WPS6546 Policy Research Working Paper 6546 How Does Risk Management Influence Production Decisions? Evidence from a Field Experiment Shawn Cole Xavier Giné James Vickery The World Bank Development Research Group Finance and Private Sector Development Team July 2013 Policy Research Working Paper 6546 Abstract Weather is a key source of income risk, especially but rainfall-sensitive cash crops, particularly among in emerging market economies. This paper uses a educated farmers. This shift in behavior occurs ex ante, randomized controlled trial involving Indian farmers to when realized monsoon rainfall is still uncertain. The study how an innovative rainfall insurance product affects results suggest that financial innovation can mitigate the production decisions. The authors find that insurance real effects of uninsured production risk. provision induces farmers to invest more in higher-return This paper is a product of the Finance and Private Sector Development Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at xgine@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team How Does Risk Management Influence Production Decisions? Evidence from a Field Experiment * Shawn Cole Harvard Business School Xavier Giné The World Bank James Vickery Federal Reserve Bank of New York * We owe a particular debt to K.P.C. Rao for his efforts in managing the field work associated with this study, and to his survey team. Outstanding research assistance was provided by Fenella Carpena, Kevin Connolly, Rajlakshmi De, Lev Menand, Veronica Postal, Jordi de la Torre, and Wentao Xiong. For valuable comments and suggestions we thank the editor Philip Strahan, two anonymous referees, several conference discussants, and many seminar and conference participants. Financial support for this project was provided by the Research Committee of the World Bank. Views expressed in this paper are those of the authors and do not necessarily reflect the opinions of the World Bank, the Federal Reserve Bank of New York or the Federal Reserve System. Supplementary data and analysis can be found on the Review of Financial Studies web site. Send correspondence to Shawn Cole, Baker Library/Bloomberg Center 271, Harvard Business School, 25 Harvard Way, Boston, MA 02163, USA; telephone: (617) 495-6525. Email: scole@hbs.edu. Small entrepreneurial firms around the world are exposed to a wide range of income risks, including recessions, demand shifts, technology shocks, weather, and natural disasters. Reflecting these risks, around one-third of US business establishments fail within two years (Puri and Zarutskie 2012). Risks associated with entrepreneurship may be even greater in volatile emerging market economies. For a risk-averse individual, these uninsured risks can be a significant disincentive to engage in entrepreneurial activity (Moskowitz and Vissing- Jorgensen 2002; Banerjee and Newman 1993). This paper studies a financial innovation designed to mitigate income risk among a sample of Indian farmers located in a semi-arid region where variation in monsoon rainfall is the dominant source of production and income risk. In this context, we study the effects on behavior of a rainfall index insurance policy, which partially insures against a poor monsoon by providing a payout contingent on low measured local rainfall. Our goal is to estimate the effects of the rainfall insurance on real production and investment decisions by farmers. Given that the decision to purchase insurance is typically endogenous, we use a randomized controlled trial (RCT) to elicit the causal effects of insurance provision. At the start of the monsoon, a randomly selected subset of farmers (the “treatment group”) is provided with 10 rainfall insurance policies with a combined market value of approximately 1,000 Indian rupees (equivalent to ca. $20 US at the time of our study). This represents a significant amount of coverage for our sample; the maximum insurance payout of 10,000 rupees (Rs.) is equivalent to about 90% of median household savings. We then study how this insurance provision influences subsequent production decisions such as crop choice and usage of agricultural inputs, compared to a control group promised a fixed payment equal to the estimated actuarial value of the insurance. We find that while insurance provision has little effect on total agricultural investments, it significantly induces farmers to invest more in riskier production activities. In particular, treated farmers increase production of the main cash crops grown in our study areas, castor and groundnut. These crops produce higher expected returns but are also more sensitive to deficient rainfall. We find that insured farmers are more likely to plant these two cash crops, sow more land with them, and devote a larger amount of agricultural inputs to them, relative 1 to uninsured farmers. Quantitatively, the fraction of farmers planting cash crops is 6 percentage points higher for the treatment group (p-value=0.041), a 12 percent increase relative to the control group, about half of whom planted cash crops in the year of our study. The evidence suggests the impact of insurance is primarily on the extensive margin: it has large effects on the decision to sow cash crops, but little effect among the subset of farmers with the highest cash crop investments. These results imply that farmers are underinsured and suggest that financial innovation that helps diversify weather risk can promote entrepreneurial production and risk-taking. We next test whether the average treatment effect varies by household characteristics and find that it is much larger for educated farmers. In contrast, the treatment effect does not vary systematically with other characteristics including farmer wealth, age, knowledge of insurance, trust in the insurance provider, or past experience with the insurance product. To investigate the role of education in more detail, we first confirm that the effect is robust to a variety of specification changes and the inclusion of a battery of additional interaction effects, including a measure of cognitive ability. Examining triple-differences, we show that the treatment effect of insurance on investment is concentrated among farmers who are both educated and expect highly variable yields. This perhaps suggests that educated farmers may be better able to think through the complex interactions between production risk, insurance and agricultural decisions. Using data on the timing of agricultural decisions, we find that the effects of insurance provision on production decisions occur ex ante, prior to the end of the insurance coverage period, when the insurance payout and monsoon rainfall are still uncertain. We also conducted a second follow-up survey in the following year, after insurance payouts were disbursed, to study how insurance payouts were ultimately spent by farmers. While the statistical power of this analysis is relatively low, our results suggest that payouts were mainly saved for the next monsoon or used to pay down high interest-rate sources of credit. 2 A. Related literature Although insurance is a key function of the global financial system, we have only a partial understanding of how insurance provision causally influences real economic behavior and risk- taking. This study is among the first in a small body of recent research which uses a RCT approach to study the causal relationship between insurance and agricultural decisions. Karlan et al. (2014) randomly allocate cash grants and discounted insurance, or both, offered by a nongovernment organization to farmers in Ghana, finding that while both increase investment, a given subsidy directed towards insurance induces a larger increase in investment than an equivalent credit subsidy. Mobarak and Rosenzweig (2013) conduct a randomized evaluation that uses subsidies to induce households to purchase insurance against late monsoon arrival. While their focus is the interaction between insurance demand and informal risk-sharing, they also find evidence that insured households plant riskier varieties of rice, although planting may happen after knowledge of the payout. In a related study, Cai et al. (2015) find evidence from China that hog insurance increases investment in hogs. Finally, Emerick et al. (2016) study the introduction of a drought-resistant rice variety that both increases average yields and reduces the sensitivity of yields to weather, finding that the rice variety increases land cultivated, fertilizer usage, and the use of a more labor intensive planting method. Several features distinguish our paper from this complementary literature. First, our design allows us to study how insurance affects the timing of decisions and to distinguish clearly between ex ante effects of insurance (effects on behavior during the insurance coverage period) and ex post effects due to the receipt of insurance payouts or the anticipation of future payouts. Second, we equate the wealth effect across experimental arms by compensating the control group so that differences in behavior between the treatment and control group are due to the state-contingent nature of the insurance. Third, we explore heterogeneity in treatment effects, documenting that behavior change is present primarily among those with relatively higher levels of education, and take care to rule out that this effect is driven by other factors, such as wealth or cognitive ability. Finally, we present evidence on how insurance payouts are used. Taken together, this nascent literature demonstrates in a variety of institutional and economic settings that access to insurance leads to an increase in risky production activities. 3 More generally, our analysis contributes to the vast literature on risk, growth, and technology adoption in emerging market economies (e.g., Acemoglu and Zilibotti 1997; King and Levine 1993; Banerjee and Newman 1993; Rosenzweig and Binswanger 1993). While technological improvements from the “Green Revolution” such as high-yield crops and chemical fertilizer has dramatically increased global agricultural productivity, traditional practices still prevail in many areas (Duflo, Kremer and Robinson 2008; Hazell 2009; Foster and Rosenzweig 2010). Our evidence suggests that limited insurance against production risk is one reason why firms limit investments that produce high expected returns but involve risk. Correspondingly, financial innovation that “completes” missing markets, like the insurance policy we study, may boost risk-taking and technology adoption. This channel may account for part of the link between finance and growth identified in prior research (Levine 2005; Beck, Levine and Loayza 2000). 1 Our evidence is also related to the financial economics literature on the link between finance and entrepreneurial activity. This literature tends to focus on developed economies and on credit rather than insurance (e.g., Adelino, Schoar and Severino 2015; Hurst and Lusardi 2004; Black and Strahan 2002). More closely related, Fan and White (2003) find evidence of greater business ownership associated with the option to declare bankruptcy, a procedure that allows entrepreneurs to shield future income and some assets from creditors, limiting the downside risk of entrepreneurship. Recent research by Campello et al. (2011) and Pérez- Gonzáles and Yun (2013) finds causal evidence that risk management affects investment and firm value, although these papers study large, financially sophisticated corporations rather than the sole proprietor farmers considered here. Finally, the household finance literature is increasingly examining the possibility that households may not make optimal decisions relating to complex financial products (e.g., Campbell 2006). Our finding that changes in behavior are concentrated among educated 1 Our analysis is also related to research studying the effects of climate change on agricultural productivity. Guiteras (2009) estimates that predicted climate change from 2010-2039 will reduce crop yields by 4.5-9 percent. While rainfall insurance cannot, of course, affect the climate, it may enable farmers to continue producing risky crops in the face of greater climate variability, mitigating the real impact of climate change. 4 farmers contributes to a growing body of evidence that education affects financial decision- making (e.g., Cole et al. 2014). 