Social Networks and the Development of Insurance Markets: Evidence from Randomized Experiments in China 1

Social Networks and the Development of Insurance Markets: Evidence from Randomized Experiments in China 1 Jing Cai 2 University of California at Berkeley Oct 3 rd, 2011 Abstract This paper estimates the role of social networks in improving the sustained demand for a new insurance product, using data from a two-phase randomized experiment in rural China. The experimental design identifies the causal effect of social networks on both short-run and long-run insurance demand and its monetary value. It also sheds light on the channels through which social networks operate, including social learning (gaining knowledge from friends who received direct financial education about the benefits of insurance), trust, imitation, informal risk sharing, and learning from friends experience with payouts. The results demonstrate that social networks have an important effect on insurance take-up, and compare favorably to direct education and subsidies. In the short-run, providing financial education to a subset of farmers has a large and positive spillover effect on other farmers. It is equivalent to 45% of the direct financial education effect and almost 20% of the subsidy effect, and is driven by diffusion of knowledge about insurance. In the long run, social networks affect insurance demand through learning about friends payout experience; this is equal to 50% of the subsidy or price reduction effect. The policy implication is that providing endless subsidies cannot generate sustainable voluntary insurance purchase. In contrast, providing financial education to a subset of households and relying on social networks to multiply its effect on others, and finding ways to increase the frequency of payouts, can effectively improve sustained insurance demand. JEL CODES: D12, D83, G22, O12, Q12 Keywords: Social network, Insurance, Demand curve, Learning 1 I am indebted to my advisors, Alain de Janvry and Elisabeth Sadoulet, for their constant support and advice. I am grateful to Michael Anderson, Michael Carter, Frederico Finan, Xavier Gine, Shachar Kariv, David Levine, Ethan Ligon, Jeremy Magruder, Craig McIntosh, Edward Miguel, Adam Szeidl, and numerous seminar participants for helpful comments and discussions. I thank the People s Insurance Company of China for their close collaboration at all stages of the project and would especially like to acknowledge the contributions of Aijun Cai, Leilao Chen, Baohua Deng, Xiaoping Fan, Xiangli Li, Zhanpeng Tao in experiment implementation and data collection. The study was funded by 3ie and ILO. All errors are my own. 2 caijing516@berkeley.edu 1

1. Introduction Social networks permeate our social and economic lives and have a fundamental role in shaping human activity. They are important in the transmission of information about job opportunities or innovations; they are the basis for the provision of mutual insurance in rural areas; they are also important in determining our decisions to purchase new financial products, as well as how we vote, how diseases spread, etc. (Jackson 2010). This paper studies social network effects on the adoption of a new insurance product. In general, poor households in rural areas are exposed to substantial weather shocks that can generate fluctuations in income and consumption if insurance markets are not complete. However, in many countries, even when such a market exists, the initial participation rate is usually very low, even with heavy government subsidies (Gine et al. 2008, 2009) and the renewal rate is not high (Stein 2010). This disappointing development of formal insurance markets has been referred to as a puzzle in need of an explanation (Cole et al. 2009; Karlan et al. 2009). This paper uses randomized experiments to study the role of social networks in improving the sustained demand for insurance by examining its effect on insurance demand in both the short and long run 3. The novel experimental design not only allows me to identify the causal effect of social networks, but also to test various channels through which social networks operate, including social learning of insurance knowledge, trust, imitation, informal risk sharing, and social learning from friends experience. In addition, a price randomization is interacted with the social network experiment to show the monetary value and relative importance of the network effect. Finally, policy recommendations are made as to how to improve sustained adoption of insurance over time. I test the role of social networks in the development of insurance market in rural China. In 2010, the People s Insurance Company of China (henceforth PICC) started to offer a new insurance policy to rice farmers in selected pilot counties. Because this product was brand new in all pilot areas, farmers and even government officials at the 3 Most papers are focused on studying ways to improve the initial participation rate, but there s no rigorous study about how to maintain good take-up rates in the long-run when subsidies are gradually removed, which is even more important for the development of insurance markets. 2

town or village level had very limited understanding of how this type of insurance works and what are the expected benefits of purchasing it. In this context, social interactions can have an important effect on adoption decisions, both in the short run and over time. Farmers may learn insurance knowledge from people who were exposed to more information or who have a better understanding of such products, or they can be influenced by other people s decisions. Moreover, as decisions to buy or renew insurance are made over time, decisions may be affected by peers experience. I use a two-phase experiment in China to answer the research question. In phase 1, which was implemented in 2010, I test the causal effect of social networks on initial insurance take-up, identify the type of information that the networks conveyed, and determine which categories of information drive the effect. This phase includes 185 villages with around 5800 households. In each village, households were randomly assigned to an early versus a late round (three days later) session, and simple sessions (introducing the insurance contract only briefly) versus intensive sessions (including financial education). I asked each household to list five of their closest friends with whom they discuss rice production and financial questions, and define the social network measure as the fraction of friends invited to the early round of intensive sessions. Since invitations to different sessions were randomized at the household level, this allows me to estimate causal effects of social networks on take-up behavior. I find that, while financial education raises take-up from 35% to 50% in the first round, having one additional friend who attended first round of financial education increases take-up of second-round participants by around 6.7%; this equals almost 45% of the direct education effect. In a subsample, a price randomization is added to identify the monetary value of the social network effect. Results show that the magnitude of the social network effect is equivalent to 18% of the price reduction effect. There are two types of information that a social network potentially can convey. First, it can diffuse knowledge about the benefits of insurance among farmers, which helps them understand the value of an insurance product and improves the take-up rate. Second, farmers can learn of each other decisions through social networks; such information can affect households decisions through mechanisms including imitation, trust (improvement of negotiating power with the insurance company), and informal risk sharing. To test 3

