pesticide use and health outcomes: evidence from agricultural water pollution … · 2017-02-07 ·...
TRANSCRIPT
Pesticide Use and Health Outcomes: Evidence from Agricultural Water Pollution in China
Wangyang Lai*1
October 29, 2016
Job Market Paper
Abstract:
By linking provincial pesticide usage reports from several Chinese statistical yearbooks
(1998-2011) with the Chinese Longitudinal Healthy Longevity Survey (1998-2011), this study
provides new evidence that pesticides adversely affect health outcomes via drinking water
exposure. I follow a difference-in-difference-in-differences framework to compare health
outcomes between people who drink surface water and ground water in regions with different
intensities of rice pesticide use before and after 2004, when China shifted from taxing
agriculture to subsidizing agricultural programs. The results indicate that a 10% increase in
rice pesticide use unfavorably alters a key medical disability index (Activities of Daily Living
or ADL) by 1% for rural residents 65 and older. This is equivalent to 2.4 and 0.72 million
dollars in medical costs and offspring’s human capital losses, respectively (in total, 0.03% of
rice production profits). Further, I provide suggestive evidence on intergenerational transfer of
caring burden by showing pesticide use reduces out-migration of the offspring in affected
households. The results are robust to a variety of robustness checks and falsification tests.
Key Words: Pesticide, Drinking Water, Public Health, Triple Difference Estimator, Medical
Costs, Human Capital Costs, China
JEL Codes: I1, Q1, Q5
* Department of Agricultural, Environmental and Development Economics, The Ohio State University, Email: [email protected]. I thank Brian Roe for comments and suggestions. I also thank Abdoul Sam, Ani Katchova, Daeho Kim, Elena Irwin, Joyce Chen, Wei Cheng, Xiaohui Tian, Wendong Zhang, Feng-an Yang, Rodrigo Perez Silva, Christopher Dunphy and seminar participants from the Department of Agricultural, Environmental and Development Economics, Ohio State University and the Heartland Environmental and Resource Workshop in University of Illinois Urbana-Champaign. All errors are my own.
1
Introduction
Agricultural non-point source pollution is a growing concern in developing countries due to
the extensive use of synthetic organic chemicals. China, in particular, experienced the fastest
increase in pesticide usage rates and now applies between 1.5 to 4 times more pesticides than
the world average (Li et al. 2014 and Zhang et al. 2015). This level of pesticide use is a
major source of pollutants found in the air, waters, soil and farm products and these
chemicals have been shown to cause both acute and chronic illnesses (Huang et al. 2005; Hu
et al. 2015; Qiao et al. 2012). Concerns about human health problems resulting from
pesticides likely contributed to China’s prioritization of environmental protection in recent
policy statements as well as the start of supply-side reform in the farm sector (No.1 Central
Document 2016; Reuters 2016).
The objective of this paper is to quantify the impact of pesticide use on health outcomes
attributable to agricultural water pollution. Previous research documents strong evidence of
the connection between contaminated water from industrial activities and various types of
diseases in China (Ebenstein 2012; Wang et al. 2008; Wu et al. 1999). However, according to
the China Pollution Source Census (2010), agriculture is now a bigger non-point source of
water contamination in China than factory effluent. Systematic analysis of this agricultural
water pollution and its possible health consequences is lacking. To investigate this question, I
link provincial data from several Chinese statistical yearbooks (1998-2011) with the Chinese
Longitudinal Healthy Longevity Survey (1998-2011).
First, I follow a difference-in-difference-in-differences (DDD) framework to compare
health outcomes between people who drink surface water versus ground water (ground water
2
suffers much less from pesticide contamination) in regions with different intensities of rice
pesticide use before and after 2004, when China shifted from taxing agricultural outputs to
subsidizing agriculture. Individual fixed effects are also implemented to account for potential
confounding factors. The parallel trend assumption is indirectly tested via graphical analysis
to establish whether high and low pesticide application regions differ in health outcomes
before 2004. I also gauge the stability of the regression results by adding or deleting a
comprehensive set of time-varying individual, household and regional characteristics.
The result shows that a 10% (2.9 yuan/mu1) increase in rice pesticide use unfavorably
alters the index of dependence known as the Activities of Daily Living2 (Katz et al. 1963) by
1% (0.07 units) for residents 65 years or older. This is consistent with the existing
epidemiological and economic literature that pesticide exposure adversely affects the human
peripheral nervous system (Ding and Bao 2014; Li et al 2014; Hu et al. 2015; Starks et al.
2012) and, therefore, can decrease the independence of older adults via degradation of this
key system. This is also consistent with the evidence that Fipronil and Imidacloprid, which
both feature insect neurotoxins, are the two most widely detected pesticides in surface water
during rice production (China Pollution Source Census, 2010). The result suggests that the
health impact of pesticide use is underestimated by previous research because degradation of
the peripheral nervous system such as losses in sensation and muscle strength, if not
accompanied with clinical symptoms, will not be attributed to pesticide exposure.
In addition, the significant results are driven by populations in rural areas where
1 1 yuan = 0.15 dollar. 1 mu = 0.067 hectares. 2.9 yuan/mu = 6.6 dollars/hectare. 2 The Activity of Daily Living (ADL) index is widely used in the medical literature to measure the dependence of older and physical inactive individuals. It is also referred as a “disability” and “functional limitations” index (Verbrugge and Jette 1994; Johnson and Wolinsky 1993).
3
millions of residents rely on surface water for their drinking water supply. This suggests an
unintentional policy side effect: the pesticides, promoted to increase farmers’ income and
welfare via enhanced agricultural productivity, are also degrading the physical independence
of older adults exposed to surface water that contains pesticide residues. Further, the study
provides suggestive evidence that pesticide use reduces out-migration of the offspring in
affected households because it increases the dependence of older household members. This
suggests that China’s aging crisis, accompanied with increasing dependence of older adults
caused by increased pesticide use, places an even bigger burden on working-age cohorts.
Several robustness checks help assess the validity of these results. A potential concern
is that, because agricultural support programs are not exclusively targeted at pesticides, the
results may be driven by other changes caused by this policy. I rule out this possibility by
conducting a falsification test in which rice pesticide use is replaced with rice fertilizer use.
The estimation would also be potentially problematic if people in regions with more
pesticide use are more likely to change drinking water sources after 2004. I address this
concern by testing whether the rates of surface water reliance for drinking sources differ
before and after 2004 in regions with different intensities of pesticide use. Another potential
concern is that, because regions with more pesticide water pollution may also have more
pesticide pollution in the air, soil and food, the results may be driven by pesticide pollution
via these other pathways. I rule out this possibility by running analogous DDD regressions
with respect to corn and wheat, which do not feature flooding as a cultural practice and, thus,
should feature less water pollution. The results are robust to a variety of robustness checks
and falsification tests.
4
Finally, I calculate the medical and human capital losses due to agricultural pesticides
by exploiting the relationship between health outcomes and medical costs and offspring’s
weekly caring hours. The results indicate that each additional unit increase in ADL increases
medical costs and offspring’s weekly caring hours by 139 yuan (20.85 dollars) and 5.95
hours, which is equivalent to approximately 42 yuan (6.3 dollars). In combination with
previous information, I am able to calculate that a 10% increase in rice pesticide use
increases medical costs and human capital losses of local populations by 15.98 and 4.83
million yuan, respectively (or 2.4 and 0.72 million dollars). This is equivalent to 0.03% of
national annual rice production profits.
This work makes several contributions. First, it is broadly related to a body of research
in economics on water pollution and health outcomes (Brainerd and Menon 2014; Cutler and
Miller 2005; Ebenstein 2012; Greenstone and Hanna 2011). Specifically, this study adds to
the existing literature by providing new evidence on the impact of water pollution
attributable to agricultural non-point pollution in China. The work most closely related to
this paper is Brainerd and Menon (2014), which assesses the impact of fertilizer runoff on
infant and child health in India from 1992 to 2005 by exploring variations from seasonal and
regional fertilizer use. They find that children exposed to higher concentrations of fertilizer
during their first month experience worse health outcomes. This paper complements their
study by emphasizing the other main agricultural chemical, i.e. pesticides, as well as another
vulnerable group, i.e. older adults. These two studies, both using data from one of the most
populated developing countries, provide a more complete quasi-experimental picture that
agricultural chemicals have led to deterioration in health outcomes. This issue has salience in
5
other developing countries that face similar challenges.
Another distinguishing facet of this paper is the emphasis on non-occupational pesticide
exposure. This is necessary because only a small proportion of the population is likely to
receive a pesticide dose via direct exposure that is high enough to cause acute symptoms.
However, many more may be at risk of developing chronic effects because of exposure to
contaminated water, air and soil, and due to ingestion of residues from farm products.
Previous economic literature, which estimates the effect of direct exposure or only deals with
“seriously poisoned” samples (Hu et al. 2015; Macharia, Mithöfer and Waibel 2013; Qiao et
al. 2012), may underestimate the broader deleterious effects of pesticide use.
A third feature that distinguishes this work from existing research is the identification
strategy. A central methodological concern is that the independent variable of interest,
pesticide use, is not randomly assigned. Most existing studies use cross-sectional data sets
(Huang et al. 2000; Qiao et al. 2012), which are subject to potential bias in ways that are not
obvious a priori. By contrast, my identification strategy first employs individual fixed effects,
so that estimates of the impact of pesticide use can be purged of the influence of unobserved
time-invariant characteristics. More importantly, the identification strategy exploits the
exogenous policy change, regional variations in pesticide use and different sources of
drinking water, which supports the view that the association between pesticide use and
health outcomes is a causal relationship rather than simply a correlation.
A fourth distinguishing aspect of this work is its reliance on more comprehensive data.
Existing work employs much more restricted data sets with respect to health outcome
information, length of time, sampling area and sample size (Huang et al. 2000; Qiao et al.
6
2012). Therefore, the current study provides a broader scope and greater statistical power
compared to previous work. In comparison with the existing research, I am able to examine a
wide range of diverse health outcomes and the span of the data allows us to consider
illnesses that may take years to manifest. In terms of economic costs, besides medical costs, I
am also able to estimate human capital losses suffered by the caretakers of those affected,
which are not addressed in previous literature. Besides, one advantage of using Chinese data
for this investigation is that population mobility was extremely low due to government
regulations and, therefore, the survey respondents’ residence at the time of observation in the
data allow for identification of their long-term water pollution exposure (Chen et al. 2013;
Ebenstein 2012).
The remainder of this paper is organized as follows. The next two sections describe
institutional background and the data. The next section examines the impact of pesticide use
on health outcomes. This section includes model specification, regression results and
robustness checks. The final section concludes.
Institutional Background
Pesticide Use in China
Food security and self-reliance remain top priorities in China. In order to feed 19% of the
world’s population with 7% of the world’s arable land, China promotes many measures to
boost crop yields, including the extensive use of pesticides. The use of pesticides in China
has increased sharply, from 0.76 million tons in 1991 to 1.8 million tons in 2011, amounting
to an average annual growth rate of 4.9%. Pesticide use per hectare is roughly 1.5 to 4 times
larger than the world average (Li et al. 2014 and Zhang et al. 2015).
7
Figure 1. Annual Chinese Grain Production and Agricultural Subsidies
Notes: Panel 1-A depicts annual grain production (rice, wheat and corn) in thousands of tons. Panel
1-B depicts annual agricultural subsidies in 100 million yuan (including a direct payment, a subsidy
for improved seed varieties, a partial rebate for farm machinery purchases and a general-input
subsidy). Sources: China Statistical Yearbook and Agricultural Cost and Benefit Yearbook.
The most recent discrete change in pesticide use comes in 2004 when China shifted
from taxing agriculture to subsidizing agricultural programs (See Gale et al. 2013 for
detailed discussions of China’s agricultural support policies). One main driver of this policy
change is concern about maintaining food security and self-reliance. As Figure 1A shows,
grain production had been stagnant from 1996 to 1999 and then decreased sharply from 1999
through 2003. Rising off-farm wages increased the opportunity cost of farm labor,
weakening incentives to engage in agricultural production. This raised great concern about
food security. In 2004, authorities began to eliminate an agricultural tax on farmers and
introduced three subsidies targeted at grain producers: a direct payment, a subsidy for
improved seed varieties, and a partial rebate for farm machinery purchases. In 2004-06,
several other support programs were introduced such as a general-input subsidy, price floors
for wheat and rice, reform of the grain marketing system and transfer payments to grain
counties. Since then, expenditures on the initial set of programs have grown rapidly (Figure
1B). This set of supporting programs has bolstered farmer’ production incentives (Chen, Fan
300
0035
000
400
0045
000
1995 2000 2005 2010
A: Grain Production (10 thousand tons)
050
01
000
1500
1995 2000 2005 2010
B: Agricultural Subsidies (100 million yuan)
8
and Gao 2010; Chen, Wu and Wang 2010; Zhou, Zhao and Zhang 2009) and crop production
increased continuously since 2004, which has been accompanied with greater pesticide use.
