the largest insurance program in history: saving one
TRANSCRIPT
1
PRELIMINARY β do not cite
The Largest Insurance Program in
History: Saving One Million Lives Per
Year in China
Jonathan Gruber, MIT and NBER
Mengyun Lin, National University of Singapore
Junjian Yi, National University of Singapore
March, 2021
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.
2
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 β 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-expectancy-
globally#:~:text=Globally%20the%20life%20expectancy%20increased,more%20than%20twice%20as%20long. 4 https://data.worldbank.org/indicator/SP.DYN.LE00.IN?locations=XL
3
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 β 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
6 Data on NCMS expenditure is from China Health Statistical Yearbook, 2008 & 2016; data on China's GDP is
from China Statistical Yearbook, 2008 & 2016.
4
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 β 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.
5
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-the-
treated 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 β there was a
larger rise in supply in areas with the smallest stock of health care facilities. We therefore
6
test whether the health outcomes were equalizing β 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 β 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 Research Report on China Healthcare Expenditure, Ministry of Health 2008. (In Chinese)
9 China Health Statistical Yearbook 2019
7
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.
ZombrΓ©, 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
China Health Statistical Yearbook 2019. 11
China Statistical Yearbook 2016.
8
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
9
program. Lei and Lin (2009) use micro-data covering the early years of the program (2000-
2006) 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 differences-
in-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
10
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 market-
oriented 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.
11
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.
12
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 6-
15% 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. 17
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.
13
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 β 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).
14
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
15
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.
16
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-longevity-survey-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.
17
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.
18
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 state-owned enterprise, town family planning stations, and drug stores.
19
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) = πΌ0 + πΌ1π΄ππππβππππ,2000 Γ πππ π‘ππ‘ + πΏππ‘β² πΌπ + πΏπ + πΏππ‘ + πΏπ‘π0,π‘ + πΏπ Γ π‘ +
ππππ‘
where ππππ‘ denotes age-adjusted mortality rate (the number of deaths per 100,000) or the
life expectancy at birth for county π in province π in year π‘; π΄ππππβππππ,2000is the share of
the population with an agricultural Hukou in county π in the year 2000; πππ π‘ππ‘ indicates that
county π has adopted NCMS in year π‘; and πππ‘ includes GDP per capita and the average
wage of urban employees.32 We also include a rich set of fixed effects: county fixed effects
(πΏπ), province-by-year fixed effects (πΏππ‘), a separate set of year fixed effects for each NCMS
timing group (i.e., NCMS-timing-by-year FE, πΏπ‘π0,π‘), and a county-specific linear time trend
(πΏπ Γ π‘).33 The variable π΄ππππβππππ,2000 is absorbed by πΏπ. With this rich specification, our
key coefficient πΌ1 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. 33
NCMS-timing fixed effects, in other words, indicates that we have separate year effects for all counties that implemented in 2004, 2005, etc.
20
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:
log(ππππ‘) = π½0 + β π½ππ{π‘ β π‘π0 = π}
π=5π=β3,πβ β1 Γ π΄ππππβππππ,2000 + πππ‘
β² π½π + πΏπ + πΏππ‘ +
πΏπ‘π0,π‘ + πΏπ Γ π‘ + ππππ‘
21
where π‘π0 is the first year when county π adopted the NCMS. The dummy variable
π{π‘ β π‘π0 = π} indicates that π‘ is π years relative to π‘π
0. 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
22
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 β 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 β 296 thousand RMB
(Hammit and Zhou, 2006) to 600 thousand β 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 Peopleβs 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.
23
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:
βπ(π|ππππ) = β0,π(π)ππ₯π(π½1ππΆπππππ +π½2ππΆππππ +π½3π΄ππππ»π’πππ’π +πΏπ
β²π½π + πΏπ‘π),
(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.