1. Background and experimental design Consumption risk-sharing, though surprisingly effective in mitigating nonsystematic income shocks (Cochrane 1991; Townsend 1994), has been found to be incomplete, particularly for spatially correlated shocks such as weather. Droughts, for example, have significant negative effects on economic well-being and health for rural households in India and other emerging market economies, suggesting that the risk of drought is underinsured (Burgess et al. 2013; Maccini and Yang 2009; Jayachandran 2006; Rose 1999; see Cole et al. 2013, for further references). When consumption cannot be fully insured against drought or other income risks, individuals may respond by smoothing income ex ante, selecting production and investment activities that generate less volatile income at the cost of lower average income (Morduch 1995; Gollier and Pratt 1996; Walker and Ryan 1990). Corporate finance research makes an analogous prediction for firms: in the presence of financial constraints, a firm facing non- diversifiable risks may invest less in additional risky projects, particularly when the project return is positively correlated with existing risk exposures (Froot and Stein 1998). Income- smoothing tactics for farmers include intercropping by drought tolerance, spatial separation of plots, shifting the timing or staggering of planting, moisture conservation measures such as bunds, furrows and irrigation, and diversifying income between agricultural and non- agricultural sources. Several papers find suggestive evidence of costly income smoothing by farmers in developing countries (Rosenzweig and Stark 1989; Morduch 1995; Dercon 1998; Dercon and Christiaensen 2011). 2 2 Rosenzweig and Stark find that farmers with more volatile profits are more likely to have a wage-earning household member. Morduch suggests that households close to subsistence devote a larger share of land to safer crop varieties. Dercon finds Tanzanian farmers with more liquid assets engage in higher-risk agricultural activities. Dercon and Christiaensen find that fertilizer purchases are lower among poorer Ethiopian households, in part due to their lesser ability to smooth shocks ex post. Rosenzweig and Binswanger (1993) estimate that a one standard deviation increase in the variability of monsoon onset would, through reduced risk-taking, reduce agricultural profits by 15 percent for the median farmer. An advantage of the present study relative to this prior research is that our RCT design exogenously varies the farmer’s exposure to rainfall risk, ameliorating concerns about omitted variables. 5 The key hypothesis tested in this paper is that the provision of insurance against rainfall risk will induce households to allocate more resources to higher-risk, higher-yield investment and production activities. To fix ideas, Section OA1 of the Online Appendix to this paper illustrates this hypothesis using a simple model in which a risk averse farmer chooses between two production activities: one safe, the other higher-yielding but risky. Insurance against production risk induces farmers to allocate more resources to the high-risk activity. While our model uses a simple CARA-normal setup, this basic prediction will apply to nearly any model with risk-averse agents, incomplete markets and production risk. We test this hypothesis in a setting where firms face a dominant, exogenous source of production risk: variation in local monsoon rainfall. Rainfall is cited as the most important source of risk by 89% of farmers in our study areas. Although these local rainfall shocks are approximately uncorrelated with aggregate asset returns, farmers in our sample have a large, non-diversified exposure to local weather risk. Recognizing the importance of rainfall risk, Indian insurers have recently developed innovative retail index insurance products designed to pay out when realized monsoon rainfall is poor. We study a particular policy designed and underwritten by ICICI Lombard, a large, privately-owned, national Indian insurance firm. Our analysis builds on a series of experiments and surveys that we have conducted since 2004 in Andhra Pradesh, India (Cole et al. 2013; Giné, Townsend and Vickery 2008). This previous work focuses on the determinants of rainfall insurance demand and the barriers to widespread insurance uptake, rather than the impact of insurance on behavior. 3 A. Crop choice and risk-taking Our empirical analysis focuses on the allocation of agricultural inputs by farmers across crop types with different levels of risk. During the main cropping season (June to November) in our study areas, farmers grow a variety of cash and subsistence crops that vary by sensitivity to low rainfall. The primary cash crops grown in the study areas are castor and groundnut, two rain-fed oilseeds, as well as paddy, which is almost exclusively irrigated and thus less subject 3 Although some of our earlier research also adopts a field experimental approach (Giné and Yang 2009; Cole et al. 2013), uptake has been too limited to allow an assessment of the impact of insurance on investment decisions. Two related laboratory experiments conducted by Lybbert et al. (2010) and Hill and Viceisza (2012) suggest that, over time, subjects learn about insurance and change behavior accordingly. 6 to rainfall risk (84% of paddy plots in our empirical sample are irrigated). The main subsistence crops grown in the area are sorghum and legumes (red gram, pigeon pea, and, to a lesser extent, green gram). Cultivation costs for the main cash crops exceed those of subsistence crops and range between Rs. 5,000 and Rs. 9,000 per hectare ($94 to $168 US) if the recommended amounts of organic and inorganic fertilizer are applied. 4 Based on the local District Handbook of Statistics, average yields for castor are 600 Kg per hectare if fertilizer is used, which would generate Rs. 10,896 in revenue at 2009 prices. Groundnut yields are 540 Kg per hectare with fertilizer, corresponding to Rs. 11,702. Sorghum yields with fertilizer are 700 Kg per hectare or Rs. 4,788, and red gram yields are 300 Kg or Rs. 5,791. Thus, expected profits for castor and groundnut are indeed higher at Rs. 2,771 and Rs. 2,951 compared to sorghum (negative Rs. 212) and red gram (Rs. 141). In terms of water requirements, castor grown in Mahbubnagar under rain-fed conditions requires 625 mm of accumulated rainfall over the season if sown around the normal planting date, while groundnut in Anantapur requires 533 mm. Red gram requires a similar amount of accumulated rainfall, 523 mm, but, in contrast, sorghum only requires 376 mm, and green gram 278 mm. 5 To summarize, castor and groundnut are more profitable on average than other crops grown in our study areas but have higher water requirements and therefore are more sensitive to drought. B. Product description The rainfall insurance policies offered in this study are an example of “index insurance,” that is, a contract whose payouts are linked to a publicly observable index like rainfall, temperature or a commodity price. Unlike traditional insurance products, index insurance is not generally subject to moral hazard or adverse selection problems, because payouts are linked to an exogenous, publicly observable variable, in this case, rainfall measured at a local rain gauge. Index insurance also involves lower administrative costs because no claims verification 4 Input recommendations (used to calculate 2009 production costs per hectare for castor, groundnut, sorghum, and red and green grams) come from the University of Agricultural Sciences in Bangalore (1999). 5 These water requirement statistics are drawn from personal communication from Dr. Bodapati Rao and Dr. Vijay Kumar, Principal Scientists at the Indian Central Research Institute for Dryland Agriculture (CRIDA). 7 process is required. However, rainfall insurance only covers rainfall-related losses and may entail significant basis risk, especially if the household is located too far from the relevant weather reference station. 6 Information frictions and high transaction costs have limited the commercial success of agricultural insurance. Insurance companies have initiated a number of index insurance pilots in recent years in the hope of developing a financially sustainable product that farmers will buy (World Bank 2005; Skees 2008). Today, rainfall insurance is one of the core product offerings of Indian agricultural insurance providers, with over 10 million farmers covered by index policies. Clarke et al. (2012) and Giné et al. (2012) provide sales data and non-technical overviews of this market as well as further institutional details. For the ICICI Lombard policies we study, payoffs are calculated based on measured rainfall at a nearby government rainfall station or an automated rain gauge operated by a third- party vendor. ICICI Lombard offers separate policies for three different contiguous phases of the monsoon, of 35-45 days in length, corresponding to sowing, flowering, and harvest. This study offered only Phase I policies, which cover the first and most critical period of the monsoon. The start date of the policy is defined as the first date at which cumulative rainfall since June 1 reaches 50 mm. The start date defaults to July 1 if June rainfall is below 50mm. Payouts are determined based on cumulative rainfall during the 35 days following the start date. The policy pays out if cumulative rainfall during this coverage period is below a threshold known as the “strike”, designed to approximate the minimum quantity of rainfall required for successful crop growth. Payouts are linear in the rainfall deficit relative to the exit or are equal to a fixed maximum amount of Rs. 1000 per policy if rainfall is below a second, lower threshold called the “exit.” As an example, Figure 1 plots cumulative measured rainfall for the insurance policy indexed to rainfall at the Naryanpet weather station. The two dotted horizontal lines represent the strike (top) and exit (bottom) levels specified in the policy. As one can see, realized rainfall 6 In our study, villages are generally located within 10 km of the reference weather station. Given the relatively flat terrain, basis risk may be relatively low for our sample, although we do not have data to test this directly. 8 was low in June, and cumulative rainfall did not reach the trigger of 50 mm. Thus, the 35 day policy coverage period started automatically on July 1. Cumulative rainfall then quickly crossed the exit (5mm) level but only reached 16 mm during the coverage period, well below the strike of 50mm. Each policy paid out Rs. 10 for each millimeter of the rainfall shortfall, or Rs. 10 x (50-16) = Rs. 340. This led to a total payout of Rs. 3,400 per farmer, since each treated farmer received ten policies. [Insert Figure 1 here] Payouts are linked to rainfall measured at a nearby gauge, and treated farmers in our study received policies linked to one of five weather stations, depending on their village. Because the 2009 monsoon turned out to be significantly below average, three of these five policies provided positive payouts ex post, with one policy providing the maximum payout of Rs. 1,000 per policy, amounting to a total payout of Rs. 10,000 for each treated farmer. See Section 2 for a table of all the realized payouts. C. The insurance experiment Our sample consists of 1,479 farmers drawn from 45 villages in two semi-arid districts of southern India, Mahbubnagar in the state of Telangana and Anantapur in the state of Andhra Pradesh. 7 Two-thirds of the sample participated in previous surveys and field experiments we conducted on rainfall insurance; these were originally selected via a stratified random sample of land-owning farmers in 37 study villages in 2004 (see Giné et al. 2008, for details). To improve statistical power for this study, an additional five hundred households were drawn from these 37 villages as well as 8 nearby villages. Figure 2 presents the timeline of events. Each farmer received a home visit from a member of a trained team of enumerators from the agricultural research organization ICRISAT between June 4 and July 13, 2009, coinciding with the onset of the 2009 monsoon season. 8 7 Both districts were part of Andhra Pradesh at the time of the study; Telangana is a new state formed in 2014. 