The Effect of Microinsurance on Economic Activities: Evidence from a

Randomized Field Experiment∗

Hongbin Cai† Yuyu Chen† Hanming Fang‡ Li-An Zhou†

April 29, 2014

Abstract

Lack of access to formal insurance market may prevent farmers from pursuing risky production

activities with potentially large returns. How does access to formal microinsurance affect production

and economic development? We report results from a large randomized field experiment conducted in

southwestern China in the context of a subsidized insurance for sows. We induce exogenous variation

in insurance coverage at the village level by randomly assigning performance incentives to the village

Animal Husbandry Worker who is responsible for signing farmers up for the sow insurance. We find that

promoting greater adoption of insurance significantly increases farmers’ tendency to raise sows, and the

short-run effect of sow insurance on sow production seems to persist in the longer run. Moreover, we

find that the increase in sow production in response to the sow insurance does not seem to be the result

of the substitution of other livestock.

Keywords: Microinsurance; Production; Field Experiment

JEL Classification Codes: C93; O12; O16

∗We would like to thank the Editor Asim Khwaja and three anonymous referees for extremely thoughtful comments and

suggestions that significantly improved the paper. We also benefited from the comments from Manuela Angelucci, Abhijit

Banerjee, Esther Duflo, Xavier Gine, Robert Miller, Jonathan Morduch, Duncan Thomas, Shing-Yi Wang, Xiaobo Zhang and

seminar/conference participants at World Bank, NBER China Conference, NBER/BREAD Conference and Cowles-PKU-UCL

Microeconometrics Conference. We also gratefully acknowledge the financial support from the Mirrlees Institute for Economic

Policy Research at Peking University. Various government agencies of Jinsha County and Bijie Prefecture in Guizhou Province

supported our field experiment, provided us with data sets and generously answered many of our questions. We are responsible

for all remaining errors.†Department of Applied Economics, Guanghua School of Management, Peking University, Beijing 100871, China. Emails

are respectively hbcai@gsm.pku.edu.cn (Cai), chenyuyu@gsm.pku.edu.cn (Chen), and zhoula@gsm.pku.edu.cn (Zhou).‡Corresponding Author. Department of Economics, University of Pennsylvania, 3718 Locust Walk, Philadelphia, PA 19104;

and the NBER. Email: hanming.fang@econ.upenn.edu

1 Introduction

Farmers in less developed economies face significant barriers in access to credit, insurance and other

financial products taken for granted in developed countries. At the same time, they typically face far more

significant risks relative to their income than their counterparts in developed economies. Lack of access to

credit is one of the binding constraints which prevent potential entrepreneurs among farmers from obtaining

the necessary capital to start or expand their business, forcing them to either stay in traditional farming or

take other less profitable paths. Lack of access to formal insurance markets can similarly prevent farmers

from pursuing risky production activities with potentially large returns.1

International aid agencies, non-governmental organizations, and profit or non-profit private banks have

devoted a large amount of resources to provide credit to residents in low income regions. The best story of

microfinance is that of Muhammad Yunus and Bangladesh’s Grameen Bank which he founded in 1976, and

was replicated in more than thirty countries from East Timor, Bosnia and even many poor neighborhoods in

the United States.2 Academically, a large empirical literature has documented the success of microfinance

programs and a theoretical literature is also developed to understand its success.3

Surprisingly, there has been much less effort, both practically and academically, devoted to provide

microinsurance to farmers in low income economies. As Morduch (2006) observed: “The prospects (of

microinsurance) are exciting, but much remains unknown. The expanding gaggle of microinsurance advo-

cates are ahead of the available evidence on insurance impacts. ... The advocates may be right, at least

in the long-term, but it is impossible to point to a broad range of great evidence on which to base that

prejudice.”

Studying the causal effect of insurance on agricultural production using observational data is a challeng-

ing task because of the problem of unobserved heterogeneity. Individuals with certain traits may self select

into some specific insurance scheme, and these unobserved traits may also affect the choice of production

technology, effort level and thus output. For instance, more risk averse farmers may prefer insurance and

at the same time devote more efforts in choosing effective technology to protect against animal diseases

and epidemics. The presence of self-selection may cause a spurious correlation between insurance coverage

and agricultural output.

To overcome the above challenge, we use in this paper an experimental approach to study the effect

of insurance access on farmers’ subsequent production decisions. Our experimental design, explained in

details in Section 3, creates an exogenous source of variation in insurance coverage across villages that is

arguably orthogonal to agricultural output, and we then use this exogenous variation to identify the causal

effect of insurance on production.4

Specifically, we report results from a large randomized field experiment conducted in southwestern

1See, e.g., Morduch (1995) for a discussion about how farm households in developing countries may seek to smooth their

consumption by altering their methods of production.2Yunus (2001) documents the origins and development of the Grameen Bank, and see Robinson (2001) for an account of

its replications around the world.3de Aghion and Morduch (2005) provide a comprehensive review. Banerjee, Duflo, Glennerster and Kinnan (2010) report

results from a large-scale randomized evaluation of the impact of introducing microcredit on various measures of economic

activities in a new market.4See Harrison and List (2004) and List (2006) for surveys and methodological discussions, including categorizations, of the

surging literature of field experiments; and see Duflo, Glennerster and Kremer (2007) for the application of the experimental

methods especially in development economics.

1

China in the context of a subsidized insurance for sows. Our study sheds light on one important question

about microinsurance: how does access to formal insurance affect farmers’ production decisions? We find

that promoting greater adoption of insurance significantly increases farmers’ tendency to raise sows, and

the short-run effect of sow insurance on sow production seems to persist in the longer run. Moreover, we

find that the increase in sow production in response to the sow insurance does not seem to be the result

of the substitution of other livestock such as pigs, sheep or cows.

There is a small existing literature in agricultural economics that examined the effect of federal crop

insurance on farmers’ decisions using non-experimental data.5 However, to the best of our knowledge

our paper is among the first to examine the causal effect of microinsurance on production behavior using

randomized field experiments.6 Related to our study, Mobarak and Rosenzweig (2012) studied the demand

for, and the effects of, offering formal index-based rainfall insurance through a randomized experiment in

an environment with existing informal risk sharing in rural India. Karlan et al. (2013) also conducted

randomized field experiments in northern Ghana in which farmers were randomly assigned to receive cash

grants, or opportunities to purchase rainfall index insurance, or a combination of the two. They found

that insurance access leads to significantly larger agricultural investment and riskier production choices.

Similarly, Cole, Gine and Vickery (2013a) randomly assigned farmers in Indian villages to rainfall insurance

or cash payments, and studied how differences in subsequent production decisions during the monsoon

between these two groups. They found that insurance provision has little effect on total agricultural

investments, but causes significant shifts in the composition of those investments, particularly among more

educated farmers. The randomizations of treatment and control in all of the above studies are at the

household level; in contrast, we randomize at the village level (see Section 3).

Also related to our study, several recent papers studied the determinants of the demand for microinsur-

ance. Gine et al. (2008) found that, for a rainfall insurance policy offered to small farmers in rural India,

the take-up is decreasing in the basis risk between insurance payouts and income fluctuations, increasing

in household wealth and decreasing in the extent to which credit constraints bind. These results match

the predictions of a simple neoclassical model augmented with borrowing constraints. However, they also

found that risk averse households are less likely to purchase insurance, and participation in village networks

and familiarity with the insurance vendor are strongly correlated with insurance take-up decisions. Closely

related, Cole et al. (2013b) documented low levels of rainfall insurance take-up, and then conducted field

experiments to understand why adoption is so low. Their experimental results demonstrated that high price

of the insurance and credit constraints of the farmers are important determinants of insurance adoption,

but they also found evidence that the endorsement from a trusted third party about the insurance policy

significantly increases the insurance take-up. In this paper we also find corroborative evidence suggesting

the importance of trust for the insurance take-up. Cai (2013) studied the role of social networks on the

decisions to purchase crop insurance in rural China for rice farmers. These studies do not examine the

causal effect of rainfall insurance on agricultural production.

Our paper is also related to the large and important literature in development economics on how poor

villagers rely on informal insurance to cope with risks. In a seminal paper, Townsend (1994) tested whether

5See, for example, Horowitz and Lichtenberg (1993), Goodwin, Vandeveer and Deal (2004) and O’Donoghue, Key and

Roberts (2007).6Innovations for Poverty Action (2009) reviews the existing and ongoing microinsurance field experiments. Most of the

experiments focus on the low take-up rate of insurance products.

2

community-based informal insurance arrangements might effectively protect the poor’s consumption levels

from unusual swings in income; he found that, among the roughly 120 households in three villages in

southern India, full insurance provides a surprisingly good benchmark although it is statistically rejected.7

Our evidence that access to formal insurance has a significant effect on the farmers’ production decisions

suggests that formal insurance can still play an important role even in areas with informal risk sharing,

possibly because formal insurance allows villagers to better insure against aggregate shocks.

The remainder of the paper is structured as follows. In Section 2 we provide the institutional back-

ground for hog production, and the insurance program for sows introduced in China in 2007. In Section

3 we describe and discuss our experimental design. In Section 4 we describe the data sets and provide

summary statistics. In Section 5 we present and discuss our experimental result that sow insurance sig-

nificantly affects farmers’ decision to raise sows in subsequent periods. Finally, in Section 6 we conclude.

Supplemental tables and omitted details are collected in an online appendix.

2 Background

Pork is an important part of the Chinese diet; about 52 million tons of pork (valued at 644.25 billion

Yuan or 101.46 billion US dollars) were produced in China in 2006.8 In China, pigs are mainly raised by

rural households in their backyard as a sideline business, and large-scale hog farms are unusual especially in

its mountainous southwestern regions. The small-scale and scattered nature of hog raising exposes farmers

to high incidence of pig diseases; and as a result, mortality rates for pigs and sows are quite high.9

Though there is no nationwide census data to estimate the annual mortality rate for sows, People’s

Insurance Company of China (PICC) estimates that the mortality rate of insured sows was about 2% in

2009.10 In Yunnan province, which neighbors our study area of Guizhou province, the mortality rate for

insured sows was estimated to be about 2.04-2.6%.11 In our own dataset described below, the sow mortality

rate is 1.92%.

2.1 The Study Area

Our field experiment was conducted in Jinsha County of Bijie Prefecture in Guizhou province. Located

in southwestern China, Guizhou is one of the poorest provinces in China and its economy relies heavily

on natural resources and agriculture. In 2007 the annual per capita net income of farmers was 2,458

Yuan (about 387 US dollars) in Bijie prefecture, and was 2,853 Yuan (449 US dollars) in Jinsha county.

Bijie prefecture has a population of 7.38 million and over 93 percent of its terrain is either highlands or

mountains with poor road conditions.

7Rosenzweig (1988), Rosenzweig and Stark (1989), Rosenzweig and Wolpin (1993), Udry (1994) and Lim and Townsend

(1998), among others, studied the various mechanisms through which villagers achieve the informal risk sharing.8The hog industry in China was valued at 644 billion Yuan in 2006, accounting for 48.4% of the total livestock industry

(Wang and Watanabe, 2007).9There are many causes for pig and sow mortality, including backward breeding technology, weak swine farm infrastructure,

poor vaccination, veterinary drug abuse and natural disasters, such as wind storm, blizzard, thunder, flooding and earthquakes.10See http://www.chinabreed.com/PIG/develop/2009/03/20090326255187.shtml.11These statistics can be found at http://www.chinabreed.com/PIG/develop/2009/03/20090326255187.shtml,

http://www.cangyuan.gov.cn/xxgk/jgzn/xzfzbm/2009-05-25/3188.shtml, and

http://pe.xxgk.yn.gov.cn/canton model2/newsview.aspx?id=1152247.

3

Pig-raising is an important source of income for farmers in our experimental area. In 2007, the market

price of a sow was about 1,500 Yuan while the annual per capita income in the experimental area was 2,853

Yuan. A sow’s market value accounts for over 10% of a household income (since the average household

has about five people). According to an industry analysis in 2009, the return from raising a poker pig was

about 17% around 2007, which was relatively high.12

2.2 The Sow Insurance Program

Infectious diseases led to large fluctuations in pork production and prices in China. For example, in

2003 a bird flu epidemic caused a sharp decline in the production of live pigs; and in the second half of

2006, a fast-spreading deadly blue ear disease brought about another shortage in the pork market, causing

pork price to be more than 60% higher in June 2007 than in June 2006. Due to the importance of pork

in the Chinese diet, the dramatic hike in pork prices led to intense public complaints and concerns about

food-price-driven inflation. As a result, the Chinese government decided to intervene and offer government

subsidy to increase pork supply. One of these government measures was to offer government-subsidized

insurance on sow deaths.13 In July 2007, the Ministry of Finance initiated a plan specifically to subsidize

the insurance of sows raised in the middle and western parts of China. Under the plan, insurance policies

for sows at a coverage of 1,000 Yuan (about 157 US dollars) in the event of death are offered at a total

annual premium of 60 Yuan, including the estimated administrative cost of 40 Yuan per policy.14 However,

the central and local governments combined would pay 80 percent (i.e. 48 Yuan) of the premium, and the

farmer only needs to pay 12 Yuan premium for the insurance. The policy will cover deaths of sows caused

by major diseases, natural disasters, and accidents.15

The Property and Casualty Company (PCC thereafter) of the PICC was designated by the central

government as the sole insurance company to underwrite the subsidized sow insurance and to settle claims.