As Figure 2 shows, annul pesticide use rates for the three main crops (rice, wheat and corn)
increased sharply after 2004.
Figure 2. Annual Pesticide Expenditures (Yuan/Mu) by Crop
Notes: Figure 2 depicts annual pesticide expenditures by rice, wheat and corn. Pesticide prices in
1997 are normalized to 1. Sources: Agricultural Cost and Benefit Yearbook.
As for the types of pesticide, organochlorine (OC) pesticides such as DDT were widely
used for agricultural pest control between the 1950s and 1980s. This extensive use of OCs
caused more than 45,000 poisoning cases each year during the 1980s and 1990s and the
number was believed to be higher before the 1980s although statistical records are lacking
(Qiao et al. 2012). In the early 1980s, the Chinese Government banned the use of OCs and
some organophosphates (OPs) in crop production because of their persistence in the
environment and their toxic properties for humans and wildlife. Consequently,
organophosphates (OPs) and pyrethroids (PYRs) have become attractive alternatives because
of their relatively lower toxicity and environmental persistence. However, in recent years,
scientists have found that even pesticides categorized as less and least toxic can cause critical
010
2030
40
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
Annual Pesticide Expenditures (Yuan/Mu) by Crop
rice wheat corn
9
chronic diseases that are not easy to diagnose and that some of these pesticides are even fatal
to humans with persistent exposure (Pingali et al., 1997; Qiao et al. 2012; Wang et al. 2012).
The detection of OPs in Arctic ice contradicts initial claims that OPs are readily degraded by
the environment (Macdonald et al. 2000; Zhang et al. 2002).
Rural Drinking Water
The safety of drinking water in rural areas also has received policy attention in China. In the
1980s, the Chinese government started the rural drinking water treatment program and spent
significant resources to improve the quality of drinking water in rural China (Zhang 2012).
However, the rural drinking water program is far from being complete. Even in 2010 there
were still 298 million rural Chinese without safe drinking water service (Yu et al. 2015).
For many decades, river pollution was mainly due to industrial activities. However,
according to the China Pollution Source Census (2010), agriculture is now a bigger source of
water contamination in China than factory effluent. In terms of pesticide pollution, numerous
studies document that various types of pesticide, including OCs, OPs and PYRs, are detected
in rivers, lakes, estuaries and coastal oceans (Guo et al. 2008; Xian et al. 2008; Zhang, Jiang
and Qu 2011). Pesticide runoff to ground water is less frequently documented and the
severity differs greatly depending on environmental factors such as soil type, rainfall amount
and the level of ground water (Sun, Wang and Jin 2009).
The China Pollution Source Census (2010) conducted pesticide runoff experiments for
both surface and ground water in 372 sites across China for a one year cycle from 2007 to
2008. For the treatment groups, they followed the local farmers’ practices of pesticide use.
That is, they used the same types and amounts of pesticides with identical spraying methods
10
and timing. For the control groups, they only used pesticides categorized as low toxicity and
high degradability. They calculated the runoff coefficients as the differences of runoff
percentage between treatment and control groups. Many types of pesticides, including
Fipronil, Imidacloprid, Dichlorvos and Butachlor, are detected in both surface and ground
water. The runoff coefficients are much higher in the surface water than in ground water. For
example, the average Fipronil runoff coefficient in the surface water was 0.15 while none
was detected in ground water. The Fipronil runoff coefficients are especially high in the
southern plains for single and double cropping rice (0.60 and 0.51 respectively).
Data and Summary Statistics
To conduct the analysis, I link provincial data from several Chinese statistical yearbooks
(1998-2011) with the Chinese Longitudinal Healthy Longevity Survey (1998-2011). The
pesticide expenditures per mu by different crops come from Agricultural Cost and Benefit
Yearbook (1998-2011) and other regional variables come from the China Statistical
Yearbook (1998-2011) and the China Environmental Statistical Yearbook (1998-2011).
The individual health data come from Chinese Longitudinal Healthy Longevity Survey
(CLHLS). CLHLS is chosen for this research because of its rich health and demographic
information, its nationwide sampling areas and length of time which covers the year of the
agricultural policy change. CLHLS was collected by the Center for Healthy Aging and
Development Studies at Peking University and co-sponsored by the US National Institute on
Aging. This survey was conducted in a randomly selected half of the counties and cities in
22 provinces (out of 31 total provinces), covering 85% of the total population in China. The
CLHLS has six waves to date (1998, 2000, 2002, 2005, 2008, and 2011). It surveyed only
11
the oldest old (aged 80 years and older) before 2002, but since 2002, the cohort aged 65–79
has also been included in the project. The survey combines an in-home interview and a basic
physical examination. Extensive information was collected on demographic characteristics,
family and household characteristics, lifestyles, diet, psychological characteristics, health,
disability, socioeconomic conditions, and so on. I only use data after 1998 because the
survey asking whether drinking water is boiled or not starts in 2000; only those who drink
boiled water are included to eliminate possible health linkages to waterborne bacteria.
China has the largest aging population and now is experiencing a profound
transformation from a young to an old society (Banister, Bloom and Rosenberg 2012). It is
projected that, by 2050, more than 30% of the population (440 millions) will be age 60 or
older (United Nations 2009). A focus on the older population is vital to economic and health
policy design. First, older residents are generally less economically productive and a larger
aging population suggests that economic growth may slow. Because a healthier population is
more productive, a reduction in the illness and disability of older adults may stimulate
economic growth. Second, relatively smaller working-age cohorts of the future will be
burdened by the need to care for and pay for the support of the elderly population. Rural
areas, where basic care and companionship are traditionally provided by family members,
face further concerns. Third, the increasing size of the older population will impose a
substantial burden on the economy as a whole because they require more medical care than
younger people. It is even worse if such an aging crisis trend is accompanied with additional
declines in health conditions caused, for example, by policies that stimulate pesticide use.
The health outcomes of infants and children, another vulnerable group, should also
12
receive top policy priority. The potential gains from extending healthy life may be greater for
them. While several large health-related datasets in China document infant and child health
outcomes, none fits the needs of this study. Two nationwide datasets (China Family Panel
Studies and Chinese Health and Retirement Longitudinal Study) include children’s
information but they were recently built and, therefore, did not capture 2004, the year of the
policy change for this quasi-experimental design. One dataset that covers 2004 is the China
Health and Nutrition Survey, but this dataset only covers five provinces and does not allow
for identification of regions. I leave the investigation of child health for future research.
Health Outcomes
Neurological abnormality is one of the most prominent health effects of pesticide exposure.
Several widely used pesticides, such as Organophosphate, Pyrethroid, Carbamate, and
Organochlorine, target the insect nervous system as their mechanism of toxicity (Casida and
Durkin 2013; Keifer and Firestone 2007). Previous epidemiological studies confirm that
pesticide exposure increases the risk of degradation in human peripheral nervous system.
Several neurological functions, such as reflex, sensation, and muscle strength, have been
detected to be adversely affected by exposure to pesticide, especially in a rural population
(Cole et al. 1998; Hu et al. 2015; Li et al. 2014). Further, there is evidence that Fipronil and
Imidacloprid, which both feature insect neurotoxins, are the two most widely detected
pesticides in surface water during rice production (China Pollution Source Census, 2010).
I use multiple indicators for the health status of the elderly that are feasibly linked to
pesticide exposure in the existing medical and toxicology literature, including the index of
Activities of Daily Living (ADL), cognitive ability, measured hypertension and self-reported
13
overall health condition.
The index of ADL is widely used to measure dependence of older and physical inactive
populations. It is also referred as a “disability” index by Verbrugge and Jette (1994) and
“functional limitations” by Johnson and Wolinsky (1993). This index is constructed by
asking participants if they needed any assistance with the following six basic activities:
bathing, dressing, eating, toileting, continence and getting out of bed and into a chair (Katz
et al. 1963). Choices for each item are ‘‘able to do without help,’’ ‘‘need some help,’’ and
‘‘need full help,’’ with scores from 1 to 3, respectively. The final score of ADL is the sum of
all 6 items and ranges from 6 to 18, with higher scores indicating more difficulty.
I am particularly interested in ADL because of its close connection with the current
research question, its serious public health consequences, and its significance in the medical
literature. First, all six measures in ADL are closely related with the integrity of the
peripheral nervous system, which is the possible target of pesticide with insect neurotoxins.
Second, physical inactivity is the fourth leading cause of death worldwide and it becomes a
major public health concern for older people (Kohl et al. 2012). Third, loss of independence
for older people places a considerable financial burden on them, their families, and the health
care system. The additional costs for those who develop a disability as measured by the ADL
total $26.1 billion in 1995 in the US (Guralnik et al. 2002; Pratt, Macera and Wang 2002).
Fourth, ADL provides a bridge that transfers life expectancy to healthy life expectancy,
which plays a significant role in measuring the quality of life expectancy (Chernew et al.
2016; Cutler, Ghosh and Landrum 2013).
Cognitive impairment is measured by the Mini-Mental State Examination (MMSE),
14
which covers the following aspects of cognitive functioning: orientation, registration,
duplication and design, calculation, recall, naming, and language (Folstein et al. 1975). The
MMSE score in this sample ranges from 0 to 30, with higher scores suggesting more
impairment. Hypertension is defined as a blood pressure more than 140/90 mm Hg and is
diagnosed by physicians. Self-reported health is assessed by a question: ‘‘In general, would
you say your health is 1) very good, 2) good, 3) so so, 4) bad, or 5) very bad?’’
Individual and Household Variables
Individual and household variables include information on socio-economic characteristics,
health behaviors and dietary pattern. Socio-economic characteristics include age, gender,
years of schooling, number of people living together and income from sons and daughters.
Health behaviors measure whether the respondents smoke, drink and exercise now (Yes or
No). Dietary patterns measure the frequency with which respondents eat meat, fish, fruit,
fresh vegetable and salt-preserved vegetable. Answers for each item are ‘‘rarely or never,’’
‘‘occasionally,’’ and ‘‘almost every day,’’ with scores from 1 to 3, respectively.
The questionnaire also asks respondents the source of their drinking water. The answers
include drinking water sources from a well; a river or lake; a spring; a pond or pool; and tap
water. I construct a binary variable water which equals one if respondents drink surface
water, i.e. water from a river or lake or a pond or pool, and equals zero otherwise.
Provincial Variables
Provincial variables include pesticide use and other control variables. Pesticide use is
measured by pesticide expenditure per mu. Other provincial variables include per capita
measures of industrial sulfur dioxide emission (SO2), industrial chemical oxygen of demand
15
(COD), GDP and the number of hospital. SO2 and COD per capita control for pollution from
industrial activities. GDP per capita measures general economy and wealth conditions. The
number of hospitals per capita controls for the level of medical service. Per capita measures
are used because they are relative indicators. All provincial variables are the average of the
last three years, which is adopted because the health data points we observe are once every
three years.
If the respondents die or migrate, they are not in the analysis samples and will be
counted in the attrition (See Appendix 1 for details). Attrition is potentially worrisome if it is
correlated with the independent variable of interest. Sample selectivity could then lead to
biased estimates. As it turns out, however, there is no evidence that attrition is correlated
with the variables of interest. I run analogous DDD regressions where the dependent variable
is the number of times this individual is sampled or whether this person has been sampled
more than once. In no regression is the coefficient on the three-way interaction term
significantly different from zero, which suggests the respondents lost to attrition are random
with respect to the triple difference design.
The Effect of Pesticide Use on Health Outcomes
Model Specification
This section investigates the health impact of pesticide use. The DDD model3 is specified
3 In this specific research, the difference-in-difference model, which limits the samples within those who drink surface water, would provide biased results. One potential concern is that, because regions with more pesticide water pollution may also have more pesticide pollution in the air, soil and food, the subsample DD results may be driven by pesticide pollution via these other pathways. Another potential concern is, if there are other unobserved changes caused by the policy, such as water availability, the DD results will also be biased.
The logic why the DDD works is that the additional noises, such as the effects of pesticide via other channels or unobserved changes caused by the policy, are common to those people who drink surface
16
as:
1 2 3 4
5
* * * * *
' (1)
it p it it p it it it p it
it it t i it
y Rpest Water Y04 Rpest Water Water Y04 Rpest Y04
Water X
where i, p and t indicate individual, province and year. yit is measured health outcomes
including ADL, cognitive ability, measured hypertension and self-reported overall health
condition. Rpest measures the intensity of rice pesticide use (yuan/mu) before the policy year
2004.4 Water equals one if the survey respondent drinks surface water and zero otherwise.
y04 equals one starting in 2004 and zero otherwise. X is a vector of control variables
including individual, household and provincial characteristics. u is a fixed effect unique to
individual i, and λ is a time effect common to all individuals in year t. is the intercept and
is an error term. Two approaches are implemented in the estimation of the standard errors
to avoid potential biases from error correlation across time and space and the most
conservative ones are reported. First, I allow for an arbitrary covariance structure within
provinces over time by computing standard errors clustered at the province level (Bertrand,
Duflo and Mullainathan 2004). Second, cluster bootstrap methods are also implemented at
the province level to reaffirm the results (Cameron and Trivedi 2010).