whether it is diffusion of insurance knowledge that is driving the network effect, I compare the effect of financial education on improving insurance take-up and understanding between two rounds. I found a smaller effect in the second round, and found that farmers understand insurance benefits better if they have more friends exposed to financial education; this means there is diffusion of insurance knowledge from first round educated farmers to second round participants. To figure out whether peers decision is the key information that networks conveyed, I estimate the effect of peers behaviors on a farmer s own decision. If I told farmers other villagers decisions, especially decisions made by their close friends, it influenced their own behaviors significantly; however, other villagers decisions had no significant effect if we did not explicitly reveal that information. This means that, while the social network can help teach financial literacy, it cannot convey decisions. This suggests that, in the short run, social networks can improve insurance take-up through the rational learning of insurance benefits. The development of insurance market is two-folded. First, improving the initial participation rate is crucial because the increased demand of individuals is a prerequisite for scaling up the program to reach and help more poor people. Second, studying sustainable ways to maintain good take-up rates in the long-run is essential. For products such as insurance, repeated take-up over time is required to generate the hoped-for impact on households and to maintain financial sustainability; because large subsidies are not likely to be provided indefinitely, it is crucial to identify a mechanism that will improve take-up even with smaller subsidies. At the same time, social networks may have important long-term effect, because over time, individuals use not only their own experience, but also the experience of others, in making purchase decisions. To study the long-term effect of social networks on insurance demand, I followed a subsample of first phase households in phase 2 (2011). The treatment in this phase is household level price randomization. Disasters happened during the first year and some households who purchased insurance in the first year received payouts. I do not only look at the general social network effect over time, which is the effect of friends decisions in phase 1 on a farmer s own demand in phase 2, but also test the effect of a specific aspect of social networks: social learning by the experience of observing friends receiving 4

payouts. To get a sense of how large the effect is, I compare it with the effect of learning by doing, which occurs when an individual receives payouts from his own insurance policy, and the price effect. Moreover, I use the regression discontinuity method to separate the learning-by-experience effect from the weather shock effect. I found that households are not influenced in the second year by their friends previous purchase behavior, but that observing more friends receiving payouts improves second year take-up at all price levels (21.7 percentage points) and makes people less sensitive to price change (offsets price effect by more than 50%). Comparing the effect of learning from others with learning from one s own experience, I find that learning from friends has about 54% of the impact of learning from one s own experience. This means that social networks affect insurance take-up over time through social learning of friends experience. This paper suggests that social networks have significant and important effect on the development of rural insurance markets. Such networks affect not only the initial participation rate, but also affect demand over time; this effect is driven by social learning of insurance benefits and by friends experiences. Several policy implications can be made based on these results. First, when doing impact evaluations, special attention should be paid to the potential spillover effect from treated to untreated households; experimental treatments should be randomized at the village level, rather than within a given village. Second, as we did not see farmers copy each others decisions over time, the commonly used practice of providing a one-time free distribution of the insurance product to a subsample cannot lead to indefinitely sustainable voluntary purchase of the insurance. In contrast, providing financial education to a subset of households and depending on social networks to extend its effect, and improving contract design to raise the potential payout rate while disseminating information about these payouts more effectively, would be good ways to support the development of insurance markets for rural households. This paper contributes to the existing literature in the following ways. First, it contributes to the social network literature by using randomized experimental methods to estimate the causal effect and monetary value of social networks on improving insurance demand 4, and explains why social networks are important. Second, it sheds light on the 4 It is challenging to identify the causal effect of social networks on adoption because it is hard to distinguish it from other factors that may give rise to similar observed outcomes such as correlated unobservable characteristics between friends (Manski 1993). Although there is a rapidly growing 5