Local branches of PCC subsequently cooperated with the local Bureau of Animal Husbandry (BAH) to

collect premium payments from pig farmers who decided to insure their sows.

2.3 Animal Husbandry Workers (AHWs)

BAH at county and township levels mobilized various resources to increase the awareness of the subsi-

dized insurance policy and its benefits via local radio and television broadcasting, but the more important

channel for the marketing campaign was through the so-called Animal Husbandry Workers (AHW here-

after). In our study area, every village has exactly one AHW, who works for the BAH on a part-time

contractor basis and is always a village resident. The AHWs serve as the bridge between formal insti-

tutions (specifically, the BAH) and rural villages for matters involving animal husbandry.16 AHWs are

12See Finance and Industry Institute of Northeast Securities (2009).13Other government measures include direct subsidy and low-interest loans for pig farmers.14Recall that the market price for a sow in our study area was around 1,500 Yuan at the time of our study.15Major diseases covered by the insurance are septicemia, blue tongue, scrapie, swine fever, hyopneumoniae, swine erysipelas,

porcine reproductive and respiratory syndrome virus (PRRSV), porcine epidemic diarrhea, streptococcus suis, and foot and

mouth disease. Natural diasters covered by the insurance are typhoon, tornado, rainstorm, lightning stroke, earthquake,

flooding, hailstorm/snowstorm, debris-flow and mountain landslide. Accidents include fire, explosion, building collapse, and

falling parts or articles from aircraft and other flying objects.16Regular obligations of the AHWs include immunization for animals, technical assistance for farmers, and the monitoring

of animal diseases and epidemics.

4

especially important in our study area because the poor transportation infrastructure makes it highly costly

for outsiders to access the villages.

PCC, in cooperation with the BAH, mobilized the AHWs to spread the words about the insurance

policies, to explain to and convince farmers about the policy’s benefits, and to act as a coordinator between

the PCC and the farmers.17 For example, the insurance policy requires that each sow be earmarked with

a unique identification number in order to be eligible for insurance, and the earmark needs to be verified

for claim purposes. The AHWs thus need to count and check all potentially eligible sows in the village

and make earmarks. The farmers are also asked to contact their village AHW to initiate a claim when an

insured sow dies. The AHW is also responsible for making sure that all relevant evidence of the loss is

preserved until the official PCC claim agent arrives to complete the claim process.

The regular income for AHWs usually includes a small fixed wage from the local BAH (which is 15 Yuan

per month in our study area) and fees for services they provide to farmers such as immunization, spay-

ing/neutering and other veterinarian treatments.18 For their participation in the sow insurance campaign,

local branches of PCC paid the AHWs an additional small lump sum to cover their food and transportation

costs. In our field experiment, which we describe in detail in Section 3, we randomly assign the AHWs into

different additional incentives for their performance in terms of the number of sow insurance purchases in

their villages. We should emphasize that when we introduced the experiment in our meetings with the

AHWs, they were simply told that we would offer them additional incentive packages. We left no hint that

we would examine the effect of the insurance or that we would run the experiment again in future.

2.4 Farmers’ Decision Regarding Raising Sows

Around a month after a female piglet is born, farmers have to decide whether it is to be raised for

breeding purposes. Female piglets not for breeding purposes need to be spayed at that time; otherwise, it

turns into a sow at around 6 months of age, when she becomes sexually active and can start breeding.

Not spaying a female piglet (and thus keeping it fertile) involves significant costs and benefits. Spaying

a piglet enables it to grow faster and produces, to most Chinese consumers, better-tasting pork.19 Thus,

on the cost side, not spaying a female piglet (thus turning it into a sow) leads to slower growth, less-tasty

and thus less valuable meat, and higher feeding costs. The benefit for raising sows is to breed piglets which

can be sold in the market. Pregnancy of a sow takes about 4 months (114 days).20 The number of piglets

born in each pregnancy varies, and it typically rises in the first three pregnancies, and reaches the peak

in the 4th to 6th pregnancy. A sow is typically kept for about 4 to 6 years and then slaughtered when

its fertility rate drops to a low level; it is rare for farmers to keep a sow for its whole life span. As we

described earlier, the insurance does not cover a sow’s natural death or if it is slaughtered by the farmer.

Recall that farmers need to contribute 12 Yuan premium toward an insurance policy that pays out

1,000 Yuan in the event of sow death due to the covered risks (which was estimated to occur at about 2%

probability). Thus, due to the government subsidy, the insurance is actuarially favorable for the farmers

for the risks covered by the insurance. However, it should be emphasized that, even with the insurance,

17The local BAHs held training sessions for the AHWs to understand the details of the sow insurance and provide basic skills

for effective promotion and persuasion, such as highlighting the convenience of the insurance purchase and claim settlement.18The average income of the AHWs from the service fees is about 3,000 Yuan per year.19The meat from a sow virtually has no market value; government regulation prohibits it from being sold in the market.20Any unsprayed female pigs older than 3-4 months of age are eligible for the coverage of the insurance policy.

5

farmers are still exposed to significant residual risks when they decide on whether to spay a female piglet

or to turn it into a sow.

First, the risks covered by the insurance, as described in Footnote 15, are only a subset of risks that

may cause sow deaths. For example, the risk that a sow dies in the birth process is not covered. Second,

farmers also face the market risks. The price for piglets – the output of a sow – is volatile. For instance, a

piglet in our study area, Guizhou province, was priced at 22.50 Yuan per kilogram in November 2007 when

our experiment was conducted, and then increased to 33.49 Yuan per kilogram in May 2008, but then it

dropped to 22.75 Yuan per kilogram in November 2008.21 That is, within a year, we see an increase of 49

percent in the piglet price in the first half of the year and then a drop of 32 percent in the second half.

Such market risks are not covered by the insurance.

Third, even ignoring the volatility of the piglet prices, it will not be worthwhile for farmers to seek the

gain of the 8 Yuan subsidy per sow by purchasing the insurance. As mentioned above, turning a female

piglet into a sow in order to be eligible for the insurance leads to a slower growth, less-tasty meat, and

more feeding costs. The pork price was about 28 Yuan per kilogram in Guizhou from November 2007 to

June 2008. The 8 Yuan subsidy can be offset by a mere 0.3 kilogram difference in the quantity of meat

production between a spayed female pig and a sow. A typical female pig weighs about 100-150 kilogram

before being sold in the market .

Finally, we should point out the 8 Yuan subsidy is only received in expectation. Farmers have to pay 12

Yuan premium up-front, and the insurance is valuable only if the sow dies for the covered causes. But as

we mentioned earlier, the market value of the sow is about 1,500 Yuan during the period of our experiment

and a farmer will still suffer significant residual financial loss if an insured sow dies.

3 Experimental Design

3.1 Randomization of Incentive Schemes

The key to obtain a consistent estimate of the effect of insurance on farmers’ production behavior is to

isolate an exogenous source of variation in insurance coverage. In the context of sow insurance as described

in Section 2, our idea is to randomize the assignment of AHWs into different incentive schemes for their

performance measured by the number of sow insurance purchases in their villages. Different incentive

schemes are expected to generate different insurance coverages across villages. Given the randomization,

the difference in incentive schemes across villages should be unrelated to the subsequent number of sows

except for the indirect effect on production through insurance coverage. In our main empirical analysis, we

will indeed use the random incentive assignment as the instrumental variable (IV) for village-level insurance

coverage and identify the causal relationship between insurance and production.

Control Group, Low Incentive Group and High Incentive Group Villages. The local government

of Jinsha County allowed us to run the experiment in 480 villages out of a total 580 villages within its

jurisdiction.22 These 480 villages are located in 27 townships. We randomly assigned the AHWs of the

21The price information is obtained from China Information Network on Husbandry at http://www.caaa.cn.22Based on information from the China Agricultural Census of 2006 (described below in Section 4), there is no systematic

difference in all economic indicators, including pig raising, between the villages in our experimental sample and the 100 left-out

villages.

6

Treatment Groups

Control Group Low Incentive Group (LIG) High Incentive Group (HIG)

Fixed Reward 50 Yuan 20 Yuan 20 Yuan

Incentives None 2 Yuan/Insured Sow 4 Yuan/Insured Sow

Number of Villages 120 120 240

Number of Covered Townships 27 27 27

Table 1: Experimental Design

480 villages into three incentive schemes. The incentives are summarized in Table 1. In the first group

of 120 villages, the AHWs were offered a fixed reward of 50 Yuan (7.87 US dollars) to participate in our

study with no additional incentives. We refer to this group as the control group villages. The AHWs in

the second group of 120 villages are offered a 20 Yuan fixed reward, and an additional payment of 2 Yuan

for each insured sow. We refer to this group as the low incentive group (LIG) villages. In the remaining

240 villages, the AHWs are offered a 20 Yuan fixed reward and an additional payment of 4 Yuan for each

insured sow.23We refer to this group as the high incentive group (HIG) villages.24 As shown in the last

row in Table 1, in each of the 27 townships there are villages assigned to all three experimental groups.

Our choices of the fixed payment and the incentives are very attractive to the AHWs. As we mentioned

in Section 2, PCC offers only a small lump sum payment to AHWs for their involvement in the sow

insurance program; moreover, the regular monthly payment from the BAH to the AHWs is only 15 Yuan.

Time Line of the Study. The time-line of our study is as follows. The government-sponsored insurance

program was initiated in July 2007, but was not implemented in our study area until the beginning of

November 2007. Our incentive experiment ran from November 21 to December 25, 2007.25 Each AHW in

our experimental village was informed about the assigned incentive plan on November 20, 2007 with the

cooperation from the local BAHs. The data on insured sows in each sampled village were collected in the

week immediately after our experiment. After our experiment, we also obtained data of total sows in each

of the villages from the local BAHs collected at two different time points, one as of the end of March 2008

and the other as of the end of June 2008.

3.2 Discussions

3.2.1 Experimental Design

Randomization at the Village vs. Household Level. As described above we implemented our ran-

domization at the village level. An alternative would be to conduct an experiment where the randomization

is at the household level. For example, we may randomly select a set of households and make available to

23We have twice as many HIG villages in our experiment as the control and LIG villages at the insistence of the local Bijie

Prefecture government officials, who a priori believed that the high incentives we offer to the AHWs would lead to more

insured sows in these villages.24Our experimental design is related to the “encouragement design” as described in Duflo, Glennerster and Kremer (2007).

The difference is that in our experiment, the incentives are provided to the AHWs, not to the farmers.25December 25, 2007 was the cut-off date for insurance purchase to be effective from January 1, 2008. Only new sows (that

were not officially registered by the AHWs by December 25, 2007) would be accepted for insurance coverage after this cut-off

date.

7

them the formal insurance option, while withholding such options to the non-selected households. However,

under such an experimental design, it is inevitable that some households in the same village have access

to formal insurance while others do not. It is impractical to refuse to cover those households who were

not offered the insurance option, but learned about it from the neighbors and would like to be insured.

Self-selection by households would contaminate the randomization in the experimental data.

Furthermore, there is another serious shortcoming of household level randomization. As we mentioned

in the introduction, there is substantial evidence that villagers in the same village are likely engaged in

informal risk sharing (see, e.g., Townsend, 1994), randomizing insurance access at the household level may

actually lead to an under-estimate of the true effect of insurance on production, due to the potential risk

shifting from those households without access to formal insurance to those with access.26

For us, randomizing at the village level has the added benefit that we do not have to collect detailed

information about each household because we have the fortunate access to the detailed pre-experiment

village level information from the China Agricultural Census (CAC) conducted in early 2007 (see Section

4.2).