The coefficient of interest is β1, the impact of pesticide use on health outcomes. Without
water and ground water, and, therefore, are differenced out by such DDD design. This shares the same spirit as Kellogg and Wolff (2008) that DDD requires milder identification assumptions to produce unbiased estimates.
Note that the assumption that DDD works is that the third difference can difference out the additional noises. In the robustness section, the analogous DDD regressions with respect to corn and wheat validate this assumption. If the additional noises are not differenced out by the third D-step, it is expected to see the significant impact of corn and wheat pesticide on ADL, but the effects are not significant. The robustness tests by adding or deleting a comprehensive set of control variables can also shed some light on this. 4 Both average and one year intensities of pesticide use before the policy change are similar. As a robustness check of different intensity measures of pesticide use, I also use total quantity per mu (tons/mu) as the main variable of interest. The results are available upon request. I cannot estimate the effect of pesticide quantities per mu by different crops because the quantity data is crop aggregated.
17
randomization of pesticide use, a major concern is that provinces that use more pesticide
could be different from provinces that use less pesticide and that these differences may be
correlated with health outcomes. For example, provinces that are stronger in agriculture and
use more pesticide may be poorer. In this case, the correlation between pesticide use and
health would be confounded with the wealth effect. For another example, individuals living
in regions with more pesticide pollution may themselves begin in different health conditions
so that any observed relationship between pesticide use and health outcomes may simply
reflect the influence of unobserved initial health conditions. These types of confounding
factors vary across provinces and individuals but are fixed over time. A common method of
controlling for time-invariant unobserved heterogeneity is to use fixed effects.5
A remaining threat is confounding factors that vary over time. For example, richer
provinces may have a rapidly increasing GDP or individuals in better health conditions may
face more slowly growing health risks. My solution is to use the
difference-in-difference-in-differences model where the key identification assumption is that
the time trends in the regions with different intensities of pesticide use are the same in the
pre-policy periods. If the time trends are the same in the pre-policy periods, then it is likely
that they would have been the same in the post-policy periods if the regions with more
5 Several other concerns can also be addressed by individual fixed effects. For example, the areas where rice pesticide use increased rapidly may be the same places with high historic use of DDT, and what I identify may be a long term lag effect of DDT use rather than the shorter run influence of modern pesticides. This is not the case because the accumulated health effects will be differenced out by individual fixed effects. As evidence, Figure 3 plots the residuals from a regression of ADL on individual fixed effects (detail explanations are on the next page). DDT was widely used in 1970s and banned in 1983. If the significant coefficients are driven by a long term lag effect of DDT, which means this effect is not differenced out by individual fixed effects, this lag effect is likely to manifest before 2005. It is less likely that the lag DDT effect from 1970s suddenly starts to manifest in 2005 as is shown in Figure 3. Further, the difference before and after 2004 in the DDD model can serve as another practice that eliminates this lag DDT effect as well as other accumulated health effects.
18
pesticide use had used less pesticide.
Figure 3. ADL Trends by Drinking Water Source for Rural and Urban Respondents.
Notes: This figure illustrates the evolution of ADL scores by drinking water sources. The plots are
residuals from regressing ADL on the individual fixed effects. Pesticide use is a continuous variable
but, for the purpose of this graph, we dichotomize pesticide use by its pre-policy mean. The blue
lines represent regions with high intensity of pesticide use and the red lines are for regions using low
intensity of pesticide. Figure 3A and 3B include both urban and rural samples and Figure 3C and 3D
only include samples in rural areas. Figure 3A and 3C represent samples drinking surface water and
Figure 3B and 3D represent samples drinking ground water. The regions using high intensity of
pesticide account for 54% of the observations. The groups drinking surface water account for 1.39%
of the observations. Specifically, for each blue line in 3A, 3B, 3C and 3D, the numbers of
observations are 366, 18781, 354 and 16251 respectively. For each red line in 3A, 3B, 3C and 3D, the
numbers of observations are 125, 15956, 115 and 11438 respectively.
While this parallel trend assumption cannot be tested comprehensively, graphical
analysis and partial tests are possible. First, I present some initial graph evidence on the
validity of the quasi-experimental design. Figure 3A and Figure 3B represent samples that
drink surface water and ground water respectively. Pesticide use is a continuous variable but,
-.5
0.5
1
2000 2005 2010
high low
A: Surface Water (urban+rural)
-.5
0.5
1
2000 2005 2010
high low
B: Ground Water (urban+rural)
-.5
0.5
1
2000 2005 2010
high low
C: Surface Water (rural)
-.5
0.5
1
2000 2005 2010
high low
D: Ground Water (rural)
19
for the purpose of this graph, I dichotomize pesticide use by its pre-policy mean. The blue
lines and red lines depict the evolution of the ADL in regions with more and less pesticide
use respectively. For those who drink surface water, the ADL levels have some fluctuations
between regions with more and less pesticide use before 2004. However, after 2004 the ADL
levels in the regions using more pesticide increase faster than those using less pesticide. As is
shown in Figure 1, this timing is commensurate with the rebound of grain production in
2004. By contrast, for those who drink ground water (Figure 3B), the ADL levels increase at
similar rate in regions with more and less pesticide use. The regression results are anticipated
in Figure 3. These graphs serve as initial evidence that pesticide use causally affects the ADL
levels through drinking water.
Formally, I test that the pre-policy time trends for the regions with different intensities
of pesticide use are not different by estimating a slightly modified version of equation (1). I
use only the observations before 2005 and modify equation (1) by excluding the policy
dummy variable and including separate year dummies. In this model, I cannot statistically
reject the hypothesis that the pre-policy year dummies are the same for regions using more
and less pesticide at conventional levels of statistical significance (Appendix 10). This also
provides initial evidence that pesticide use causally affects the ADL levels through drinking
water.6
Further, a potential violation of parallel trends would be if people in regions with more
and less pesticide use are systemically different in control variables from individual,
household and regional levels, and health outcomes would have varied according to these
6 As one reviewer points out, it could also be that the statistical power is not enough to identify any different trends in two regions.
20
control variables regardless of pesticide use. For example, rural-urban migration owing to
industrialization and urbanization may increase pesticide use because of rural labor shortages;
this may also worsen health conditions for those 65 and older as a result of less attendance
and help. In this case, adding or deleting the number of people living together and industrial
pollution are likely to change the coefficients if this migration effect is not differenced out by
the DDD estimator. Therefore, examining whether coefficient estimates on the three-way
interaction term change when the individual, household and regional characteristics are
excluded in the regression can shed light on whether changes in outcome variables related to
these characteristics are correlated with the three-way interaction term, constituting another
partial test of the parallel trend assumption.
Results
This subsection reports the impact of pesticide use on health outcomes. In Table 2, the first
two columns report results with only rural samples and the last two columns report results
with full samples. In Model 1, I find that a one yuan per mu increase in expenditures on rice
pesticides increases the index of ADL by 0.024 units. That is, a 10% (2.9 yuan/mu) increase
in rice pesticide use increases the ADL score by 1% (0.07 units) for rural residents 65 and
above. Model 2 presents results without the inclusion of any other right-hand-side control
variables. Exclusion of the control variables makes little difference to the coefficients. These
results are consistent with the received epidemiological and economic literature that
pesticide exposure adversely affects the peripheral nervous system (Ding and Bao 2014; Li
et al 2014; Hu et al. 2015; Starks et al. 2012) and, therefore, can increase the dependence of
older adults via degradation of this key system.
21
In addition, this effect is mainly driven by populations in rural areas where millions of
rural residents rely on surface water for their drinking supply. I include both urban and rural
samples in Model 3 and find a modest decline in the coefficients but a large decrease in the
precision of measurement. Exclusion of the control variables makes little difference to the
coefficient (Model 4). The results suggest an unintentional policy side effect that mitigates
the policy objectives of increasing farmers’ income and welfare.
Note that the increase in the ADL score need not be pathological. Another possible
explanation for these results is that pesticide exposure causes illnesses with clinical
symptoms that increase ADL. However, this hypothesis finds less support because no
symptomatic diseases are directly linked to pesticide exposure when I run DDD models with
the self-reported diseases as dependent variables.7 This insignificance may come from the
inaccuracy of self-reported diseases. It may also come from my narrow focus on the drinking
water pathway and pesticides. It may also due to the limited focus on the non-occupational
population, whereas the literature that identifies various acute and chronic diseases resulting
from pesticide use usually focus on occupationally exposed samples.8 Overall, these results
suggest that the health impact of pesticide use is underestimated because subclinical neuro
electrophysiological9 changes such as degradation of sensation and muscle strength, if not
7 Besides cognitive ability, measured hypertension and self-reported health condition, we also test other self-reported diseases. Self-reported diseases include cardiovascular diseases, gastrointestinal diseases, respiratory diseases, cancer and central neurological diseases. The information is measured by a survey question that asks whether the respondent suffers from each of these diseases. Choices for each item are “Yes” and “No”. Self-reported diseases are less accurate, which is the reason we do not use these as the main dependent variables. 8 I am not ruling out direct exposure from farm work involving agrochemicals. However, the sample only includes people older than 65. These people are not likely to work on the land during the era with increased pesticide use. As evidence, in the questionnaire, respondents are asked their age when they ended physical labor. Only 2% of the respondents report they are still engaged in any type of physical labor now (not necessarily farm work). 9 Subclinical denotes a disease that is not severe enough to present definite or readily observable
22
accompanied with clinical symptoms, will not be attributed to pesticide exposure (Ding and
Bao 2014; Hu et al. 2015; Pingali et al. 1997).
Robustness
I conduct several robustness checks to help assess the validity of the results. A potential
concern is that, because agricultural support programs are not exclusively targeted at
pesticide, the results may be driven by other changes caused by this policy, such as fertilizer
use, offspring migration and residential water availability.10 I rule out these possibilities by
first using rice fertilizer use as a falsification test and then further exploring the relationship
between the policy change and these potential confounding factors.
Fertilizer Use. The logic that fertilizer use can serve as a placebo test is due to its
similarity and difference with pesticide use. Fertilizer, as another main agricultural chemical,
is also widely used in crop production and its application also increases sharply after the
2004 policy change. As with pesticides, fertilizer can enter human bodies through air, waters,
soil and farm products. However, unlike pesticides, there is no evidence in medical literature
that fertilizer exposure can degrade the human peripheral nervous system.
I replicate the analysis in Equation (1) using fertilizer use in the pre-policy year as the
variable of interest. If it is some other unobserved factors that drive the significance of the
results, we should also observe a significant impact of fertilizer use on ADL. In Table 3, I
find in no model specification that fertilizer use is positively associated with ADL (Model 1
and 2). I also add annual fertilizer use into Equation (1) as a control variable and find little
symptoms. Electrophysiology is the study of the electrical properties of biological cells and tissues. In neuroscience, it includes measurements of the electrical activity of neurons, and particularly action potential activity. Recordings of large-scale electric signals from the nervous system are useful for electrodiagnosing and monitoring potential activity of neurons. 10 I thank anonymous reviewers for numerous comments and suggestions on these robustness checks.
23
effect on the estimated coefficient β1 (Models 3 and 4). These results have two implications.
First, fertilizer itself is not the unobserved factor that drives the significant impact of
pesticide use on ADL, which is consistent with the medical literature that fertilizer does not
cause degradation of the human peripheral nervous system. Second, the positive effect of
pesticide use on ADL is not driven by other confounding factors such as offspring migration
and residential water availability. Otherwise, we should also observe significant impact of
fertilizer use on ADL. To further support my conclusion, I explore the relationship between
the policy change and offspring migration and residential water availability.
Offspring Migration. Ebenstein et al (2011) find that labor out-migration and fertilizer
use per hectare are positively correlated, arguing that as rural workers leave their villagers,
those remaining behind in farming areas are forced to compensate for labor scarcity by using
more fertilizer. Similarly, pesticides may also compensate for the lack of labor on weed
management and pest control. One potential concern that may arise here is that the increase
of fertilizer and pesticide use caused by agricultural support programs may be positively
correlated with offspring out-migration. Therefore, instead of neurotoxic effects of pesticides,
it is the offspring out-migration and consequent reduction in caring for older adults that leads
to their diminished health and increased dependence. I rule out this possibility by showing
the policy change does not predict the change in the migration patterns in regions with
different intensities of pesticide use. Additionally, I provide suggestive evidence, consistent
with the main findings in this study, that pesticide use may reduce the offspring migration in
the affected households because it increases the dependence of older adults.
