, this program was highly cost-effective. 34 We estimate that the NCMS reduced the mortality rate by 2.9% - 12.7% between 2004 and 2007, and 13.1% from 2008 onwards. We then calculate the number of lives saved per year during these two periods with the baseline mortality rate of 582 per 100,000 in 2003. 35 We estimate that the NCMS extended life expectancy by 2.7% by 2010, which was 1.94 years based on the life expectancy of 72.4 in 2003. 36 We use life expectancies from the 2000 and 2010 China Population Census and 2005 1% Sampling Census, and compute the life expectancies in other years between 2000 and 2010 by linear projection. 37 This standard of compensation is set by the Supreme Court in the Interpretation of the Supreme People’s Court on Several Issues of the Applicable Law in Hearing Cases of Compensation for Personal Injury (inforce since 2004). The per capita disposable income rural households in 2008 is 6.1 thousand (in 2010 RMB) according to China Statistical Yearbook 2009. 38 Regulations on Compulsory Insurance for Traffic Accident Liability of Motor Vehicles. 22 Part V: Effects on Utilization and Other Health Outcomes In this section, we dive into more details on the impact of the NCMS on health care utilization and on other measures of health outcomes. This allows us to confirm that our mortality outcomes reflect increased utilization of health care resources, and broader improvements in underlying health. CLHLS Results The 1998-2014 waves of the CLHLS offer us several advantages in exploring the mechanisms behind our effects. Most notably, we actually measure NCMS enrollment as an outcome variable. This allows us to estimate treatment-on-the-treated IV models, where we instrument NCMS enrollment by an interaction term between NCMS passage in that county and individuals' agricultural hukou status. However, the CLHLS does not ask about the deceased's NCMS enrollment status. We assume their NCMS status remains the same as it was in the last wave when they were alive. Since the hukou type is not available in CLHLS, we follow the literature and assign an agricultural hukou to an individual if he or she lives in rural areas and has no access to government-sponsored welfare scheme for residents with urban hukou (Huang and Zhang 2020). 39 We begin by confirming our mortality results using these microdata. In particular, we estimate a mortality hazard model of the form: ′ ℎ𝑐 (𝑚|𝑆𝑖𝑐𝑚 ⁡) = ℎ0,𝑐 (𝑚)𝑒𝑥𝑝(𝛽1 𝑁𝐶𝑀𝑆𝑖𝑐𝑚 + 𝛽2 𝑁𝐶𝑀𝑆𝑐𝑚 + 𝛽3 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 + 𝑿𝒊 𝛽𝑘 + 𝛿𝑡𝑚 ), (3) 39 According to the population census 2000, the correlation between agricultural hukou status and residing in rural areas is as high as 0.99. 23 where ℎ𝑐 (𝑚|𝑆𝑖𝑐𝑚 ⁡)⁡is the hazard of death for individual 𝑖 with characteristics 𝑆𝑖𝑐𝑚 in county 𝑐 in year 𝑚 relative to the first year when the individual 𝑖 participated in the survey; ℎ0,𝑐 (𝑚) is the baseline hazard of death, which is assumed to be constant within the same county to account for all unobserved county-specific health-related heterogeneities; 𝑆𝑖𝑐𝑚 includes NCMS enrollment (𝑁𝐶𝑀𝑆𝑖𝑐𝑚 ), an indicator for NCMS county (𝑁𝐶𝑀𝑆𝑐𝑚 ), a dummy for having an agricultural Hukou (𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 ), calendar year fixed effects (𝛿𝑡𝑚 ), and individual characteristics (𝑿𝒊⁡ ) . These characteristics include dummies for age, gender, minority status, household size, housing income per capita, drinking, smoking, and an indicator of getting adequate medical service when in serious illness in childhood. All these characteristics are held constant at their baseline values. The NCMS enrollment status might be endogenous because participating in the program is voluntary. For example, health endowments and risk preferences, which are unobserved by researchers, may differ between enrollees and non-enrollees. To address this concern, we use the county NCMS adoption interacted with one’s agricultural hukou status to instrument for the individual NCMS enrollment status, 𝑁𝐶𝑀𝑆𝑖𝑐𝑚 , in equation (3). This instrument is valid because it utilizes the staggered adoption of NCMS across counties, and a pre-determined eligibility criterion of Hukou status. The 2SLS estimates, however, are inconsistent, because equation (3) is nonlinear. Instead, we use the control function (CF) method (Wooldridge, 2015). Similar to the 2SLS method, the CF method has two stages. The first-stage specification is 𝑁𝐶𝑀𝑆𝑖𝑐𝑚 = 𝛼0 + 𝛼1 𝑁𝐶𝑀𝑆𝑐𝑚 × 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 + 𝛼2 𝑁𝐶𝑀𝑆𝑐𝑚 + 𝛼3 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 + 𝑿′𝒊 𝛼𝑘 + 𝛿𝑐 ⁡ + ⁡ 𝛿𝑡𝑚 + 𝜀𝑖𝑐𝑚⁡ , (4) where 𝑁𝐶𝑀𝑆𝑖𝑐𝑚 indicates that individual 𝑖 in county 𝑐 has enrolled in the NCMS in year 𝑚 relative to the first year when the individual 𝑖 participated in the survey; 𝑁𝐶𝑀𝑆𝑐𝑚 indicates 24 that county 𝑐 has adopted the NCMS in year 𝑚, and 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 indicates that individual 𝑖 has an agricultural Hukou. We control for the same set of controls as in equation (3) along with county fixed effects (𝛿𝑐 ).40 We predict the residual 𝜀̂𝑖𝑐𝑚 after estimating equation (4). We then estimate the second-stage equation (equation (3)), additionally controlling for 𝜀̂𝑖𝑐𝑚 . Based on Wooldridge (2015), the estimates of 𝛽1 is consistent using this CF method. Because 𝜀̂𝑖𝑐𝑚 is predicted in the first stage, the standard errors for the estimates in the second stage are bootstrapped. Our first stage estimate is shown in Table A2 of Appendix 4. We find that there is a very strong first-stage effect, implying a takeup rate of roughly 80% among those made eligible for the program. This estimate is very robust to varying the set of controls that are included in the model. Table 5 shows our mortality results. The first column shows our baseline hazard without correcting for the endogeneity of NCMS enrollment.41 We find a significant reduction in the hazard rate of dying of almost 50%. In the second column, we include the control function to control for potential endogeneity of NCMS enrollment. Our estimate falls to about 30%, suggesting positive selection into the NCMS program. This is consistent with the fact that more educated, and therefore healthier, individuals were more informed about the program in its early years, so they were the first to enroll. The estimate on the first-stage residual is less than 1, which is consistent with the positive selection. How does this result compare to our aggregate estimates? Given that one-year mortality rate before NCMS in our sample is 0.163, a 30 % reduction in mortality hazard 40 Note that county fixed effects are implicitly included in the hazard specification through the use of a countyspecific baseline hazard. 41 All estimates of the mortality hazard model are reported after taking exponential to represent proportional increases or decreases in mortality rates relative to one. 25 indicates a 34% reduction in mortality rate after NCMS enrollment. 42 Considering the average takeup rate of 80% of NCMS at the county level, our aggregate estimates suggest a decrease of 24% in aggregate mortality rates for a county with 100 percent NCMS enrollees. The CLHLS results show a larger mortality reduction among the elderly. This is consistent with the fact that the elderly is more vulnerable to health-service amenable mortality than the younger cohort and therefore benefit more from the health insurance expansion (Nolte and McKee, 2012). The rich data available in the CLHLS allow us to go beyond mortality to look at other health outcomes as well. In particular, we have five other health outcome measures in a sample of survivors. To do so, we move from our hazard framework to individual IV methods at the individual level: 𝑌𝑖𝑐𝑡 = 𝛽0 + 𝛽1 𝑁𝐶𝑀𝑆𝑖𝑐𝑡 + 𝑋𝑖′ 𝛽𝑘 + 𝛿𝑐𝑡 + 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑡 + 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑐 + 𝜖𝑖𝑐𝑡 , (5) where 𝑌𝑖𝑐𝑡 is a measure of health outcome of individual 𝑖 in county 𝑐 in wave 𝑡; and 𝑁𝐶𝑀𝑆𝑖𝑐𝑡 indicates that individual 𝑖 has enrolled in the NCMS in wave 𝑡. Because enrolling in the NCMS is voluntary, 𝑁𝐶𝑀𝑆𝑖𝑐𝑡 is endogeneous in equation (5). We use the IV method to address the endogeneity, and the first stage specification is: 𝑁𝐶𝑀𝑆𝑖𝑐𝑡 = 𝛼0 + 𝛼1 𝑁𝐶𝑀𝑆𝑐𝑡 × 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 + 𝑋𝑖′ 𝛼𝑘 + 𝛿𝑐𝑡 + 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑡 + 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑐 + 𝜏𝑖𝑐𝑡⁡ . (6) Compared equation (6) with (5), the excluded instrument for 𝑁𝐶𝑀𝑆𝑖𝑐𝑡 is the interaction term between NCMS passage in that county and individual 𝑖's Hukou status (𝑁𝐶𝑀𝑆𝑐𝑡 × 42 Given the estimated hazard ratio of 0.694, we have h(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 =1) h(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 =0) = 0.694. Based on the 𝑡 relationship between mortality hazard h(𝑡) and mortality rate⁡𝐹(𝑡) in − ln(1 − 𝐹(𝑡)) = ∫0 h(𝑢)𝑑𝑢 , we have −ln⁡(1−𝐹(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 =1) −ln⁡(1−𝐹(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡=0) = 𝑡 ∫0 h(𝑢|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 =1)𝑑𝑢 𝑡 ∫0 h(𝑢|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 =0)𝑑𝑢 = 0.694. We then back out 𝐹(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 = 1) since 𝐹(𝑡|𝑁𝐶𝑀𝑆 𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡𝑡 = 0) = 0.163 in our sample. 