Randomized Incentives to the AHWs vs. Randomized Phase-in. At the village level, a most

obvious alternative research design is randomized phase-in (see, e.g., Miguel and Kremer, 2004).27 We

initially pursued this idea, but the Bijie Prefecture government insisted that preventing some randomly

selected villages from accessing the partially subsidized sow insurance was impractical. We then debated

alternative ways to randomly generate differential access to insurance. We believe that, by randomly

allocating incentives to the AHWs, we can generate de facto differential access to the insurance product

in different villages. Indeed the first-stage result reported in Table 5 below confirmed that incentives we

provided to the AHWs led to substantial differences in the number of insured sows.28

3.2.2 AHWs’ Performance Incentives and Farmers’ Take-up of Sow Insurance

We now describe several channels through which our performance incentives to the AHWs may affect

farmers’ take-up of sow insurance, and discuss several potential threats to the validity of our incentive

assignment as instruments for sow insurance.

Information Channel. Insurance is new to the farmers in this area and they are unlikely to know

how it works. Our performance incentives to the AHWs can lead them to work harder to provide the

farmers with information about insurance in general and the potential benefits of the sow insurance in

particular, thus increasing take-up. It is possible that the AHWs may be led by the performance incentives

to exaggerate the benefits of the sow insurance scheme in order to persuade the farmers to enroll. However,

this type of exaggeration should not directly affect the farmers’ subsequent sow production decisions.

26Angelucci and De Giorgi (2009) and Angelucci et al. (2009) made similar observations in their study of the indirect effects

of Progresa transfers on family members. They show that cash transfers to households eligible for Progresa transfers indirectly

increase the consumption of ineligible households living in the same villages because ineligible households receive more gifts

and loans from eligible housesholds.27Duflo, Glennerster and Kremer (2007) provided a general discussion about different field experiment designs.28With a random phase-in, villages will either have or not have access to the insurance option, the experimental variation

in access to insurance option is restricted to a 0/1 dichotomous variation. In our experimental design, we can in principle

generate a much richer variation in insurance access because we can potentially provide a large variety of incentives to AHWs.

8

Sharing the Performance Incentives with Farmers. The AHWs may offer to share part of

the performance incentive payments to the farmers in order to induce them to enroll in the insurance

program. For example, the AHWs in the high incentive group villages receive 4 Yuan per insured sow. If

the AHW shares a fraction of the 4 Yuan incentive payments to the farmers, more farmers may purchase

the insurance because the kickback reduces the effective price of the insurance. However, this channel

should not directly affect the farmers’ subsequent sow production decisions, except possibly through an

income effect which we assume to be small.29

Building Trust. In the microinsurance setting, farmers are required to pay their insurance premium

up front (although just a small fraction of the entire premium in our case), before receiving any potential

benefit from this policy in the event of a sow death.30 As a result, farmers may be seriously concerned about

whether they are able to get the payment as promised in the insurance contract if covered contingencies

occur, and/or whether the government would continue to offer and subsidize similar sow insurance in

the future.31, 32 Importantly, if a local government fails to deliver its promises in the contract, there is

virtually no way for farmers to sue the government in court. This lack of trust can lead to low take-up

of the subsidized sow insurance. Our performance incentives may lead the AHWs to exert more effort to

improve the farmers’ trust in the insurance scheme, and change their subjective beliefs that their claims

will be honored in the event of a loss. Our evidence in Section 5.4 can be interpreted as evidence for

the importance of trust. This channel should not directly affect the farmers’ subsequent sow production

decisions.

We now briefly discuss several prominent mechanisms that may potentially threat the instrument

validity. We argue that they are unlikely to cause major problems in the interpretation of our estimated

effect of insurance on subsequent sow production reported in Section 5.

The first threat is that higher performance incentives to the AHWs may lead them to promise the

farmers that they will deliver their on-site extension services such as immunization, neutering/spaying,

and veterinary care etc. at lower costs or more expediently, only if the farmers enroll in insurance. In

this channel, high performance incentives to the AHWs lead to higher enrollment in the sow insurance,

but at the same time also directly affects farmers’ subsequent sow production decisions, thus invalidating

the incentive assignments as the IV for insured sows.33 We discuss in Section 5.4 evidence from a rare

unexpected snow storm that occurred right after our experiment, which suggests that this is unlikely an

important channel for our performance incentives to affect insurance enrollment.

A second concern is that the random incentive assignments we gave to the AHWs in our experiment

may directly affect how the numbers of sows in March and June 2008, which we use as the measure for

post-experiment level of production, would be counted. This concern is unwarranted, however, because the

AHW and the local BAH staff counting the number of sows and pigs in each village are typically not the

29Our experimental design does not allow us to distinguish the income effect from the effect of insurance.30In contrast, in micro-credit programs farmers receive money from the government or financial institutions up front. Trust

of the farmers about the government or financial institution is not important for farmers’ participation in micro-credit settings.31The issue of trustworthiness of government policies is particularly relevant in China since governments at all levels often

renege on their policy promises, and from the viewpoint of Chinese farmers, local bureaucrats at townships are always searching

for “legitimate” reasons to ask them for money and sometimes even cheat them into paying unnecessary fees.32Indeed, though we did not anticipate it, the government decided not to offer such sow insurance to farmers in 20008 due

to substantial drop in pork prices during the period of May and November 2008.33Similarly, our IV strategy will be threatened if the villagers perceive the more-aggressively marketed insurance as a sign

that the government would offer other types of incentives related to sow production.

9

same person. Local BAH has its own staff to conduct regular agricultural statistics (including counting

the sow population).34 Moreover, the counting of the numbers of sows in March and June 2008 occurred

after our incentives ended on December 25, 2007.

A third concern is that the AHWs and the farmers may collude to either fleece the insurance company

or extract more bonus payments from us. For example, an AHW and a farmer may fake a spayed pig

as a sow and enroll her in the insurance, and then slaughter the pig but at the same time try to claim

death benefits; or an AHW and a farmer may insure a sow and then fake its death to obtain payment

from the insurance company. These are difficult to implement, however, because the insurance company

always sends claim staff to the scene to verify all evidence; thus for this type of collusion to work, the claim

agent from PCC must also be involved. Another worry is that the AHWs and the farmers may fake the

number of insured sows in order to receive more bonus payments from us. However, as we detail below

in Section 4, the actual numbers of insured sows we use to decide how much bonus an AHW will receive

from us are obtained from the insurance company. Thus the AHW receives bonus payments from us only

if the 12 Yuan premium is paid to the insurance company and when an actual insurance policy is issued.

Because the 12 Yuan premium exceeds even our high bonus amount, this collusion scheme is not profitable.

Importantly, however, any of the potential collusions mentioned above, if they occur at all, would weaken

our instruments and make it harder for us to find the effects of insurance on subsequent sow production.

4 Data

4.1 Data from the Experiment

The data collected from our experiment are at the village level. For each village, we obtained from the

insurance company the total number of insured sows, including the identification number of the insured

sows; we also collected information about a list of the AHW characteristics, including his/her name, age,

gender, education and so on. We also recorded the total payment received by the AHW in each village.

4.2 Other Data Sets

We match the data collected during the experiment with two other data sources: the China Agricultural

Census (CAC) of 2006, and the detailed sow death records and sow productions in 2008 from the local

BAH.

China Agricultural Census of 2006. The China Agricultural Census (CAC) was conducted by the

National Bureau of Statistics of China between January and February of 2007. The CAC covers 250 million

rural households in 640 thousand villages and 35 thousand townships in China, and it collects detailed

information about agricultural production and services in farming, forestry, husbandry and fishery as of

the fourth quarter of 2006.35 We obtained the detailed CAC data for all villages in our study area, Bijie

Prefecture in Guizhou Province. The CAC has several components, including one that is filled out by

34See Articles 5 and 20 in “Regulations for Chinese Agricultural Statistical Data Collection” at web link

http://www.gov.cn/zwgk/2006-08/31/content 374576.htm) for reference. These articles state that the statistical staff member

is typically the head or the accountant of the village.35For more detailed information about this census, see “The Action Plan of the Second National Agricultural Census” at

http://www.stats.gov.cn/zgnypc/.

10

village leaders regarding village characteristics such as registered population, villagers working as migrant

workers elsewhere, total farm land area, basic infrastructure (such as paved road, water treatment facility,

schools, etc.) and village government financial information. The main component of the CAC data,

however, is collected at the household level. Household heads are asked to enter information for every

member of their households. We observe from the household component detailed household information

including how many individuals reside in the household, their relationship to the household head, their age

and gender composition, the amount of contract land, the amount of land in use, ownership of housing,

the self-estimated value of house(s), ownership of durable goods, the availability of electricity, water and

other amenities, the number of household members that receive government subsidies, and engagement in

various agricultural activities including the number of sows and number of pigs raised in the household.

We aggregate up the relevant household data to the village level and then match it, together with the

village component of the CAC, to our experimental data using the unique village identification number

common to CAC and our experimental data.

Data from the Local BAH. We obtained data from the agricultural statistics collected by the local

BAH. In particular, we obtained the counts of the number of sows in each village tabulated by the BAH

at the end of the third and fourth quarters of 2007, as well as the tabulations at the end of the first two

quarters of 2008. We also obtained the sow death records from the BAH. When a sow dies, the village

AHW initiates a call to the insurance company to start the claim process, records the death, and collects

claim evidence, in particular, the number on the ear of the sow that uniquely identifies the sow. The

insurance company, PCC, then sends its claim staff to check and confirm the death and its causes. If the

death is confirmed to be covered by the policy, the company makes compensation payment to the farmer.

The AHW then submits the list of identification ear numbers of the dead sows to the BAH at the township

level and then up to the local BAH.

4.3 Descriptive Statistics

Table 2 presents the basic summary statistics of the key variables used in our analysis, both for the

whole sample and separately by experimental groups. An observation is a village that participated in our

experiment. In the last column of Table 2, we report the p-value for the hypothesis that the means of

the three groups (control, LIG and HIG) are equal. It shows that for all the pre-experimental variables,

the hypotheses that the means are equal across the villages assigned in the three experimental groups

can not be rejected.36 The pre-experiment variables are characteristics of the villages collected before our

experiment period (November 21 to December 25, 2007), mostly from the Chinese Agricultural Census.

The average number of sows in December 2006 was 16.3 across all 480 villages;37 and it was 17.9, 13.2

and 16.9 respectively among the control, LIG and HIG villages. Even though the means are different, a

formal test cannot reject the null that the means of the three groups are equal (with a p-value of 17%).

36We also conducted more formal tests of the quality of randomization underlying our experiment. We regress the probability

of being assigned to one of the three experimental groups on a list of pre-experiment village-level variables using linear

probability, multinomial Probit and and multinomial Logit specifications. These tests overwhelmingly show that none of the

included variables predict the experimental group assignment. The details are reported in the online appendix.37The ratio of the numbers of pigs and sows for a typical village is about 20. Given the fact that a sow typically gives birth

to about 8-12 piglets per litter and can have two litters each year, such a ratio is not too surprising.

11

Wh

ole

By

Gro

up

Sam

ple

Contr

ol

LIG

HIG

p-v

alu

e

Var

iab

les

Mea

nS

DM

ean

SD

Mea

nS

DM

ean

SD

Pre

-Exp

erim

ent

Var

iab

les:

No.

ofS

ows

inD

ec.

2006

16.3

21.4

17.9

26.5

13.2

14.2

16.9

21.5

0.1

7

No.

ofS

ows

inS

ept.

2007

29.1

31.8

28.8

43.1

28.1

20.5

29.8

29.4

0.9

0

No.

ofS

ows

inD

ec.

2007

31.2

34.5

32.5

46.5

26.4

23.7

32.5

31.6

0.4

3

No.

ofP

igs

inD

ec.

2006

356.2

228.4

363.9

248.3

338

.3228.1

361.3

218.4

0.6

1

Vil

lage

Pop

ula

tion

1029.1

677.8

1048.7

654.2

1017.9

672.0

1025.0

694.5

0.9

3

No.

ofV

illa

gers

asM

igra

nt

Wor

kers

196.0

116.4

188.8

127.1

193

.7103.1

200.8

117.5

0.6

4

Ave

.V

illa

ger

Age

33.2

2.1

33.1

2.1

33.4

2.3

33.2

1.9

0.4

5

Ave

.V

illa

ger

Ed

uca

tion

(Yea

rs)

5.9

50.7

56.0

00.7

86.0

00.7

75.9

00.7

30.4

0

Fra

ctio

nM

ale

inV

illa

ge0.5

40.0

30.5

40.0

30.5

50.0

30.5

40.0

20.3

1

Lan

dp

erH

ouse

hol

d(M

u)

4.3

11.9

74.0

91.9

14.2

81.9

54.4

32.0

00.3

1

Log

Hou

seV

alu

e9.8

3.6

39.8

7.6

19.7

5.6

39.8

4.6

30.3

0

No.

ofS

urn

ames

inth

eV

illa

ge5.3

62.6

85.4

13.0

85.1

92.4

75.4

22.5

70.7

2

No.

ofV

illa

gers

inN

ewM

edic

alC

oop

.S

chem

e551.5

300.7

560.3

322.3

532

.1299.2

556.9

291.2

0.7

2

No.

ofH

ouse

hol

ds

Rec

eivin

gG

ov.