24
Figure 4: Number of People Living Together
Notes: This figure depicts the number of people living together in regions using high and low
intensity of agricultural chemicals (blue lines and red lines respectively). Figure 4A is by pesticide
levels and Figure 4B is by fertilizer levels.
Figure 4A depict the evolution of number of people living together in regions with
higher and lower pesticide use intensity respectively. Pesticide use is a continuous variable
but, for the purpose of this graph, I dichotomize pesticide use by its pre-policy mean.
Together with Figure 2, I do observe that, as the pesticide use increase each year, the number
of people living together decreases, which suggests that agricultural chemicals are positively
correlated with out-migration. However, I do not observe different migration pattern
between regions using more and less pesticide intensity before and after 2004, which
suggests that the offspring migration and possible changes in levels of care provided to older
adults are not the factors that drive the impact of pesticide use on ADL. Formally, I follow
one reviewer’s suggestion to test whether the policy change predicts the number of people
living together in regions with different intensities of pesticide use. In no specification
pesticide use significantly predicts the number of people living together before and after the
policy change (Table 4). I also allow the effect of pesticide use to vary each year by adding
separate year dummies. The numbers of people living together are not significantly different
2.5
33.
54
2000 2005 2010
high low
A: # of People Living Together (by pesticide)
2.5
33.
54
2000 2005 2010
high low
B: # of People Living Together (by fertilizer)
25
each year between regions using more and less pesticide. As a comparison, I replicate these
analyses using rice fertilizer use as the variable of interest. In no specification the relevant
coefficients are significant (Models 3 and 4 in Table 4). If the significant impact of pesticide
use on ADL is driven by offspring out-migration, it is expected that policy change predicts
the number of people living together in regions with different intensities of agricultural
chemicals.
Furthermore, several additional observations based on the previous analysis suggest that
it is the pesticide neurotoxin that increases the dependence of older adults and, therefore,
reduces the offspring’s out-migration. Figure 4B depicts the evolution of number of people
living together in regions with high and low intensity of fertilizer use, but different from 4A,
the two lines in Figure 4B have the tendency to be close after the policy change, i.e. regions
with high intensity of fertilizer, compared with regions with low intensity, are associated
with more out-migration after the policy change. This is consistent with the finding from
Ebenstein et al (2011) that labor out-migration and fertilizer use are positively correlated but
also raises the question of why we do not observe this pattern between regions using more
and less pesticide. A conjecture consistent with my main findings is that the increased
dependence of older adults caused by pesticide increases the number of people living
together and, therefore decreases out-migration.
As predicted by these two graphs, I can also find support from the regression results in
Table 5. I replicate the equation (1) with number of people living together as outcome
variables for both pesticide and fertilizer. The results show that pesticide use significantly
increases the number of people living together but fertilizer use does not. All these analyses
26
together suggest pesticide uses increases the dependence of older adults and, therefore, may
reduce offspring out-migration. Note that those people living together are not necessarily
their offspring and those people not living together do not necessarily migrate out. Because
this dataset lacks detail information on offspring’s migration, this evidence can only be
considered suggestive with more thorough analysis left for future research.
Water Availability. Another potential concern comes from water availability after the
2004 agricultural policy change. If the water available for household use is reduced in
regions with increased pesticide use because of increased agricultural water use resulting
from policies supporting agricultural expansion, the results will be biased. Similar with the
previous robustness checks, I rule out this possibility by showing that the household water
consumption in regions using more and less pesticide do not differ significantly before and
after 2004, and that adding or deleting the amount of water consumed by residence in
equation (1) has no effect on the coefficients of interest (Table 6).
Types of drinking water. The estimation would also be potentially problematic if people
in regions with more pesticide use are more likely to change drinking water sources after
2004. To address this concern, I test whether reliance upon surface water for drinking
sources differ before and after 2004 in regions with more and less pesticide use. That is, I run
a difference-in-difference regression with the type of drinking water as the dependent
variable and the interaction of pesticide use and year 2004 as the variable of interest. I also
allow the effect of pesticide use on the type of drinking water to change each year by adding
separate year dummies. All relevant coefficients are clearly insignificant, suggesting regions
with more and less pesticide exposure do not differ with respect to reliance on drinking
27
surface water before and after 2004 (Table 6).
Different Pollution Channels. A potential concern is that, because regions with more
water pollution containing pesticides may also have more pesticide pollution in the air, soil
and food, the results may be driven by pesticide pollution via these other pathways. To rule
out alternative pathway, I run analogous DDD regressions with respect to corn and wheat,
which do not feature flooding as a cultural practice and, thus, should create less water
pollution. If it is rice pesticides that affect health outcomes through surface water, the results
should not be driven by production technologies with less water pollution. The coefficients
linking health outcomes and pesticide use from corn and wheat production are clearly
insignificant (Table 7). Additionally, in previous analyses, I present results adding and
deleting the samples in urban areas where treated tap water is provided. If it is through
surface water that pesticides affect health outcomes, the significant results should not be
driven by urban residents who do not drink untreated surface water. I find that the significant
results are not driven by urban samples.
Monetary Losses
This section further explores these results by translating estimated effects into expected
monetary losses. I do this by exploiting the relationship between ADL and medical costs and
offspring’s weekly caring hours.11 As Figure 7 shows, the medical costs and offspring’s
11 Note that the relationship between ADL and medical cost/caring time is not causal. These results should be interpreted with caution. Both the medical cost and offspring caring time in CLHLS is self-reported. I acknowledge that they are not as accurate as administrative records. However, the medical cost in CLHLS seems to be reliable when compared with the medical costs in another nationwide survey, Chinese Health and Retirement Longitudinal Study (CHARLS). In CHARLS, the annual medical cost is 1289.62 yuan (804.96 out-patient and 484.66 in-patient), which is very close to 1378.69 yuan in CLHLS.
28
weekly caring hours increase with the ADL score. Formally, the model is specified as:
1ADL 'it it it t i ity X (2)
where yit represents medical costs and offspring’s weekly caring hours. Other variables are as
defined as before and the most conservative standard errors are reported.
Figure 7. Annual Medical Expenditures and Offspring’s Weekly Caring Hours by ADL Score
I find positive and significant effects of ADL on medical cost and offspring’s weekly
caring hours (Table 8) and the coefficients are little changed when adding or deleting control
variables. The estimates suggest that each additional unit increase in ADL increases medical
costs and offspring’s weekly caring time by 139 yuan12 and 5.95 hours, respectively.
According to China Statistical Yearbook 2012, the average annual income per capital is
about 14,582 yuan, which is about 7 yuan per hour. That is, a one unit increase in ADL
increases offspring’s human capital losses by approximately 42 yuan.
In combination with previous information and 1.65 million aging population who
drinks surface water, I am able to calculate that a one yuan per mu increase in expenditures
on rice pesticides increases medical costs and human capital losses of local populations by
12 1 Yuan = 0.15 US Dollar
01
,00
02
,00
03
,00
04
,00
0
6 7 8 9 10 11 12 13 14 15 16 17 18
Annual Medical Expenditures (Yuan) by ADL Score
02
04
06
08
01
00
6 7 8 9 10 11 12 13 14 15 16 17 18
Offspring's Weekly Caring Hours by ADL Score
29
5.51 and 1.67 million yuan ($0.83 and $0.25 million), respectively.13 That is, a 10% increase
in rice pesticide use increases medical costs and human capital losses of local populations by
15.98 and 4.83 million yuan ($2.4 and $0.72 million), respectively. In sum, a 10% increase
in rice pesticide use increases monetary losses by 20.81 million yuan ($3.12 million), which
is equivalent to 0.03% of national annual rice production profits. 14 The results still
underestimate the real costs of pesticide use because of the narrow focus on the aging
population and the water exposure channel.
Conclusion
Pesticides make a significant contribution to maintaining world food production, but this is
accompanied with numerous forms of environmental degradation as well as elevated threats
on human health. This paper seeks to strengthen our understanding on the tradeoff between
greater use of pesticide to increase yields and the deterioration in human health. By linking
provincial data from several Chinese statistical yearbooks (1998-2011) with the Chinese
Longitudinal Healthy Longevity Survey (1998-2011), I compare health outcomes between
people who drink surface water and ground water in regions with different intensities of rice
pesticide use before and after 2004, when China shifted from taxing agriculture to
subsidizing agricultural programs.
13 According to the China’s population census in 2010, the population aged 65 and above is 118.9 million. The CLHLS survey shows that 1.39% population aged 65 and above drink surface water. These suggest the population affected by pesticide polluted surface water is 1.65 million (118.9×1.39%). Medical costs are computed by 0.024×139×1.65 million yuan and human capital costs are computed by 0.024×42×1.65 million yuan. These calculations only account for those 65 and older who drink surface water. More details are in Appendix 2. 14 The average rice production profits per mu from 2000 to 2011 were 194.84 yuan and the data was from Agricultural Cost and Benefit Yearbook. The average sown area of rice from 2000 to 2011 was 399.6 million mu and the data was from China Statistics Yearbook. The average rice production profits were 194.84×399.6=77858 million yuan.
30
The results indicate that a 10% (2.9 yuan/mu) increase in rice pesticide use unfavorably
alters the index of dependence in Activities of Daily Living (Katz Index of ADL) by 1%
(0.07 units) for rural residents 65 and above. This is equivalent to 15.98 and 4.83 million
yuan ($2.4 and $0.72 million) in medical costs and offspring’s human capital losses,
respectively (in total, 0.03% of rice production profits). This result suggests that hidden
health impact such as losses in sensation and muscle strength, if not accompanied with
clinical symptoms, will not be attributed to pesticide exposure. The result also suggests an
unintentional policy side effect that mitigates the policy objectives of increasing farmers’
income and welfare.
The findings may shed light on the tradeoff faced in agricultural and environmental
policy design. For example, the estimated monetary losses due to pesticide usage can be used
as a guide to choosing subsidization levels for crop insurance and for funding farmers’
pesticide training programs given that previous studies show that risk aversion surrounding
farm production outcomes and the lack of pesticide knowledge contribute to overuse of
pesticides (Alam and Wolff 2016; Chen, Huang and Qiao 2013; Liu and Huang 2013). These
issues have salience in other developing countries that face similar challenges to sustainable
development, particularly where countries seek greater food sovereignty. Although I provide
evidence that pesticides affect health outcomes through a drinking water channel, other
channels deserve future research.
31
Reference:
Alam, S. and Wolff, H., 2016. Do Pesticide Sellers Make Farmers Sick? Information, Health, and
Adoption of Technology in Bangladesh. Journal of Agricultural and Resource Economics
Banister, J., Bloom, D.E. and Rosenberg, L., 2012. Population aging and economic growth in China,
The Chinese Economy. Springer 114-149
Bertrand, M., Duflo, E., Mullainathan, S., 2004. How much should we trust
differences-in-differences estimates? Quarterly Journal of Economics 119(1), 249-275
Brainerd, E. and Menon, N., 2014. Seasonal effects of water quality: The hidden costs of the Green
Revolution to infant and child health in India, Journal of Development Economics 107,
49-64
Cameron, A. and Trivedi, P., 2010. Microeconometrics: Using Stata, Revised Edition, Chapter 11
and 24, Stata Press, TX, USA
Casida, J.E. and Durkin, K.A., 2013. Neuroactive insecticides: targets, selectivity, resistance, and
secondary effects, Annual Review of Entomology 58, 99-117.
China, the State Council. 2016. The No. 1 Central Document. Retrieved online Mar. 10, 2016 from
http://www.xinhuanet.com/fortune/cjzthgjj/158.htm
Chernew, M., Cutler, D.M., Ghosh, K. and Landrum, M.B., 2016. Understanding the Improvement in
Disability Free Life Expectancy in the US Elderly Population. National Bureau of Economic
Research.
Chen, Y., Ebenstein, A., Greenstone, M., Li, H., 2013. Evidence on the impact of sustained exposure
to air pollution on life expectancy from China’s Huai River policy. Proceedings of the
National Academy of Sciences 110, 12936-12941
32
Chen, R., Huang, J., Qiao, F., 2013. Farmers' knowledge on pest management and pesticide use in Bt
cotton production in china. China Economic Review 27, 15-24
Chen, H., Wu, L., Wang, Y., 2010. The effect of subsidies on grain production in China. Journal of
Agrotechnical 4, 100-106
Chen, F., Fang, Q., Gao, M., 2010. Agricultural policies and the adjustment of grain production.
Economic Research Journal 11, 101-114.
Editorial Committee, 2011, Data Summary of the First China Pollution Source Census, China
Environmental Science Press
Cole, D., Carpio, F., Julian, J. and Leon, N. 1998. Assessment of peripheral nerve function in an
Ecuadorian rural population exposed to pesticides, Journal of Toxicology and Environmental
Health (United Kingdom).