24
where βπ(π|ππππ)is the hazard of death for individual π with characteristics ππππ in county
π in year π relative to the first year when the individual π participated in the survey; β0,π(π)
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; ππππ includes
NCMS enrollment (ππΆπππππ), an indicator for NCMS county (ππΆππππ), a dummy for
having an agricultural Hukou (π΄ππππ»π’πππ’π), calendar year fixed effects (πΏπ‘π), and individual
characteristics (πΏπ) . 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 oneβs agricultural hukou
status to instrument for the individual NCMS enrollment status, ππΆπππππ, 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
ππΆπππππ = πΌ0 + πΌ1ππΆππππ Γ π΄ππππ»π’πππ’π + πΌ2ππΆππππ + πΌ3π΄ππππ»π’πππ’π +πΏπβ²πΌπ +
πΏπ + πΏπ‘π + νπππ, (4)
where ππΆπππππ indicates that individual π in county π has enrolled in the NCMS in year π
relative to the first year when the individual π participated in the survey; ππΆππππ indicates
25
that county π has adopted the NCMS in year π, and π΄ππππ»π’πππ’π indicates that individual π
has an agricultural Hukou. We control for the same set of controls as in equation (3) along
with county fixed effects (πΏπ).40 We predict the residual νοΏ½ΜοΏ½ππ after estimating equation (4).
We then estimate the second-stage equation (equation (3)), additionally controlling for
νοΏ½ΜοΏ½ππ. Based on Wooldridge (2015), the estimates of π½1 is consistent using this CF method.
Because νοΏ½ΜοΏ½ππ 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 county-specific 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.
26
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:
ππππ‘ = π½0 + π½1ππΆπππππ‘ + ππβ²π½π + πΏππ‘ + π΄ππππ»π’πππ’π Γ πΏπ‘ + π΄ππππ»π’πππ’π Γ πΏπ + ππππ‘ ,
(5)
where ππππ‘ is a measure of health outcome of individual π in county π in wave π‘; and
ππΆπππππ‘ indicates that individual π has enrolled in the NCMS in wave π‘. Because enrolling in
the NCMS is voluntary, ππΆπππππ‘ is endogeneous in equation (5). We use the IV method to
address the endogeneity, and the first stage specification is:
ππΆπππππ‘ = πΌ0 + πΌ1ππΆππππ‘ Γ π΄ππππ»π’πππ’π + ππβ²πΌπ + πΏππ‘ + π΄ππππ»π’πππ’π Γ πΏπ‘ +
π΄ππππ»π’πππ’π Γ πΏπ + ππππ‘. (6)
Compared equation (6) with (5), the excluded instrument for ππΆπππππ‘ is the interaction
term between NCMS passage in that county and individual π's Hukou status (ππΆππππ‘ Γ
42 Given the estimated hazard ratio of 0.694, we have
h(π‘|ππΆππ ππππππππππ‘π‘=1)
h(π‘|ππΆππ ππππππππππ‘π‘=0)= 0.694. Based on the
relationship between mortality hazard h(π‘) and mortality rateπΉ(π‘) in β ln(1 β πΉ(π‘)) = β« h(π’)ππ’ π‘
0, we have
βln(1βπΉ(π‘|ππΆππ ππππππππππ‘π‘=1)
βln(1βπΉ(π‘|ππΆππ ππππππππππ‘=0)=
β« h(π’|ππΆππ ππππππππππ‘π‘=1)ππ’ π‘0
β« h(π’|ππΆππ ππππππππππ‘π‘=0)ππ’π‘0
= 0.694. We then back out
πΉ(π‘|ππΆππ ππππππππππ‘π‘ = 1) since πΉ(π‘|ππΆππ ππππππππππ‘π‘ = 0) = 0.163 in our sample.
27
π΄ππππ»π’πππ’π). In both equations (5) and (6), the vector of covariates, ππ, 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 (πΏππ‘) to account for unobserved time-varying characteristics that are specific to a
county, hukou-by-wave fixed effects (π΄ππππ»π’πππ’π Γ πΏπ‘) to account for unobserved time-
varying characteristics that are specific to a hukou type, and hukou-by-county fixed effect
(π΄ππππ»π’πππ’π Γ πΏπ) 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.
28
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.
29
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.
30
Part VI: Mechanisms β 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.
31
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-of-
pocket 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
32
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
33
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.
34
of the same supply. 47 The results strongly confirm the equalizing impact of the NCMS-
induced 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 Γ 0.656 Γ 0.577).
35
instrument supply by πππ π‘ππΆππ Γ π΄ππππβπππ2000 Γ π΅ππ ππππππ π’ππππ¦ (the measure
included in Panel C of Table 8). We add as a control to the regression πππ π‘ππΆππ Γ
π΄ππππβπππ2000, 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 π΅ππ ππππππ π’ππππ¦ and
π΄ππππβπππ2000 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 ππΆππππππππππππ‘ (which continues to be instrumented by πππ π‘ππΆππ Γ
π΄ππππ»π’πππ’) and ππ’ππππ¦.
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 β 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 β 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.