Su

bsi

dy

182.2

92.2

183.9

90.4

178

.899.6

183.0

89.5

0.9

0

Pos

t-E

xp

erim

ent

Var

iab

les:

No.

ofIn

sure

dS

ows

22.6

726.8

815.4

710.6

921.

51

19.4

826.8

534.0

30.0

0

No.

ofS

owD

eath

sin

Sn

owS

torm

0.1

90.8

30.0

80.4

00.1

00.4

40.2

81.0

90.0

6

No.

ofS

ows

inM

arch

2008

38.4

34.3

32.5

24.2

35.7

31.4

42.1

38.7

0.1

0

No.

ofS

ows

inJu

ne

2008

42.9

37.7

36.6

26.6

39.

835.8

46.9

42.1

0.1

1

Tab

le2:

Su

mm

ary

Sta

tist

ics

ofK

eyV

illa

ge-L

evel

Var

iab

les.

12

It is interesting to note that the average number of sows across all villages increased by almost 80 percent

from 16.3 in December 2006 to 29.1 in September 2007, right before we conducted our study. In fact, the

average numbers of sows in September 2007 were very close across the three experimental groups, with

means being 28.8, 28.1 and 29.8 respectively for the control, LIG and HIG villages. The number of pigs in

each village is about 350 in December 2006. The average population is about 1,000 with an average age of

about 33, and about 20% of the villagers work elsewhere as migrant workers.

Table 2 also reports the summary statistics of several post-experimental variables. The average number

of insured sows across the villages is 22.67. If we use the number of sows in September 2007 as the

actual number of sows eligible for insurance, the aggregate take-up rate is about 78%. However, there is

substantial variation in insurance take-up rates across the experimental groups. Among the control group

villages, an average of 15.47 sows out of an average of 28.8 available sows were insured; in LIG villages, an

average of 21.51 sows out of an average of 28.1 available sows were signed up for insurance; and in HIG

villages, an average of 26.85 out of 29.8 sows were signed up for insurance.

We also report the number of sow deaths during the deadly snow storm between January 12 and

February 25, 2008 (see Section 5.4 for details). On average, 0.19 sows died in one village. However, the

average number of sow deaths in the control villages is 0.08, lower than that of 0.10 in LIG villages and

that of 0.28 in HIG villages. This is consistent with possible moral hazard among farmers with insured

sows. In addition, the last two rows of Table 2 show that the number of sows in March 2008 and June

2008 continues to rise from the levels in September 2007.

4.4 Test for Parallel Trend Between the Control and Treatment Villages

Our IV estimator below of the effect of insurance access on production also relies on the assumption

that there is no systematic difference in the trend of sow production between control and treatment villages.

In Figure 1, we plot the average number of sows by the three experimental groups for the entire period

for which we have data. It shows that prior to our experimental intervention, the number of sows grew

similarly in the three groups, especially between the control and HIG, but afterwards the HIG grew more

quickly.38

5 Results on the Effect of Insurance on Production

In this section, we report the results on the effect of insurance on subsequent sow production. We

first report in Section 5.1 the Ordinary Least Square (OLS) results where we regress the number of sows

measured in March 2008 and June 2008, about three and six months after our experiment respectively, on

the number of sows insured during our experimental period. However, in order to estimate a causal effect of

insurance on production, one needs to exploit some exogenously induced variations in insurance coverage.

In Section 5.3, we use the random experimental group assignment as instruments for the number of insured

38We also conducted a more formal test of the parallel trend assumption where we regress the number of sows measured at

the different points in time on time dummies and the interaction of the treatment dummies and the time dummies. The formal

test results suggest that there is a parallel trend between the control and treatment group villages before our experiment, but

there is a different trend between the control and the HIG groups after our experiment. The details are reported in the online

appendix.

13

8

13

18

23

28

33

38

43

48

53

Dec‐06

Feb‐07

Apr‐07

Jun‐07

Aug‐07

Oct‐07

Dec‐07

Feb‐08

Apr‐08

Jun‐08

Control

LIG

HIG

Figure 1: The Time Trend of the Number of Sows for Different Incentive Groups.

sows in order to recover the overall causal effect of insurance access on subsequent sow production at the

village level.39, 40

5.1 Results from the OLS Regressions

Table 3 reports results from the following OLS specifications:

Yi = α0 + α1Insured Sowsi + α2Sows2006i + Township Dummies + εi, (1)

where Yi represents the number of sows in village i measured in March 2008 (Panel A) or June 2008

(Panel B), “Insured Sows” represents the number of insured sows in village i by the end of the fourth

quarter of 2007, and “Sows2006” represents the number of sows measured in December 2006, and a set of

Township dummies are included in some specifications in order to control for the effects of township-specific

characteristics on sow-raising.41 Robust standard errors are reported in parenthesis in Table 3.

39Since the goal of the subsidized sow insurance was to increase pork production, it is useful to mention that there is a

strong positive correlation between the number of sows and the number of slaughtered porker pigs. Using data from China

Husbandry Yearbook in 2008-2011, we find that a 1 percentage-point increase in the number of sows is associated with a 1.22

percentage-point increase in the number of the slaughtered porker pig.40The overall effect we estimate is the sum of both the extensive or the intensive margins, where the extensive margin refers

to the previous sow farmers increasing the number of sows, and the intensive margin refers to the production of new sow

farmers. We do not have the necessary household level data to distinguish the two margins.41We choose to include the number of sows measured in December 2006 instead of in September 2007 for two reasons. The

pork price spike occurred in early 2007 somewhat unexpectedly, so the sows in December 2006 were raised without the effect

from the pork price spike. Second, using sows measured at December 2006 also mitigates the effect of potential behavioral

change in anticipation of the possible government subsidized sow insurance. However, if we were to replace Sows2006 by

Sows2007, the estimated coefficient on Insured Sows barely changes both qualitatively and quantitatively. Results are

available from the authors upon request.

14

Panel A: Number of Sows in March 2008 Panel B: Number of Sows in June 2008

Variables (1) (2) (3) (4) (5) (6)

No. of Insured Sows1.215***

(.110)

1.239***

(.078)

1.093***

(.150)

1.323***

(.088)

1.314***

(.090)

1.158***

(.169)

No. of Sows in

Dec. 2006

.298*

(.164)

.318*

(.190)

Constant11.14***

(2.768)

16.79***

(3.906)

15.91***

(3.604)

13.21***

(1.911)

17.91***

(4.276

16.96***

(3.99)

Township Dummies No Yes Yes No Yes Yes

Adjusted-R2 .67 .78 .79 .66 .76 .78

Table 3: OLS Regression Results on the Relationship between Sow Insurance and Subsequent Sow Pro-duction.Notes: (1) Robust standard errors are in parentheses; (2) *, **, *** denote significance at 10%, 5% and 1%, respectively.

In order to examine the causal effect of the access to the formal insurance on sow ownership, ideally we

need a proper measure of insurance access. However, such a measure of insurance access is unavailable in

our data set. As a result, we use the number of insured sows in each village as a proxy for accessibility.42

The number of insured sows contains information on availability, but it may be contaminated by other

factors. In the next section, we will report the reduced-form effects of the treatments both on insured sow

ownership and total sow ownership

Focusing on the specifications with controls of both Township dummies and Sows2006 reported in

Column (3) of Panel A and Column (6) of Panel B, we see that insuring one more sow in the fourth

quarter of 2007 is associated with 1.093 more sows raised in March 2008 and 1.158 more sows in June 2008,

after controlling for the number of sows in the village at the end of 2006 and the Township dummies. Both

coefficient estimates are strongly statistically significant with p-value close to 0.

However, the variation in the number of insured sows used in the above OLS regressions includes not

only the exogenous variation induced by the randomly assigned AHW incentives, but also endogenous

variations across villages that may result from selection on unobserved heterogeneity across villages. Thus

the OLS estimate cannot be interpreted as causal effects of insurance on subsequent production. For

example, it could be that a village where more farmers are contemplating raising more sows are more

likely to purchase sow insurance when such option is presented. This would lead to an upward bias in the

estimated effect of insurance on production.

5.2 Results from Reduced-Form Regressions

Table 4 reports the reduced-form regression results, examining how the treatment dummies affect the

subsequent sow productions. We see that high-incentive group villages have higher level of sows in both

March and June of 2008 than low-incentive groups, and two incentive groups generate more sows than the

42In order to measure the access or take-up of the insurance plan, an alternative and seemingly natural measure will be the

percentage of sows insured. However, we have controlled for the number of sows in the end of 2006 for all regressions, and

the estimated coefficient on the number of insured sows should be interpreted as the effect of the insurance on the subsequent

number of sows raised at the village. For robustness check, we also used percentage of sows insured over the sows in 2006 as

the measure for the access, the basic results remain qualitatively similar.

15

Panel A: Number of Sows in March 2008 Panel B: Number of Sows in June 2008

Variables (1) (2) (3) (4)

Low Incentive Group4.895

(4.478)

7.110**

(3.477)

5.146

(5.055)

7.497*

(4.043)

High Incentive Group9.196**

(3.894)

7.491**

(3.182)

9.876**

(4.283)

8.066**

(3.538)

No. of Sows in Dec. 20061.120***

(.152)

1.189***

(0.166)

Constant20.954***

(5.262)

13.280***

(4.175)

22.236***

(6.270)

14.089***

(5.041)

Township Dummies Yes Yes Yes Yes

Adjusted-R2 .274 .546 .293 .547

Table 4: The Effect of Incentive Group Assignments on Subsequent Sow Production: Reduced-FormResults.Notes: (1) Robust standard errors are in parentheses; (2) *, **, *** denote significance at 10%, 5% and 1%, respectively.

control group villages. The coefficients on the two incentive groups, reported in Columns (2) and (4), are

both statistically significant at least 10 percent level.

5.3 Results from IV Regressions

In order to identify the causal effect of insurance coverage on sow production, we need to isolate the

exogenous variation in insurance access induced by the randomly assigned incentives we provide to the

AHWs. In this section, we use the experimental group assignment as instruments for insured sows in

estimating regression equation (1).

First-Stage Results A valid instrument variable for Insured Sows in Equation (1) requires that it be

orthogonal to the error term ε and that it be significantly correlated with Insured Sows when all other

relevant independent variables are controlled. Since the assignments of experimental group to villages were

random and should be unrelated to the sow production at the village level, as demonstrated in Table 2,

the first requirement for experimental group assignment as valid IVs for Insured Sows is automatically

satisfied.43

Now we report the first-stage result that shows that the second requirement for valid IVs is also satisfied.

Table 5 reports the result from regressing the number of insured sows on the experimental groups (Low

Incentive Group and High Incentive Group, with the Control group as the default category), controlling

for the number of sows measured in December of 2006 and a set of Township dummies.44 Overall we find

a very strong and significant effect of AHW performance incentive assignment on the insurance coverage.

According to the estimates in the preferred specification (Column 3), moving from the control group with

43Indeed Hansen’s J-statistics from the IV regression reported in Table 6 is only 0.068; thus the over-identification test does

not reject the null that all instruments are valid (with a p-value of 0.7938).44We have also run speficications with the number of sows measured in September 2007 as additional controls. The coefficient

estimates on the group assignments do not change, both qualitatively and quantitatively. These results are available from the

authors upon request.

16

Number of Insured Sows

Variables (1) (2) (3)

Low Incentive Group6.075***

(2.027)

6.606***

(2.072)

9.622***

(2.064)

High Incentive Group11.433***

(2.405)

11.336***

(2.687)

12.007***

(2.418)

No. of Sows in Dec. 2006.670***

(.189)

Constant15.433***

(.977)

18.538***

(4.268)

2.134

(5.778)

Township Dummies No Yes Yes

Adjusted-R2 .031 .20 .46

Table 5: The Effect of Group Assignments on the Number of Insured Sows: First-Stage Results.Notes: (1) Robust standard errors are in parentheses; (2) *, **, *** denote significance at 10%, 5% and 1%, respectively.

fixed compensation to the low incentive treatment group results in nearly 9.6 additional insured sows.