puzzle of why weather insurance has very low take-up. Although existing works have tested the possible explanations of trust, credit constraints and ambiguity aversion (Gine et al. 2009, Cole et al. 2008, Bryan 2010), social interactions remain less explored as a possible determinant. I provide evidence that social networks can improve the initial participation rate and insurance demand over time. Third, this paper contributes to financial education literature. Existing literature on financial education seems to show little or no effect on individual decisions 5. In a context where insurance is new, and farmers have relatively low levels of general education, my results show that lack of financial education is a major constraint on the demand for insurance, and that modest financial training can improve take-up rates significantly. The rest of the paper is organized as follows. Section 2 describes the background and the insurance contract. Section 3 presents the experimental design. Section 4 presents the results, and section 5 concludes. 2. Background Rice is the most important food crop in China. Nearly 50% of the farmers produce rice, and more than 60% of the Chinese people use rice as their staple food. In order to maintain food security and shield farmers from negative weather shocks, in 2009, the Chinese government designated the People s Insurance Company of China (PICC) to design and offer the first rice production insurance program to rural households in 31 pilot counties 6. The program was extended to 62 counties in 2010 and 99 counties in 2011. Our experimental sites are two rice production counties included in the second round pilots in Jiangxi province, which is one of China s major rice bowls. In these two counties, rice production is the main source of income for more than 80% farmers. No households had ever purchased such insurance before, and most of them have never interacted with PICC literature that uses experimental methods (Duflo and Saez 2003; Miguel and Kremer 2004; Duflo, Kremer, and Robinson 2010; Oster and Thornton 2009), longitude data analysis (Conley and Udry 2010), or a structural approach (Banerjee, Chandrasekhar, Duflo and Jackson 2010) to solve this endogeneity issue, there s no conclusive result. 5 For example, Duflo and Saez (2003) found that a benefits information fair increased enrollment in retirement plans by 1.25 percentage points after 11 months, a small effect in absolute terms. 6 Before 2009, although there was no insurance, if big natural disasters happened, governments issued subsidies to households whose productions were seriously hurt. However, the level of subsidy was usually very limited and far from enough for farmers to restart production. 6

before. As a result, farmers, and even government officials at the village or town level, have very limited knowledge of agricultural insurance products and the insurance company as the provider. The insurance contract is as follows. The fair price is 12 RMB per mu per season 7. The government gives a 70% subsidy on the premium, so farmers only pay the remaining 30%, which is 3.6 RMB per mu. The insurance covers natural disasters including heavy rain, flood, windstorm, extremely high or low temperature, and drought. If any of the above disasters happened and led to 30% or more loss in yield, farmers are eligible to receive payments from the insurance company. The indemnity rule is illustrated in Figure 1. The payout amount increases linearly with the loss rate in yield, with a maximum payout of 200 RMB. The rate of loss in yield will be determined by a group of insurance agents and agricultural experts, who will come to the village to estimate the rice yield in different plots and calculate the loss rate 8. Since the average gross income from cultivating rice is between 700 RMB to 800 RMB per mu, and the production cost is around 300 RMB to 400 RMB per mu, this insurance program provides a partial insurance which covers 25% - 30% of the gross income or 50% - 70% of the production cost. Based on communications with local government officials and farmers, the actual probability of disasters which can cause 30% or more loss in yield is estimated to be around 12%, so the fair price of this product, which is the price that makes the insurance company break even, should be higher than the 3.6 RMB/mu paid by farmers and lower than the 12 RMB/mu received by the insurance company 9. As a result, the insurance company can earn profit and survive if the fixed cost is not too large, and the expected benefit of purchasing insurance is positive for farmers, implying that it is optimal for all farmers who cultivate rice to purchase it. 7 1 RMB = 0.15 $, 1 mu = 0.067 hectare, and in each year, farmers produce two or three seasons of rice. 8 For example, consider a farmer who has 5 mu in rice production. If the normal yield per mu is 500kg and because of a windstorm, the farmer s yield decreased to 250kg per mu, then the loss rate is 50% and he is supposed to get 200*50% = 100 RMB per mu from the insurance company. 9 The insurance company profit from insuring 1 mu of rice equals: premium probability of disaster * indemnity fixed cost. In our case, probability of 30% disaster * indemnity = 12% * 200 * 30% = 7.2 RMB. 7

3. Experimental Design and Data We use a two-phase experimental design to identify the role of social networks in the development of rural insurance markets. The first phase was carried out in spring 2010, which includes 185 natural villages with around 5,300 households. We use data from the first phase to test the short-run impact of social networks on insurance. The second phase was implemented in spring 2011, in which we followed 1,871 households covered in phase 1, and examines the effect of social networks overtime. 3.1. Phase 1: Identify social network effect on initial year demand In this phase, we answer three questions: first, what is the causal effect of social networks on initial insurance demand? Second, what is the monetary value of the social network effect? Third, in order to provide more concrete policy suggestions, we ask what kind of information did people learned through social networks that drive the effect? Insurance knowledge or others purchase decisions? If network effect exists because farmers care about friends decisions, it means the mechanisms of social network effect can be imitation, trust, or informal risk sharing. If farmers actually learned insurance knowledge from friends, which improves their own take-up, it means it is social learning of insurance benefits that drives the social network effect. Two types of information sessions were offered to households in this phase. The simple session took around 20 minutes, during which we only introduced the contract 10. The intensive session took around 45 minutes and covered all information provided during simple sessions, plus a financial education to help farmers understand how insurance works and the benefit of purchasing it 11. In each village, we held two rounds of sessions to introduce the rice insurance program. The second round sessions were held 3 days after we finished the first round to allow late participants communicate with early participants. During each round, there were 10 We explain terms in the contract including the insurance premium, the amount of subsidy provided by the government, the response of the insurance company, the maximum payout, the period of responsibility, rules of loss checking, and the procedures for making payouts. 11 Sample questions included in financial education including: How does the insurance program differ from a government subsidy? How much payout can you get under different conditions? What is the expected benefit of purchasing insurance for 5 continuous years depending on different disaster frequencies and levels to see whether you can gain or loss from taking it, etc. 8