8 Although we planned to distribute all insurance policies before the start of the insurance coverage period, delays in the shipping of policy certificates from ICICI headquarters in Mumbai resulted in 40 percent of the initial visits occurring on or after the policy activation date. Distribution occurred close to the start of the activation period in these cases, however, within five days on average, and only six percent of farmers had started planting by the time of the initial household visit and insurance assignment. The small amount of monsoon investments occurring before the distribution of insurance may mean that our results are slightly attenuated relative to the case where policies were 9 During the visit, the enumerator first conducted a short baseline survey, collecting demographic data and other information. They then explained the recommended fertilizer dosages for castor and groundnut, the two main rain-fed cash crops in the area, as well as the concept of insurance, and gave specific details about the policies offered by ICICI Lombard. [Insert Figure 2 here] The farmer then received a scratch card (similar to the format of a scratch-off lottery ticket sold in the United States), revealing treatment assignment. The key treatment for the purposes of this paper is the assignment of the farmer to either an insurance group (treatment) or a control group. Farmers in the treatment group received a certificate for 10 Phase-I weather insurance policies, similar to those sold in the region in previous years. “Control” farmers received a post-dated check for Rs. 200, equal to our estimate of the actuarially fair value of these 10 policies based on calculations using historical rainfall data. 9 This check could be cashed at the local branch of BASIX, the microfinance institution that sells ICICI Lombard rainfall insurance in our study villages. The control group payment was provided to ensure that differences in behavior between the insurance and control groups reflect the state-contingent nature of the insurance, rather than a wealth effect. The insurance has the same expected value as the fixed payment received by the control group; the key difference is that the realized insurance payout is contingent on low realized rainfall. By design, the date when the check could be cashed also coincided with the expected timing of insurance payouts. This was to ensure that there were no differences in behavior between the treatment and control groups induced by earlier relaxation of liquidity constraints in one group compared to the other. distributed earlier, as earlier distribution would have given farmers more time to adjust behavior in response to receiving insurance coverage. 9 Actuarial values are estimated using historical rainfall data from two Indian Meteorological Department (IMD) weather stations, one in each district, using the approach described in Giné et al. (2008). Historical rainfall data from the other three weather stations are not available because these gauges are maintained by a private vendor rather than the IMD and were only recently installed. The market price of ten insurance policies in our study areas ranges from Rs. 800 to Rs. 1100, significantly exceeding our estimate of average actuarial value. This high markup likely reflects the administration and transaction costs of offering insurance; it is also inclusive of taxes and a loading factor paid to the sales agent. In previous work, we argue that high transaction costs and prices represent one of several barriers to higher index insurance adoption (Cole et al. 2013). 10 A second independent treatment, also provided via the scratch card, involved coupons for discounts on locally appropriate inorganic fertilizer (DAP in Anantapur, NP fertilizer in Mahbubnagar). Unfortunately the implementation of this treatment was largely unsuccessful, due to operational difficulties and the fact that the subsidies did not substantially affect fertilizer purchase behavior. 10 For that reason, we do not study it here, although we always control for the household’s fertilizer discount treatment status in our empirical analysis. 11 Treatments were assigned randomly and independently across households. The use of scratch cards ensured that neither the respondent nor the enumerator had prior information about the household’s treatment status. Farmers also had the option to purchase additional insurance policies from BASIX, although few did so in practice. In October and November 2009, after the growing season, the ICRISAT team revisited each farmer to conduct a follow-up survey, collecting information on agricultural investments and production decisions during the monsoon as well as asset data (e.g., livestock, and financial assets such as savings, loans, and insurance), risk-coping behavior, additional demographic information, and attitudes and expectations regarding weather and insurance payouts. Although payouts had not been made by the time of the follow-up survey, because of the poor monsoon in 2009, 93% of the farmers in the treated group reported in the follow-up survey that they expected to receive a payout. Roughly the same percentage expected final crop yields to be below average. Payouts to the insurance and control group were made in December 2009 and January 2010. This timing is well after one might have expected, given that policies indicate a settlement date of “thirty days after the data release by data provider and verified by Insurer.” However, the timing was relatively consistent with previous years. The long timeframe reflected both slow release of rainfall data and slow processing by ICICI Lombard. 10 The number of subsidized fertilizer bags was calibrated to fertilizer usage from a survey conducted in 2006. According to that survey, 70 percent of farmers in Mahbubnagar and 34 percent in Anantapur had used fertilizer, and users generally purchased at most two bags. However, follow-up data collected in November 2009 revealed significantly higher fertilizer usage than suggested by the earlier survey. 11 In practice, our results are almost identical whether or not we control for the fertilizer treatments, not surprisingly given that the two treatments are statistically independent by design. 11 In mid-2010, we conducted a second follow-up survey to measure how realized insurance payouts were used by farmers. Although not the focus of this paper, our analysis of these data is summarized briefly in Section 3.G and discussed in more detail in Section OA2 of the Online Appendix. 2. Summary Statistics Table 1 reports baseline summary statistics about farmers’ household characteristics, education, insurance knowledge, trust, expectations, credit and assets (panels A-D), and statistics on agricultural investments during the 2009 monsoon (panel E). Baseline statistics are drawn from the initial baseline survey whenever possible. Since logistical constraints limited the length of the baseline survey, a subset of variables were collected using recall questions in the first follow-up survey conducted just after the 2009 monsoon. Respondents in the follow-up survey were asked to report fixed characteristics (e.g., years of schooling) and provide recall data on the value of land and other assets as of June 2009. [Insert Table 1 here] Panel A presents demographic data. The average household has 5.15 members with a 50-year old household head that is usually (91%) male. Panel B reports measures of education for the household head. On average, heads have obtained 3.75 years of schooling, with slightly more than half (54%) self-reporting being “unschooled.” Literacy is low, with only 43 percent and 40 percent of heads self-reporting being able to read and write, respectively. These statistics are similar to those reported in our previous work (e.g., see statistics in Cole et al. (2013), which are based on a 2006 survey instrument). Online Appendix Table OA1 reports additional summary statistics for savings, credit, and assets. Given that insurance provision was randomized, we should not observe systematic differences in baseline characteristics between the treatment and control groups. We confirm this in Online Appendix Table OA2, for demographic characteristics, the household head’s education, knowledge, trust, expectations, financial assets and credit, livestock and other assets including land, and agricultural investments in the prior monsoon. Validating the 12 randomization, we find a statistically significant difference between the two groups at the ten percent level or lower for only one out of 59 variables (the use of non-traditional savings). An F-test of the null hypothesis that all average characteristics are the same for the treatment and control groups cannot be rejected (p-value = 0.67). Panel E presents control group summary statistics for agricultural investments during the 2009 monsoon, drawn from the first follow-up survey. We collected data on area of land sown with cash crops (castor or groundnut) and all crops, the timing of planting, and amounts spent and used for different agricultural inputs. For a subset of inputs, we also measured input usage for cash crops, in addition to total usage. A very high share (97%) of farmers planted some crop, and roughly half (47%) planted cash crops. Fewer farmers planted cash crops in 2009 than 2008, reflecting the poor 2009 monsoon. Also reflecting the poor monsoon, 15% of farmers abandoned their crop during the 2009 monsoon season. Online Appendix Table OA3 presents disaggregated statistics for usage and spending on individual inputs, including seeds, fertilizer, manure, pesticide, irrigation and hired labor. Table 2 summarizes contract details and realized payouts for the five insurance policies (recall that farmers received policies linked to different rainfall stations, depending on their village location). Three of the five policies realized a positive payout, and the 242 treated farmers with insurance indexed to Hindupur station rainfall received the maximum payout of Rs. 1,000 per policy (Rs. 10,000 in total). In Section 3, we use variation in payouts across rainfall stations to distinguish between ex ante and ex post effects of insurance provision. In Section OA2 of the Online Appendix, this variation in payouts is used to help identify how insurance payouts were ultimately used by farmers. [Insert Table 2 here] 3. Estimation results A. Insurance treatment effects Table 3 presents the estimated average treatment effects of insurance provision on farmers’ agricultural decisions during the 2009 monsoon season. We analyze five outcome variables: (i) a dummy equal to one if any agricultural inputs were used during the monsoon, (ii) the log 13 of the acres of land sown, (iii) the log of the market value of agricultural inputs used, and two shares, (iv) the share of total cultivated land and (v) the share of market value of agricultural inputs devoted to cash crops. Panel A (B) reports the estimates for all (cash) crops, respectively. [Insert Table 3 here] Each outcome variable is regressed on a dummy for whether the farmer received the insurance treatment (the key variable of interest), a set of village dummies, a dummy for each fertilizer treatment, and, in some regressions, a set of household characteristics 12. Column 2 reports the mean of the dependent variable for the control group, while Column 3 reports the percentage of observations equal to zero. To conserve space, only results for the key coefficient on the insurance treatment dummy are reported. Coefficients are presented in columns 4 and 5, reported as marginal effects. In Panel A, we find a positive, although not statistically significant, effect of the insurance treatment on the quantity of inputs used or the area of land cultivated. However, when the analysis is restricted to castor and groundnut investments in Panel B, the treatment effects become much larger and are also statistically significant at the 5% level or lower in each specification. Quantitatively, assignment to the insurance treatment group increases the probability of planting cash crops by 6 percentage points (or 12 percent). We estimate an increase in ln(1+land planted for cash crops) of 0.086 (Tobit marginal effect). 13 As shown in the “cash crop shares” results in panel B, assignment to the insurance treatment group increases the share of total cultivated land devoted to cash crops by 4.7 percentage points and the share of inputs (measured by market value) by 3.4 percentage points. 12 We use probit for binary outcomes and tobit for censored outcomes. As a robustness check, we also estimated Table 3 using a linear probability model (see Online Appendix Table OA4). Results are similar. Note that Fernandez-Val (2009) shows that estimates of marginal effects based on a probit are either unbiased or exhibit negligible bias, even though probit and tobit estimators generally produce biased estimates of the structural model parameters in the presence of fixed effects. 13 Online Appendix Table OA5 also reports regression results for cash crop usage split up by individual agricultural inputs type (hybrid seeds, improved seeds, fertilizer, etc.). The disaggregated treatment effects are positive in each regression, although because of the much lower power, not usually statistically significantly different from zero. 14 Column 5 of Table 3 reports results controlling for three household characteristics: age, years of education, and wealth. Adding these controls has little effect on the estimates, consistent with the random assignment of farmers to the treatment and control groups. To summarize, we find significant increases in the quantity of cash crop investments by farmers randomly assigned to receive rainfall insurance policies and the share of total investments directed to cash crops. The effects on total agricultural investments, while positive, are not statistically significant. This latter result could be consistent with the presence of fixed short-run production factors (e.g. a given amount of arable land owned by the farmer, which cannot be easily adjusted in the short run) or the presence of financial constraints. 