36
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
37
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
38
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)
39
Bibliography
Baicker, K., Taubman, S. L., Allen, H. L., Bernstein, M., Gruber, J. H., Newhouse, J. P., ... &
Finkelstein, A. N. (2013). The Oregon experimentβeffects of Medicaid on clinical
outcomes. New England Journal of Medicine, 368(18), 1713-1722.
Babiarz, K. S., Miller, G., Yi, H., Zhang, L., & Rozelle, S. (2010). New evidence on the impact of
Chinaβs New Rural Cooperative Medical Scheme and its implications for rural primary
healthcare: multivariate difference-in-difference analysis. Bmj, 341, c5617.
Bai, C. E., & Wu, B. (2014). Health insurance and consumption: evidence from Chinaβs new
cooperative medical scheme. Journal of Comparative Economics, 42(2), 450-469.
Blumenthal, D., & Hsiao, W. (2015). Lessons from the EastβChina's rapidly evolving health care
system. New England Journal of Medicine, 372(14), 1281-1285.
Burns, L. R., & Liu, G. G. (Eds.). (2017). China's healthcare system and reform. Cambridge University
Press.
Chen, Y., & Jin, G. Z. (2012). Does health insurance coverage lead to better health and educational
outcomes? Evidence from rural China. Journal of health economics, 31(1), 1-14.
Cheng, L., Liu, H., Zhang, Y., Shen, K., & Zeng, Y. (2015). The impact of health insurance on health
outcomes and spending of the elderly: evidence from China's new cooperative medical
scheme. Health economics, 24(6), 672-691.
Curtin, L. R., & Klein, R. J. (1995). Direct standardization (age-adjusted death rates): US Department
of Health and Human Services. Public Health Service, Centers for Disease Control and
Prevention, National Center for Health Statistics, 6.
Du, L.,& Zhang, W. (2007) The Development on China's Health, No. 3, Social Science Academic
Press: Beijing, China.
Ehrlich, I., & Becker, G. S. (1972). Market insurance, self-insurance, and self-protection. Journal of
political economy, 80(4), 623-648.
Gertler, P., Locay, L., & Sanderson, W. (1987). Are user fees regressive?: The welfare implications of
health care financing proposals in Peru. Journal of econometrics, 36(1-2), 67-88.
Goldin, J., Lurie, I. Z., & McCubbin, J. (2021). Health insurance and mortality: Experimental evidence
from taxpayer outreach. The Quarterly Journal of Economics, 136(1), 1-49.
Gruber, J., Hendren, N., & Townsend, R. M. (2014). The great equalizer: Health care access and
infant mortality in Thailand. American Economic Journal: Applied Economics, 6(1), 91-107.
Huang, W., & Zhang, C. (2020) The Power of Social Pensions: Evidence from Chinaβs New Rural
Pension Scheme. American Economic Journal: Applied Economics. (Forthcoming).
Lei, X., & Lin, W. (2009). The New Cooperative Medical Scheme in rural China: does more coverage
mean more service and better health? Health economics, 18(S2), S25-S46.
Liang, L., & Langenbrunner, J. C. (2013). The long march to universal coverage: lessons from China.
Litvack, J. I., & Bodart, C. (1993). User fees plus quality equals improved access to health care:
results of a field experiment in Cameroon. Social science & medicine, 37(3), 369-383.
Liu, K. (2016). Insuring against health shocks: Health insurance and household choices. Journal of
health economics, 46, 16-32.
40
Liu, X., & Cao, H. (1992). China's cooperative medical system: Its historical transformations and the
trend of development. Journal of Public Health Policy, 13(4), 501-511.
Long, Q., Klemetti, R., Wang, Y., Tao, F., Yan, H., & Hemminki, E. (2012). High Caesarean section
rate in rural China: is it related to health insurance (New Co-operative Medical
Scheme)?.Social science & medicine, 75(4), 733-737.
Mao, Z. (2005). Pilot program of Ncms in China: system design and progress. Background paper for
the World Bank China Rural Health Study. World Bank Health. Nutrition and Population-China
Rural.
Milcent, C. (2018). Healthcare reform in China: From violence to digital healthcare. Springer.
Miller, Sarah, Norman Johnson, and Laura Wherry. βMedicaid and Mortality: New Evidence From
Linked Survey and Administrative Data.β December 17, 2020.
Nolte, E., & McKee, C. M. (2012). In amenable mortalityβdeaths avoidable through health careβ
progress in the US lags that of three European countries. Health Affairs, 31(9), 2114-2122.