two sessions, one simple and one intensive. Each household was invited to one of these four sessions. The experimental design in phase 1 is illustrated in figure 2.1 and 2.2. There are four randomizations in this experiment, two on village level and two on household level. Village level randomization is shown in figure 2.1. First, we randomly divided villages into type I and type II. In type I villages, all households face the same final price 3.6 RMB, while in type II villages, we randomly assigned seven different prices ranging from 1.8 RMB to 7.2 RMB to different households in 2 nd round sessions. Type II villages are used to measure the monetary value of social network effect. The second village-level randomization is only within type I villages and what we randomized is the default option in 1 st round sessions. If the default was BUY, then you needed to sign off if you did not want to purchase the insurance, while if the default was NOT BUY, then you had to sign 12 on if you decided to buy the insurance. Default options were the same in the two firstround sessions within each village. The objective of using different default options was to generate exogenous variations in the take-up rate in first-round sessions across villages 13. The within village randomization is presented in figure 2.2. First, in both type I and type II villages, households were randomly assigned to exactly one of the four sessions T1 ~ T4. Because we test social network effect by looking at whether 2 nd round participants 14 are more likely to buy insurance if they have more close friends exposed to formal financial education in T2, this randomization is used to get exogenous variations in the fraction of friends exposed to 1 st round financial education for 2 nd round participants, and hence identify the social network effect within villages. 12 During sessions where default = BUY, before we asked farmers to make decisions, we told them the following: We think that this is a very good insurance product, and we believe that most farmers will choose to buy it, so it is more convenient for us to record who does not buy it rather than who buys it. So if you decided to buy it, there is nothing you need to do, the premium will be deducted from your agricultural card automatically; if you do not want to buy it, then please come here and sign. 13 According to existing literature (Laibson et al., 2008), default options can influence households financial decisions significantly. The reason of why default option matters can be because households found it too complicate to make a decision by themselves, or they think the option is set as the dfault because it is a good choice. For more details, refer to Laibson et al (2008). 14 In each village, we invited household heads to attend one of the four sessions. No one could attend more than one session. For all household level randomizations in this experiment, we stratified the sample according to village, household size, and average rice production per member in the most recent year before randomization. Similar to what we did in Experiment #1, we asked village leaders to inform and invite household heads to attend these sessions in order to maintain a high attendance rate. 9

Second, in type I villages where no price randomization was implemented 15, for each 2 nd round sessions, after the presentation and before participants made final decisions, we randomly divided them into three groups and led different groups to separate rooms, and then we disseminated different additional information to the different groups of participants. Specifically, we did not give farmers in groups U1 and U4 any additional information and directly asked them to make decisions, so these two groups receive exactly the same information from us as T1 and T2; for farmers in groups U2 and U5, we told them the overall take-up rate at the two 1 st round sessions in the village that they belong to, so they knew the number of people who attended previous sessions and how many of them had purchased the insurance; in groups U3 and U6, we showed farmers the detailed decision list at the two 1 st round sessions, so they knew specifically who purchased the insurance and who did not. This randomization helps us to answer whether the network effect is driving by learning of insurance knowledge or peers decisions. In all cases, households make take-up decisions individually at the end of our visit. Moreover, all households were asked to respond to a household survey. The survey is mainly composed of six parts: first, household background including household size, age and education of the household head, rice production and sales, household income, borrowing, etc.; second, natural disasters experienced in recent three years and loss rate in yield; third, experience in purchasing any insurance and reimbursement received; fourth, risk attitude and perception of future disasters 16 ; fifth, ten questions which test farmers understanding of information provided during sessions; and sixth, general and detailed social network questions which ask each household to rank and list five of their most closely related friends with whom they frequently discuss rice production and financial related problems, and ask what specific topics do they usually discuss. 15 In type II villages, 2 nd round sessions T3 and T4 are exactly the same as 1 st round sessions T1 and T2, respectively. No additional information was provided after the presentation. 16 Risk attitudes were elicited by asking sample households to choose between increasing amounts of certain money (riskless option A) and risky gambles (risky option B) in table A1. The number of riskless options was then used as a measure of risk averse. The perceived probability of future disasters was elicited by asking what do you think is the probability of a disaster that leads to more than 30% loss in yield next year? 10