14 It may also simply reflect our limited statistical power. The estimates in Table 3 represent local average treatment effects. Figure 3 instead plots the cumulative distribution function of investment in cash crops by insurance treatment status. This plot suggests that the effect of the insurance treatment is quite non-linear. Insurance causes a sizeable number of farmers to switch on the extensive margin from not growing cash crops into growing cash crops, consistent with the probit regression estimates. But for farmers in the top part of the distribution of cash crop investments, insurance provision has little or no effect on cash crop inputs used. In other words, the provision of insurance appears to primarily affect the extensive margin of investment decisions. [Insert Figure 3 here] Figure 3 also reveals that there is a discrete jump in the level of cash crop investment once the farmer decides to invest a positive amount. This points to the presence of scale economies; farmers do not sow a given crop below a minimum scale. Around this decision threshold, the provision of insurance against income risk makes farmers more willing to invest a positive amount in castor and/or groundnut. According to our data, the minimum area cultivated under cash crops is 0.5 acres, accounting for 10 percent of average landholdings. 14 Table OA6 of the Online Appendix analyzes the effect of the insurance treatments on various measures of the take- up of credit during the 2009 monsoon, when agricultural investment decisions were made. We find that treated farmers are no more likely to take up credit (of any form) than farmers in the control group, suggesting that being insured did not relax credit constraints. 15 For farmers planting cash crops, the median area under cash crops cultivation is 3 acres (70 percent of landholdings). B. Heterogeneous treatment effects In Table 4, we test for heterogeneity in the insurance treatment effect along several dimensions, including household wealth, age, education, knowledge, trust and payouts. We estimate regressions of the form: outcome = f(a + b. insurance + c. characteristic + d. insurance x characteristic + … + e), where “insurance” is a dummy equal to one if the farmer was assigned to the insurance treatment group, and “characteristic” is the source of heterogeneity of interest (e.g. wealth, age, education etc.). Our primary interest is the coefficient d on the interaction term. Table 4 presents results using the dummy variable indicating whether the farmer plants cash crops as the outcome variable. Column 1 of Table 4 reproduces the estimate from Table 3. Columns 2 – 7 include the six different characteristics, one at a time, and their interaction with the insurance treatment dummy. Column 8 includes all the characteristics and interactions together. The top part of the table reports the estimated interaction effects, while the uninteracted effects are reported below. As before, coefficients are reported as marginal effects. [Insert Table 4 here] In Column 2, we study how the insurance treatment effect varies with wealth, measured by an index constructed as the first principal component of asset holdings (see the Appendix for details of variable construction for this index and selected other variables used in this study). It is unclear theoretically what effect to expect. On one hand, wealthy farmers may already have informal insurance arrangements or the ability to smooth temporary income shocks (e.g., because they have sufficient liquid assets or access to credit), reducing their need for rainfall insurance, as in Mobarak and Rosenzweig (2013), or may be locally less risk averse, if utility exhibits decreasing absolute risk aversion. On the other hand, wealthy farmers may find it easier to adjust agricultural practices in response to a shift in the risk-return frontier due 16 to insurance (e.g., because they are less financially constrained). Empirically, we find that the treatment effect is decreasing in wealth, but the relationship is not statistically significant. The direct effect of wealth is, however, positive and statistically significant at the 1 percent level; in other words, as expected, wealthy farmers are much more likely to invest in cash crops. Column 3 considers heterogeneity by age of the household head. The interaction term is economically small and not statistically significant. Column 4 considers heterogeneity in treatment effects by educational attainment, measured by years of education of the household head. Strikingly, we find positive, economically important, statistically significant interaction effects, implying that the treatment effect of insurance provision is concentrated amongst educated farmers. Quantitatively, an additional year of education increases the effect of the insurance treatment on the probability of planting cash crops by 1.8 percentage points, significant at the 1 percent level. Furthermore, the treatment effect is not statistically or economically different from zero for farmers without formal education. We further investigate the role of education in shaping farmers’ response to insurance in Section 3.C below. Column 5 examines whether the treatment effect varies with the farmer’s understanding of how insurance works, measured by the number of correct answers to questions on the circumstances under which a payout would be received, and awareness of the product, while Column 6 tests whether the treatment effect varies with the farmer’s trust in Basix, the insurance vendor. Interestingly, unlike education, neither the knowledge or trust interaction terms are statistically significant. Column 7 uses ex post realized payouts as the interaction variable. This provides a test of whether farmers’ investment responses might reflect their expectation of receiving a high payout in the future (e.g. because of early information that the monsoon is likely to be poor). If true, this would change the interpretation of our results, since it would imply that our treatment effect reflects factors other than just the hedging benefits of insurance. This interaction variable is quantitatively small and not statistically significant, implying that the investment response is not driven by this anticipation effect. 17 Finally, Column 8 includes all the interaction variables jointly rather than one at a time. Consistent with the results from columns 2 – 7, only the heterogeneous treatment effect by educational attainment is statistically significant. The coefficient on years of education x treatment increases slightly compared to Column 4, from 0.018 to 0.024, and remains statistically significant at the 1 percent level. We repeat this analysis using two other dependent variables, the log area of land planted with cash crops and the log value of the investment in cash crops. To conserve space, results are presented in tables OA7 and OA8 of the Online Appendix. Results are very similar to Table 4. The treatment effect of insurance is economically much larger for educated farmers, statistically significant at the 1 percent level. None of the other interaction variables are statistically significant We also analyze heterogeneity in the insurance treatment effects by other characteristics measured at or before baseline (results reported in Table OA9 of the Online Appendix), including current or recent indebtedness (as a measure of financial slack); wealth and wealth squared (to measure potential nonlinear or threshold heterogeneous treatment effects by wealth); landholdings (as an alternative measure of wealth); a threshold level of land (to test whether farmers with small plots may be less able to switch to cash crops because of minimum scale effects); two measures of experience with insurance: a dummy for whether the farmer previously purchased rainfall insurance and an instrumented dummy for whether the farmer purchased insurance in 2006 using the randomized treatments from Cole et al. (2013) as instruments for purchase; and a measure of the subjective variance of expected agricultural yields as self-reported by the farmer. None of these interaction effects is statistically significant, either when included one at a time or all together. Summing up, the main source of heterogeneity that we are able to identify given the power of our statistical tests is the farmer’s level of educational attainment. Consistent with this result, Karlan et al. (2014) finds in a different setting that insurance has larger effects on investment when the household head can read. Related, some previous research finds evidence that education is positively correlated with take-up of rainfall index insurance or an insured 18 loan (Cole et al. 2013; Giné and Yang 2009). 15 We note that while the insurance treatment is randomly assigned, education, of course, is not. Thus, our results could reflect omitted variables that are correlated with educational attainment but not captured by age, wealth, trust, or the other variables included in Table 4 or Table OA9. Although we cannot fully rule out the presence of such omitted variables, we do view the education result as quite striking, given its significance and robustness to which other interaction terms are included in the specification. We turn to a more detailed analysis below. C. Further Analysis of Education Heterogeneous Treatment Effects Table 5 presents further analysis to investigate why education is so important in shaping farmers’ responses to the insurance treatment. We test whether the heterogeneous treatment effects (HTE) by education are limited to particular subsets of the sample (e.g., farmers with more land or with bank debt), investigate the effects of adding interactions between the insurance treatment and other measures of cognitive skills, and study the functional form of the education HTE. We regress a dummy equal to one if the farmer planted cash crops on a set of education variables and interaction terms of interest, controlling for all the farmer characteristics and interaction terms from the multivariate regression in Column 8 of Table 4 (e.g., wealth, treatment x wealth, age, treatment x age, and so on). We do this to minimize concerns about omitted variable biases. Our main results are robust to whether or not these additional controls are included. [Insert Table 5 here] Column 1 of Table 5 reproduces Column 8 of Table 4 for ease of reference. Column 2 examines functional form: it adds to the specification from Column 1 two “step function” dummies measuring categories of educational attainment (1-5 years of education and 6+ years of education) and their interactions with the insurance treatment. The linear education interaction variable (years of education x treatment) remains statistically significant, while the 15 An insured loan in Giné and Yang (2009) is a debt contract bundled with an insurance policy with a maximum payout equal to the principal and interest to be repaid. 19 two additional interaction terms are not significant, either individually or jointly. In other words, we cannot reject the null that the effect of insurance on cash crop investments increases linearly with years of education. While our statistical power is limited, this result speaks against the hypothesis that our effect only appears beyond a minimum threshold level of education, especially because the point estimate on the 6+ years dummy is negative. Columns 3 and 4 include interactions between the insurance treatment and two other measures of cognitive skills: a dummy for whether the farmer self-reports that they can read and the farmer’s score in a short three-question Raven test, a well-known nonverbal test of analytic intelligence based on pictograms (Raven 2000; Pind, Gunnarsdottir and Johannesson 2003; Carpenter, Just and Shell 1990). The goal of this analysis is to test whether our result is driven by a particular facet of education, either literacy or analytic reasoning. In both cases, the additional interaction variable is positive, although small and not statistically significant, while the coefficient on education x treatment barely changes in size. This suggests that the education HTE is not driven narrowly by either of these dimensions of cognitive skill. Columns 5, 6 and 7 include triple-interaction terms to study whether education is a necessary but not sufficient condition for insurance provision to induce changes in production behavior. We split the sample in turn by: (i) landholdings (above vs below median), (ii) usage of bank credit at baseline (any vs none), and (iii) farmers’ self-reported estimates of the standard deviation of agricultural yields, a measure of how risky the farmer perceives the environment to be (above vs below median). 