26 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 ). In both equations (5) and (6), the vector of covariates, 𝑋𝑖 , includes dummies for age, gender, minority status, household size, housing income per capita, drinking, smoking, and an indicator of getting adequate medical service when in serious illness in childhood. These covariates are held constant at their baseline values. This is a very rich completely saturated model: we control for county-by-wave fixed effects (𝛿𝑐𝑡 ) to account for unobserved time-varying characteristics that are specific to a county, hukou-by-wave fixed effects (𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑡 ) to account for unobserved timevarying characteristics that are specific to a hukou type, and hukou-by-county fixed effect (𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢𝑖 × 𝛿𝑐 ) to control for unobserved time-invariant characteristics of individuals with the same hukou type in a county. By including this full set of interactions in both the first and second stages, we are running a triple-difference IV model.43 The standard errors are clustered at the county level. The first two outcomes are indicators for being seriously ill over the past two years and for limited activities of daily living. The third variable is mental health, an indicator for a positive state of mind. We also have two dummies of health status, one that is self-reported and one that is interviewer reported. The two measures are correlated with a correlation coefficient of 0.378. The results for these other health measures are shown in the third column of Table 5, using our IV model. We find a significant reduction in the probability of being seriously ill and limitations of activities of daily living, and a significant improvement in mental health; these effects amount to 50% to 70% of the sample mean, in line with the one-third reduction in the mortality hazard we observe in these data. The effects on both self- 43 Our estimation results remain robust when we additionally control for individual fixed effects, which are reported in Tables A3 - A5 in Appendix 5. 27 reported health and reporter-reported health are significant, showing an effect which is roughly 85% as large as the share of population not in good health. To assess mechanisms, the final column of Table 5 extends our results to examine several measures of health care utilization available in the CLHLS, continuing to control for selection. The first is a dummy variable for respondents reporting that they received "adequate medical service". Only 8% of the sample reports that they do not receive adequate medical service, and this gap is closed under NCMS. We then show that the odds of receiving any medical spending rises by 11.8 percentage points if individuals are covered by NCMS, which is over 50% of the share of those who report not having positive medical spending. Among those with spending, there is a 42% increase in spending. 44 The CLHLS data, therefore, usefully confirms our finding of significant mortality reductions and shows that these are reflected in other measures as health as well. Moreover, the results show a substantial increase in health care spending - consistent with these improvements in health. CHNS Results The other micro-data set we use is the CNHS, which is a community-based survey of 225 communities from 54 counties in five waves from 2000 to 2011. We estimate models of the same specification as equations (5) and (6). Due to data availability, the vector of covariates includes gender, age, minority status, household size, years of schooling and household income per capita, and a set of variables at the community level to measure 44 The CLHLS did not ask about the precise medical spending until wave 2005. Therefore, we use wave 2005, 2008, 2011 and 2014 to do the analysis. Individuals pay the total costs of their medical care and are then reimbursed, allowing survey respondents to report their total spending. 