Since the sample mean of the insurance coverage is 22.6, the increase of 9.6 sows represents about 43%

of the sample mean, which is an economically significant effect. Moreover, as expected, we find that this

incentive effect is stronger for the high incentive group. When Township dummies are included, the whole

set of independent variables can explain more than 45 percent of the total variation in the number of

insured sows.45

Second-Stage Results. Table 6 reports the second-stage regression results. It shows that when we only

use the exogenous variation in insurance coverage induced by the variations in the random assignment of

incentives to AHWs, the estimated effects of insurance on subsequent production are smaller than those

from the OLS regression reported in Table 3; but nonetheless, the effects of the number of insured sows in

the fourth quarter of 2007 on the number of sows measured in March and June 2008 are both statistically

and economically significant. In the preferred specifications, Columns (3) and (6), one additional insured

sow in the fourth quarter of 2007 increases the number of sows by 0.78 by March 2008 and by 0.83 by June

2008. Both coefficient estimates are significant at 1% level.

These effects are very large. From the first-stage result reported in Table 5, we know that the low and

high incentive group villages insured about 9.6 and 12.0 more sows, respectively, than the control group

villages. These increases in insured sows, according to the estimates in Table 6, led to about 7.5 and 9.4

more sows being raised by March 2008 in the low and high incentive group villages respectively than in the

control group villages.46 Note from Table 2, however, the actual difference in the number of sows in March

2008 between the low incentive group villages and the control group villages is only 3.2, suggesting that if

the extra incentives to the AHW workers were not provided in the low incentive villages, there would have

been 4.3 fewer sows in these villages than in the control villages because, after all, the control villages had

45The partial R2 from the first-stage is about 0.0625, and the first-stage F -statistics for the significance of the excluded

instrument is 13.49.46Note that these magnitudes are remarkably close to the reduced-form impact of the incentive treatment group on the

number of sows in March and June of 2008, as reported in Table 4.

17

more sows in both December 2006 and September 2007. The difference in the number of sows between the

high incentive group villages and control group villages should be understood analogously.

It is worth mentioning that there seems to be some suggestive evidence that the effect of insurance on

the number of sows is larger when the number of sows was measured in June 2008 (6 months after the

insurance was provided) than in March 2008. This is true both for the OLS result in Table 3 and the IV

results in Table 6. This most likely reflects the fact that farmers have had longer time by June 2008 to

react to the sow insurance by turning young female piglets into sows.

Making Sense of the Effect of Insurance on Sow Production. The effect of sow insurance on

the number of sows we estimated in Table 6 is likely the combination of two types of actions by the

farmers: first, the farmers may increase the total number of sows by purchasing additional fertile female

pigs; second, the farmers keep as sows some of their own female piglets that would otherwise be spayed.47

Our experiment took place in December 2007, and the first post-experiment measurement of the number

of sows occurred in March 2008, only 3 months later. Our results show a large increase in reported sow

numbers by this time. However, according to the time-line of sow production described in Section 2.4, it

takes about six months to raise a sow. There was not enough time to have such a large natural increase

in the number of sows due to the insurance. The only way to rationalize such a large increase would be

that there were a lot of piglets close to one month in age right in December 2007 available for farmers to

turn into sows. Although we don’t have data on the number of young female piglets per village at the

end of 2007, a back-of-the-envelope calculation indicates that it is plausible to see about 10-sow difference

between control and treatment groups reported in Table 5. As shown in Table 2, there were about 356

pigs and 16 sows per village in December 2006, and 29 sows per village in September 2007. Suppose the

sow-to-pig ratio is constant over time, then there would be 645 pigs in September 2007. Even without any

growth of pigs from September to December in 2007, there should be around 322 female pigs per village.

Farmers typically raise meat pigs to be around 6-8 months old (12 months at most) and then sell them

to the market. If the distribution of female pigs at different ages is uniform, the number of young female

piglets around one-month age in December 2007 was at least 26.

5.4 The Effect of An Unexpected Ice and Snow Storm

In this section, we briefly describe some additional evidence from an unexpected, and severe, ice and

snow storm, which occurred in early 2008, just a month and a half after our field experiment. The ice and

snow storm hit southern and southwestern China, and Guizhou was one of the most affected provinces.

This storm began in mid-January and lasted for a month, and its scope and severity were unprecedented

in at least the last fifty years. Since snow storms in general are rare in this part of China, let alone one

with such severity, many sows and pigs died during the snow storm especially for those sows raised in the

backyard of village households which lacked necessary facilities.48 News report indicated that there were

a total of 5,973 sows that died during the storm in Guizhou province.49 The ice and snow storm of such a

47Unfortunately, we are unable to disentangle them because we do not have data on the total number of female pigs.48According to Wang and Watanabe (2007), summer months are the most deadly months for pigs in general.49See the newsreport on Xinhuanet (Feburary 20, 2008), “Guizhou Fully Made the Compensation for the Insured Sows

Died during the Snowstorm”, at www.gz.xinhuanet.com/xwpd/2008-02/20/content. While this snowstorm caused a lot of sow

deaths in Guizhou province, the damage on sows in our experimental area was somehow modest. As shown in Table 2, the

18

Number of Sows in March 2008 Number of Sows in June 2008

Variables (1) (2) (3) (4) (5) (6)

No. of Insured Sows.904***

(.241)

.854***

(.252)

.775***

(.244)

.962***

(.266)

.916***

(.281)

.829***

(.279)

No. of Sows in

Dec. 2006

.537***

(.206)

.566**

(.230)

Constant18.116***

(5.530)

41.106***

(14.595)

2.933

(3.534)

21.316***

(6.090)

53.740***

(16.180)

2.760

(3.728)

Township Dummies No Yes Yes No Yes Yes

Adjusted-R2 .624 .73 .77 .61 .72 .76

Table 6: IV Regression Results on the Effect of Sow Insurance on Subsequent Sow Production.Notes: (1) Robust standard errors are in parentheses; (2) The Instruments for the No. of Insured Sows are the group

assignments; (3) *, **, *** denote significance at 10%, 5% and 1%, respectively.

scale was totally unexpected, and the amounts of snowfall in the villages in our sample were uncorrelated

with the village characteristics.

In the online appendix, we report results from IV regressions similar to those in Table 6 except that

we now add the interaction of “No. of Insured Sows” and “No. of Sow Deaths in Snow Storm,” using

our experimental group assignments as the instruments for “No. of Insured Sows,” and instrument for

the interaction terms involving “No. of Insured Sows” by the corresponding interaction terms with our

experimental group assignments. Interestingly, we find that the coefficients on the interaction terms are

positive and significant at least at 15% significance level.

We consider this positive interaction effect between the sow death in the storm and insurance as sup-

portive evidence that trust plays an important role in the farmers’ insurance take-up decisions. As we

mentioned in Section 3.2.2, when farmers do not have complete trust on whether the insurance product is

genuine, the insurance policy itself becomes a risk. Nothing is more convincing to the villagers that the

government subsidized sow insurance is for real than actually paying out the promised damage compen-

sation in this unusual ice and snow storm.50 In villages with more sow deaths and more insured sows,

there would be more positive cases that the sow insurance contracts are honored. Such positive cases of the

insurance contracts being honored would raise the villagers’ trust for the sow insurance program. Thus this

mechanism will predict that the effect of sow insurance for subsequent sow production should be stronger

in such villages.51, 52

number of sow deaths per village is 0.19, which accounts for 0.7 percent of the number of sows per village in September 2007.50Indeed, as reported by Xinghua News Agency and Financial Times (Chinese), following government directives, the insur-

ance company quickly dispatched work teams to remote villages to deal with claim evaluations and settlements. See “Guizhou

Province Made Full Compensations on Lactating Sows Which Were Insuranced and Died of the Ice and Freeze Storm,”

Xinghua News Agency, Feburary 20, 2008; and “Insurance Industry Meets with the Ice and Freeze Disaster in the Special

Way” (Financial Times, Chinese, Feburary 27, 2008).51In Table 7 of Cai et al. (2009), we also document that the number of insured sows is significantly associated with the

coverage of new rural cooperative medical schemes in the village as well as the coverage of government subsidy in the village,

further suggesting the importance of trust for the government in insurance take-up. Details are in the online appendix.52Our examination of the role of trust for the villagers’ demand of insurance echoes that of Cole et al. (2013b) in their study

of rainfall insurance. We should emphasize that in their setting the rainfall insurance was offered by a for-profit insurance

company without premium subsidy. Thus, the trust examined in their setting is the trust for insurance products offered

19

The positive interaction effect between the sow death in the storm and insurance also casts doubt on

the hypothesis that the effect of insurance coverage is mainly driven by the channel of lowering the future

cost of veterinary services in exchange for the farmers’ insurance enrollment (see Section 3.2.2). Under

such a mechanism, we would see the positive effect of insurance coverage on sow production, but we should

not see any significant effect of the interaction between the severity of the storm and insured sows.

5.5 Longer Run Effects and Spillover on Other Livestock

In this subsection, we present some additional results regarding the longer run effects of the sow

insurance on sow production and the spillover effects of sow insurance on other livestock. For this purpose,

we collected additional village-level data on the numbers of sows and pigs measured in December 2009 and

December 2010, as well as the numbers of sheep and cows. Because the numbers of sheep and cows were

collected less frequently than that of sows and pigs, we only have their information for December 2008.

Table 7 presents IV regression results of the number of sows, pigs, sheep and cows on the number of

insured sows during our experiment period (the end of 2007), after controlling for the corresponding stocks

at the end of December 2006 from China Agricultural Census (see Section 4). In Columns (1) and (2),

we find that the number of insured sows by December 2007 has a positive effect on the number of sows in

2009 and 2010, though the effect is significantly only at 20% level. However, if we take the average number

of sows measured at the end of the three year period 2008-2010, we find in Column (3) that the number

of insured sows in 2007 has a positive effect that is statistically significant at 1% level. We similarly find

in Columns (4)-(5) that the variation in the number of insured sows induced by our experimental group

assignment is positively related to the number of pigs (for pork production) in 2009, 2010, as well as the

average number of pigs in the three-year period of 2008-2010. These effects are significant at 15% to 20%

level. In Columns (7) and (8), we also find that the number of insured sows has a positive, but generally

insignificant effect on the number of sheep and cows in the villages. These results suggest that the positive

and significant short-run effect of the number of insured sows on sow production, as reported in Table 6,

seems to have some persistence in the longer run; moreover, the increase in sow production in response to

the sow insurance does not seem to be the result of the substitution of other livestock such as pigs, sheep

or cows. In fact, if anything, the spillover effects of sow insurance on the production of other livestock

seem to be positive, though the effects are only marginally significant.

6 Conclusion

In this paper, we report results from a large randomized field experiment that evaluates the effect

of microinsurance on subsequent production. Making use of a heavily-subsidized insurance for sows, we

randomize the incentive schemes offered to AHWs, which generates plausible exogenous variations in the

effective insurance access across 480 villages in our experimental sample. This allows us to use the random

incentive scheme assignment as the instrumental variable for insurance access to recover the causal effect of

insurance access on production. Our results indicate that promoting greater adoption of formal insurance

significantly leads to subsequent increase in sow-raising. To the best of our knowledge this is one of the

commercially, while in our setting the trust is for government-sponsored, partially subsidized insurance products. We should

also note that we did not randomize trust in our experimental design, while Cole et al. (2013b) did in their study. It is

interesting to note that our evidence strongly corroborates their findings.

20

Sow

sP

igs

Sh

eep

Cow

s

2009

2010

Ave

.2008-1

02009

2010

Ave.

2008-1

02008

2008

Var

iab

les

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

No.

ofIn

sure

dS

ows

.291

†

(.20

6)

.291

†

(.207)

.452***

(.175)

1.2

84†

(.950)

1.4

86†

†

(1.0

07)

1.3

46†

(.956)

.592†

†

(.373)

.490

(.497)

No.

ofS

ows

inD

ec.

2006

.164

(.19

1)

.165

(.19

0)

.288*

(.151)

No.

ofP

igs

inD

ec.

2006

-.079†

(.059)

-.010†

†

(.061)

-.085

†

(.059)

No.

ofS

hee

pin

Dec

.20

06-.

011

(.024)

No.

ofC

ows

inD

ec.