Cutler, D. and Miller, G. 2005. The role of public health improvements in health advances: the
twentieth-century United States, Demography 42, 1-22.
Cutler, D.M., Ghosh, K. and Landrum, M.B. 2013. Evidence for significant compression of
morbidity in the elderly US population. National Bureau of Economic Research
Ding, G., Bao, Y., 2014. Revisiting pesticide exposure and children's health: Focus on China.
Science of the Total Environment 472, 289-295
Ebenstein, A., 2012. The consequences of industrialization: evidence from water pollution and
digestive cancers in China. Review of Economics and Statistics 94, 186-201
Folstein, M.F., Folstein, S.E., McHugh, P.R., 1975. “Mini-mental state”: a practical method for
grading the cognitive state of patients for the clinician. Journal of psychiatric research 12,
189-198
33
Gale, H.F., 2013. Growth and evolution in China's agricultural support policies. USDA-ERS
Economic Research Report
Guo, L., Qiu, Y., Zhang, G., Zheng, G.J., Lam, P.K., Li, X., 2008. Levels and bioaccumulation of
organochlorine pesticides (OCPs) and polybrominated diphenyl ethers (PBDEs) in fishes
from the Pearl River estuary and Daya Bay, South China. Environmental Pollution 152,
604-611
Guralnik, J.M., Alecxih, L., Branch, L.G. and Wiener, J.M., 2002. Medical and long-term care costs
when older persons become more dependent, American Journal of Public Health 92,
1244-1245.
Hu, R., Huang, X., Huang, J., Li, Y., Zhang, C., Yin, Y., Chen, Z., Jin, Y., Cai, J., Cui, F., 2015.
Long-and short-term health effects of pesticide exposure: a cohort study from China. PloS
one 10, e0128766
Huang, J., Hu, R., Rozelle, S., Pray, C., 2005. Insect-resistant GM rice in farmers' fields: assessing
productivity and health effects in China. Science 308, 688-690
Huang, J., Qiao, F., Zhang, L., Rozelle, S., 2000. Farm pesticide, rice production, and human health.
CCAP's Project Report 11
Johnson, R.J. and Wolinsky, F.D., 1993. The structure of health status among older adults: disease,
disability, functional limitation, and perceived health, Journal of health and social
behavior105-121
Katz, S., Ford, A.B., Moskowitz, R.W., Jackson, B.A., Jaffe, M.W., 1963. Studies of Illness in the
Aged. JAMA: The Journal of the American Medical Association 185, 914-919
Kellogg, R. and Wolff, H. 2008. Daylight time and energy: Evidence from an Australian experiment,
Journal of Environmental Economics and Management 56, 207-220.
34
Keifer, M.C. and Firestone, J., 2007. Neurotoxicity of pesticides, Journal of Agromedicine 12, 17-25
Kohl, H.W., Craig, C.L., Lambert, E.V., Inoue, S., Alkandari, J.R., Leetongin, G., Kahlmeier, S. and
Group, L.P.A.S.W., 2012. The pandemic of physical inactivity: global action for public
health, The Lancet 380, 294-305
Li, H., Zeng, E.Y., You, J., 2014. Mitigating pesticide pollution in China requires law enforcement,
farmer training, and technological innovation. Environmental toxicology and chemistry 33,
963-971
Liu, E.M., Huang, J., 2013. Risk preferences and pesticide use by cotton farmers in China. Journal of
Development Economics 103, 202-215
Macdonald, R., Barrie, L., Bidleman, T., Diamond, M., Gregor, D., Semkin, R., Strachan, W., Li, Y.,
Wania, F., Alaee, M., 2000. Contaminants in the Canadian Arctic: 5 years of progress in
understanding sources, occurrence and pathways. Science of the Total Environment 254,
93-234
Macharia, I., Mithöfer, D., Waibel, H., 2013. Pesticide handling practices by vegetable farmer in
Kenya. Environment, development and sustainability 15, 887-902
Pingali, P.L., Hossain, M., Gerpacio, R.V., 1997. Asian rice bowls: the returning crisis? Int. Rice Res.
Inst.
Pratt, M., Macera, C.A. and Wang, G., 2000. Higher direct medical costs associated with physical
inactivity, The Physician and Sports Medicine 28, 63-70
Qiao, F., Huang, J., Zhang, L., Rozelle, S., 2012. Pesticide use and farmers' health in China's rice
production. China Agricultural Economic Review 4, 468-484
Reuters, 2016. “China farm sector needs supply-side reform: agriculture minister”. Retrieved online
35
Mar. 10, 2016 from
http://www.reuters.com/article/us-china-agriculture-idUSKCN0W90GG?utm_content=
articles&utm_campaign=NLCampaign&utm_source=Newsletter&utm_term=newsletterediti
on&utm_medium=email
Starks, S.E., Hoppin, J.A., Kamel, F., Lynch, C.F., Jones, M.P., Alavanja, M.C., Sandler, D.P., Gerr,
F., 2012. Peripheral nervous system function and organophosphate pesticide use among
licensed pesticide applicators in the Agricultural Health Study. Environmental health
perspectives 120, 515
Sun, X., J. Wang and Y. Jin., 2009. Advances in Research on Pesticide Pollution to the Aquatic
Environment and Health Impact in China. Journal of Environmental Health 7, 649-652
Verbrugge, L.M. and Jette, A.M., 1994. The disablement process, Social Science & Medicine 38,
1-14
Wang, M., Webber, M., Finlayson, B., Barnett, J., 2008. Rural industries and water pollution in
China. Journal of Environmental Management 86, 648-659
Wang, P., Tian, Y., Wang, X.-J., Gao, Y., Shi, R., Wang, G.-Q., Hu, G.-H., Shen, X.-M., 2012.
Organophosphate pesticide exposure and perinatal outcomes in Shanghai, China.
Environment international 42, 100-104
Wu, C., Maurer, C., Wang, Y., Xue, S., Davis, D.L., 1999. Water pollution and human health in
China. Environmental Health Perspectives 107, 251
Xian, Q., Ramu, K., Isobe, T., Sudaryanto, A., Liu, X., Gao, Z., Takahashi, S., Yu, H., Tanabe, S.,
2008. Levels and body distribution of polybrominated diphenyl ethers (PBDEs) and
hexabromocyclododecanes (HBCDs) in freshwater fishes from the Yangtze River, China.
36
Chemosphere 71, 268-276
Yu, X., Geng, Y., Heck, P., Xue, B., 2015. A review of China’s rural water management.
Sustainability 7, 5773-5792
Zeng, Y., Vaupel, J., Xiao, Z., Liu, Y., Zhang, C., 2015. Chinese Longitudinal Healthy Longevity
Survey (CLHLS), 1998-2012. ICPSR36179-v1. Ann Arbor, MI: Inter-university Consortium
for Political and Social Research [distributor]. http://doi.org/10.3886/ICPSR36179.v1
Zhang, C., Hu, R., Shi, G., Jin, Y., Robson, M.G., Huang, X., 2015. Overuse or underuse? An
observation of pesticide use in China. Science of the Total Environment 538, 1-6
Zhang, J., 2012. The impact of water quality on health: Evidence from the drinking water
infrastructure program in rural China, Journal of health economics 31, 122-134
Zhang, W., Jiang, F., Ou, J., 2011. Global pesticide consumption and pollution: with China as a focus.
Proceedings of the International Academy of Ecology and Environmental Sciences 1, 125
Zhou, Y., Zhao, W., Zhang, X., 2009. Impact evaluation of recent agricultural support policies.
Issues in Agricultural Economy 5, 4-11.
Table 1. Variable Description and Summary Statistics
Variable Description Obs. Mean Std. Dev.
Household Variables
Age Age in years 35228 86.17 11.22
Education Years of schooling 35121 2.17 3.51
Male Male=1; Female=0 35228 0.45 0.50
Live_with Number of people living together 35228 2.96 1.99
Income_sd Annual income from son and daughter (Yuan) 28446 1538.39 2307.83
Care_time Total weekly caring hours of all offspring
(children, grandchildren and their spouses)
7742 34.04 47.70
Health Behaviors
Smoke Currently smoke (Yes=1; No=0) 35228 0.19 0.39
Drink Currently drink (Yes=1; No=0) 35228 0.20 0.40
Exercise Currently exercise (Yes=1; No=0) 35228 0.32 0.46
Dietary Pattern
Staple Daily staple foods consumed (gram) 35228 2.98 1.28
Meat The frequency of meat consumption 35228 2.19 0.68
Fish The frequency of fish consumption 35228 1.88 0.63
Fruit The frequency of fruit consumption 35228 1.87 0.59
Vegetable The frequency of vegetable consumption 35228 2.55 0.56
Salt_vege The frequency of salt-preserved vegetable
consumption
35228 1.81 0.77
Water Drinking water source
(Surface water=1; Ground water=0)
35228 0.01 0.12
Provincial Variables
Rpest Rice pesticide intensity (Yuan/Mu/Year) 35228 28.14 14.18
Rfer Rice fertilizer intensity (Yuan/Mu/Year) 35228 86.13 23.01
Cpest Corn pesticide intensity (Yuan/Mu/Year) 27203 6.60 1.99
Wpest Wheat pesticide intensity (Yuan/Mu/Year) 19424 9.17 4.03
SO2 Provincial industrial sulfur dioxide emission
(10000 Tons/10000 persons/Year)
35228 0.0144 0.0039
COD Provincial industrial COD discharge
(10000 Tons/10000 persons /Year)
35228 0.0052 0.0032
GDP Provincial Gross Domestic Product
(100 million Yuan/10000 persons /Year)
35228 0.6669 0.2476
Hospnum Number of hospital per province 35228 2.9893 1.9015
(10000 persons)
Health Indicators
Variable Description Obs. Mean Std. Dev.
ADL Activities of Daily Living score
(lower=better)
35228 6.99 2.33
Cability Cognitive ability, MMSE score (lower=better) 35223 22.49 8.94
Hyp Health worker diagnosis of hypertension
(Yes=1; No=0)
34534 0.49 0.49
Selfhealth Self-reported health condition (lower=better) 32696 2.53 0.93
Medcost Medical cost (Yuan) 20433 1378.69 2719.93
Notes: (1) All individual and household level variables are from CLHLS and they represent their
situations in the survey year. All provincial variables are from statistics yearbooks and they are the
average of the last three years. This average is adopted because the health data points we observe are
once every three years. (2) The annual income from son and daughter, income_sd, was not collected
in the first two waves. The annual medical costs, medcost, and offspring’s weekly caring hours,
care_time, were not collected in the first three waves. (3) In the dietary panel, the answers for each
consumption frequency are ‘‘rarely or never,’’ ‘‘occasionally,’’ and ‘‘almost every day,’’ with scores
from 1 to 3, respectively. (4) ADL ranges from 6 to 18, with higher scores indicating more dependent.
Cognitive impairment is measured by the Mini-Mental State Examination (MMSE). It ranges from 0
to 30, with higher scores suggesting more impairment. Hypertension is defined as a blood pressure
more than 140/90 mm Hg and is diagnosed by physicians (Yes=1; No=0). Self-reported health is
assessed by a question: ‘‘In general, would you say your health is 1) very good, 2) good, 3) so so, 4)
bad, or 5) very bad?’’ (5) 1 Yuan = 0.15 US Dollar. 1 Mu = 0.067 Hectare.
Table 2. The Effect of Rice Pesticide Use on ADL
(1) (2) (3) (4) VARIABLES ADL ADL ADL ADL Rpest*Water*Y04 0.024* 0.034** 0.016 0.025 (0.068) (0.033) (0.218) (0.148) Observations 28,158 28,158 35,228 35,228 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows the effect of rice pesticide use on the dependence of adults 65 years and
older (measured in ADL) in the DDD model, accounting for individual and time fixed effects. The
control variables include socioeconomic characteristics, dietary patterns and provincial variables
such as industrial pollution, GDP and hospital numbers. The first two models include only samples in
the rural areas and the next two models include all samples. Models (1) and (3) include all the control
variables and Models (2) and (4) are without control variables. Complete regression results are in the
Appendix 3. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Table 3. The Effect of Rice Fertilizer Use on ADL
(1) (2) (3) (4) VARIABLES ADL ADL ADL ADL Rfer*Water*Y04 0.018 0.008 (0.295) (0.547) Rpest*Water*Y04 0.024* 0.017 (0.064) (0.201) Observations 28,158 35,228 28,158 35,228 Control Variables Yes Yes Yes Yes Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table reports results testing whether the significant impact of pesticide use on ADL is
driven by fertilizer use. The first two models run the analogous DDD models with rice fertilizer use
as the variable of interest, accounting for individual and time fixed effects. The control variables
include socioeconomic characteristics, dietary patterns and provincial variables such as industrial
pollution, GDP and hospital numbers. The next two models add annual fertilizer use as a control
variable in Equation (1) to test the stability of the pesticide coefficient. Models (1) and (3) only
include samples in rural areas and Models (2) and (4) include all samples. Complete regression
results are in the Appendix 4. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, *
p<0.1.