16 We then interact the education interaction term with this additional variable; e.g., Column 5 includes treatment x education x landholdings > median and treatment x education x landholdings < median. Specifications also include all the relevant non-interacted variables and single and double interactions. Looking at columns 5 and 6, the point estimates on the “above median landholdings” and “any bank credit usage” triple interaction variables are positive and significant, respectively, while the corresponding coefficients for the “below median land” and “no bank credit” groups are about half as large and not statistically significant. However, our estimates 16 We compute this variable by eliciting from each farmer a histogram of the distribution of agricultural yields. Specifically, each farmer was asked to arrange a set of 10 stones across different ranges for agricultural yields to indicate the relative likelihood of different outcomes. See the Appendix for more details. 20 are not precise enough in either case to reject the null that the coefficients on the two triple interaction terms are equal, as reported at the bottom of Table 5. In other words, the heterogeneity in treatment effects by education is more pronounced for farmers with more land or access to credit but not statistically significantly so. However, in Column 7, we find evidence that the insurance treatment effects are concentrated in the subset of farmers that are both educated and believe that agricultural yields are highly volatile. For such farmers, an additional year of education increases the probability of planting cash crops by 4.4%, statistically significant at any conventional level (p < 0.001). This is more than five times the coefficient for the “low variance of yields” group of 0.7%. Unlike columns 5 and 6, we can reject the null that the treatment effect is the same for the two groups; the difference between these two coefficients is statistically significant at the 1 percent level, as shown at the bottom of Table 5. Although we do not want to over-interpret this finding, it is suggestive that educated farmers may be better able to “think through” the relationship between insurance and riskiness of agricultural investments and the implications for optimal risk-taking. To conserve space, Table 5 focuses only on one dependent variable: a dummy for whether the farmer made any cash crop investments. Tables OA10 and OA11 of the Online Appendix repeat the analysis using the two other dependent variables from Table 3: the area of land planted with cash crops and the value of the investment in cash crops. Results are similar to those presented above. Taken together, our results imply that education more broadly, rather than just literacy or analytic intelligence or even prior knowledge about rainfall insurance, is important for determining whether farmers change production behavior when insured. Our data do not allow us to pin down the exact mechanism within education. We speculate that education may help teach farmers to solve problems and evaluate unfamiliar situations and that a well-educated farmer, even if unfamiliar with a specific financial product, will be better able to learn about the product as needed once they receive it and to logically reason how access to the product should influence other decisions. Our evidence that the insurance treatment effect is concentrated among farmers who are both educated and view yields as highly volatile suggests 21 that education helps farmers think through the complex relationship between production risk, insurance, and agricultural decisions. It would, of course, be interesting to conduct similar analysis in other settings, both to test the external validity of our findings and to further “unpack” the mechanisms underlying the education results presented here. Furthermore, we find that subjective perceptions of agricultural risk affect educated farmers’ responses to insurance provision; since these perceptions may reflect a combination of objective information and individual beliefs, it would be informative to test whether similar results hold using objective measures of production risk (e.g., exploiting differences in topography or weather volatility across regions). If results are indeed similar, it would strengthen the case that insurance provision promotes risk-taking in environments in which underlying income risks are extreme. More generally, our finding that education affects farmers’ responses to insurance provision has potentially interesting implications for the distributional effects of financial innovation. Specifically, new financial instruments that change the tradeoffs between risk and return may increase income inequality by educational attainment, at least during a transition period, if educated households are more likely to change behavior in response to the change in the feasible set of risk-return outcomes. 17 D. Timing Figure 4 presents evidence on how the insurance treatment affects the timing of cash crop investments. This figure is constructed by estimating regressions similar to Table 3, tracing out how the insurance treatment affects the probability of planting cash crops by different points in the monsoon season. Specifically, each point on the graph represents the marginal effect from a probit regression, where the dependent variable is equal to 1 if the farmer had planted cash crops by date t. The explanatory variables are the insurance treatment dummy and the other controls from Column 4 of Table 3. Vertical lines indicate the period in which insurance policies were distributed and the end of the time period covered by any of the five insurance policies. 17 See Townsend and Ueda (2006) for a model-based quantitative evaluation of the relationship between economic growth, financial deepening, and inequality in an emerging market context (the Thai economy between 1976-1996). 22 [Insert Figure 4 here] As expected, the insurance treatment effect is close to zero at the point when the insurance policies are randomly allocated to farmers. The cumulative treatment effect by date then rises sharply, becoming statistically significant by the end of the insurance coverage period. It subsequently flattens out and converges to the point estimate from the average treatment effect regression in Table 3. To summarize, this analysis shows that the effects of the insurance treatment on behavior occur during the planting season, prior to the end of the insurance coverage period and several months before the insurance payout is received. This suggests that insurance induced farmers to take riskier production decisions during the planting season in the knowledge that they would be partially hedged in the event of a poor monsoon. An alternative hypothesis is that our results are responses to the ex post receipt of insurance payouts or the anticipation of future payouts. This hypothesis is not consistent with the timing of the behavioral response in Figure 4, however; also speaking against this explanation, our analysis in Table 4 finds no correlation between treatment effects of insurance and realized ex post payouts. In other words, the insurance treatment effects we observe appear to reflect the relaxation of a risk constraint, rather than a wealth effect or relaxation of credit constraints. E. Qualitative self-reported changes in behavior Complementing this statistical analysis, the follow-up survey conducted after the 2009 monsoon simply asked farmers in the insurance treatment group to self-report whether and how the provision of insurance affected their investment behavior. We asked farmers whether the knowledge of being insured led to an increase, decrease, or no change in the amount of fertilizer, seeds, and other inputs they used, and whether it influenced decisions about planting, replanting and/or abandoning crops. Responses are summarized in Table 6. [Insert Table 6 here] Between 36-52% of farmers report not changing their behavior, depending on the question. Among the remainder, a significantly larger fraction reported increasing agricultural input usage as opposed to reducing it. This was true for five of six inputs; e.g., 50% reported 23 using more fertilizer, while only 14% reported using less. The exception was bullock labor (23% more, 29% less). Farmers also report that it influenced them towards planting earlier (26%, compared to 5% who report being influenced to plant later) and against abandoning crops. We view this evidence as suggestive at best, given the well-known biases associated with responses to subjective survey questions (Bertrand and Mullainathan 2001) and the difficulty farmers may have in introspectively assessing what their behavior would have been in the counterfactual situation in which they did not have insurance. Bearing these caveats in mind, the direction of farmers’ responses seem consistent with our econometric evidence that insurance induces investment in risky agricultural activities. F. Additional robustness checks Three additional sets of robustness checks are reported in the Online Appendix (see tables OA12, OA13 and OA4). First, to test the sensitivity of our results to the transformation ln(1+variable) used for two of the dependent variables in Table 3, we re-estimate these results using alternative transformations of the form ln(x+variable), where x is set to 10, 0.1, 0.01 instead of 1, as well as using an inverse hyperbolic sine transformation instead of a log transformation. Our results on cash crop investments are robust to these alternative transformations. For a few of the alternative log transformations, insurance also has a marginally statistically significant positive effect on total investment; these effects are positive, although not statistically significant, in Table 3. Second, to test for the influence of outliers, we re-estimate our main results from Table 3 after winsorizing the top and bottom 2% of all continuous variables. Results are almost unchanged, suggesting that the logarithmic and share transformations already do a good job of limiting the influence of extreme observations. Third, as mentioned earlier, we find similar estimated marginal effects of the insurance treatment if we estimate the specifications from Table 3 using a linear probability model, rather than tobit and probit estimators. G. Impact of Insurance Payouts 24 Unexpectedly, southern India experienced a severe drought during the monsoon season of the year of our experiments. Reflecting the low realized rainfall during the rainfall insurance coverage period, many insured farmers received significant cash payouts ranging from Rs. 2100 (ca. $42) to Rs. 10,000 (ca. $200), as indicated in Table 2. Section OA2 of the Online Appendix presents evidence on what farmers did with these payouts, based on data from a second follow-up survey conducted in mid-2010. We briefly summarize this evidence below. 18 We first tabulate farmers’ self-reported accounting of how payouts were used. Farmers report that roughly half of the funds were used for consumption or agricultural investments, with the rest saved, used to retire debt, or given away. Only around one-tenth of funds were given away, and such gifts were generally restricted to the farmer’s family. Like our earlier evidence on self-reports, these results should be treated with caution: Karlan, Osman and Zinman (2016), for example, show in the context of micro-credit that individuals do not accurately self-report expenditures when receiving credit. Fungibility may be one source of confusion: farmers may report the proximate use of funds rather than their incremental spending relative to the counterfactual of not having received a payout. Bearing these concerns in mind, farmers’ responses taken at face value suggests payouts are retained and spent by the recipient, rather than being socialized within the village. In the more formal part our analysis, we regress a range of ex post outcomes on insurance treatment status and the size of the insurance payout as well as village dummies and other controls. Because all insured farmers in a given village received the same payout, identification comes from within-village differences between treated and untreated farmers in villages where payouts were high relative to the corresponding difference in villages where the insurance did not pay out. We find farmers who received large payouts subsequently reported greater trust in the insurance provider, perceived less basis risk in the insurance product, and were more likely to have paid down high interest rate debt. We do not find systematic differences in subsequent investment, labor supply, or asset values, perhaps reflecting the late survey timing and low 18 This analysis is omitted from the main text in part because our statistical power is low; in addition, this evidence is less novel than our analysis of the ex ante effects of insurance, as there is already a significant body of well-identified research on how households use unexpected cash windfalls (e.g., de Mel, McKenzie and Woodruff 2008). 