Souteyrand, Y. P., Collard, V., Moatti, J. P., Grubb, I., & Guerma, T. (2008). Free care at the point of
service delivery: a key component for reaching universal access to HIV/AIDS treatment in
developing countries. Aids, 22, S161-S168.
State Council, 2002. "The Decision of the State Council on Strengthening Rural Healthcare." (in
Chinese)
Sun, X., Jackson, S., Carmichael, G., & Sleigh, A. C. (2009). Catastrophic medical payment and
financial protection in rural China: evidence from the New Cooperative Medical Scheme in
Shandong Province. Health economics, 18(1), 103-119.
Wagstaff, A., Lindelow, M., Jun, G., Ling, X., & Juncheng, Q. (2009). Extending health insurance to
the rural population: an impact evaluation of China's new cooperative medical
scheme. Journal of health economics, 28(1), 1-19.
Wagstaff, A. (2009). Reforming China's rural health system. World Bank Publications.
Wooldridge, J. M. (2015). Control function methods in applied econometrics. Journal of Human
Resources, 50(2), 420-445.
You, X., & Kobayashi, Y. (2009). The new cooperative medical scheme in China. Health policy, 91(1),
1-9.
Yi, J., Zhao, S. & Zou, H. (2021). Estimating an Equilibrium Model of Health Insurance and Healthcare
Expenditure. Working paper.
Zhou, M., Liu, S., Kate Bundorf, M., Eggleston, K., & Zhou, S. (2017). Mortality in rural China declined
as health insurance coverage increased, but no evidence the two are linked. Health
Affairs, 36(9), 1672-1678.
ZombrΓ©, D., De Allegri, M., & Ridde, V. (2017). Immediate and sustained effects of user fee
exemption on healthcare utilization among children under five in Burkina Faso: a controlled
interrupted time-series analysis. Social Science & Medicine, 179, 27-35.
41
Figure 1. NCMS Rollout
Notes: This figure plots the fraction of counties implementing the NCMS over time. The number of NCMS
counties is obtained from 2005 -2011 China Health Statistical Yearbook and the total number of counties from
2011 China Civil Affairs' Statistical Yearbook.
42
Figure 2. Distribution of the share of agricultural hukou out of the population
Note: This figure plots the distribution of the share of population with an agricultural hukou from China's 2000
Population Census.
43
Panel A. Mortality rate
Panel B. Life expectancy at birth
Figure 3. Mortality rate and Life expectancy around NCMS adoption: subsamples by agricultural share
Notes: This figure plots the average value of the residuals from regressing Mortality rate (Panel A) and Life
expectancy (Panel B) on NCMS-timing-by-year FE and county-specific linear trend. Counties in the sample are
split by the median value of the share of population with an agricultural hukou. Observations more than 3
years (5 years) before (after) the NCMS are binned into groups. Sample period is from 2004 to 2010.
44
Panel A. Mortality rate
Panel B. Life expectancy at birth
Figure 4. Mortality rate and Life expectancy around NCMS adoption
Notes: This figure plots the estimated coefficients on the interactions between year-to-NCMS dummies and
2000 agricultural share from the regression model specified in equation (2). The year before NCMS adoption is
omitted, so the estimates are normalized to zero in that year. Observations more than 3 years (5 years) before
(after) the NCMS are binned into groups. Results are weighted by county population. Sample period is from
2004 to 2010. The dashed lines represent the 95 percent confidence intervals based on standard errors
clustered at the county level.
45
Figure 5. OOP spending vs. Total spending by NCMS enrollment
Notes: This figure plots the OOP spending against the total spending from the CHNS. The sample is restricted
to those with positive medical spending both before and after enrolling in the NCMS.
46
Table 1. Summary of public health insurance programs in developing countries
Year Country Name of Insurance Scheme Number of People (year)
1990 Brazil Sistema Γnico de SaΓΊde (SUS) 200 million (2018)
1992 Vietnam Social Health Insurance 80 million (2018)
1993 Colombia Regimen Subsidiado (SR) 34 million (2016)
1995 Philippines Philippines Health Insurance Corporation (PhilHealth) up to 104 million (2018)
1999 Rwanda Mutuelles de SantΓ© 12 million (2018)
2002 Thailand Universal Coverage Health Scheme (UCS) 53 million (2018)
2003 China New (Rural) Cooperative Medical Scheme (NCMS or NRCMS) 670 million (2015)
2005 Ghana National Health Insurance Scheme (NHIS) 11 million (2014)
2018 India Pradhan Mantri Jan Arogaya Yojana (PM-JAY) 103 million (2019)
Notes: Authors review of material on developing country health insurance expansions. Data sources listed in Appendix 1. For the Philippines, there is larger controversy over the reach of the program, so we list an upper bound here.