Table 4. The Effects of Pesticide and Fertilizer Use on Number of People Living Together
(1) (2) (3) (4) VARIABLES Live_with Live_with Rpest*Y04 0.005 Rfer*Y04 0.002 (0.164) (0.297) Rpest*y2002 0.001 Rfer* y2002 -0.002 (0.888) (0.483) Rpest*y2005 0.005 Rfer* y2005 0.001 (0.252) (0.782) Rpest*y2008 0.006 Rfer* y2008 0.001 (0.158) (0.773) Rpest*y2011 0.006 Rfer* y2011 -0.002 (0.340) (0.635) Observations 28,158 28,158 28,158 28,158 Control Var. Yes Yes Yes Yes Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows results testing whether offspring out-migration is the unobserved factor that
drives the significant impact of pesticide use on ADL. We estimate the effects of rice pesticide and
fertilizer use on the number of people living together in a difference-in-differences model, accounting
for individual and time fixed effects. The control variables include socioeconomic characteristics,
dietary patterns and provincial variables such as industrial pollution, GDP and hospital numbers.
Models (1) and (3) test whether the policy change predict the number of people living together in
regions with different intensities of pesticide and fertilizer use respectively. Models (2) and (4) allow
the effects of chemical uses to vary each year by adding separate year dummies. Complete regression
results are in the Appendix 5. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, *
p<0.1.
Table 5. The Effects of Pesticide and Fertilizer Use on Number of People Living Together
(5) (6) (7) (8) VARIABLES DDD DDD DDD DDD Rpest*Water*Y04 0.044** 0.042** (0.021) (0.035) Rfer*Water*Y04 0.023 0.022 (0.568) (0.597) Observations 28,158 28,158 28,158 28,158 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows effects of pesticide and fertilizer use on the number of people living together
in the DDD models, accounting for individual and time fixed effects. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Models (1) and (3) include all the control variables and Models (2) and
(4) are without control variables. Complete regressions are in Appendix 6. Robust p-values are
reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Table 6. Robustness Checks of Water Availability and Types of Drinking Water
(1) (2) (3) (4) VARIABLES Water
Consumption ADL Water Water
Rpest*Y04 -0.037 Rpest*Y04 0.000 (0.573) (0.370) Rpest*Water*Y04 0.023* Rpest*y2002 0.000 (0.073) (0.915) Rpest*y2005 0.000 (0.565) Rpest*y2008 0.000 (0.712) Rpest*y2011 0.000 (0.704) Observations 204 28,158 28,158 28,158 Control Var. Yes Yes Yes Yes Year FE Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes
Notes: This table reports results of robustness checks on water availability and types of drinking
water. The first model tests whether household water consumption in regions with different
intensities of pesticide differ significantly before and after 2004, accounting for provincial and year
fixed effects. The consumption of water is measured in million cubic meters at the provincial level.
The control variables for the first model include provincial variables such as industrial pollution,
GDP and hospital numbers. The second model tests whether adding or deleting residential water
consumption affects the coefficient of pesticide use in the DDD model. The third model tests whether
the pattern of drinking water differs before and after the policy change in regions with higher and
lower intensity of pesticide use. The last model allows the effect of pesticide use on the type of
drinking water to change each year by adding separate year dummies. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Complete regression results are in the Appendix 7. Robust p-values are
reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Table 7. The Effects of Wheat and Corn Pesticide Use on ADL
(1) (2) (3) (4) VARIABLES ADL ADL ADL ADL (a) Wpest*Water*Y04 0.110 0.161 0.057 0.107 (0.794) (0.677) (0.891) (0.780) Observations 15,388 15,388 19,424 19,424 (b) Cpest*Water*Y04 0.096 0.187 0.021 0.111 (0.651) (0.422) (0.904) (0.611) Observations 21,341 21,341 27,203 27,203 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows results of two falsification tests. Panel (a) reports results of wheat pesticide
use on ADL in the DDD model, accounting for individual and time fixed effects. Panel (b) reports
results of corn pesticide use on ADL in an analogous DDD model. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. The first two models only include samples in rural areas and the last two
models include full samples. Model (1) and (3) include all the control variables and Model (2) and (4)
are without control variables. Complete regression results are in the Appendix 8. Robust p-values are
reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Table 8: The Effect of ADL on Medical Costs and Offspring’s Weekly Caring Time
Medical Costs Offspring's Weekly Caring Hours
VARIABLES (1) (2) (3) (4)
ADL 138.896*** 144.291*** 5.947*** 6.303*** (0.000) (0.000) (0.002) (0.001)
Observations 20,433 20,433 7,742 7,742 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table reports the results of the effect of ADL on medical costs and offspring weekly
caring hours, accounting for individual and year fixed effects. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Model (1) and (3) include all control variables and Model (2) and (4) are
without control variables. Note that the relationship between ADL and medical cost/caring time is
not causal. Complete regression results are in appendix 9. Robust p-values are reported in
parentheses *** p<0.01, ** p<0.05, * p<0.1.
Appendix 1:
The Chinese Longitudinal Healthy Longevity Survey (CLHLS) was collected by the Center for Healthy Aging and Development Studies at
Peking University and co-sponsored by the US National Institute on Aging. The baseline survey and the follow-up surveys with replacement for
deceased elders were conducted in a randomly selected half of the counties and cities in 22 of China’s 31 provinces in 1998, 2000, 2002, 2005,
2008 and 2011. The surveyed provinces are Liaoning, Jilin, Heilongjiang, Hebei, Beijing, Tianjing, Shanxi, Shaanxi, Shanghai, Jiangsu,
Zhejiang, Anhui, Fujian, Jiangxi, Shangdong, Henan, Hubei, Hunan, Guangdong, Guangxi, Sichuan, and Chongqing. The population in the
survey areas constitutes about 85 percent of the total population in China.
The CLHLS surveyed only the oldest old (aged 80 years and older) before 2002, but since 2002, the cohort aged 65–79 has also been
included in the project. In 2002 and 2008, this survey also surveyed one of the elderly interviewees’ adult children aged 35–65, but this part of
data was not released. The survey combines an in-home interview and a basic physical examination. The questionnaire design was based on
international standards and was adapted to the Chinese cultural/social context and carefully tested by pilot studies and interviews. The data
collected includes demographic characteristics, family and household characteristics, lifestyles, diet, psychological characteristics, health,
disability, socioeconomic conditions, and so on. Questions that could not be reliably answered by older adults were minimized. Those who were
too sick to participate were not interviewed.
Table A1. CLHLS Sample Distributions of Age and Gender
Age Followed Samples Newly Added Total
Male Female Total Male Female Total Male Female Total
1998 Baseline Survey 80-89 --- --- --- 1,787 1,741 3,528 1,787 1,741 3,528 90-99 --- --- --- 1,299 1,714 3,013 1,299 1,714 3,013 100+ --- --- --- 481 1,937 2,418 481 1,937 2,418 Total --- --- --- 3,567 5,392 8,959 3,567 5,392 8,959 2000 80-89 996 1,048 2,044 1,471 1,403 2,874 2,467 2,451 4,918 90-99 720 907 1,627 925 1,260 2,185 1,645 2,167 3,812 100+ 262 891 1,153 256 1,022 1,278 518 1,913 2,431 Total 1,978 2,846 4,824 2,652 3,685 6,337 4,630 6,531 11,161 2002 35-65 --- --- --- 3,132 1,346 4,478 3,132 1,346 4,478 65-79 --- --- --- 2,456 2,438 4,894 2,456 2,438 4,894 80-89 1,454 1,411 2,865 673 672 1,345 2,127 2,083 4,210 90-99 948 1,236 2,184 590 858 1,448 1,538 2,094 3,632 100+ 277 917 1,194 442 1,685 2,127 719 2,602 3,321 Total 2,679 3,564 6,243 7,293 6,999 14,292 9,972 10,563 20,535 2005 35-65 2,183 820 3,003 --- --- --- 2,183 820 3,003 65-79 1,849 1,805 3,654 959 935 1,894 2,808 2,740 5,548 80-89 1,147 1,231 2,378 741 748 1,489 1,888 1,979 3,867
90-99 522 824 1,346 943 1292 2,235 1,465 2,116 3,581 100+ 158 600 758 360 1462 1,822 518 2,062 2,580 Total 5,859 5,280 11,139 3,003 4,437 7,440 8,862 9,717 18,579 2008 35-65 --- --- --- 1,909 1,894 3,803 1,909 1,894 3,803 65-79 1,437 1,417 2,854 822 610 1,432 2,259 2,027 4,286 80-89 975 1,011 1,986 1,175 1,117 2,292 2,150 2,128 4,278 90-99 775 1,066 1,841 1,137 1,643 2,780 1,912 2,709 4,621 100+ 169 625 794 498 2,086 2,584 667 2,711 3,378 Total 3,356 4,119 7,475 5,541 7,350 12,891 8,897 11,469 20,366 2011 35-65 1 0 1 270 238 508 271 238 509 65-79 1,493 1,325 2,818 219 112 331 1,712 1,437 3,149 80-89 1,165 1,188 2,353 151 136 287 1,316 1,324 2,640 90-99 918 1,275 2,193 99 141 240 1,017 1,416 2,433 100+ 225 815 1,040 62 355 417 287 1,170 1,457 Total 3,802 4,603 8,405 801 982 1,783 4,603 5,585 10,188 Six Survey in Total (1998、2000、2002、2005、2008、2011) 35-65 2,184 820 3,004 5,311 3,478 8,789 7,495 4,298 11,79365-79 4,779 4,547 9,326 4,456 4,095 8,551 9,235 8,642 17,87780-89 5,737 5,889 11,626 5,998 5,817 11,815 11,735 11,706 23,44190-99 3,883 5,308 9,191 4,993 6,908 11,901 8,876 12,216 21,092100+ 1,091 3,848 4,939 2,099 8,547 10,646 3,190 12,395 15,585Total 17,674 20,412 38,086 22,857 28,845 51,702 40,531 49,257 89,788
Notes: (1) This table reports the sample distribution of age and gender in CLHLS. (2) In 2002 and 2008, this survey also surveyed one of the elderly
interviewees’ adult children aged 35–65, but this part of data was not released. (3) Sources: Zeng et al. 2015.
Table A1 reports the sample distributions of age and gender in the CLHLS. The full sample size is 89788. In this paper, in order to
implement individual fixed effects, the data is limited to those who are sampled more than once, resulting to a sample size of 56399.
Additionally, I limit the samples to the years after 2000, to those who drink boiled water and to those with non-missing control variables,
resulting in a sample size of 39986. Further combinations with crop production result in sample sizes of 35228, 19424 and 27203 for rice,
wheat and corn productions respectively.
The attrition rates at the individual level are 46.15%, 44.06%, 49.33%, 52.01% and 49.26% in 2000, 2002, 2005, 2008 and 2011
respectively.15 The attribution rates are higher than usual because the survey samples are older adults. The lost samples were replaced by new
interviewees of the same sex and age (or within the same 5-year age group). On average, one individual was sampled 2.39 times. If the
respondents die or migrate, they will be counted in the attrition. Attrition is potentially worrisome if it is correlated with the independent
variable of interest. I run analogous DDD regressions where the dependent variable is the number of times this individual is sampled or whether
this person has been sampled more than once. In no regression (Table A2) is the coefficient on the three-way interaction term significantly
different from zero, which suggests the respondents lost to attrition are random with respect to the triple difference design.