25 power of our estimates. We do find that farmers in the insurance treatment group who received large payouts report statistically significantly lower consumption than the control group (i.e., we can reject the joint hypothesis that treatment and treatment x payout are zero). While this result may seem counterintuitive, it may simply reflect the fact that treated farmers rationally chose to take on more risk ex ante, resulting in lower drought income that was not fully compensated by insurance payouts. It is not necessarily evidence of ex ante “mistakes.” 4. Conclusions We find that the provision of insurance against rainfall risk influences production decisions among a sample of Indian farmers. In particular, insured farmers increase agricultural investments in higher-return but rainfall-sensitive cash crops. This shift in behavior is concentrated on the extensive margin and among more-educated farmers. Investigating the timing of the change in behavior, we show that it occurs ex ante, before the resolution of uncertainty about the timing of the monsoon. These findings, as well the results of other recent complementary research, imply that farmers are underinsured and that insurance arrangements that “fill in” missing markets have significant effects on entrepreneurial production and risk-taking. Financial innovations that help pool and diversify risk may thus play a significant role in boosting growth and real incomes in emerging market economies. Such a role relies, however, on financial deepening and further improvements to product design and delivery that mitigate “barriers” to insurance uptake, including high prices and transaction costs, basis risk, limited familiarity and trust, and financial constraints (Cole et al. 2013; Clarke 2016; Rampini and Viswanathan 2015). In some cases, public insurance provision may be warranted if barriers to individual adoption are high due to market failures. From a broader international perspective, the availability of insurance to mitigate income risk for entrepreneurs (and would-be entrepreneurs) varies widely across countries and over time. Examples include access to health insurance, unemployment insurance, bankruptcy protection, and social security. Analyzing the effects of insurance arrangements on entrepreneurial activity and risk-taking is a promising area for future research. 26 References de Mel, S., D. McKenzie, and C. Woodruff. 2008. Returns to capital: Results from a randomized experiment. Quarterly Journal of Economics 123: 1329-72. Acemoglu, D., and F. Zilibotti. 1997. Was Prometheus unbound by chance? Risk, diversification, and growth. Journal of Political Economy 105: 709-51. Adelino, M., A. Schoar, and F. Severino. 2015. House prices, collateral and self-employment. Journal of Financial Economics 117: 288-306. Banerjee, A. V., and A. F. Newman. 1993. Occupational choice and the process of development. Journal of Political Economy 101: 274-298. Beck, T., R. Levine, and N. Loayza. 2000. Finance and the sources of growth. Journal of Financial Economics 58: 261-300. Bertrand, M., and S. Mullainathan. 2001. Do people mean what they say? Implications for subjective survey data. American Economic Review 91: 67-72. Black, S. E., and P. E. Strahan. 2002. Entrepreneurship and bank credit availability. Journal of Finance 57: 2807-33. Burgess, R., O. Deschenes, D. Donaldson, and M. Greenstone. 2013. The unequal effects of weather and climate change: Evidence from mortality in India. Working Paper, MIT. Cai, H., Y. Chen, H. Fang, and L. Zhou. 2015. The effect of microinsurance on economic activities: Evidence from a randomized natural field experiment. Review of Economics and Statistics 97: 287-300. Campbell, J. 2006. Household Finance. Journal of Finance 61: 1553-1604. Campello, M., C. Lin, Y. Ma, and H. Zou. 2011. The real and financial implications of corporate hedging. Journal of Finance 66: 1615-1647. Carpenter, P., M. Just, and P. Shell. 1990. What one intelligence test measures: A theoretical account of the processing in the Raven Progressive Matrices test. Psychological Review 97: 404-31. Clarke, D. J. 2016. A theory of rational demand for index insurance. American Economic Journal: Microeconomics 8: 283-306. Clarke, D., O. Mahul, K. Rao, and N. Verma. 2012. Weather based crop insurance in India. World Bank Policy Research Working Paper 5985. 27 Cochrane, J. H. 1991. A simple test of consumption insurance. Journal of Political Economy 99: 957-76. Cole S., X. Giné, J. Tobacman, P. Topalova, R. Townsend, and J. Vickery. 2013. Barriers to household risk management: Evidence from India. American Economic Journal: Applied Economics 5: 104-35. Cole, S., A. Paulson, and G. Shastry. 2014. Smart money? The effect of education on financial outcomes. Review of Financial Studies 27: 2022-51. Dercon, S. 1998. Wealth, risk and activity choice: Cattle in western Tanzania. Journal of Development Economics 55: 1-42. Dercon, S., and L. Christiaensen. 2011. Consumption risk, technology adoption and poverty traps: Evidence from Ethiopia. Journal of Development Economics 96: 159-73. Duflo, E., M. Kremer, and J. Robinson. 2008. How high are rates of return to fertilizer? Evidence from field experiments in Kenya. American Economic Review Papers and Proceedings 98: 482-88. Emerick, K., A. de Janvry, E. Sadoulet, and M. H. Dar. 2016. Technological innovations, downside risk and the modernization of agriculture. American Economic Review 106: 1537- 61. Fan, W., and M. J. White. 2003. Personal bankruptcy and the level of entrepreneurial activity. Journal of Law and Economics 46: 543-567. Fernández-Val, I. 2009. Fixed effects estimation of structural parameters and marginal effects in panel probit models. Journal of Econometrics 150: 71-85. Foster, A., and M. Rosenzweig. 2010. Microeconomics of technology adoption. Annual Review of Economics 2: 395-424. Froot, K. A., and J. C. Stein. 1998. Risk management, capital budgeting and capital structure policy for financial institutions: An integrated approach. Journal of Financial Economics 47: 55-82. Giné, X., L. Menand, R. Townsend, and J. Vickery. 2012. Microinsurance: A case study of the Indian rainfall index insurance market. In Handbook of the Indian Economy, ed. C. Ghate, 167-194. Oxford, UK: Oxford University Press. Giné, X., R. Townsend, and J. Vickery. 2008. Patterns of rainfall insurance participation in rural India. World Bank Economic Review 22: 539-566. 28 Giné, X., and D. Yang. 2009. Insurance, credit, and technology adoption: Field experimental evidence from Malawi. Journal of Development Economics 89: 1-11. Gollier, C., and J. W. Pratt. 1996. Risk vulnerability and the tempering effect of background risk. Econometrica 64: 1109-23. Guiteras, R. 2009. The impact of climate change on Indian agriculture. Working Paper, University of Maryland. Hazell, P. B. R. 2009. Transforming agriculture: The green revolution in Asia. In Millions Fed, ed. D. Spielman, and R. Pandya-Lorch, 25-32. Washington, DC: IFPRI. Hill, R. V., and A. Viceisza. 2012. An experiment on the impact of weather shocks and insurance on risky investment. Experimental Economics 15: 341-371. Hurst, E., and A. Lusardi. 2004. Liquidity constraints, household wealth, and entrepreneurship. Journal of Political Economy 112: 319-47. Jayachandran, S. 2006. Selling labor low: Wage responses to productivity shocks in developing countries. Journal of Political Economy 114: 538-75. Karlan, D., R. Osei, I. Osei-Akoto, and C. Udry. 2014. Agricultural decisions after relaxing credit and risk constraints. Quarterly Journal of Economics 129: 597-652. Karlan, D., A. Osman, and J. Zinman. 2016. Follow the money not the cash: Comparing methods for identifying consumption and investment responses to a liquidity shock. Journal of Development Economics 121: 11-23. King, R. G., and R. Levine. 1993. Finance, entrepreneurship and growth. Journal of Monetary Economics 32: 513-42. Levine, R. 2005. Finance and growth: Theory and evidence. In Handbook of Economic Growth 1, ed. P. Aghion, and S. Durlauf, 865-934, Elsevier. Lybbert, T. J., F. B. Galarza, J. McPeak, C. B. Barrett, S. R. Boucher, M. R. Carter, S. Chantarat, A. Fadlaoui, A. Mude. 2010. Dynamic field experiments in development economics: Risk valuation in Morocco, Kenya, and Peru. Agricultural and Resource Economics Review 39: 1–17. Maccini, S., and D. Yang. 2009. Under the weather: Health, schooling and economic consequences of early-life rainfall. American Economic Review 99: 1006-26. Mobarak, A. M., and M. Rosenzweig. 2013. Selling formal insurance to the informally insured. Working Paper, Yale University. 29 Morduch, J. 1995. Income smoothing and consumption smoothing. Journal of Economic Perspectives 9: 103-114. Moskowitz, T. J., and A. Vissing-Jørgensen. 2002. The returns to entrepreneurial investment: A private equity premium puzzle? American Economic Review 92: 745-78. Pérez-González, F., and H. Yun. 2013. Risk management and firm value: Evidence from weather derivatives. Journal of Finance 68: 2143–76. Pind, J., E. Gunnarsdottir., and H. Johannesson. 2003. Raven’s Standard Progressive Matrices: New school age norms and a study of the test’s validity. Personality and Individual Differences 34: 375-86. Puri, M., and R. Zarutskie. 2012. On the life cycle dynamics of venture-capital- and non- venture-capital-financed firms. Journal of Finance 67: 2247-93. Rampini, A., and S. Viswanathan. 2015. Household risk management. Working Paper, Duke University. Raven, J. 2000. The Raven's Progressive Matrices: Change and stability over culture and time. Cognitive Psychology 41: 1-48. Rose, E. 1999. Consumption smoothing and excess female mortality in rural India. Review of Economics and Statistics 81: 41-49. Rosenzweig, M. R., and H. P. Binswanger. 1993. Wealth, weather risk and the profitability of agricultural investment. Economic Journal 103: 56-78. Rosenzweig, M. R., and O. Stark. 1989. Consumption smoothing, migration, and marriage: Evidence from rural India. Journal of Political Economy 97: 905-26. Skees, J. 2008. Innovations in index insurance for the poor in lower income counties. Agricultural and Resource Economics Review 37:1-15. Townsend, R. 1994. Risk and insurance in village India. Econometrica 62: 539-91. Townsend, R., and K. Ueda. 2006. Financial deepening, inequality and growth: A model-based quantitative evaluation. Review of Economic Studies 73: 251-293. Walker, T. S., and J. G. Ryan. 1990. Village and Household Economics in India’s Semi-Arid Tropics. Baltimore, MD: John Hopkins University Press. World Bank. 2005. Managing agricultural production risk: innovations in developing countries. Washington, DC: World Bank Agriculture and Rural Development Department, World Bank Press. 30 Figure 1: Example of Calculation of Rainfall Insurance Payout Chart plots cumulative rainfall and rainfall insurance policy parameters for the policy indexed to the Naryanpet weather station. 31 Figure 2: Timeline 32 Figure 3: Cumulative density, log investment in cash crops The y- axis plots the natural log of 1 + the amount invested in cash crops (in Rs.) for the treatment and control groups. Farmers in each group are sorted in increasing order of cash crop investments. 33 Figure 4: Effect of insurance treatment status on timing of cash crop investments The x-axis of the figure plots the passage of time in 2009, measured in terms of “kartis” (a kartis is a variable period of time approximately two weeks in length). The label below each tick mark reports the first date of the kartis. The y-axis plots the effect of insurance treatment status on the probability of having planted cash crops by the kartis in question. The shaded region indicates the period during which the insurance policies were distributed to treated farmers. The dashed vertical line indicates the end of the coverage period for any of the insurance policies. The solid line displays the point estimate of the insurance treatment effect; the dashed lines above and below the solid line represent a 95% confidence interval. 