47
Table 2. Summary statistics of DSP data
DSP sample National statistics
Mean S.D. Obs
Age-adjusted mortality rate 546.298 126.882 847 568
Life expectancy at birth 75.325 3.642 847 74
Share of Agricultural population in 2000 0.707 0.254 121 0.7
GDP per capita 22,663 14945 847 22,072
Wage 26,705 10851 847 27,102
Notes: a. DSP sample: Age-adjusted death rate is measured as the number of deaths per 100 thousand people. The age adjustment is based on the China's 2000 Population Census. Share of agricultural population in 2000 is obtained from China's 2000 Population Census. GDP per capita and wage are obtained from China City Statistical Yearbook 2005-2011 and measured in 2010 RMB. The statistics are weighted by the county population. b. National statistics: Mortality rate and life expectancy at birth are statistics in 2007, calculated using the China's 2000 and 2010 Population Census. Share of agricultural population in 2000 is obtained from China's 2000 Population Census. GDP per capita and wage are the average values over 2004 through 2010 and measured in 2010 RMB, obtained from 2011 China Statistical Yearbook.
48
Table 3. Summary statistics of the main variables from the CLHLS and the CHNS
Mean S.D. Obs
Panel A: CLHLS individual sample
Health outcomes Suffering from serious illness (past two years, Dummy) 0.183 0.387 53,224
Limited activities of daily living (ADL) (Dummy) 0.252 0.434 53,224
Mental health (Dummy) 0.228 0.419 53,224
Self-reported health (Dummy) 0.853 0.354 53,224
Interviewer-reported health (Dummy) 0.862 0.345 53,224
Healthcare utilization Adequate medical service (at present, Dummy) 0.920 0.271 58,565
Positive medical spending (the previous year, Dummy) 0.791 0.406 31,378
Total medical spending (the previous year, 1,000 RMB) 2.733 6.064 24,832
Financial difficulty in getting medical service (at present, Dummy) 0.059 0.236 52,046
Health behaviours Smoking (Dummy) 0.177 0.382 58,589
Alcohol consumption (Dummy) 0.191 0.393 58,589
Panel B: CHNS individual sample
Health outcomes
Cognitive deficit (Dummy) 0.509 0.500 5,942
Limited activities of daily living (ADL) (Dummy) 0.333 0.471 5,942
Diabetes (Dummy) 0.019 0.137 41,491
Hypertension (Dummy) 0.093 0.290 41,491
Healthcare utilization Healthcare use (past 4 weeks, Dummy) 0.109 0.311 46,519
Inpatient care use (past 4 weeks, Dummy) 0.009 0.096 46,519
Outpatient care use (past 4 weeks, Dummy) 0.095 0.293 46,519
Preventive care use (past 4 weeks, Dummy) 0.037 0.190 46,519
Total medical spending (past 4 weeks, 1,000 RMB) 1.537 6.546 4,005
Out-of-pocket spending (past 4 weeks, 1,000 RMB) 1.112 3.807 680
Out-of-pocket /Total spending (past 4 weeks) 0.952 0.168 680
Health behaviours Smoking (Dummy) 0.241 0.428 40,909
# Cigarettes per day 16.183 9.079 9,624
Alcohol consumption (Dummy) 0.289 0.454 40,909
Consumption of alcohol per day (liang, 50mg) 1.551 1.979 10,888
Panel C. CHNS community sample
# Public hospitals (per 10,000 people) 0.543 1.329 790
# Public clinics (per 10,000 people) 1.119 2.499 790
# Private clinics (per 10,000 people) 0.343 1.412 790
# Clinic (per 10,000 people) 1.462 2.803 790
# Healthcare staff in medical facilities (per 1,000 people) 4.921 15.819 790
# Beds in medical facilities (per 1,000 people) 5.006 14.922 790
49
Notes: This table reports the summary statistics of the main variables from the CLHLS and the CHNS. Panel A and Panel B show the individual-level health outcomes, healthcare utilization, and health behaviors from the CLHLS and the CHNS, respectively. The monetary values are measured in 2010 RMB. Panel C summarizes variables used in the analysis of community healthcare supply from the CHNS, weighted by community population.