15 The attrition rates are adjusted by not including the cohort aged 35-65 because this part of data is not released.
Table A2. Tests of Attrition Bias
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES All All Rural Rural All All Rural Rural Rpest*Water*Y04 0.008 0.005 0.003 -0.001 -0.001 -0.002 -0.002 -0.004 (0.563) (0.684) (0.885) (0.929) (0.883) (0.656) (0.639) (0.382) Rpest*Water 0.008 0.008 0.011 0.011 0.004*** 0.004*** 0.005** 0.005*** (0.209) (0.208) (0.162) (0.167) (0.001) (0.002) (0.018) (0.007) Water*Y04 -0.384 -0.273 -0.259 -0.116 -0.017 0.014 0.018 0.063 (0.389) (0.483) (0.649) (0.814) (0.899) (0.905) (0.904) (0.626) Rpes*Y04 0.001 0.003 0.004 0.004 0.000 0.000 0.001 0.000 (0.857) (0.630) (0.530) (0.474) (0.850) (0.823) (0.563) (0.809) Water -0.009 -0.039 -0.075 -0.105 -0.048 -0.056 -0.075 -0.089 (0.966) (0.841) (0.770) (0.670) (0.189) (0.112) (0.249) (0.127) Y04 0.183 0.179 0.213 0.095 0.085 0.163*** 0.128 0.154*** (0.524) (0.224) (0.511) (0.531) (0.374) (0.000) (0.223) (0.000) Age -0.022*** -0.024*** -0.007*** -0.008*** (0.000) (0.000) (0.000) (0.000) Smoke -0.005 -0.004 0.000 -0.003 (0.843) (0.863) (0.947) (0.663) Drink 0.021 0.020 0.010 0.010 (0.253) (0.375) (0.133) (0.155) Exercise 0.213*** 0.191*** 0.064*** 0.060*** (0.000) (0.000) (0.000) (0.000) Live_with 0.009 0.011* 0.002 0.001 (0.182) (0.098) (0.126) (0.169) Staple 0.001 0.000 0.000 0.001 (0.714) (0.879) (0.662) (0.508)
Fish -0.047*** -0.055*** -0.018*** -0.020*** (0.001) (0.000) (0.000) (0.000) Fruit 0.020 0.029 -0.005 -0.001 (0.245) (0.116) (0.206) (0.833) Vegetable 0.021* 0.017 0.002 -0.002 (0.060) (0.146) (0.817) (0.865) Meat -0.050*** -0.055*** -0.009** -0.009** (0.000) (0.000) (0.027) (0.046) Salt_vege 0.022* 0.017 0.007* 0.005 (0.089) (0.216) (0.086) (0.382) SO2 -0.467 -14.875 0.687 -1.466 (0.985) (0.610) (0.916) (0.846) COD 27.568 23.900 6.011 5.033 (0.539) (0.678) (0.642) (0.741) GDP -0.037 -0.630 -0.180 -0.546 (0.976) (0.650) (0.691) (0.281) Hospnum 0.019 0.012 0.027 0.026 (0.627) (0.776) (0.135) (0.192) Constant 4.157*** 2.475*** 4.942*** 2.472*** 1.387*** 0.733*** 1.679*** 0.729*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Observations 46,879 46,879 37,318 37,318 46,879 46,879 37,318 37,318 R-squared 0.097 0.039 0.104 0.041 0.082 0.034 0.084 0.034 Control Variables Yes No Yes No Yes No Yes No Year FE Yes Yes Yes Yes Yes Yes Yes Yes Province FE Yes Yes Yes Yes Yes Yes Yes Yes
Notes: This table reports results of testing attrition bias in the DDD models, account province and year fixed effects. The dependent variable of the first four
models is the number of times this individual is sampled and the dependent variable of the last four models is whether this person has been
sampled more than once. The control variables include socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Model (1), (2), (5) and (6) include all samples and Model (3), (4), (7) and (8) include only rural samples. Model (1), (3), (5) and
(7) include all control variables and Model (2), (4), (6) and (8) are without control variables. Robust p-values are reported in parentheses *** p<0.01, **
p<0.05, * p<0.1.
Appendix 2:
In this appendix, I provide more details on translating estimated effects into expected
monetary losses. In the section of “The Effect of Pesticide Use on Health Outcomes”, the
result shows that a one yuan per mu increase in expenditures on rice pesticides increases the
index of ADL by 0.024 units. In the section of “Monetary Loss”, the result shows that a one
unit increase in ADL increases medical costs and offspring’s weekly caring time by 139 yuan
and 42 yuan, respectively. In sum, a one yuan per mu increase in expenditures on rice
pesticides increases medical costs and human capital losses by 3.34 and 1.01 yuan
respectively (Equation A1).
( )MonetaryCosts ADL MedicalCosts HumanCapitalCosts
Pesticide Pesticide ADL
(A1)
According to the China’s population census in 2010, the population aged 65 and above
is 118.9 million. The CLHLS survey shows that 1.39% population aged 65 and above drink
surface water. This suggests the population affected by pesticide polluted surface water is
1.65 million (118.9×1.39%). In combination with previous information, I am able to
calculate that a one yuan per mu increase in expenditures on rice pesticides increases
medical costs and human capital losses of local populations by 5.51 and 1.67 million yuan
($0.83 and $0.25 million), respectively. That is, a 10% (2.9 yuan/mu) increase in rice
pesticide use increases medical costs and human capital losses of local populations by 15.98
and 4.83 million yuan, respectively ($2.4 and $0.72 million). In sum, a 10% increase in rice
pesticide use increases monetary losses by 20.81 million yuan ($3.12 million).
According to the Agricultural Cost and Benefit Yearbook, the average rice production
profits per mu from 2000 to 2011 were 194.84 yuan. According to the China Statistics
Yearbook, the average sown area of rice from 2000 to 2011 was 399.6 million mu. The
average rice production profits were 194.84×399.6=77858 million yuan. This suggests that
the monetary losses from a 10% increase in rice pesticide use are equivalent to 0.03% of
national annual rice production profits.
Appendix 3:
Table A3. The Effect of Rice Pesticide Use on ADL
(1) (2) (3) (4) VARIABLES ADL ADL ADL ADL Rpest*Water*Y04 0.024* 0.034** 0.016 0.025 (0.068) (0.033) (0.218) (0.148) Rpest*Water -0.037*** -0.043*** -0.026*** -0.032*** (0.001) (0.000) (0.007) (0.002) Water*Y04 -0.325 -0.595 -0.199 -0.449 (0.488) (0.288) (0.685) (0.472) Rpest*Y04 0.003 -0.000 0.004 0.001 (0.696) (0.968) (0.613) (0.873) Water 0.905*** 1.090*** 0.675*** 0.849*** (0.002) (0.000) (0.008) (0.004) Age 0.215** 0.192*** (0.017) (0.002) Smoke -0.310*** -0.313*** (0.000) (0.000) Drink -0.083 -0.098* (0.232) (0.092) Exercise -0.569*** -0.636*** (0.000) (0.000) Live_with 0.021** 0.022** (0.032) (0.044) Staple -0.065*** -0.070*** (0.000) (0.000) Fish -0.052** -0.045* (0.026) (0.076) Fruit -0.077*** -0.069** (0.008) (0.046) Vegetable -0.236*** -0.265***
(0.000) (0.000) Meat -0.040 -0.071** (0.201) (0.013) Salt_vege -0.095*** -0.101*** (0.000) (0.000) SO2 13.910 26.812 (0.645) (0.287) COD 41.685 24.975 (0.393) (0.656) GDP -1.833 -1.376 (0.163) (0.207) Hospnum 0.059 0.071** (0.169) (0.012) Constant -8.969 6.604*** -7.182 6.667*** (0.216) (0.000) (0.146) (0.000) Observations 28,158 28,158 35,228 35,228 R-squared 0.712 0.699 0.710 0.695 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows the effect of pesticide use on the dependence of older adults (measured in
ADL) in the DDD model, accounting for individual and time fixed effects. The control variables
include socioeconomic characteristics, dietary patterns and provincial variables such as industrial
pollution, GDP and hospital numbers. The first two models include only samples in the rural areas
and the next two models include all samples. Models (1) and (3) include all the control variables and
Models (2) and (4) are without control variables. Robust p-values are reported in parentheses ***
p<0.01, ** p<0.05, * p<0.1. Appendix 4:
Table A4. The Effect of Rice Fertilizer Use on ADL
(1) (3) (5) (7) VARIABLES ADL ADL ADL ADL Rfer*Water*Y04 0.018 0.008 (0.295) (0.547) Rfer*Water -0.023 -0.016 (0.164) (0.195) Rfer*Y04 0.006** 0.008*** (0.036) (0.006) Water*Y04 -1.028 -0.315 (0.415) (0.752)
Rpest*Water*Y04 0.024* 0.017 (0.064) (0.201) Rpest*Water -0.037*** -0.026*** (0.001) (0.005) Water*Y04 -0.326 -0.200 (0.489) (0.681) Rpest*Y04 0.003 0.004 (0.689) (0.584) Rfer1 0.006 0.009 (0.417) (0.140) Water 1.691 1.218 0.906*** 0.675*** (0.203) (0.211) (0.002) (0.007) Age 0.228*** 0.207*** 0.226** 0.208*** (0.009) (0.001) (0.013) (0.000) Smoke -0.308*** -0.310*** -0.312*** -0.316*** (0.000) (0.000) (0.000) (0.000) Drink -0.086 -0.101* -0.083 -0.099* (0.211) (0.079) (0.228) (0.088) Exercise -0.572*** -0.638*** -0.571*** -0.638*** (0.000) (0.000) (0.000) (0.000) Live_with 0.021** 0.021** 0.021** 0.021** (0.036) (0.047) (0.037) (0.049) Staple -0.065*** -0.069*** -0.065*** -0.070*** (0.000) (0.000) (0.000) (0.000) Fish -0.051** -0.044* -0.053** -0.046* (0.031) (0.083) (0.027) (0.067) Fruit -0.079*** -0.070** -0.077*** -0.069** (0.006) (0.038) (0.008) (0.044) Vegetable -0.238*** -0.269*** -0.236*** -0.267*** (0.000) (0.000) (0.000) (0.000) Meat -0.043 -0.074** -0.040 -0.071** (0.172) (0.012) (0.196) (0.013) Salt_vege -0.098*** -0.104*** -0.095*** -0.101*** (0.000) (0.000) (0.000) (0.000) SO2 5.867 14.869 10.151 19.022 (0.856) (0.558) (0.747) (0.453) COD 46.767 34.964 49.319 38.153 (0.240) (0.388) (0.293) (0.464) GDP -2.266* -1.872* -2.019 -1.682 (0.084) (0.087) (0.155) (0.135) Hospnum 0.054 0.064** 0.058 0.069*** (0.199) (0.023) (0.174) (0.010) Constant -9.669 -8.002* -10.197 -8.987* (0.170) (0.098) (0.172) (0.066)
Observations 28,158 35,228 28,158 35,228 R-squared 0.712 0.711 0.712 0.710 Control Variables Yes Yes Yes Yes Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table reports results testing whether the significant impact of pesticide use on ADL is
driven by fertilizer use. The first two models run the analogous DDD models with fertilizer use as the
variable of interest, accounting for individual and time fixed effects. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. The next two models add fertilizer use as a control variable in Equation
(1). Models (1) and (3) only include samples in rural areas and Models (2) and (4) include all
samples. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, * p<0.1. Appendix 5: Table A5. The Effects of Pesticide and Fertilizer Use on Number of People Living Together
(1) (2) (3) (4) VARIABLES DD DD DD DD Rpest*Y04 0.005 (0.164) Rpest*y2002 0.001 (0.888) Rpest*y2005 0.005 (0.252) Rpest*y2008 0.006 (0.158) Rpest*y2011 0.006 (0.340) Rfer*Y04 0.002 (0.297) Rfer* y2002 -0.002 (0.483) Rfer* y2005 0.001 (0.782) Rfer* y2008 0.001 (0.773) Rfer* y2011 -0.002 (0.635) Age 0.024 0.021 0.012 0.110 (0.692) (0.850) (0.829) (0.356)
Smoke 0.074* 0.074* 0.074* 0.072* (0.083) (0.083) (0.082) (0.092) Drink 0.015 0.015 0.016 0.013 (0.742) (0.739) (0.733) (0.777) Exercise 0.108*** 0.108*** 0.109*** 0.108*** (0.000) (0.000) (0.000) (0.000) Staple 0.007 0.007 0.007 0.006 (0.231) (0.245) (0.236) (0.266) Fish -0.014 -0.014 -0.015 -0.015 (0.629) (0.626) (0.611) (0.612) Fruit 0.002 0.002 0.002 0.002 (0.938) (0.940) (0.948) (0.932) Vegetable 0.028 0.028 0.028 0.029 (0.390) (0.394) (0.388) (0.369) Meat 0.065*** 0.065*** 0.064*** 0.065*** (0.000) (0.000) (0.000) (0.000) Salt_vege 0.040** 0.040** 0.039** 0.040** (0.012) (0.010) (0.013) (0.011) SO2 11.739 13.100 8.268 5.749 (0.504) (0.504) (0.665) (0.775) COD -26.240 -27.386 -22.357 -18.896 (0.469) (0.506) (0.578) (0.663) GDP -0.174 -0.234 0.260 0.008 (0.869) (0.847) (0.798) (0.994) Hospnum 0.025 0.026 0.025 0.015 (0.484) (0.557) (0.466) (0.674) Constant 0.908 1.119 1.595 -6.104 (0.847) (0.902) (0.729) (0.529) Observations 28,158 28,158 28,158 28,158 R-squared 0.720 0.720 0.720 0.721 Control Variables Yes Yes Yes Yes Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows results testing whether offspring out-migration is the unobserved factor that
drives the significant impact of pesticide use on ADL. We estimate the effects of rice pesticide and
fertilizer use on the number of people living together in a difference-in-differences model, accounting
for individual and time fixed effects. The control variables include socioeconomic characteristics,
dietary patterns and provincial variables such as industrial pollution, GDP and hospital numbers.