34 Table 1: Summary Statistics Variable Mean St. dev. p10 p50 p90 A. Demographic Characteristics Household size 5.15 2.05 3 5 8 Age of household head (Years) 49.6 12.4 35 50 65 Gender of household head (1="Male") 0.91 0.28 1 1 1 B. Education of Household Head Highest level of schooling completed by head (Years) 3.75 4.76 0 0 11 Household head is unschooled (1 = "Yes") 0.54 0.50 0 1 1 Household head has 1 to 5 years of schooling (1 = "Yes") 0.15 0.35 0 0 1 Household head has more than 6 years of schooling (1 = "Yes") 0.31 0.46 0 0 1 Household head able to read (1 = "Yes") 0.43 0.50 0 0 1 Household head able to write (1 = "Yes") 0.40 0.49 0 0 1 Raven's test of analytic intelligence, household head (0-3) 1.58 0.97 0 2 3 C. Knowledge, Trust and Expectations of Household Head Insurance knowledge index (0-5) 1.73 2.12 0 0 5 Head has heard of rainfall insurance (1 = "Yes") 0.42 0.49 0 0 1 Trust in BASIX (1 if trust > 4/10, 0 otherwise) 0.40 0.49 0 0 1 St. dev. of expected cash crop yield (kg/acre) 46.2 38.1 14.1 35.4 88.3 D. Credit and Assets Bank credit (1= "Yes") 0.71 0.46 0 1 1 Any Credit (1="Yes") 0.91 0.29 1 1 1 Total area of agricultural land (Acres) 5.37 5.47 1.75 4 10 Wealth Index (Principal Component) 0.00 1.70 -2.1 0.03 2.07 E. Land Use and Agricultural Investments in 2009 Monsoon All crops Positive investment/Any agricultural input used? (1 = "Yes") 0.97 0.17 1 1 1 Total cultivated land (Acres) 3.98 3.69 1 3 8 In which Kartis did farmer plant? 15.73 2.83 13 16 19 Market value of agricultural inputs used (Rs.) 22934 22169 5500 16550 46000 Cash crops Positive invesment/Any agricultural input used? (1 = "Yes") 0.47 0.50 0 0 1 Total cultivated land (Acres) 1.80 2.93 0 0 5 In which Kartis did farmer plant? 15.25 2.41 13 15 18 Did farmer replant crop in 2009 monsoon? (1="Yes" ) 0.05 0.23 0 0 0 Did farmer abandon crop in 2009 monsoon? (1="Yes") 0.15 0.36 0 0 1 Market value of agricultural inputs used (Rs.) 6195 12174 0 0 17700 Share of total cultivated land devoted to cash crops 0.39 0.45 0 0 1 Share of market value of ag. inputs devoted to cash crops 0.36 0.43 0 0 1 Notes: Summary statistics for the sample of 1479 individuals that participated in both the baseline and follow-up surveys. Sections A through D report baseline characteristics at the start of the monsoon for the entire sample. Section E reports ex post investment variables for the control group (the sample of 736 individuals that did not receive the insurance treatment). See Appendix for variable definitions. Table 2. Policy Details Realized Number of Reference Start End Strike Exit Per mm Maximum Payout per Treatment Station Date Date (mm) (mm) (Rs.) Payout Policy (Rs.) Farmers Atmakur June 12 July 16 45 5 10 1000 0 38 Narayanpet July 1 August 4 50 5 10 1000 341.5 170 Mahbubnagar June 6 July 10 70 10 10 1000 0 112 Hindupur July 1 August 4 25 0 10 1000 1000 242 Anantapur July 1 August 4 30 5 10 1000 210 175 Notes: This table reports insurance policy details and payouts for the study year. The “Strike” level is the rainfall threshold below which the policy begins to pay; the policy pays the amount indicated in “Per mm” for each mm of shortfall below this threshold. The “Exit” level is the rainfall threshold below which the policies pays the “Maximum Payout.” The “Realized Payout per policy” reports the payout received by insured farmers in 2009. Each farmer in the treatment group received 10 insurance policies. The Mahbubnagar and Anantapur stations are owned and operated by the Indian Meteorological Department. The other three weather stations were installed in 2005 by INGEN, a private company, for the purposes of the rainfall insurance program. Table 3. Effects of insurance on agricultural investments Mean of Dep. Percent of Household covariates included? Dependent Variable: Estimator Variable for Control Control Group Group = 0 No Yes (1) (2) (3) (4) (5) A. Investments in all crops Positive investment/Any ag. Probit 0.97 3 0.012 0.009 inputs used (1 = Yes) (0.011) (0.008) ln(1+total cultivated land, acres) Tobit 1.39 7 0.028 0.039 (0.034) (0.031) ln(1+market value of ag. inputs Tobit 9.45 3 0.082 0.110 used, Rs.) (0.087) (0.083) B. Investments in cash crops Positive investment/Any ag. Probit 0.47 53 0.060** 0.064** inputs used (1 = Yes) (0.029) (0.030) ln(1+total cultivated land, acres) Tobit 0.66 54 0.086** 0.093*** (0.037) (0.036) ln(1+market value of ag. inputs used, Rs.) Tobit 4.24 53 0.451** 0.485** (0.218) (0.216) Cash crop shares Share of total cultivated land Tobit 0.39 52 0.047** 0.048** planted with cash crops (0.021) (0.021) Share of market value of ag. Tobit 0.37 51 0.034* 0.035* inputs devoted to cash crops (0.019) (0.019) Notes: This table reports the marginal effect of the insurance treatment dummy on various measures of monsoon agricultural investments; each row presents a different dependent variable. The first three dependent variables relate to investments in all crops. Dependent variables in the next three regressions relate to investments in cash crops only. The final two specifications consider the share of total agricultural inputs used for growing cash crops. Two versions of each model are presented, one without additional household covariates (Column 4) and one with (Column 5). These household covariates are Age of Head, Education of Head, and the Wealth Index. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table 4. Heterogeneous Effects of Insurance Treatment Dependent variable: Positive investment in cash crops (1=Yes) Baseline model Single interactions of treatment with: Multivariate [as Table 3] Wealth Age Education Knowledge Trust Payouts (1) (2) (3) (4) (5) (6) (7) (8) Insurance treatment dummy 0.060** 0.062** 0.014 -0.004 0.067* 0.064* 0.033 -0.196 (0.029) (0.029) (0.123) (0.037) (0.039) (0.038) (0.045) (0.143) Interaction effects: treat x Wealth Index -0.012 -0.022 (0.018) (0.020) treat x Age of Head 0.001 0.004 (0.002) (0.003) treat x Education of Head (years) 0.018*** 0.024*** (0.006) (0.007) treat x Insurance Knowledge Index (0-5) 0.022 0.013 (0.034) (0.036) treat x Head has heard of -0.104 -0.088 rainfall insurance (0.143) (0.151) treat x Trust BASIX (1= yes) -0.008 -0.001 (0.061) (0.068) treat x Payout (1,000 Rs.) 0.059 0.033 (0.073) (0.078) Uninteracted household characteristics: Wealth Index 0.041*** 0.048*** (0.013) (0.014) Age of Head -0.002 -0.004** (0.002) (0.002) Education of Head (years) -0.005 -0.013** (0.005) (0.005) Insurance Knowledge Index -0.065** -0.063** (0.026) (0.027) Head has heard of rainfall 0.292*** 0.286** insurance (1=yes) (0.110) (0.112) Trust BASIX (1=yes) 0.022 -0.001 (0.044) (0.048) Payout (1,000 Rs.) na na Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy as well as interactions between the treatment dummy and various household characteristics. Dependent variable is equal to 1 if the farmer invested resources in planting cash crops. Since the "knowledge of insurance" questions were only asked of farmers that were aware of insurance, the specification including the knowledge index also includes a dummy for whether the farmer had heard of insurance. No direct effect of insurance contract payout on investment reported since payout only varies by village, and is absorbed by village dummies. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Probit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table 5. Education and the Effects of Insurance on Agricultural Investments Dependent variable: Positive investment in cash crops (1=Yes) Education Interaction Variable(s): Education Additional Literacy + Raven's score + Education interacted with: (years) [as education years of years of Land- Bank credit St. dev. of Tab 4 Col 8] dummies education education holdings usage expected ag yield (1) (2) (3) (4) (5) (6) (7) Insurance treatment dummy -0.196 -0.184 -0.197 -0.210 -0.231 -0.170 -0.122 (0.143) (0.146) (0.144) (0.152) (0.151) (0.155) (0.150) Interaction effects: treat x Education of Head (years) 0.024*** 0.048** 0.022* 0.024*** (0.007) (0.020) (0.013) (0.007) treat x Education = 1-5 years -0.075 (0.120) treat x Education = 6 + years -0.258 (0.177) treat x Head can read 0.013 (0.120) treat x Raven's Test score 0.005 (0.032) treat x Education x landholdings > 0.029*** median (0.009) treat x Education x landholdings < 0.014 median (0.011) treat x Education x bank credit 0.027*** (0.008) treat x Education x No bank credit 0.014 (0.013) treat x Education x st dev of exp yield 0.044*** > median (0.010) treat x Education x st dev of exp yield 0.007 < median (0.009) Uninteracted education measures Education of Head (years) -0.013** -0.042*** -0.019** -0.013** (0.005) (0.015) (0.009) (0.005) Education = 1-5years 0.144* (0.086) Education = 6+ years 0.314** (0.139) Head can read 0.071 (0.084) Raven's Test score -0.033 (0.024) Education x landholdings > median -0.023*** (0.007) Education x landholdings < median 0.001 (0.007) Education x bank credit -0.016*** (0.006) Education x no bank credit -0.004 (0.009) Education x st. dev. of exp yields > -0.024*** median (0.007) Education x st. dev. of exp yields < -0.004 median (0.006) Landholdings dummies no no no no yes no no Bank credit dummies no no no no no yes no St dev of exp yield dummies no no no no no no yes Other covars. from Table 4? yes yes yes yes yes yes yes Hypothesis tests on interaction variables Joint significance: 1-5 and 6+ yrs education 0.387 Joint significance: literacy and education 0.003 Joint significance: Raven's and education 0.003 Equality: high vs low landholding 0.272 Equality: Bank credit vs no credit 0.381 Equality: High SD vs Low SD of exp yields 0.004 Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy and various interactions between the insurance treatment and measures of education and cognition. All specifications include other covariates from Table 4 (wealth, age, knowledge of insurance/heard of insurance, trust, payouts) and their interactions with the treatment dummy. “Landholdings dummies” refers to dummy for landholdings > median and the interaction term landholdings > median x treatment. “Bank credit dummies” and “St dev of exp yield dummies” are similarly defined. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Probit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table 6: Self-Reported Effects of Rainfall Insurance on Agricultural Investments Effect of rainfall insurance on: The amount used of: More No Change Less Fertilizer 50% 36% 14% Seeds 41% 43% 16% Pesticides 32% 41% 27% Bullock labor 23% 48% 29% Hired labor 35% 42% 23% Funds borrowed to finance agricultural inputs 26% 52% 22% The timing of initial planting Earlier 26% No change 69% Later 5% The decision of whether to abandon crops Against 26% No change 67% Towards 7% Notes: This table tabulates the self-reported effects of rainfall insurance on investment decisions, as reported by 743 farmers in the treatment group. This information was collected during the first follow-up survey, conducted in late 2009. Appendix: Selected Variable Definitions Variable Descriptive information Survey A. Demographic Characteristics Age of Household 2009 Age of the household head, in years. Head Follow-Up B. Education of Household Head The household head's highest level of schooling completed, measured in years. 5 is equivalent to primary school completion, 7 is secondary school completion, 12 is Education of 2009 high school completion. 13-16 correspond to Diploma/vocational course, Bachelor Household Head Follow-Up degree (3 years), Professional Bachelors degree (4 years), and Masters degree, respectively. Literacy (Farmer Dummy variable indicating that household head self-reports that they are able to 2009 can read) read and understand a newspaper. Follow-Up Score (0-3) based on the number of correct responses to three analytic intelligence test questions. In each question, a a patterned picture with a piece missing is 2010 Raven's Test score displayed, and respondents had to select the correct piece from six choices, after Follow-Up being shown an example. C. Knowledge, Trust and Expectations of Household Head Individuals are asked to calculate, given a set of assumptions, whether they would Insurance get an insurance payout and how large would that payout be. Five questions are 2009 Knowledge Index asked, with one point is assigned to each 'good' response. The index is the sum of Baseline correct responses [0-5]. Dummy variable equal to one if the respondent has not heard of rainfall insurance. 