2006

-.331††

(.203)

Con

stan

t14

.514

***

(4.1

82)

13.9

35***

(3.8

11)

10.4

60***

(3.1

67)

540.1

8***

(28.7

5)

545.9

0***

(28.9

6)

538.7

15***

(19.6

0)

33.2

91

(35.1

7)

184.8

9***

(16.5

7)

Tow

nsh

ipD

um

mie

sY

esY

esY

esY

esY

esY

esY

esY

es

Ad

just

ed-R

2.4

94.4

86

.612

.851

.782

.801

.486

.674

Tab

le7:

IVE

stim

ate

sof

the

Lon

ger-

run

Eff

ects

ofS

owIn

sura

nce

onS

owan

dP

igP

rod

uct

ion

and

the

Sp

illo

ver

Eff

ects

onO

ther

Liv

esto

cks

(Sh

eep

an

dC

ows)

.Notes:

(1)

Robust

standard

erro

rsare

inpare

nth

eses

;(2

)T

he

inst

rum

ents

for

“N

o.

of

Insu

red

Sow

s”are

the

gro

up

ass

ignm

ents

,and

the

inst

rum

ents

for

the

inte

ract

ion

term

sin

volv

ing

“N

o.

of

Insu

red

Sow

s”are

the

corr

esp

ondin

gin

tera

ctio

nte

rms

wit

hth

egro

up

ass

ignm

ents

;(3

)***,

**,

*,††

,†

den

ote

signifi

cance

at

1%

,5%

,10%

,15%

and

20%

resp

ecti

vel

y.

21

first large-scale experimental study on the effect of microinsurance on farmer production behavior. Our

finding suggests that microinsurance may be as important as microfinance in increasing production, and

microinsurance can supplement and strengthen the effects of microfinance by protecting the farmers from

the inherent risk of entrepreneurial activities.53 Our experimental design, as well as corroborating evidence

from the effect of an unexpected snow storm, suggests that the documented correlation between the sow

insurance coverage and the subsequent sow production may be causal. We also find some suggestive

evidence that the positive and significant short-run effect of sow insurance on sow production seems to

have some persistence in the longer run; moreover, the increase in sow production in response to the

sow insurance does not seem to be the result of the substitution of other livestock such as pigs, sheep or

cows. Our evidence from the effect of an unexpected snow storm reported in Section 5.4 also suggests that

trust, or lack thereof, for government-sponsored insurance products acts a significant barrier for farmers’

willingness to participate in the insurance program. This finding is consistent with those of Cole et al.

(2013b). We believe that overcoming the issue of the lack of trust should be a crucial consideration in the

next wave of microinsurance revolution. It is important to note that in this study we are unable to address

several important questions due to data limitations. First, how would providing access to microinsurance

complement the effectiveness of micro-credit programs?54 Second, how does the increase in sow production

affect the farmers’ well-beings as measured by consumption, for example? There are fruitful avenues for

future research.

References

Angelucci, Manuela and Giacomo De Giorgi, “Indirect Effects of an Aid Program: How do Cash

Transfers Affect Non-Eligibles Consumption,” American Economic Review, 2009, 99, 486–509.

, , Marcos Rangel, and Imran Rasul, “Insurance and Investment Within Family Networks,”

mimeo, 2009.

Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan, “The Miracle of

Microfinance: Evidence from a Randomized Evaluation,” MIT Working Paper, 2010.

Cai, Hongbin, Yuyu Chen, Hanming Fang, and Li-An Zhou, “Microinsurance, Trust and Economic

Development: Evidence from a Randomized Natural Field Experiment,” NBER Working Paper 15396,

2009.

Cai, Jing, “Social Networks and the Decision to Insure: Evidence from Randomized Experiments in

China,” Working Paper, University of Michigan, 2013.

53It is important to note, as we mentioned in Section 2.4, that the sow insurance studied in our paper only insures farmers

against loss from the death of the sows. To the extent that the value of raising sows also depends on the price of the pork and

thus the price of piglets, whose fluctuations are not insured by the sow insurance considered in this study, even the farmers

who purchased the sow insurance still face significant risks. We would expect that the effect of insurance on sow production

will be even larger if the farmers also have access to additional insurance against price fluctuations.54Gine and Yang (2007) study whether the provision of rainfall insurance induces farmers to borrow to invest in new varieties.

Karlan, Kutsoati, McMillan and Udry (2010) study how farmer risk aversion affects investment decisions, and whether a loan

product with an insurance component encourages farmers to take and benefit from credit.

22

Cole, Shawn, Xavier Gine, and James Vickery, “How Does Risk Management Influence Production

Decisions? Evidence from a Field Experiment,” Harvard Business School Working Paper 13-080, 2013.

, , Jeremy Tobacman, Petia Topalova, Robert Townsend, and James Vickery, “Barriers to

Household Risk Management: Evidence from India,” American Economic Journal: Applied Economics,

2013, 5 (1), 104–135.

de Aghion, Beatriz Armendariz and Jonathan Morduch, The Economics of Microfinance, The

MIT Press, Cambridge: MA, 2005.

Duflo, Esther, Rachel Glennerster, and Michael Kremer, “Using Randomization in Development

Economics Research: A Toolkit,” in Handbook of Development Economics, Vol. 4, Elsevier-North Hol-

land, 2007, pp. 3895–3962.

Finance and Industry Institute of Northeast Securities, “A Research Report on the Pig-Raising

Industry (in Chinese),” http://wenku.baidu.com/view/ec642e1ffc4ffe473368ab15.html, 2009.

Gine, Xavier and Dean Yang, “Insurance, Credit and Technology Adoption: Field Experiment Evi-

dence from Malawi,” Policy Research Working Paper Series 4425, World Bank, 2007.

, Robert Townsend, and James Vickery, “Patterns of Rainfall Insurance Participation in Rural

India,” The World Bank Economic Review, 2008, 22 (3), 539–566.

Goodwin, Barry K., Monte L. Vandeveer, and John L. Deal, “An Empirical Analysis of Acreage

Effects of Participation in the Federal Crop Insurarance Program,” American Journal of Agricultural

Economics, 2004, 86(4), 1058–1077.

Harrison, Glenn W. and John A. List, “Field Experiments,” Journal of Economic Literature, 2004,

42 (4), 1009–1055.

Horowitz, John K. and Erik Lichtenberg, “Insurance, Moral Hazard, and Chemical Use in Agricul-

ture,” American Journal of Agricultural Economics, 1993, 75, 926–935.

Innovations for Poverty Action, “Addressing Agricultural Risks: Review of Ongoing Microinsurance

Field Experiments,” mimeo, 2009.

Karlan, Dean, Edward Kutsoati, Margaret McMillan, and Christopher Udry, “Crop Price

Indemnified Loans for Farmers:A Pilot Experiment in Rural Ghana,” Journal of Risk and Insurance,

2010, forthcoming.

, Robert Osei, Isaac Osei-Akoto, and Christopher Udry, “Agricultural Decisions after Relaxing

Credit and Risk Constraints,” Working Paper, Yale University, 2013.

Lei, Xiaoyan and Wanchuan Lin, “The New Cooperative Medical Scheme in Rural China: Does More

Coverage Mean More Service and Better Health?,” mimeo, 2008.

Lim, Youngjae and Robert Townsend, “General Equilibrium Models of Financial Systems: Theory

and measurement in Village Economies,” Review of Economic Dynamics, 1998, 1, 59–118.

23

List, John A, “Field Experiments: A Bridge between Lab and Naturally Occurring Data,” The B.E.

Journal of Economic Analysis & Policy, 2006, 6 (2 - Advances), Article 8.

Miguel, Ted and Michael Kremer, “Worms: Identifying Impacts on Education and Health in the

Presence of Treatment Externalities,” Econometrica, 2004, 71(1), 159–218.

Mobarak, A. Mushfiq and Mark Rosenzweig, “Selling Formal Insurance to the Informally Insured,”

Yale Economic Growth Center Discussion Paper No. 1007, 2012.

Morduch, Jonathan, “Income Smoothing and Consumption Smoothing,” Journal of Economic Perspec-

tives, 1995, 9 (3), 103–114.

, “Micro-insurance: The Next Revolution.,” in Abhijit Banerjee, Roland Benabou and Dilip Mookherjee

ed., What Have We Learned About Poverty? Oxford University Press, 2006.

O’Donoghue, Erik, Nigel Key, and Michael J. Roberts, “Does Risk Matter for Farm Businesses?

The Effect of Crop Insurance on Production and Diversification,” mimeo, Economic Research Service,

USDA, 2007.

Robinson, Marguerite S., The Microfinance Revolution: Sustainable Finance for the Poor., Washington,

D.C.: World Bank, 2001.

Rosenzweig, Mark., “Risk, Implicit Contracts and the Family in Rural Areas of Low-Income Countries.,”

Economic Journal, 1988, 98, 1148–1170.

Rosenzweig, Mark and Kenneth Wolpin, “Credit Market Constraints, Consumption Smoothing and

the Accumulation of Durable Production Assets in Low-Income Countries: Investment in Bullocks in

India,” Journal of Political Economy, 1993, 101, 223–244.

and Oded Stark, “Consumption-Smoothing, Migration and Marriage: Evidence from Rural India,”

Journal of Political Economy, 1989, 97, 905–926.

Townsend, Robert, “Risk and Insurance in Village India,” Econometrica, 1994, 62, 539–592.

Udry, Christopher, “Risk and Insurance in a Rural Credit Market: An Empirical Investigation in

Northern Nigeria,” Review of Economic Studies, 1994, 61, 495–526.

Wang, Jimin and Mariko Watanabe, Pork Production in China: A Survey and Analysis of the Industry

at a Lewis Turning Point, Institute of Development Economics, Japan., 2007.

Yunus, Muhammad, Banker to the Poor: The Autobiography of Muhammad Yunus, Founder of the

Grameen Bank, New York: Oxford University Press, 2001.

24

Online Appendix:The Effect of Microinsurance on Economic Activities: Evidence from a Randomized Field

Experiment

Hongbin Cai

Yuyu Chen

Hanming Fang

Li-An Zhou

In this appendix, we include some details omitted in the main text. Appendix A presents the formal test

of the quality of randomization referenced in Footnote 36; Appendix B presents the regression based formal

test of the parallel trend between the control and treatment villages referenced in Footnote 38. Appendix

C collects the omitted details of our analysis using the unexpected snow storm, which we referred to in

Section 5.4 in the main text. Appendix D contains the details of Table 7 in our NBER Working Paper No.

15396 which we referenced in Section 5.4.

A Test of Randomization

In Table 2 we showed that for almost all the pre-experiment variables, their means are equal across the

villages assigned in the three experimental groups. Table A reports a more formal test of the quality of

randomization underlying our experiment. It regresses the probability of being assigned to the three exper-

imental groups on a list of pre-experiment village-level variables. We report the coefficient estimates from

the linear probability model, as well as the multinomial Probit and Logit models. Table A overwhelmingly

shows that none of the included variables predict the experimental group assignment. In the whole Table

A, which reports 72 coefficient estimates, only two are marginally significant at 10% level. Also, note that

the adjusted R2 for the linear probability models and the pseudo-R2 for the Logit model are both less than

0.017, suggesting that the experimental group assignments are very much random. We also run a different

regression (without reporting) with each baseline covariate as dependent variable and treatment dummies

as explanatory variable (standard errors are clustered at the township level), and find the similar result:

the experimental treatments are orthogonal to the pre-test characteristics of villages.

B Test for Parallel Trend Between the Control and Treatment Villages

Our IV estimator below of the effect of insurance access on production also relies on the assumption that

there is no systematic difference in the trend of sow production between control and treatment villages.

Here we report a formal test of this parallel trend assumption. We have the information about the

number of sows measured at three different time points (December 2006, September 2007 and December

2007) before our experimental treatment and at two different time points (March 2008 and June 2008)

after our experimental treatment.1 Table B reports the fixed effect regression results when we regress

the number of sows measured at the different points in time on time dummies and the interaction of the

1The sow counts for December 2007 were collected right when our experiment was conducted. As such, we treat it as the

pre-experiment counts.

1

Linear Probability Probit Logit

Variables LIG HIG LIG HIG LIG HIG

No. of Sows in Dec. 2006-.0017

(1.685)

.0004

(.24)

-.0110

(1.59)

-.0027

(.50)

-.0149

(1.64)

-.0038

(.60)

No. of Pigs in Dec. 2006-9.9e-06

(.07)

.0001

(.41)

.0001

(.12)

.0002

(.31)

.0002

(.16)

.0003

(.33)

No. of Villagers in New Medical

Cooperative Scheme (NMCS)

-.0001

(1.16)

.0001

(.98)

-.0003

(.62)

.0003

(.55)

-.0004

(.52)

.0004

(.57)

No. of Households Receiving

Gov. Subsidy

.0006

(1.48)

-.0006

(1.22)

.0018

(.0021)

-.0011

(.55)

.0022

(.80)

-.0014

(.55)

Ave. Villager Age.005

(.46)

.0013

(.12)

.0319

(.0471)

.0198

(.50)

.044

(.71)

.027

(.53)

Ave. Villager Education (Years).042

(1.13)

-.056*

(1.76)

.076

(.40)

-.150

(1.00)

.109

(.41)

-.176

(.90)

Fraction Male in Village1.13

(1.12)

-.46

(.43)

5.00

(1.20)

.920

(.26)

7.38

(1.26)

1.80

(.41)

Village Population/1,000.009

(.33)

-.008

(.18)

.024

(.16)

-.005

(.03)

.041

(.21)

-.007

(.03)

No. of Villagers as

Migrant Workers

1.7e-06

(.01)

.0002

(.87)

.0004

(.36)

.0007

(.90)

.0007

(.46)

.0009

(.92)

No. of Surnames in the Village-.009

(.90)

.005

(.38)

-.030

(.53)

.003

(.05)

-.048

(.63)

-.0006

(.01)

Whether Village is the Township

Government Location?