28 socioeconomic, transportation, and sanitation conditions. Standard errors are clustered at the county level. There are two key differences between the CLHLS and CHNS results to keep in mind in comparing the results. First, the CLHLS data includes only older individuals, while the CHNS includes the non-elderly as well. Second, the CLHLS measures medical utilization over the past year, while the CHNS measures it only over the previous four weeks. Table 6 presents our key results for the CHNS. In the first column, we examine other measures of health outcomes. In particular, we focus on dummy variables for having a cognitive deficit or limitations in activities of daily living, which are measured for those age 55+ only, and on having diabetes or hypertension, which are measured at all ages above 12. We see significant and sizeable reductions in all health measures. The odds of a cognitive deficit fall by almost half, while the odds of a limitation in activities of daily living falls by more than half. There is a four-fifths reduction in the incidence of diabetes and a two-thirds reduction in hypertension (albeit at very low full population means). The second column examines measures of health care utilization. We find that health care usage rises by 8.4 percentage points. This is composed of a 1.9 percentage points rise in use of inpatient care and a 7.2 percentage points rise in outpatient care. The survey also asks separately about receipt of preventive care, and this rises by 5.4 percentage points. Total medical spending increases by 39% among those with positive spending. But the estimate is not precise due to a relatively small sample size. Therefore, both of these valuable micro-data surveys confirm the transformational results of the NCMS program in a treatment-on-the-treated framework. There are dramatic improvements in health care utilization, as well as improvements in health outcomes along a wide variety of dimensions. 29 Part VI: Mechanisms – OOP Spending, Supply Responses, and Ex-Ante Moral Hazard The impressive effects of the NCMS raise the natural question of mechanisms of response. One mechanism is the protection against out-of-pocket risk provided by insurance, which increases utilization and improves health outcomes. Another mechanism highlighted by Gruber et al. (2014) is the role of induced supply-side effects. By raising demand for care, public programs can expand the supply of care available. The Chinese context offers a potential multiplier effect for this channel through the simultaneous liberalization of private investment in public health care facilities. Using the CHNS, we can examine both potential explanations. Importantly, this survey includes not only individuals surveyed in those communities but also a survey of all medical facilities in the communities, including public hospitals, public clinics, and private clinics.45 This part of the survey is answered by the government official who is in charge of the bureau of health at the community level. Out-of-Pocket Spending Table 7 shows the impact of the NCMS on out-of-pocket medical costs, using our treatment-on-the-treated framework in both the CLHLS and CHNS data. The first column uses a dependent variable from a question asked in the CLHLS about whether the respondent refuses any necessary medical service due to financial difficulty. We find a highly significant reduction in this measure of 5.1 percentage points, which is 85% of the population mean. Unfortunately, the CLHLS does not ask about out of pocket medical spending, but the CHNS does. 45 The community module records a roster of all medical facilities available to the residents in the community, including facilities locating both inside and outside the community. We only consider the facilities locating inside the community in our analysis. 30 The second column shows that, in fact, we find an insignificant effect on (the log of) out-of-pocket spending. This is surprising given the insurance protection provided by the program, but it also reflects the imprecision in our estimate. This is because only 330 persons in our data have out-of-pocket medical spending in the month before the survey, both before and after the NCMS becomes available. In addition, the drop in OOP costs is smaller than what might be expected since total medical spending is rising so rapidly. The third column changes the dependent variable to the ratio of OOP spending to total spending. This shows a significant reduction of 29 percentage points in the share of medical costs borne out of pocket. Finally, the NCMS program is clearly reducing the risk of extreme levels of out-ofpocket medical costs. Figure 5 shows the relationship between total medical spending and OOP spending before and after enrollment in NCMS. Before the NCMS, there is an essentially 1-1 correspondence between total spending and out of pocket spending. But after NCMS, there