3.2. Phase 2: Identify social network effect over time In a subset of natural villages (72 out of 185, around 2,000 households) covered in phase 1, a follow-up experiment was conducted one year after the first time provision of the rice insurance program. All households in these villages and have been surveyed in phase 1 were sampled for the follow-up experiment, regardless of whether they purchased insurance or not in phase 1. Combining data of both phases, we test the long-run effect of social networks, especially the impact of observing friends receiving payouts, on second year insurance demand curve. In phase 2, what we randomized is the level of subsidy. The subsidy level ranged from 90% to 40%, and the corresponding final prices faced by households varied from 1.2 RMB to 7.2 RMB. Except for the final prices, everything else remained the same in the contract as in phase 1. In total, there are eight different prices offered, but within each village, only two or three prices were assigned 17. For example, if the price set is {1.8, 2.6, 5.4}, all sample households in the village were randomly assigned with one of these three prices. To do randomization of price sets on the village level, the sample was stratified according to village size (total number of households) and first year village level payout ratio. For price randomization on the household level, the sample was stratified according to rice production area. With the price randomization, we can test the long-run social network effect on both levels of demand and price sensitivity (monetary values of longrun social network effect). In each village, we gather households assigned with the same price levels together and hold meetings for different price groups simultaneously. During the meeting, we briefly repeat items in the insurance contract and announce the list of people who purchased insurance and have received payout last year 18, so all households know who in the village received payout, and the amount of payout. This helps us to test the effect of an important type of social network effects: social learning of friends experience. Specifically, we calculate the fraction of close friends (listed in phase 1 household survey) who purchased insurance and received payout to use that as the measure of social learning of friends 17 For price sets with two prices {P1, P2}, P1 <= 3.6 and P2 > 3.6; for price sets with three prices {P1, P2, P3}, P1 < 3.6, P2 = {3.6, 4.5}, and P3 > 4.5. 18 After the insurance was offered in April 2010, low temperature disaster happened in October 2010, just before the harvest of the late season rice, which lead to yield loss for most farmers. 11

experiences, and test its impact on second year demand curve. Households make second year purchase decisions individually right after the meeting. Similar to what we did in phase 1, all households were asked to respond to a household survey. The survey is composed of parts: first, household background such as household size, age and education of the household head, rice production, etc.; second, natural disasters experienced in last year, loss rate in yield, and methods of coping with losses; third, purchase decisions last year and payout experience; forth, questions testing households trust on the insurance company of the loss checking and payout issuing process; fifth, the same ten questions we asked in phase 1 which test farmers insurance knowledge; sixth, willingness to pay for this rice insurance; and seventh, risk altitude and perception of future disasters. 4. Estimation Strategies and Results 4.1 Effect of social networks on initial year insurance demand This section tests the role of social networks in driving insurance demand when it was offered at the first time and explains what information and mechanism is driving the effect using data collected in phase 1. 4.1.1. Estimation of the social network effect In figure 3, we draw the average take-up rate in different information sessions. It shows that while the difference between the two 1 st round sessions is substantial, there s almost no difference between the two 2 nd round sessions. Moreover, the take-up rate of second round sessions is higher than that of 1 st round simple sessions. The above evidence suggests that financial education provided during 1 st round intensive session improved farmers take-up rates, and during the three days between the two round sessions, there s information diffusion from 1 st round to 2 nd round participants, i.e. the social network effect. To estimate the effect of social networks on insurance take-up, refer to figure 2, we use type I villages (no price randomization) only. First, we estimate the financial education effect using the sample of 1 st round participants (T1 vs. T2); second, we focus on households who were assigned to 2nd round session groups U1 and U4 (no take-up 12

information revealed) to test whether they are more likely to buy insurance if they have more close friends invited to 1st round intensive session (financial education). The financial education effect is estimated using the following equation:!"#!"#!" =!! +!!!"#$"%&'!!" +!!!!" +!! +!!" (1) where!"#$%!!" is an indicator of the purchase decision made by household! in village!, which takes a value of one if the household decided to buy the insurance and zero otherwise.!"#$"%&'!!" is a dummy variable equal to one if household! was invited to one of the two intensive sessions in village! and zero otherwise.!!" includes household characteristics such as gender, age, production size, etc., and!! includes village dummies. According to results in table 1, the take-up rate of 1 st round intensive sessions is 14 percentage points (42%, from 35% to 50%) higher than that of 1 st round simple sessions, suggesting a large and significantly positive financial education effect in the 1 st round. To test the social network effect, i.e. the spillover effect from 1 st round financially educated farmers to 2 nd round households, we estimate the following equation based on second round participants in groups U1 and U4:!"#$%!!" =!! +!!!"#$%&!!" +!!!!" +!! +!!" (2) where the social network measure is defined as by the fraction of friends in your social network who have been invited to first round intensive session 19. Since households are more likely to be exposed to information provided during the financial education if they have more friends attended, we expect a positive social network effect. Estimation results are listed in table 2. According to column (1), households who were not included in 1 st round sessions are significantly influenced by the number of close friends who were invited to 1 st round intensive sessions, and the magnitude of effect is around 33.7 percentage points, which means having one additional close friend attending 1 st round intensive session (raises the network measure by 20% because each household lists five friends) increases your own take-up rate by 33.7*0.2 = 6.74 percentage points, which catches around 45% of the effect of having yourself receiving the financial education (column (1) in table 1). In column (2), we add control variables in the regression. It shows 19 For example, if a household listed 5 friends, and 2 of them were invited to 1 st round intensive session, then the social network measure equals 0.4. 13