50
Table 4. Impacts of NCMS adoption on aggregate mortality
Baseline Removing covariates Sample from 2004 to 2007 Placebo test
(1) (2) (3) (4) (5) (6) (7) (8)
Age-adjusted
Mortality Rate Life
Expectancy Age-adjusted
Mortality Rate Life
Expectancy Age-adjusted
Mortality Rate Life
Expectancy Age-adjusted
Mortality Rate Life
Expectancy
Post NCMS Γ AgriShare2000 -0.195** 0.040** -0.196** 0.040*** -0.254** 0.051** -0.007 0.003
(0.082) (0.015) (0.081) (0.015) (0.120) (0.026) (0.204) (0.036)
County FE Yes Yes Yes Yes Yes Yes Yes Yes
Province-by-year FE Yes Yes Yes Yes Yes Yes Yes Yes
NCMS-timing-by-year FE Yes Yes Yes Yes Yes Yes Yes Yes
County-specific linear trend Yes Yes Yes Yes Yes Yes Yes Yes
Covariates Yes Yes No No Yes Yes Yes Yes
Obs 847 847 847 847 484 484 704 704
Adjusted R-squared 0.758 0.705 0.758 0.705 0.735 0.634 0.765 0.712
Mean 546.298 75.325 546.298 75.325 550.041 75.271 836.383 70.798
Notes: This table shows the estimated coefficients from equation (1) in various specifications. Columns (1) and (2) report the baseline results from equation (1) using DSP sample from 2004 to 2010. Columns (3) and (4) report results from specifications removing economic covariates. Columns (5) and (6) restrict the sample period to 2004 -2007. Columns (7) and (8) report the results from a placebo test using the sample from 1994 to 2000. Age-adjusted mortality rate and life expectancy are in logarithm. Results are weighted by county population. Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
51
Table 5. Impacts of NCMS enrollment on health outcomes and healthcare utilization using CLHLS
Mortality hazard Other health outcomes Healthcare Utilization
(1) (2) (3)
(4) Baseline Control function
NCMS enrollment 0.554*** 0.694***
(0.025) (0.062)
1st
-stage Residuals
0.775***
(0.069)
Obs 78,446 78,446
Suffering from serious illness
-0.129**
(0.063)
Obs
53,224 Mean
0.183
Limited ADL
-0.125**
(0.063)
Obs
53,224 Mean
0.252
Mental health
0.111**
(0.054)
Obs
53,224 Mean
0.228
Self-reported health
0.123**
(0.051)
Obs
53,224 Mean
0.853
Interviewer-reported health
0.122**
(0.050)
Obs
53,224 Mean
0.862
Adequate medical service
0.089**
(0.043)
Obs
58,565
Mean
0.920
Positive medical spending
0.118*
(0.067)
Obs
31,378
Mean
0.791
Total medical spending (log)
0.418*
(0.222)
52
Obs
24,832
Mean 2733.296
Year FE Yes Yes No No
County stratification Yes Yes No No
Covariates Yes Yes Yes Yes
County-by-wave FE No No Yes Yes
Hukou-by-wave FE No No Yes Yes
Hukou-by-County FE No No Yes Yes
Notes: This table reports the estimation results from equations (3) - (6) using the CLHLS. Column (1) shows how NCMS enrollment is related to one's mortality hazard based on equation (3). Column (2) presents the results from the control function method, with πππ π‘ππΆππ Γ π΄ππππ»π’πππ’ as an instrument. All estimates are reported after taking exponential to represent proportional increases or decreases in mortality rates relative to one. Standard errors in parentheses are computed based on 1,000 bootstraps and converted via the delta method. Columns (3) and (4) report the results on other health outcomes and healthcare utilization in the sample of survivors. Each cell presents the IV coefficients on ππΆππππππππππππ‘ based on equations (5) and (6). Dependent variables in these regressions are shown in the first column. Information on medical spending is available after 2005. Monetary values are in 2010 RMB. Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
53
Table 6. Impacts of NCMS enrollment on health outcomes and healthcare utilization using CHNS
(1) (2) (3)
Health outcomes Healthcare utilization Medical spending
Cognitive Deficit -0.220*