estimated a positive relationship but much below 1, so that for those with the highest total spending, they are bearing a much smaller share of spending in OOP costs. Supply Expansions A different mechanism for our effects could be the expansion in supply. To examine this, we model community-level provider supply, which is captured for all communities included in the CHNS. We estimate difference-in-differences models similar to equation (1) but at the community/wave level. The dependent variable is the log of the number of facilities per 10,000 population. We measure this separately for public hospitals and clinics, as well as for private clinics. As 31 noted earlier, beginning in 2003, the private sector was allowed to invest in public health facilities. Since the NCMS only reimbursed medical expenses in public facilities, these facilities served more and more patients and thus were attractive to private-sector investor. The model also includes community fixed effects, province-by-wave fixed effects, NCMS timing-specific wave effects, and a community-specific linear trend, as well as the time-varying measures that evaluate each community on the quality of economic outcomes (e.g., daily wage), market outcomes (e.g., number of supermarkets), transportation outcomes (e.g., distance to bus stop and train stop) and housing outcomes (e.g., access to electricity and tap water). Regressions are weighted by community population and standard errors are clustered at the county level. We also examine the log of the number of healthcare staff/beds per 1,000 population using the same specification. Panel A of Table 8 shows the results for a variety of supply measures. We find large supply responses in aggregate. Compared with a community with no agricultural population, a community with the median agricultural share in our sample (75%) experiences a 35% rise in public hospitals and a 47% rise in public clinics. On the other hand, the number of private clinics falls by 40%, which is consistent with the fact that the NCMS program only reimburses public providers. The rise in total clinics is surprisingly modest, given the larger increase in public clinics and the baseline higher level of those clinics; we explain this further below. We also find a dramatic rise in the total number of healthcare workers practicing in the area, and in the total number of beds available. This is a substantial response. Applied to the entirety of China, this implies that by 2010, the NCMS resulted in over 4 thousand 32 more public hospitals, 57 thousand more public clinics, and 24 thousand more total clinics nationwide. 46 These striking results raise the concern that private investment decisions drove the decisions of localities to adopt NCMS. To address this, in panel B of Table 8, we rerun the regression specification to include the two-year lead of NCMS interacted with baseline share of agricultural population in our supply equations. We see an insignificant coefficient, suggesting that NCMS adoption was not following expansions of private capital. This substantial response also raises the natural question: are the impressive health benefits that we see driven, at least partially, by more access to medical facilities? It is impossible to disentangle through our existing approach since we only have one instrument (NCMS availability) and two mechanisms (OOP costs and supply increases). But we can disentangle them to some extent by noting that the expansion of health care supply in response to the NCMS was highly equalizing. This equalization is well expected due to two reasons. First, in areas with lower initial supply, price elasticity of demand for medical service is larger (Yi et al., 2021). Therefore, the policy triggers a more pronounced increase in demand for healthcare and induces a larger increase from the supply side. Second, the policy-induced increase in demand is shared between both existing and newly-established facilities. Thus, there are more investment opportunities regarding new facilities in areas with lower initial supply. Panel C of Table 8 uses the same dependent variables as in the first row but include two terms: our instrument and an interaction of that instrument with the ex-ante measure 46 Given the agricultural share of 72% nationwide in 2000, our estimates suggest an increase of 33% in the number of public hospitals (per 10,000), which is 0.12 based on the corresponding value of 0.09 in 2003. Considering the scaling of population growth, the increase in the number of public hospitals is over 4,000. Similarly, we can find an increase of over 57,000 in the number public clinics and over 24,000 in the number of clinics of all types. 