that the magnitude of social network effect does not change much with additional controls. In addition, it suggests that elder people or more educated farmers are more likely to buy the insurance. Moreover, households who are more risk averse, or those who predict higher probability of natural disasters in the following year, tend to purchase insurance. In column (3), we test whether the magnitude of social network effect depends on whether you received financial education or not. According to the coefficient of the interaction term between the network measure and intensive session, the social network effect is larger in 2 nd round simple sessions, which means that households are less influenced by their friends when they have better understanding of insurance. While having friends exposed to financial education improves your own take-up, the effect can be nonlinear. It is possible that having just one friend receiving financial education is enough, or having more friends financially educated can generate larger effects. We test the nonlinearity in two ways as shown in table 3. In column (1), we include both the social network measure and its square. The coefficient of the square term is negative, but it is not significant, which provides suggestive evidence that the relationship between the social network measure and insurance take-up is concave. In column (2), we directly test the effect of having different numbers of friends on take-up. The result shows that having more than one friends exposed to 1 st round financial education improves your own take-up by 20 percentage points on average, which is significantly larger than the influence of having only one friend receiving financial education (6.23 percentage points). As a result, the magnitude of social network effect increases with the number of friends receiving financial education. In order to understand the importance of the social network effect better, we estimate the monetary value of it using type II villages in which price randomization was implemented. Specifically, we estimate the effect of removing subsidies on insurance demand, and test whether households are less sensitive to price increase if they have more friends exposed to financial education. The intuition is that households may value the insurance product more, if they have more friends financially educated, as we see a positive spillover effect above, and in this case, the purchase decision could be less influenced by price. The estimation equation is as follows:!"#$%!!" =!! +!!!"#$!!" +!!!"#$%&!!" +!!!"#$!!"!"#$%&!!" +!!!!" +!! +!!" (3) 14

where!"#$!!" is the final price assigned to household! in village! after government subsidy, which ranges from 1.8 RMB to 7.2 RMB, with a total of seven different prices. The coefficient of the interaction term tells us whether social network effect can substitute part of the subsidy effect. Results are presented in table 4. Column (1) shows that reducing subsidies by 10% (1.2 RMB) decreases take-up by around 10 percentage points. However, according to column (2), the interaction term between social network and price is significantly negative, which suggests that households are less sensitive to subsidy reduction if they have more friends receiving financial education. Specifically, having one additional close friend receiving financial education mitigates the price effect by around 0.152*0.2/0.166 = 18%. It is easier to see this effect by comparing the demand curve between households who have more friends receiving financial education (measure of social network higher than sample median) and those who have less friends financially educated (measure of social network lower than sample median) in figure 3. It is clear that the insurance demand curve is higher and flatter, especially under high prices, when you have a high proportion of friends exposed to financial education in intensive sessions. In summary, the above results tell us that providing financial educations to introduce insurance knowledge improves initial insurance take-up significantly. More importantly, we find a large and significant social network effect on insurance adoption, having one more friends attending financial educations catches almost 45% of the first order education effect, and mitigates around 18% of the negative price effect. 4.1.2. What information did social networks conveyed? A natural question to ask now is why do social networks matter? What did people actually learned from their friends? As indicated by results in the above section, the social network effect on improving short-run insurance take-up is large, so separating different channels is crucial for us to provide more concrete policy recommendations on how to improve the initial participation rate. Generally speaking, social networks may influence technology or financial product adoption because individuals care about other people s decisions (Bandiera and Rasul, 2006; Banerjee, 1992; Rogers, 1995; Ellison and Fudenberg, 1993), or because people learn about how to use the technology from their friends (Duflo and Saez 2003; Munshi and Myaux 2006; Miguel and Kremer 2007, Oster 15