-.031

(.35)

.165

(1.60)

.380

(.81)

.759*

(1.76)

.504

(.77)

.948

(1.64)

Land per Household (Mu)-.0019

(.14)

.0130

(1.37)

.0305

(.38)

.0546

(.99)

.0397

(.36)

.712

(.99)

(Pseudo-) R2 .0167 .0156 ... .0162

Table A: Test for the Quality of Randomization for the Field Experiment.Notes: (1) Absolute values of t−statistics are reported in parentheses. Robust standard errors clustered at the township

level are used in calculating the t statistics; (2) All regressions include an unreported constant term; (3) *, **, *** denote

significance at 10%, 5% and 1%, respectively.

2

treatment dummies and the time dummies. The coefficient estimates for “LIG × September 2007”, “LIG ×December 2007”, “HIG × September 2007” and “HIG × December 2007”, namely the interactions between

the treatment dummies and the two pre-experiment periods, are not statistically significant, while those

for the interaction between HIG treatment dummies and post-experiment period dummies are significant

at close to 5% level. These suggest that there is a parallel trend between the control and treatment group

villages before our experiment, but there is a different trend between the control and the HIG groups after

our experiment.2

C The Effect of An Unexpected Snowstorm: Details

Here we provide the details of the omitted results referred to in Section 5.4 in the main text.

Ice and Snow Storm in Early 2008. In early 2008, just a month and a half after our field experiment,

a severe ice and snow storm hit southern and southwestern China and Guizhou was one of the most

affected provinces. This storm began in mid-January and lasted for a month, and its scope and severity

were unprecedented in at least the last fifty years. Since snow storms in general are rare in this part of

China, let alone one with such severity, many sows and pigs died during the snow storm especially for those

sows raised in the backyard of village households which lacked necessary facilities.3 News report indicated

that there were a total of 5,973 sows that died during the storm in Guizhou province.4 This unexpected

event offers us a rare opportunity to test whether the incentives facing the AHWs might affect farmers’

sow production directly, as discussed above, rather than through the insurance. The ice and snow storm

of such a scale was totally unexpected, and the amounts of snowfall in the villages in our sample were

uncorrelated with the village characteristics. If the insurance purchases are mainly driven by the AHWs

promising to offer cheaper and/or additional veterinary services to the farmers, we would see the positive

effect of insurance coverage on sow production, but we would not see any significant effect of the interaction

between the severity of the storm and insured sows. The presence of the significant effect of the interaction

will cast doubt on the hypothesis that the effect of insurance coverage is mainly driven by the channel of

lowering the future cost of veterinary services in exchange for the farmers’ insurance enrollment.

Results. Tables C1 and C2 report the IV estimates of the effect of the insured sows and the sow deaths

during the snow storm on subsequent sow productions as of March and June 2008 respectively. In both

tables, we use our experimental group assignments as the instruments for “No. of Insured Sows,” and

instrument for the interaction terms involving “No. of Insured Sows” by the corresponding interaction

terms with our experimental group assignments.

2In alternative specifications, we grouped the LIG and HIG villages as a single “Treatment” group. We found that the

coefficients for the interaction terms of “Treatment × March 2008” and “Treatment × June 2008” are significant at close to

5% level, while the coefficients for “Treatment × September 2007” and “Treatment × December 2007” are insignificant.3According to Wang and Watanabe (2007), summer months are the most deadly months for pigs in general.4See the newsreport on Xinhuanet (Feburary 20, 2008), “Guizhou Fully Made the Compensation for the Insured Sows

Died during the Snowstorm”, at www.gz.xinhuanet.com/xwpd/2008-02/20/content. While this snowstorm caused a lot of sow

deaths in Guizhou province, the damage on sows in our experimental area was somehow modest due to the uneven distribution

of snowstorm. As shown in Table 2, the number of sow deaths per village is 0.19, which accounts for 0.7 percent of the number

of sows per village in September 2007.

3

Variables Number of Sows

September 20078.18***

(3.74)

December 200714.41***

(5.80)

March 200820.52***

(8.21)

June 200824.66***

(9.87)

LIG × September 20074.33

(1.39)

LIG × December 20071.01

(0.28)

LIG × March 20083.41

(0.96)

LIG × June 20083.39

(0.96)

HIG × September 20072.03

(0.75)

HIG × December 20072.93

(0.97)

HIG × March 20085.42*

(1.80)

HIG × June 20086.08**

(2.03)

Constant15.85***

(20.96)

R2 .341

Table B: Testing for the Parallel Trend in the Number of Sows Between the Treatment and Control Villages.Notes: (1) Absolute values of t−statistics are reported in parentheses. Robust standard errors clustered at the township

level are used in calculating the t statistics; (2) The omitted time period is December 2006; (3) *, **, *** denote significance

at 10%, 5% and 1%, respectively.

4

The specifications in Column (1) of the two tables are similar to those in Table 6 except that we now

add “No. of Sow Deaths in Snow Storm.” The estimates in Column (1) show that, without any other

interactions, villages that lost more sows to the storm actually ended up with more sows just a few months

later. Since there is a mechanical negative effect of the sow deaths during the snow storm on the number

of sows, this strongly suggests some large positive offset.

In Column (2) of Tables C1 and C2, we include the interaction of “No. of Insured Sows” and “No.

of Sow Deaths in Snow Storm.” Interestingly, the coefficients on the interaction terms are positive and

significant at least at 15% significance level. As we argued before, this result suggests that even though

there may be some pathways for our incentive assignment to the AHWs to affect sow production directly,

they are likely unimportant or at least not dominant.

How should we interpret the positive and significant effect on the interaction between the sow death

in the storm and insurance reported in Column (2) of Tables C1 and C2? We would like to interpret this

positive interaction effect between the sow death in the storm and insurance as the effect of trust-building

behind the insurance purchase and claim settlement (see the third channel in Section 3.2.2). When farmers

do not have complete trust on whether the insurance product is genuine, the insurance policy itself becomes

a risk. Nothing is more convincing to the villagers that the government subsidized sow insurance is for real

than actually paying out the promised damage compensation in this unusual event. Indeed, as reported

by Xinghua News Agency and Financial Times (Chinese), following government directives, the insurance

company quickly dispatched work teams to remote villages to deal with claim evaluations and settlements.5

It is useful to emphasize that, for the trust effect of sow insurance to operate, the death and subsequently

time payment of loss for even one insured sow would be powerful.

From this perspective, the ice and snow storm and the subsequent compensation from the insurance

company for insured sow deaths has two possible effects. On the one hand, it probably led farmers to have

a higher awareness of the riskiness of the environment; on the other hand, the insurance company’s prompt

claim processing provided farmers a unique opportunity to learn about the credibility of the insurance

product. The first effect, in the absence of insurance options, may lead to fewer sows in the future.6 The

second effect implies that, in villages with more sow deaths and more insured sows, there would be more

positive cases that the sow insurance contracts are honored. Such positive cases of the insurance contracts

being honored would raise the villagers’ trust for the sow insurance program. Thus this mechanism will

predict that the effect of sow insurance for subsequent sow production should be stronger in such villages.7

Thus indeed, Table C1 provides support for our hypothesis that villages where gains in trust for the

government sponsored sow insurance programs are greater do experience a larger production response to

the access to insurance.

5See “Guizhou Province Made Full Compensations on Lactating Sows Which Were Insuranced and Died of the Ice and

Freeze Storm,” Xinghua News Agency, Feburary 20, 2008; and “Insurance Industry Meets with the Ice and Freeze Disaster in

the Special Way” (Financial Times, Chinese, Feburary 27, 2008).6However, when farmers have access to formal insurance that does not adjust premium for the changes in perceived death

risks, the effect on future sow production is ambiguous.7Our examination of the role of trust for the villagers’ demand of insurance echoes that of Cole et al. (2013b) in their study

of rainfall insurance. We should emphasize that in their setting the rainfall insurance was offered by a for-profit insurance

company without premium subsidy. Thus, the trust examined in their setting is the trust for insurance products offered

commercially, while in our setting the trust is for government-sponsored, partially subsidized insurance products. We should

also note that we did not randomize trust in our experimental design, while Cole et al. (2013b) did in their study. It is

interesting to note that our evidence strongly corroborates their findings.

5

Variables (1) (2) (3) (4) (5)

No. of Insured Sows.771**

(.304)

.716**

(.318)

-2.730

(4.120)

-.792

(2.366)

-1.903

(2.662)

No. of Sows in Dec. 2006.507**

(.227)

.520**

(.229)

.507**

(.229)

.499**

(.229)

.648***

(.237)

No. of Sow Deaths in Snowstorm5.13**

(2.35)

2.46***

(.57)

-52.81

(49.90)

-67.39

(57.75)

-57.05*

(28.86)

No. of Insured Sows × No. of Sow Deaths in Snowstorm.074*

(.041)

.094**

(.043)

.832

(.815)

1.068††

(.655)

Log Housing Value-3.429

(6.097)

-.833

(4.833)

No. of Insured Sows × Log Housing Value.346

(.420)

.147

(.238)

No. of Sow Death in Snowstorm × Log Housing Value5.273

(4.805)

6.673

(5.538)

No. of Insured Sows × No. of Sow Death

in Snowstorm × Log Housing Value

-.071

(.076)

Education-6.184

(8.143)

No. of Insured Sows × Education.396

(.445)

No. of Sow Death in Snowstorm × Education9.206*

(4.684)

No. of Insured Sows × No. of Sow Death in Snowstorm

× Education

-.152

(.103)

Constant-.976

(3.38)

9.646

(9.714)

62.82

(63.80)

33.96

(51.17)

40.35

(48.79)

Township Dummies Yes Yes Yes Yes Yes

Adjusted-R2 .789 .807 .823 .838 .801

Table C1: IV Estimates of the Effect of Sow Deaths During the Snow Storm on Sow Production in March2008.Notes: See notes for Table C2.

6

Variables (1) (2) (3) (4) (5)

No. of Insured Sows.823**

(.368)

.773*

(.389)

-2.623

(5.107)

.267

(3.090)

-2.799

(3.278)

No. of Sows in Dec. 2006.536*

(.265)

.548*

(.269)

.545*

(.279)

.529*

(.277)

.655**

(.255)

No. of Sow Deaths in Snowstorm5.230*

(2.671)

2.768***

(.857)

-54.67

(58.70)

-78.88

(71.65)

-66.843

(34.843)

No. of Insured Sows × No. of Sow Deaths in Snowstorm.069††

(.045)

.092††

(.060)

1.299*

(.735)

1.436***

(.518)

Log Housing Values-2.924

(7.587)

.827

(6.347)

No. of Insured Sows × Log Housing Value.339

(.524)

.043

(.3156)

No. of Sow Death in Snowstorm × Log Housing Value5.475

(5.657)

7.778

(6.865)

No. of Insured Sows × No. of Sow Death

in Snowstorm × Log Housing Value

-.116*

(.065)

Education-8.740

(9.578)

No. of Insured Sows × Education.555

(.546)

No. of Sow Death in Snowstorm × Education10.615*

(5.395)

No. of Insured Sows × No. of Sow Death in Snowstorm

× Education

-.211**

(.082)

Constant-1.22

(3.49)

13.63

(12.02)

70.53

(77.88)

28.50

(65.71)

55.653

(57.260)

Township Dummies Yes Yes Yes Yes Yes

Adjusted-R2 .773 .788 .804 .823 .788

Table C2: IV Estimates of the Effect of Sow Deaths During the Snow Storm on Sow Production in June2008.Notes: (1) Robust standard errors clustered at the township level are in parentheses; (2) The instruments for “No. of Insured

Sows” are the group assignments, and the instruments for the interaction terms involving “No. of Insured Sows” are the

corresponding interaction terms with the group assignments; (3) ***, **, *, ††, † denote significance at 1%, 5%, 10%, 15%

and 20% respectively.