(0.126)
Obs 5,942
Mean 0.509
Limited ADL -0.257**
(0.128)
Obs 5,942
Mean 0.333
Diabetes -0.015*
(0.009)
Obs 41,491
Mean 0.019
Hypertension -0.065**
(0.030)
Obs 41,491
Mean 0.093
Healthcare use 0.084**
(0.040)
Obs 46,519
Mean 0.109
Inpatient care 0.019*
(0.010)
Obs 46,519
Mean 0.009
Outpatient care 0.072**
(0.035)
Obs 46,519
Mean 0.095
Preventive care 0.054**
(0.026)
Obs 46,519
Mean 0.037
Total medical spending (log) 0.387
(0.638)
Obs 4,005
Mean 1537.011
54
Covariates Yes Yes Yes
County-by-wave FE Yes Yes Yes
Hukou-by-wave FE Yes Yes Yes
Hukou-by-County FE Yes Yes Yes
Notes: This table shows the results about health outcomes, healthcare utilization and medical spending from CHNS. Each cell presents the IV coefficients on NCMS enrollment based on equations (5) and (6). Dependent variables in these regressions are shown in the first column. πΏππππ‘πππ΄π·πΏ and πΆπππππ‘ππ£ππππππππ‘are measured for those age 55+ only and the information is available in waves 2000 - 2006. π·πππππ‘ππ and π»π¦ππππ‘πππ πππ are measured at all ages above 12. Sample in Column (3) is restricted to those with positive medical spending. Monetary values are in 2010 RMB. Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
55
Table 7. Impacts of the NCMS on financial protection
(1) (2) (3)
Financial difficulty Log (OOP spending) OOP/Total spending
NCMS enrollment -0.051** -0.429 -0.289**
(0.026) (1.163) (0.135)
Covariates Yes Yes Yes
County-by-wave FE Yes Yes Yes
Hukou-by-wave FE Yes Yes Yes
Hukou-by-County FE Yes Yes Yes
Obs 52,046 680 680
Mean 0.059 1112.364 0.952
Notes: This table reports the estimation results from equations (5) and (6) and the dependent variables are shown in the first row. Column (1) uses the CLHLS sample and πΉπππππππππππππππ’ππ‘π¦ is an indicator for not seeking any medical service due to lack of money. This information is available after 2000. Columns (2) and (3) use the CHNS sample and the sample is restricted to those with positive OOP spending both before and after the NCMS initiation in their communities. Monetary values are in 2010 RMB. Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
56
Table 8. Healthcare supply
(1) (2) (3) (4) (5) (6)
Public hospitals Public clinics Private clinics Clinics Healthcare workers Beds
Panel A. NCMS effects on healthcare supply (log)
Post NCMS Γ AgriShare2000 0.461*** 0.620*** -0.528** 0.081 1.031** 0.819**
(0.159) (0.232) (0.249) (0.284) (0.405) (0.389)
Panel B. Controlling for lead NCMS
Post NCMS Γ AgriShare2000 0.454*** 0.618** -0.514** 0.092 1.041** 0.808*
(0.162) (0.245) (0.253) (0.285) (0.422) (0.408)
Post NCMS (2-yr lead) Γ AgriShare2000 0.050 0.017 -0.091 -0.078 -0.066 0.077
(0.183) (0.247) (0.208) (0.262) (0.455) (0.455)
Panel C. Interaction effects with baseline supply (log)
Post NCMS Γ AgriShare2000 0.673*** 1.451*** -0.121 1.070*** 1.855*** 1.444***
(0.147) (0.285) (0.199) (0.342) (0.435) (0.400)
Post NCMS Γ AgriShare2000 Γ Baseline supply (log) -0.656*** -0.844*** -1.470*** -0.792*** -0.916*** -0.874***
(0.245) (0.177) (0.202) (0.171) (0.210) (0.244)
P50 vs. P75 of baseline supply (percentage points) 28 114 65 45 68 76
Obs 790 790 790 790 790 790
Mean 0.543 1.119 0.343 1.462 4.921 5.006