and Thornton 2009), or because social networks affect individual perceptions about the values or benefits of a product (Miguel and Kremer, 2007; Kohler, Behrman, and Watkins, 2001; Oster and thornton, 2009). In our case, since insurance is a financial product rather than a technology, people do not need to learn how to use it, and thus we do not have to consider the learn to use channel and only need to consider the importance of two types of information which can be possibly conveyed by social networks and influence people s behavior: insurance knowledge and purchase decisions. If social network improves take-up because it helped diffusing insurance knowledge, it means insufficient knowledge or understanding of insurance impairs adoption, then providing financial education would be crucial; if the network effect is driven by the fact that farmers are influenced by their friends behavior, which can be because of trust (famers have larger negotiating power with insurance company if more of them purchase together), imitation (farmers want to act like each other), or informal risk-sharing (your decision depends on whether households from whom you borrow or households to whom you lend to purchased it), then using some lowcost marketing strategies to guarantee a high adoption rate on pilot clients could significantly improve take-up of follow-up customers. First, to test the insurance knowledge mechanism, i.e. social networks helped diffuse insurance knowledge, which improves farmers understanding of insurance benefits, and thus raises the take-up rate, we compare the magnitude of the financial education effect on insurance take-up and knowledge between 1 st round (T1 v.s. T2) and 2 nd round sessions (U1 v.s. U4, no take-up information revealed). Intuitively, if late participants can learn insurance knowledge from early participants during the time interval between the two round sessions, regardless of whether they were assigned to the simple or intensive session in the 2 nd round, we should see a smaller effect of financial education in the second round. Estimation equations are as follows:!"#$%!!" =!! +!!!"#$"%&'!!" +!!!"#$%!!" +!"#$%#&'(!!"!"#$%!!" +!!!!" + +!! +!!" (4)!"#$%&'(!!" =!! +!!!!"#!$%&!!" +!!!"#$%!!" +!!!"#$"%&'!!"!"#$%!!" +!!!!" +!! +!!" (5) 16

Where!"#$%!!" is a dummy variable which equals one if a household is assigned to one of the two 2 nd round sessions, and!"#$%&'(!!" is a measure of insurance knowledge, which equals the score that a household got in the ten questions that we asked during the household survey to test farmers insurance knowledge. Results are presented in table 5. Columns (1) and (2) show that while financial education raises take-up rate significantly in the 1 st round (14 percentage points), it makes almost no difference in the second round. Similar patterns on insurance knowledge can be found in columns (3) and (4). Moreover, while the levels of insurance take-up and insurance knowledge between the two second round sessions are very similar, it is significantly higher than that of 1 st round simple sessions. Specifically, to test whether households who have more friends attending 1 st round financial education have better insurance knowledge, we estimate the following equation:!"#$%&'(!!" =!! +!!!"#$%&!!" +!!!!" +!! +!!" (6) According to column (5), having one additional friend assigning to 1 st round intensive session improves your own level of insurance knowledge by 7 percentage points. These results support the argument that during the time interval between the two round sessions, farmers communicated the insurance knowledge they learned, so that 2 nd round participants already know what will be presented before they come to the session, and such informal learning of insurance knowledge increases insurance take-up significantly. Second, to estimate whether social networks conveyed the purchase decision, we test the effect of other people s decisions, i.e., overall take-up rate in 1 st round sessions, and your networks take-up rate in 1 st round sessions, on 2 nd round participants behavior. Look at the effect of 1 st round overall take-up first:!"#$%!!" =!! +!!!"#$%&'"(!! +!!!"#"$!!" +!!!"#$%&'"(!!!"#"$!!" +!!!!" +!!" (7) Where!"#$%&'"(!! is the overall take-up rate in 1 st round sessions (T1 and T2) in village!, which is a continuous variable ranging from 0 to 1, and!"#"$!!" is an indicator of whether or not we have explicitly revealed the overall 1 st round take-up rate. The hypothesis here is that individuals are more likely to purchase insurance if they see higher take-up rates in previous sessions. However, OLS estimation cannot give a consistent 17

estimation because unobservable variables such as social norms may affect both!"#$%&'"(!! and!"#$%!!". As shown in table 6, default options in 1 st round sessions give significant and substantial variations in 1 st round take-up rates: the average take-up rate of default = BUY sessions is around 12 percentage points higher than that of default = NOT BUY sessions. As a result, we can use default options in 1 st round sessions as the IV for 1 st round overall take-up rates. OLS and IV estimation results are presented in table 7. According to columns (1) and (2), farmers are more likely to buy insurance when the overall 1 st round take-up rate is higher, but the effect is much smaller if we did not explicitly revealed that information. To see this more clearly, we break down the sample and re-estimate the influence of 1 st round overall take-up rate in columns (3) (6). According to columns (3) and (4), if farmers did not received the take-up information from us, they are not influenced by decisions made by 1 st round participants. However, columns (5) and (6) show that increasing the 1 st round take-up rate by 10% and explicitly reveal it to 2 nd round participants can raise their take-up rate by around 4.3%, which is almost half of the first-order effect. To see if results are similar when we focus on effect of close friends behaviors, we estimate the following equation using the sample of 2 nd round participants who did not received take-up information or who saw the detailed decision list from us (U1, U3, U4, and U6 in figure 2.2):!"#$%!!" =!! +!!!"#$%&'"(!! +!!!"#$%&'"($)$(*+,!!" +!!!"#"$!!" +!!!"#$%&'"(!!!"#"$!!" +!!!"#$%&'"($)$(*+,!!"!"#"$!!" +!!!!" +!!" (8) Where TakeupRateNetwork ij represents the take-up rate among close friends of household! who attended first-round sessions in village!. Similar to what we discussed in the last section, both!"#$%&'"(!! and!"#$%&'"(!"!#$%&!!" are endogenous. While we still use 1 st round default option as the IV for overall 1 st round take-up rate, we use two IVs for!"#$%&'"($)$(*+,!!" : Default*Fraction of network in first-round sessions (1 st round take-up rates matter more to you if more of your friends are included), and Fraction of first- 18