7

Robustness Checks.8 While the snowstorm was unexpected and the snowfall amount was random, its

impact may be correlated with the physical village characteristics. For example, the snowstorm may have

caused more damages in villages with lower income and wealth. If it is true, the positive interaction effects

with the snowstorm variable shown in Column (2) of Tables C1 and C2 may simply reflect some other

differential treatment effect across villages. In Column (3) of Tables C1 and C2, we include the log of the

village housing values, and its interactions with the number of insured sows as well as the number of sow

deaths in the snow storm. Note that the inclusion of these additional controls largely does not change

either the magnitude or the statistical significance of the coefficient estimates on the interaction term of

“No. of Insured Sows” and “No. of Sow Deaths in Snow Storm.”

However, our interpretation of the results in Column (2) of Tables C1 and C2 as evidence for the effect

of trust still encounter several potential challenges. The first concern is that, adding log house values in

levels as we did in Column (2) may not adequately control for the wealth effect. To more adequately

disentangle the trust effect from the wealth effect, in Column (4) of Tables C1 and C2, we include a triple

interaction term between log of the village housing value, the number of sow deaths in the snow storm and

the number of insured sows, while at the same time controlling for a set of double interaction terms between

these three variables. The coefficient on the interaction term between sow deaths in the snowstorm and

the number of insured sows still has the expected positive effect on the number of sows in both March and

June 2008, and it is significant for June 2008.

Second, because as we described in Section 2 the insurance payment is less than the market value of a

sow, it is possible that the more likely confounding explanation is a liquidity effect as supposed to a wealth

effect. Unfortunately we do not have direct observations on either total village savings account balances or

income to control for the liquidity effect. In Column (5) of Tables C1 and C2, we proxy income or liquid

assets using average level of education in the village since it is plausible to believe that higher education is

correlated to higher income. Specifically, we include an interaction term between average village education,

the number of sow deaths in the snow storm and insured sows, while at the same time controlling for a

set of double interaction terms between these three variables. Results in Column (4) indicate that the

triple interaction is negative (and also statistically significant for June 2008), which implies that fewer

new sows in response to sow deaths when the village has more liquidity. This suggests that liquidity,

or something correlated with liquidity, is indeed part of the explanation for our findings. However, the

coefficient estimates of the interaction term between the number of sow deaths in the snow storm and the

number of insured sows still have a positive sign in both cases and statistically significant at 12% level for

March 2008 and 1% level for June 2008. This suggests that something about the insurance payouts other

than liquidity is having an effect.

D Table 7 in Cai et al (NBER Working Paper 15396)

While we have argued that the snowstorm effect reported in Tables C1 and C2 is consistent with the

trust story, we do not have direct evidence on the effect of trust on subsequent insurance take-up. As a

partial remedy for the lack of direct evidence on the effect of trust on the insurance take-up, we would

like to discuss some additional evidence presented in an earlier version of our paper (Cai et al. (2009)).

In Table 7 of Cai et al. (2009), we document that the number of insured sows is significantly associated

8We are grateful to an anonymous referee for suggesting these robustness checks.

8

with the coverage of new rural cooperative medical schemes in the village as well as the coverage of

government subsidy in the village. The new rural cooperative medical scheme was launched by Chinese

government in 2003 and both central and local governments provided subsidies to the premium. Even

with heavy subsidies and reasonable price, the take-up of this scheme surprisingly varied across villages.

In our village sample, the average take-up rate was only 50%. One of the obstacles was the trust on the

ability of government to deliver the promise. Against this background, we hypothesize that if trust is also

an important determinant for villagers’ purchase of the sow insurance, we would expect to see that there

would be a positive correlation between a village’s take-up of the medical scheme and its take-up of the

sow insurance. The positive correlation between government subsidy coverage and sow insurance can be

interpreted as evidence for the hypothesis that villagers who have been receiving government subsidies

tend to have a higher trust for government sponsored program in general, and thus the subsidized sow

insurance in particular. While this additional evidence is still subject to other interpretations and thus

is not perfect in supporting the role of trust, it may help rule out wealth effects in picking up the whole

story. In particular, those people who received government subsidy tended to be poorer than those who

didn’t, but they may have higher trust in government-sponsored program.

D.1 The Relationship Between Participation in the New Cooperative Medical Scheme

and Insurance Purchase

Our first evidence that trust, or lack therefor, for government sponsored programs may prevent farmers

from purchasing the heavily subsidized sow insurance is to show that villages where the farmers have pre-

viously demonstrated a higher level of participation in another government sponsored voluntary insurance

program – the New Rural Cooperative Medical Scheme (NCMS) – are also more likely to purchase the sow

insurance.

New Cooperative Medical Scheme.9 The original Cooperative Medical Scheme (CMS), introduced in

rural China in the 1950s, was dismantled with the collapse of the collective economy in the early 1980s. As

a result, by 1985 only about 5% of rural counties in China had any form of health care insurance. Chinese

government launched the New Cooperative Medical Scheme in 2003, aiming to provide health coverage

for the nation’s entire rural population by 2010. Both the central and provincial governments provide

subsidies to the premium. For Guizhou Province, in 2003 when NCMS was initiated, the central and

provincial government respectively subsidized the program at the rate of 10 Yuan and 40 Yuan per enrollee

annually; and the subsidy amounts were increased to 20 Yuan and 50 Yuan respectively in 2006. Each

enrollee is only required to contribute 24 Yuan annually. Somewhat surprisingly, the heavily subsidized

and reasonably priced NCMS was not an immediate success; at the national level, the percentage of rural

residents covered under the NCMS increased from 3% in 2004 to 40.57% in 2006. In our sample of villages,

as described in Table 2, about slightly over 50% of the villagers signed up the NCMS, with quite sizeable

variations in the coverage rates across villages.

Empirical Result. While there might be a multitude of reasons for the variation in the NCMS coverage

across villages, in this section we entertain the hypothesis that one of these sources is the trust for govern-

ment sponsored programs. Under this hypothesis, if trust is also an important determinant for villagers’

9See Lei and Lin (2008) and references cited therein for more background information about the NCMS.

9

purchase of the sow insurance, we would then expect to see that there would be positive correlation between

a village’s take-up of the NCMS and its take-up of the sow insurance.

Columns (1)-(4) in Table D reports the relationship between the number of insured sows and the number

of villagers who participated in the NCMS. Results from the preferred specification, which controls for the

number of sows in December 2006, as well as the experimental group assignment and Township dummies,

are reported in Column (4). The estimated coefficient on the number of villagers in NCMS is positive and

significant at the 10% level, thus suggesting that there is some underlying common unobservable factor that

leads the villagers to more likely join the NCMS and purchase the sow insurance. At the point estimate,

a one standard deviation increase in the number of villagers in the NCMS, which is about 300 from Table

2, will lead to an increase of 4.5 additional insured sows, which represents about 20% of the mean number

of insured sows (about 22.7 from Table 2).

While we do not have direct evidence of what the common unobservable factors are, there are several

candidates. One candidate common unobservable factor is trust, or lack thereof, for government sponsored

programs. Villages that have a higher trust for government programs in general will both have a higher

participation rate in the NCMS, and are more likely to purchase the subsidized sow insurance. In villages

assigned to the high incentive group, the AHWs can increase sow insurance coverage by spending more

time to dispel the doubts that farmers may harbor toward the government sponsored insurance product.

The fact that AHWs are the residents of villages and their personal reputation is held as a “hostage” if

they cheat their fellow villagers helps AHWs do their work well.

Another candidate for the unobservable factor is that villages may differ in how easy the information

about the government sponsored programs spread. For example, some villages may have more close-knit

social networks that words about the government sponsored programs, whether it is the NCMS or the

sow insurance, will quickly spread, leading to higher participation rates in both the NCMS and the sow

insurance. This channel, however, does not seem to be very plausible because the NCMS has been in

place for almost four years, and it is unlikely that any villager has not heard about the NCMS. Yet,

the participation rate for the NCMS is still low in the villages with low participation rate for the sow

insurance.10

D.2 The Relationship Between the Number of Households Receiving Government

Subsidies and the Purchase of Sow Insurance

Our second piece of suggestive evidence that trust for government sponsored programs may be the

main reason for the seemingly low take-up rate of the sow insurance is the systematic relationship between

the number of households receiving government subsidies and the purchase of sow insurance, reported

in Columns (5)-(8) in Table D. Here we hypothesize that villagers who have been receiving government

subsidies tend to have a higher trust for government sponsored programs in general, and thus the subsidized

sow insurance product in particular. The preferred specification with all the controls is reported in Column

(8) and the coefficient estimate on the number of households receiving government subsidy is positive and

significant at 1%. At the point estimate of the coefficient, a one standard-deviation increase in the number

10Moreover, if closer knit social network is the reason for better information spreading for the government sponsored

programs, it would at the same time also lead to less, not more, need for formal insurance.

10

No.

of

Insu

red

Sow

s

Var

iab

les

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

No.

ofV

illa

gers

inN

ew

Coop

.M

edic

alSch

eme

.012

4*

(1.8

0)

.0124*

(1.8

6)

.0149*

(1.8

1)

.0146*

(1.8

5)

No.

ofH

ouse

hol

ds

Rec

eivin

g

Gov

.S

ub

sid

y

.0565***

(2.9

2)

.0563***

(2.9

9)

.0610***

(2.7

7)

.0605***

(2.8

6)

No.

ofS

ows

inD

ec.

2006

.676

2**

(2.7

1)

.6836***

(2.8

7)

.6004**

(2.2

9)

.6109**

(2.4

3)

.6510**

(2.6

2)

.6584***

(2.7

8)

.5856**

(2.2

4)

.5954**

(2.3

8)

Low

Ince

nti

veG

rou

p9.9

37***

(3.9

6)

9.5

26***

(3.4

9)

9.7

54***

(3.8

3)

9.3

18***

(3.3

8)

Hig

hIn

centi

ve

Gro

up

12.3

40***

(3.9

2)

11.7

73***

(3.9

3)

12.3

25***

(3.9

0)

11.8

19***

(3.8

9)

Con

stan

t5.

135*

(1.7

6)

-3.6

69

(.91)

-.258

(.10)

-8.4

97**

(2.2

3)

2.1

02

(.72)

-6.6

11

(1.5

6)

-3.0

01

(1.1

6)

-11.2

18***

(2.8

9)

Tow

nsh

ipD

um

mie

sN

oN

oY

esY

esN

oN

oY

esY

es

Ad

just

ed-R

2.3

549

.3903

.4485

.4794

.3710

.4062

.4613

.4923

Tab

leD

:T

he

Rel

ati

on

ship

Bet

wee

nth

eN

um

ber

ofV

illa

gers

Par

tici

pat

ing

inth

eN

ewR

ura

lH

ealt

hC

oop

(Col

um

ns

1-4

)an

dth

eN

um

ber

ofH

ou

seh

old

sR

ecei

vin

gG

over

nm

ent

Su

bsi

die

s(C

olu

mns

5-8)

and

the

Pu

rch

ase

ofth

eSow

Insu

ran

ce:

Som

eP

reli

min

ary

Evid

ence

for

the

Role

of“T

rust

”fo

rG

over

nm

ent.

Notes:

(1)

Abso

lute

valu

esoft−

stati

stic

sare

rep

ort

edin

pare

nth

eses

.R

obust

standard

erro

rscl

ust

ered

at

the

tow

nsh

iple

vel

are

use

din

calc

ula

ting

thet

stati

stic

s;(2

)

*,

**,

***

den

ote

signifi

cance

at

10%

,5%

and

1%

,re

spec

tivel

y.

11

of households receiving government subsidy, which is 92 from Table 2, is associated with an increase of 5.4

insured sows, representing about 25% of the mean number of insured sows (about 22.7 from Table 2).

Potential lack of information about the government sponsored programs can not the common unob-

servable factor explaining the positive association between households receiving government subsidies and

the number of insured sows reported above. The reason is that any eligible households with per capita

income below a government specified threshold will automatically receive government subsidies; there is no

need for filling out an application and thus there can not be an effect of the knowledge about the existence

of the subsidy programs.

It is also worth pointing out that villages with more households receiving government subsidies tend to

be poorer, and as a result farmers in such villages may be more risk averse. However, this channel would

not have explained why the different incentives we provide to AHWs had a large positive effect on the

insurance purchase; after all, the villages are randomly assigned to the experimental groups, and as Table

2 and Table A showed clearly, villages assigned to different experimental groups do not seem to exhibit

any difference in the numbers of households receiving government subsidies.

12