Notes: This table presents the results about the supply-side response to NCMS using the CHNS. Dependent variables are shown in the first row. The number of hospitals and clinics are measured as the number per 10,000 people. The number of healthcare staff and beds are measured as the number per 1,000 people. All supply measures are in logarithm. Panel A shows the estimated coefficients on πππ π‘ππΆππ Γ π΄ππππβπππ2000, Panel B includes the two-year lead of NCMS interacted with π΄ππππβπππ2000, and Panel C includes the interactions of πππ π‘ππΆππ Γ π΄ππππβπππ2000with the same supply in the first row. π50π£π . π75 (Baseline supply) shows the differences in the percentage changes of supply for communities at the median of baseline supply relative to those at the top quartile, assuming a median share of agricultural population (0.75). Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
57
Table 9. Impacts of healthcare supply on healthcare utilization and health outcomes
(1) (2) (3) (4)
Public hospitals Clinics
Healthcare workers
Beds
Panel A. Healthcare use NCMS enrollment 0.118* 0.110* 0.132* 0.116*
(0.063) (0.063) (0.069) (0.069)
Healthcare supply (log) 0.029 0.032 0.029 0.029
(0.023) (0.031) (0.019) (0.036)
Obs 41,335 41,335 41,335 41,335
Mean 0.092 0.092 0.092 0.092
Panel B. Limited ADL NCMS enrollment -0.378 -0.351 -0.572 -0.534
(0.295) (0.230) (0.654) (0.994)
Healthcare supply (log) -0.065 -0.089 -0.170 -0.179
(0.231) (0.137) (0.341) (0.783)
Obs 4,671 4,671 4,671 4,671
Mean 0.335 0.335 0.335 0.335
Panel C. Cognitive Deficit NCMS enrollment -0.478 -0.591** -0.595 -1.264
(0.426) (0.243) (0.514) (2.288)
Healthcare supply (log) 0.165 0.030 0.013 -0.502
(0.319) (0.125) (0.236) (1.759)
Obs 4,671 4,671 4,671 4,671
Mean 0.534 0.534 0.534 0.534
Panel D. Diabetes NCMS enrollment -0.036 -0.002 -0.037 -0.037
(0.035) (0.049) (0.035) (0.037)
Healthcare supply (log) -0.008 -0.029 -0.005 -0.007
(0.014) (0.023) (0.009) (0.019)
Obs 36,517 36,517 36,517 36,517
Mean 0.018 0.018 0.018 0.018
Panel E. Hypertension NCMS enrollment -0.062 -0.010 -0.072 -0.093
(0.055) (0.074) (0.064) (0.117)
Healthcare supply (log) -0.017 -0.040 -0.024 -0.067
(0.026) (0.033) (0.022) (0.079)
Obs 36,517 36,517 36,517 36,517
Mean 0.092 0.092 0.092 0.092
Covariates Yes Yes Yes Yes
County-by-wave FE Yes Yes Yes Yes
Hukou-by-wave FE Yes Yes Yes Yes
Hukou-by-County FE Yes Yes Yes Yes
58
Notes: This table shows the results of how NCMS enrollment and local healthcare supply affects individuals' healthcare utilization and health outcomes. We re-estimate equations (5) and (6), additionally including local supply (log) shown in the first row. We instrument ππΆππππππππππππ‘ and π»ππππ‘βπππππ π’ππππ¦(πππ) with πππ π‘ππΆππ β π΄ππππ»π’πππ’, and πππ π‘ππΆππ β π΄ππππβπππ2000 β π΅ππ ππππππ π’ππππ¦(πππ) as shown in the first row. In addition to controls in equations (5) and (6), we also control for πππ π‘ππΆππ β π΄ππππβπππ2000, π΅ππ ππππππ π’ππππ¦(πππ), and π΄ππππβπππ2000. Each panel shows the IV estimates on NCMS enrollment and healthcare supply for different outcome measures. Standard errors clustered at the county level are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.
59
Table 10. Ex-ante moral hazard: smoking and alcohol consumption
CLHLS sample CHNS sample
(1) (2) (3) (4) (5) (6)
Smoking Drinking Smoking # cigarettes (log) Drinking Amount alcohol (log)
NCMS enrollment -0.029 0.011 -0.057 0.098 -0.054 -0.172
(0.034) (0.046) (0.035) (0.141) (0.044) (0.228)
Covariates Yes Yes Yes Yes Yes Yes
County-by-wave FE Yes Yes Yes Yes Yes Yes
Hukou-by-wave FE Yes Yes Yes Yes Yes Yes
Hukou-by-County FE Yes Yes Yes Yes Yes Yes
Obs 58,589 58,589 40,909 9,624 40,909 10,888
Mean 0.177 0.191 0.241 16.183 0.289 1.551
Notes: This table shows the impacts of the NCMS on smoking and alcohol consumption based on equations (4). Results in Columns (1) and (2) rely on the CLHLS, and results in Columns (3) - (6) rely on the CHNS and restricted to age above 12. Dependent variables are in the first row. Standard errors clustered at the county in parentheses. * p<0.10, ** p<0.05, *** p<0.01.