round network in intensive sessions 20 (since we expect that intensive sessions raise the takeup rate relative to simple sessions, your friends included in first-round sessions should have a higher overall take-up rate if more of them were assigned to the intensive session). According to results in table 8, we found similar patterns: decisions made by friends in your social network dos not influence your own decision (column (5)). This is not because it is not important as it has a large and significant influence if we explicitly reveal that information (columns (8)), but because natural communication cannot convey such information. As a result, households care about other villagers decisions, especially behaviors of their friends, when they make their own purchase decisions. However, such information cannot be conveyed by social networks, which means that while there s diffusion of insurance knowledge between farmers, farmers cannot learn each others purchase decisions through communication. In summary, the above results suggest that the short-run social network effect is mainly driven by the diffusion of insurance knowledge, as opposed to mechanisms as trust, imitation, or informal risk sharing. 4.2 Effect of social networks on long-run insurance demand Until now we find that social networks improve short-run insurance demand significantly through diffusing of insurance knowledge. However, because insurance is a type product that requires continuous purchase over time, even if subsidies are gradually removed, to generate the hoped-for significant impact on households, it s important to estimate the social network effect over time, and to see whether it helps mitigate price effect in the long-run. In this section, we estimate the impact of social networks, especially social learning of friends experience, on second year insurance demand curve based on survey data from both phases, and payout data from the insurance company. 4.2.1. Social network effect over time In this part, we estimate the overall social network effect in phase 2, which is one year after phase 1. Specifically, we look at the effect of close friends take-up decisions in 20 If among the five friends you listed, 3 of them were invited to first-round sessions and 2 of them were assigned to first-round intensive sessions, then the fraction of network in first-round = 2/5= 0.4, and the fraction of first-round network in intensive sessions = ½ = 0.5. 19

phase 1 on your own insurance demand curve in phase 2. The estimation equation is as follows:!"#$%!!"! =!! +!!!"#$!!"! +!!!"#$%&'()'"*!!"! +!!!"#$!!"!!"#$%&'()'"*!!"! +!!!!" +!! +!!" (9) Where!"#$%!!"! is a dummy equal to 1 if household! in village! bought the insurance in phase 2;!"#$!!"! is the final price faced by household! in village! in phase 2, which ranges from 1.2 RMB to 7.2 RMB;!"#$%&'()'"*!!"! is defined as the proportion of close friends who have purchased insurance in phase 1 21 ;!!" includes household characteristics such as age, gender, education of household head, household size, rice production area, and measure of insurance knowledge, risk attitude, and perceived probability of future disasters in phase 2.!! includes village dummies. However, OLS estimation cannot give consistent estimations of the above equation because friends decision in phase 1 is endogenous. According to results reported in section 4.1, default options and 1 st round financial education influence purchase decisions a lot. As a result, here we use two IVs for friends decision in phase 1: %friends in 1 st round sessions * default option, and %friends in 1 st round intensive sessions 22. Table 9 presents the estimation results. First, raising prices (removing subsidies) has a significantly negative effect on insurance demand in phase 2: on average, removing subsidies by 10% decreases insurance demand by around 5.4 percentage points. Second, OLS estimation results in columns (1) and (2) show that farmers are more likely to purchase insurance in phase 1 if more of his friends purchased it in phase 1: having one additional friend buy insurance in phase 1 increases the probability that you purchase it in phase 2 by around 2.3 percentage points, and this result is barely significant. Turning to the IV result, columns (3) show that the two IVs influences friends decisions in phase 1 significantly, however, according to columns (4) and (5), there s no significant effect of social networks in phase 2. Moreover, as presented in column (6), the interaction term between price and network measure is not significant, which means phase 1 decisions 21 For example, if a household listed 5 close friends during phase 1 household survey, and 3 of them purchased insurance in phase 1, then the variable is defined as 3/5 = 0.6. 22 One concern of using %friends in 1 st round intensive sessions as the IV is that this may affect friends financial knowledge, and your own knowledge through information diffusion in networks, this will further affect your phase 2 demand curve. To deal with this, we controlled for households insurance knowledge in phase 2 in this estimation. 20