33 of the same supply. 47 The results strongly confirm the equalizing impact of the NCMSinduced expansion, with the main effect rising and a strong negative interaction. In particular, these findings help explain why our total clinics estimate in the first panel is so small; there is clearly a strong correlation between total clinic expansion and baseline clinic availability. The magnitudes are economically sizable. Take public hospitals as an example. Assuming a median share of the agricultural population, a community at the median of the ex-ante supply experiences an increase in the number of public hospitals that is 28.4 percentage points higher than a community at the top quartile. 48 To more clearly interpret this interaction, we present this calculation (increase at the median relative to the top quartile) under the result. This mean reversion in area supply potentially allows us to assess the role that supply is playing in driving our utilization and outcome results. If the effects are driven solely by increased access to a given supply, then they should be stronger in areas with more initial supply but should not rise as supply grows. If, on the other hand, the results are driven solely by the reform-induced expansion in supply, then we should see stronger impacts in the areas where supply increased the most. Of course, both effects could be present, so this means that our estimates understate the role of supply in driving outcomes. To assess the role played by supply, we incorporate NCMS-induced supply increases into our individual outcome regressions. That is, we add to equations (5) and (6) above a measure of local supply, for each of our supply measures. Since supply is endogenous, we 47 The ex-ante supply is measured using the last wave of the CHNS before the NCMS is adopted in the given community. 48 Compared to a community at the top quartile of the ex-ante supply, a community at median has 57.7% less public hospitals before NCMS adoption (log of difference is 0.577). We then calculate the interaction effects arising from the difference of baseline supply are 28 percentage points (= 75 × 0.656 × 0.577). 34 instrument supply by 𝑃𝑜𝑠𝑡⁡𝑁𝐶𝑀𝑆 × 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒2000 × 𝐵𝑎𝑠𝑒𝑙𝑖𝑛𝑒⁡𝑠𝑢𝑝𝑝𝑙𝑦 (the measure included in Panel C of Table 8). We add as a control to the regression 𝑃𝑜𝑠𝑡⁡𝑁𝐶𝑀𝑆 × 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒2000, to allow for other NCMS-induced effects on more agricultural areas; recall that the existing model only controls for agricultural hukou (an individual measure) interacted with survey wave, so this additional control captures any effects on the individual arising from changes in generally more rural areas. We also include 𝐵𝑎𝑠𝑒𝑙𝑖𝑛𝑒⁡𝑠𝑢𝑝𝑝𝑙𝑦 and 𝐴𝑔𝑟𝑖𝑆ℎ𝑎𝑟𝑒2000 to control for other health-related factors that are varying with the initial supply and share of agricultural population in the community. The results of this regression are shown in Table 9. We show the results for four different measures of supply (public hospitals, total clinics, healthcare workers and beds), and for a set of outcome variables from Table 6. For each outcome variable, we show the IV coefficients on 𝑁𝐶𝑀𝑆⁡𝑒𝑛𝑟𝑜𝑙𝑙𝑚𝑒𝑛𝑡 (which continues to be instrumented by 𝑃𝑜𝑠𝑡⁡𝑁𝐶𝑀𝑆 × 𝐴𝑔𝑟𝑖𝐻𝑢𝑘𝑜𝑢) and 𝑆𝑢𝑝𝑝𝑙𝑦. The first panel shows the results for the healthcare use dummy. We find that for all measures of health care supply, there is a positive coefficient on supply, consistent with some of our finding being driven by the supply channel. But the estimate is imprecise. The next four panels show results for our four health outcome measures in the CNHS. For limited ADLs, we see a similar pattern to health care use – a generally significant coefficient on enrollment, with a right-signed but insignificant coefficient on supply. For cognitive deficit, however, the supply coefficient is wrong-signed. For diabetes and hypertension, the coefficient is right-signed and of a more similar magnitude to the enrollment effect – but now both the enrollment and supply effects are insignificant. Overall, these results are at best suggestive that supply is playing a role in driving our responses. 