Models (1) and (3) test whether the policy change predict the number of people living together in
regions with different intensities of pesticide and fertilizer use respectively. Models (2) and (4) allow
the effects of chemical uses to vary each year by adding separate year dummies. Robust p-values are
reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Appendix 6: Table A6. The Effects of Pesticide and Fertilizer Use on Number of People Living Together
(5) (6) (7) (8) VARIABLES DDD DDD DDD DDD Rpest*Water*Y04 0.044** 0.042** (0.021) (0.035) Rpest*Water -0.033 -0.032 (0.101) (0.125) Rpest*Y04 0.004 0.004 (0.209) (0.215) Water*Y04 -1.218** -1.188* (0.041) (0.063) Rfer*Water*Y04 0.023 0.022 (0.568) (0.597) Rfer*Water -0.028 -0.026 (0.468) (0.498) Rfer*Y04 0.002 0.003 (0.313) (0.200) Water*Y04 -1.801 -1.704 (0.556) (0.580) Water 0.788 0.754 2.051 1.923 (0.133) (0.162) (0.470) (0.500) Age 0.023 0.013 (0.695) (0.826) Smoke 0.074* 0.076* (0.088) (0.080) Drink 0.015 0.015 (0.742) (0.741) Exercise 0.110*** 0.110*** (0.000) (0.000) Staple 0.007 0.007 (0.230) (0.232) Fish -0.014 -0.015 (0.624) (0.603) Fruit 0.002 0.001 (0.923) (0.963) Vegetable 0.028 0.029 (0.383) (0.369) Meat 0.065*** 0.064*** (0.000) (0.000)
Salt_vege 0.040** 0.039** (0.012) (0.012) SO2 12.192 8.883 (0.493) (0.643) COD -25.786 -22.870 (0.473) (0.567) GDP -0.169 0.275 (0.874) (0.784) Hospnum 0.025 0.025 (0.477) (0.465) Constant 0.907 3.188*** 1.573 3.183*** (0.848) (0.000) (0.731) (0.000) Observations 28,158 28,158 28,158 28,158 R-squared 0.721 0.720 0.721 0.720 Control Variables Yes No Yes No Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table shows effects of pesticide and fertilizer use on the number of people living together
in the DDD models, accounting for individual and time fixed effects. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Models (1) and (3) include all the control variables and Models (2) and
(4) are without control variables. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05,
* p<0.1. Appendix 7:
Table A7. Robustness Checks of Water Availability and Types of Drinking Water
(1) (2) (3) (4) VARIABLES Water
Consumption ADL Water Water
Rpest*Y04 -0.037 Rpest*Y04 0.000 (0.573) (0.370) Rpest*Water*Y04 0.023* Rpest*y2002 0.000 (0.073) (0.915) Rpest*y2005 0.000 (0.565) Rpest*y2008 0.000 (0.712) Rpest*y2011 0.000 (0.704)
Age 0.207** 0.004 0.004 (0.021) (0.144) (0.350) Smoke -0.311*** 0.004 0.004 (0.000) (0.267) (0.267) Drink -0.083 -0.003 -0.003 (0.229) (0.273) (0.256) Exercise -0.572*** 0.001 0.001 (0.000) (0.756) (0.756) Live_with 0.021** -0.001 -0.001 (0.034) (0.231) (0.232) Staple -0.065*** -0.000 -0.000 (0.000) (0.376) (0.379) Fish -0.052** 0.001 0.001 (0.028) (0.738) (0.762) Fruit -0.077*** -0.004** -0.004** (0.007) (0.016) (0.017) Vegetable -0.235*** 0.000 0.001 (0.000) (0.797) (0.784) Meat -0.041 -0.003* -0.003* (0.196) (0.056) (0.057) Salt_vege -0.095*** 0.001 0.001 (0.000) (0.655) (0.645) SO2 -0.057** 20.353 0.873 0.608 (0.029) (0.542) (0.477) (0.688) COD 0.115 25.309 3.199 3.354 (0.128) (0.626) (0.248) (0.281) GDP 0.004*** -1.802 -0.019 -0.005 (0.000) (0.185) (0.767) (0.950) Hospnum -0.000 0.058 -0.001 -0.001 (0.324) (0.180) (0.591) (0.669) Constant -44.623 -8.607 -0.321 -0.295 (0.359) (0.236) (0.116) (0.315) Observations 204 28,158 28,158 28,158 R-squared 0.985 0.712 0.591 0.591 Control Var. Yes Yes Yes Yes Year FE Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes
Notes: This table reports results of robustness checks on water availability and types of drinking water. The first model tests whether household water consumption in regions with different intensities of pesticide differ significantly before and after 2004, accounting for provincial and year fixed effects. The consumption of water is measured in million cubic meters at the provincial level. The control variables for the first model include provincial variables such as industrial pollution, GDP and hospital numbers. The second model tests
whether adding or deleting residential water consumption affects the coefficient of pesticide use in the DDD model. The third model tests whether the pattern of drinking water differs before and after the policy change in regions with higher and lower intensity of pesticide use. The last model allows the effect of pesticide use on the type of drinking water to change each year by adding separate year dummies. The control variables include socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution, GDP and hospital numbers. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Appendix 8:
Table A8. The Effects of Wheat and Corn Pesticide Use on ADL
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES ADL ADL ADL ADL ADL ADL ADL ADL Wpest*Water*Y04 0.110 0.161 0.057 0.107 (0.794) (0.677) (0.891) (0.780) Wpest*Water -0.190 -0.226 -0.145* -0.180 (0.261) (0.182) (0.065) (0.168) Water*Y04 -0.847 -1.354 -0.197 -0.718 (0.778) (0.628) (0.947) (0.795) Wpest*Y04 0.014 0.002 0.018 0.011 (0.847) (0.965) (0.773) (0.720) Cpest*Water*Y04 0.096 0.187 0.021 0.111 (0.651) (0.422) (0.904) (0.611) Cpest*Water -0.245 -0.289 -0.174 -0.222 (0.163) (0.154) (0.186) (0.180) Cpest*Y04 0.079 0.054 0.097** 0.071** (0.139) (0.105) (0.045) (0.040) Water*Y04 -0.183 -0.712 0.191 -0.349 (0.899) (0.649) (0.867) (0.802) Water 1.629 2.019* 1.120* 1.496 1.354 1.619 0.976 1.275 (0.178) (0.100) (0.057) (0.122) (0.212) (0.183) (0.200) (0.174) Age 0.240* 0.205** 0.219* 0.212** (0.081) (0.030) (0.083) (0.012) Smoke -0.215*** -0.247*** -0.293*** -0.322*** (0.005) (0.000) (0.000) (0.000)
Drink -0.104 -0.134 -0.102 -0.115* (0.364) (0.122) (0.249) (0.089) Exercise -0.552*** -0.647*** -0.598*** -0.682*** (0.000) (0.000) (0.000) (0.000) Live_with 0.024 0.032 0.027* 0.029* (0.220) (0.154) (0.054) (0.056) Staple -0.068*** -0.072*** -0.066*** -0.071*** (0.000) (0.000) (0.000) (0.000) Fish -0.023 -0.046* -0.028 -0.028 (0.442) (0.051) (0.284) (0.368) Fruit -0.058** -0.062 -0.091*** -0.084** (0.017) (0.208) (0.007) (0.049) Vegetable -0.232*** -0.272*** -0.235*** -0.265*** (0.000) (0.000) (0.000) (0.000) Meat -0.057 -0.098** -0.054 -0.085** (0.244) (0.031) (0.156) (0.013) Salt_vege -0.082** -0.073*** -0.095*** -0.099*** (0.013) (0.001) (0.001) (0.000) SO2 29.579 34.026 22.239 36.562* (0.333) (0.105) (0.482) (0.099) COD 68.693 48.151 41.092 20.360 (0.481) (0.631) (0.366) (0.706) GDP -2.145 -1.231 -1.840 -1.588 (0.376) (0.571) (0.332) (0.293) Hospnum 0.058 0.063 0.060 0.061 (0.360) (0.159) (0.298) (0.102) Constant -11.106 6.632*** -8.412 6.694*** -9.506 6.570*** -8.869 6.635*** (0.299) (0.000) (0.255) (0.000) (0.343) (0.000) (0.178) (0.000)
Observations 15,388 15,388 19,424 19,424 21,341 21,341 27,203 27,203 R-squared 0.708 0.696 0.705 0.689 0.708 0.694 0.708 0.692 Control Variables Yes No Yes No Yes No Yes No Year FE Yes Yes Yes Yes Yes Yes Yes Yes Individual FE Yes Yes Yes Yes Yes Yes Yes Yes
Notes: This table shows results of two falsification tests. The first four models report results of wheat pesticide use on ADL in the DDD model, accounting for individual and time fixed effects. The last four models report results of corn pesticide use on ADL in an analogous DDD model. The control variables include socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution, GDP and hospital numbers. Model (1), (2), (5) and (6) include only rural samples and Model (3), (4), (7) and (8) include full samples. Model (1), (3), (5) and (7) include all the control variables and Model (2), (4), (6) and (8) are without control variables. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Appendix 9:
Table A9. The Effect of ADL on Medical Costs and Offspring’s Weekly Caring Time
(1) (2) (3) (4) VARIABLES Medical Costs Offspring’s Weekly Caring Time
ADL 138.896*** 144.291*** 5.947*** 6.303*** (0.000) (0.000) (0.002) (0.001) Age 162.837 -3.770 (0.143) (0.365) Smoke -291.402*** -8.939 (0.006) (0.165) Drink -201.650 -5.751 (0.157) (0.350) Exercise -71.783 -7.525 (0.271) (0.100) Live_with 36.367* -2.746** (0.093) (0.033) Staple -42.989* -1.070 (0.054) (0.177) Fish 11.926 -2.528 (0.896) (0.247) Fruit 35.988 1.401 (0.462) (0.606) Vegetable 110.988** 2.118 (0.015) (0.425) Meat -31.588 0.958 (0.719) (0.700) Salt_vege 3.673 -3.509 (0.927) (0.136) SO2 24,348.678 115.484 (0.622) (0.959) COD -23,935.541 -5,127.403 (0.876) (0.417) GDP 2,302.048 43.773 (0.321) (0.514) Hospnum 26.172 0.348 (0.682) (0.907) Constant -15,514.677* 1,115.708*** 326.237 -35.729*** (0.090) (0.000) (0.400) (0.000) Observations 20,433 20,433 7,742 7,742 R-squared 0.647 0.645 0.880 0.878 Control Variables Yes No Yes No
Year FE Yes Yes Yes Yes Individual FE Yes Yes Yes Yes
Notes: This table reports the results of the effect of ADL on medical costs and offspring weekly
caring hours, accounting for individual and year fixed effects. The control variables include
socioeconomic characteristics, dietary patterns and provincial variables such as industrial pollution,
GDP and hospital numbers. Model (1) and (3) include all control variables and Model (2) and (4) are
without control variables. Note that the relationship between ADL and medical cost/caring time is
not causal. Robust p-values are reported in parentheses *** p<0.01, ** p<0.05, * p<0.1.
Appendix 10:
Table A10. Tests of Parallel Trends
(1) (2) VARIABLES Rural Rural Rpest*Water*d_year3 -0.051 -0.041 (0.184) (0.291) Rpest*Water -0.009 -0.019 (0.778) (0.545) Water*d_year3 1.899 1.580 (0.113) (0.193) Rpest*d_year3 0.002 0.000 (0.769) (0.996) Rpest*Water*d_year4 -0.024 -0.005 (0.509) (0.892) Water*d_year4 1.069 0.422 (0.388) (0.746) Rpest*d_year4 0.003 0.002 (0.783) (0.805) Water -0.248 0.098 (0.803) (0.927) Age 0.159* (0.090) Smoke -0.326*** (0.000) Drink -0.100 (0.341) Exercise -0.486*** (0.000) Live_with 0.013 (0.292) Staple -0.068*** (0.000) Fish -0.037 (0.218) Fruit -0.031 (0.569) Vegetable -0.185*** (0.000) Meat -0.045 (0.283) Salt_vege -0.035 (0.210)
SO2 -28.180 (0.498) COD 19.131 (0.759) GDP 1.877 (0.604) Hospnum 0.297** (0.031) Constant -7.029 7.360*** (0.412) (0.000) Observations 16,994 16,994 R-squared 0.781 0.773 Control Variables Yes No Year FE Yes Yes Individual FE Yes Yes
Notes: This table reports result of testing parallel trends by using rural samples before 2005. The
control variables include socioeconomic characteristics, dietary patterns and individual variables
such as industrial pollution, GDP and hospital numbers.. Model (1) include all control variables and
Model (2) are without control variables. Robust p-values are reported in parentheses *** p<0.01, **
p<0.05, * p<0.1.