35 Ex-Ante Moral Hazard Finally, we consider a potentially offsetting effect: ex-ante moral hazard: the fact that medical spending is now more financially protective may lead individuals to take more risks. As first laid out by Ehrlich and Becker (1972), this topic has fascinated health economists for decades. And the results are consistent, with very few studies finding evidence of ex-ante moral hazard. For example, recent evidence in Baicker et al. (2013) shows no effect of randomized health insurance provision of activities such as smoking. We assess the role of ex-ante moral hazard in our context by using data on drinking and smoking from both the CLHLS and the CHNS. We re-estimate equation (5) with dependent variables for any drinking or smoking, and for the log of the number of cigarettes or the log of amount of alcohol in liang (50gm). The amount of daily consumption is only available in the CHNS. The results, shown in Table 10, suggest that in fact, there may be less smoking and drinking among those enrolling in the NCMS, but the coefficient is not significant, as are none of the remaining coefficients on the amount of consumption conditional on current smokers and drinkers. We can certainly rule out large effects, allowing us to reject a rise in the incidence of smoking or in the incidence of drinking. Part VII: Conclusions The rise of modern medicine in the developing world has brought with it the need to address high medical costs relative to individual resources. A natural approach to addressing these costs is through expanded public insurance systems. A variety of countries around the world have introduced such systems in recent decades, with the most notable 36 recent addition being the PM-JAY program in India which is ultimately designed to cover the bottom 40% of the population, or almost 500 million persons. The largest of these expansions in terms of persons covered is the NCMS program introduced in rural China in 2003. Previous research on this program has delivered mixed and generally muted evidence of success from this program, particularly in terms of health outcomes. Our research using aggregate health data delivers a very different conclusion: the NCMS was an unqualified success. The program saved more than one million lives per year, explaining nearly four-fifths of the rise in life expectancy during its first eight years. Moreover, the program was highly cost effective, with a cost per life saved well below common benchmarks. Given these striking results from aggregate data, we explore the leading sources of health micro-data for China as well to confirm and extend our findings. We find from two different surveys that the NCMS program induced major increases in health care utilization, as well as improvements in a variety of other health measures, ranging from self-reported health to activities of daily living and the incidence of disease. Finally, we explore the mechanisms through which NCMS created these effects. We show modest reductions in out-of-pocket costs overall, but very sizeable reductions at the top of the spending distribution. And we show that the extra demand induced by NCMS, along with liberalization of rules around private investments in the health care sector, induced a massive increase in health care supply. Our results suggest that this increase in supply had an effect on outcomes above and beyond the demand effect, but they are imprecise. The fact that this incredible health policy success has gone largely underappreciated by the health policy community is striking and suggests the potential value of revisiting the 37 impacts of major health expansions in the developing world. Indeed, these results are quite consistent with findings for the health benefits of the Affordable Care Act (ACA) in the U.S., where recent quasi-experimental studies find sizeable impacts on mortality.49 Comparable well-identified studies across a broader set of nations could allow us to understand under which conditions health insurance expansions can be most effective. That said, there remain many questions to answer about the NCMS. In particular, we were unable to convincingly disentangle the demand and demand-induced supply side channels for health improvements, which would be critical for nations where market signals due to demand expansions are not enough to call forth more supply. And there much more to do to understand the heterogeneity of these effects, and how it depended on local decisions of features such as the nature of outpatient coverage. Finally, as China urbanizes, there are questions about the ongoing value of programs that are differentiated between urban and rural areas; comparable analysis of the urban insurance expansion that began in 2008 would help inform this issue. 49 Miller et al. (2020); Goldin et al. (2021) 38