2009 Heard of Insurance This is used due to the skip pattern in the Insurance Knowledge Index variable Baseline (those questions are only asked if the farmer is aware of insurance). Dummy variable equal to one if trust in the insurance vendor BASIX is greater 2009 Trust than 4, based on the question: "on a scale of 0-10, how trustworthy do you think Baseline the BASIX organization is?", otherwise zero. Each respondent reports the expected minimum and maximum yields that could be realized from one acre of land, assuming that rains are “very poor” and “very good”, respectively. (This is done for castor yields in Mahbubnagar, and groundnut St. dev. of yields in Anantapur, and under the assumption that fertilizer is not used). The expected yield enumerator computes the midpoint M between these two yields, as well as 2009 (kg/acre) of cash additional midpoints halfway between the minimum and M (m1) and between M Baseline crops without and the maximum (m2), resulting in a 5-point support. The respondent is then fertilizer asked to distribute 10 beans according to the likelihood that yields will be between the minimum and m1, between m1 and M, between M and m2 and between m2 and the maximum. The standard deviation of expected yield is computed as the standard deviation of the data in this histogram. D. Credit and Assets Dummy variable equal to one if at the time of the 2009 baseline survey, the 2009 Bank Credit respondent either has outstanding loans from a bank, and/or has been approved for Baseline credit from a bank at least once since the end of 2008 monsoon. Appendix (Continued): Selected variable definitions Variable Descriptive information Survey Dummy variable equal to one if, at the time of the 2009 baseline survey, the respondent eitherhas outstanding loans from a bank, family and friends, 2009 Any Credit microfinance institutions (BASIX), moneylender, or other, or equal to one if Baseline indicates that has applied for credit from any of those sources since the end of 2008 monsoon and was approved at least once. Total area of agricultural land that belonged to the household as of the beginning 2009 Landholdings of the Mrigashira kartis (June 8, 2009), in acres. Follow-Up First component of a principal components analysis (PCA). Variables includes a dummy if the household owns different specific types of livestock as well as the 2009 log total value of livestock, a dummy if the household has any type of credit, a Wealth index: Baseline & dummy if the household has any type of savings, the log of total amount of PCA 2009 savings and credit, the house type, the number of rooms in the house, the total area Follow-Up of agricultural land, the log of the house value, the log of the land value, and the log of the value of other assets. E. Land Use and Agricultural Investments in 2009 Monsoon All crops: Any ag. Dummy variable equal to one if any agricultural inputs were used for the following 2009 inputs used (1 = categories: Hybrid seeds, improved seeds, fertilizer, manure, pesticide, irrigation, Follow-Up Yes) hiring tractors or other implements, manual labor, and bullock labor. All crops: Total 2009 cultivated land, Total cultivated land used towards all crops, in acres. Follow-Up acres All crops: Market Market value used on inputs (Rs.) towards all crops for the following categories: 2009 value of ag. inputs Hybrid seeds, improved seeds, fertilizer, manure, pesticide, irrigation, hiring Follow-Up used, Rs. tractors or other implements, manual labor, and bullock labor. Cash crops: Any Dummy variable equal to one if any agricultural inputs were used for cash crops 2009 ag. inputs used (1 for only the following categories: Hybrid seeds, improved seeds, fertilizer, manure, Follow-Up = Yes) and pesticide. Cash crops: Total 2009 cultivated land, Total cultivated land used towards cash crops, in acres. Follow-Up acres Cash crops: Market value of Market value used on inputs (Rs.) towards cash crops for only the following 2009 ag. inputs used, categories: Hybrid seeds, improved seeds, fertilizer, manure, and pesticide. Follow-Up Rs. Share of total cultivated land Share of total cultivated land used towards cash crops relative to total cultivated 2009 devoted to cash land used towards all crops. Follow-Up crops Appendix (Continued): Selected variable definitions Variable Descriptive information Survey Share of market Share of market value used on inputs towards cash crops for the categories of val. of ag. inputs 2009 hybrid seeds, improved seeds, fertilizer, manure and pesticide relative to market devoted to cash Follow-Up value used on inputs towards all crops for those same categories. crops F. Policy Details Admin Payout (Rs.) Ex post payouts in Rupees. Reported in the text in units of 000s. Data Online Appendix for “How Does Risk Management Affect Production Decisions? Evidence from a Field Experiment” by Shawn Cole, Xavier Giné and James Vickery NOT FOR PRINT PUBLICATION Contents: Page 1: OA1. Simple Theoretical Framework Page 5: OA2. Analysis of Impact of Rainfall Insurance Payouts Page 12: OA3. Additional Tables (Tables OA1 to OA18) Section OA1. Theoretical Framework This section of the Online Appendix presents a simple illustrative model of a farmer’s entrepreneurial decisions to highlight the interaction between insurance access and production behavior. The key result illustrated by the model is that for a risk-averse farmer, investment in risky production activities is increasing in access to insurance against production risk. Although we assume a very simple setting to highlight the basic intuition, this result extends to a much more general class of models, as discussed in the main text. A. Basic setup and timing Consider a one-period model of a farmer with initial wealth W0 and constant absolute risk aversion (CARA) utility. The farmer has access to a risky production activity or project (e.g., sowing cash crops, or applying fertilizer), and decides at the start of the period what fraction of their wealth (α) to devote to this risky activity. The net return on investment (per rupee invested) � is the expected return and e is a zero-mean normally distributed error � + e, where is given by term: e ∼ N(0, σ2e). The remainder of their wealth is invested in a safe activity, which for simplicity is assumed to produce a real return of zero. The farmer can partially hedge the production risk associated with the risky activity by purchasing insurance. We denote the amount spent on insurance premia by ϕ. The insurance payout is negatively correlated with the return on investment, but not perfectly (i.e. there is some basis risk). Net of the initial premium, the net payout on the insurance (per rupee of premium) is given by: -e + u - µ, where u ∼ N(0, σ2u). The higher is σ2u, the greater the basis risk. We generally assume that µ > 0, which means that the expected insurance payout net of the premium is negative (i.e. the insurance is not actuarially fair). 1 Summary of timing: At the start of the period the farmer chooses how much to invest ( ) and how much insurance to purchase (ϕ). At the end of the period, the return on the risky production activity and the insurance payout are realized. The farmer then consumes their initial wealth W0 plus their net income from the investment and from insurance. 1 This could be due to a combination of imperfect competition amongst insurers, administrative costs of providing the insurance, or compensation for the risk borne by the insurer (to the extent the insurer themselves is not fully diversified). 1 We assume there is an interior solution (i.e. the fraction of their wealth invested in the risky project, inclusive of any insurance purchased, is between zero and one), and that µ is large enough so that insurance demand is positive in equilibrium. B. Optimal investment in the presence of insurance The farmer’s objective is to maximize expected end-of period utility E[u(W1)]. End of period wealth (W1) is given by the law of motion: End of period wealth (W1 ) = initial wealth (W0 ) + investment return (Y) + insurance payout (IP) = W0 + α(R + e) + ϕ (-e + u - μ) Given our exponential-normal setup, and denoting the farmer’s coefficient of absolute risk aversion by γ, the farmer’s problem can be written as: max α, ϕ E[u(W1)] = max α, ϕ {E(W1) - ½γ var(W1)} [OA.1] where: � - ϕµ E(W1) = W0 + α var(W1) = (α - ϕ)2σ2e + ϕ2σ2u Taking first order conditions of [OA.1] with respect to α and ϕ, and solving the resulting simultaneous equations, the optimal investment level is given by the following expression: � − 1 � ∗ = � 2 + 2 � [OA.2] An alternative and similar expression can be derived if we assume that the level of insurance ϕ is assigned exogenously to the household, rather than being a decision variable. (This is perhaps the setting that corresponds most exactly to the design of our field experiment). In this case, optimal investment is given by the simpler expression: � 1 ∗ = 2 + [OA.3] 2 C. Comparative statics Inspecting expression [OA.2] yields the following comparative statics results for the farmer’s equilibrium level of investment in the risky production activity: Proposition: The farmer’s equilibrium investment in the risky activity (α*) is: A. decreasing in the expected per-unit net cost of insurance (µ). B. decreasing in the basis risk of the insurance (σ2u) C. decreasing in the variance of investment returns (σ2e) D. decreasing in risk aversion (γ) �) E. increasing in the expected return on investment ( Proof: By taking first derivatives of [OA.2] with respect to each parameter. The same comparative statics results apply to the alternative expression for optimal investment assuming that insurance is assigned exogenously [equation OA3]. The only difference is that part A of the Proposition instead states that investment in the risky production activity (α*) is increasing in the exogenously determined level of insurance (ϕ), rather than being decreasing in the cost of insurance. The key result of the model is that an improvement in access to insurance – either an increase in the amount of exogenously provided insurance, a reduction in the cost of the insurance, or an improvement in the quality of the insurance while keeping the cost fixed – increases investment in the risky activity. The simple intuition for these results is that the farmer’s optimal level of investment trades off the high expected return of the investment against its risk. Improving access to insurance against production risk allows the farmer to reduce the background risk associated with any given investment level (i.e. to shift this risk-return frontier outwards), allowing the farmer to invest more in equilibrium. Given these results, it is also straightforward to verify that the farmer’s expected income and expected utility are decreasing in the expected per-unit net 3 cost of insurance (µ), and the basis risk of the insurance (σ2u), so that improving access to insurance increases expected income and welfare. Since we assume exponential utility, there are no wealth effects in the model. In reality, insurance provision may affect behavior both through its risk-management benefits and because it increases household wealth. To control for this, in our field work we compare two groups, one of which receives insurance for free, the other of which is promised the actuarial value of the insurance for free. In other words our design attempts to hold fixed the ex ante expected wealth of the household between the treatment and control groups. 4 Section OA2. Analysis of Impact of Rainfall Insurance Payouts This section reports in detail on how treatment individuals who received a payout reported spending the insurance payouts, and how the treatment and payouts affected consumption, investment, and attitudes. During the 2009 monsoon season, India experienced a drought during the normal planting period followed by heavy rains during crop growth and harvest. Nationally, accumulated rainfall during the monsoon months was 79% of normal, defined as a 50-year average by the Indian Meteorological Department. Rainfall during the critical early planting period was very low in the two districts where the experiment was conducted (65.1% of normal in Mahbubnagar in June; and 16.8% for Anantapur in June). Although total rainfall recovered (rainfall for the entire growing season was 77.6% of normal in Mahbubnagar, and 117.6% in Anantapur, due to high rainfall in August), this low early rainfall affected yields of the main cash crops, especially groundnut in Anantapur. According to district-level data from the Ministry of Agriculture, groundnut yields in Anantapur were only 42% of the 10 year average, while castor yields in Mahbubnagar were 95% of the 10 year average. 2 Reflecting this low rainfall during the coverage period, most (but not all) insured farmers received cash payouts, ranging from Rs. 2100 (ca. $42) to Rs. 10,000 (ca. $200), as shown in Table 3 of the main text. A. 2010 Follow-up survey The second follow-up survey collected data on a range of issues including agricultural production, investment in the dry season (Rabi), the use of payouts (if any) and household savings and consumption. Owing to operational constraints this survey was not presented to farmers until April-June 2010, significantly after payouts were received in December 2009. The delay in the data collected may have contributed to poor recall and the lack of statistical power in detecting impacts despite the magnitude of the payouts. Table OA14 presents follow-up summary statistics about the effects of treatment and payouts on attitudes towards insurance and financial outcomes (Panel A) and on real outcomes (Panel B). Table OA15 contain the definition of the variables. 2 Data from the Directorate of Economics and Statistics of the Ministry of Agriculture can be accessed at http://eands.dacnet.nic.in/. 5 According to Table OA14, about 60% of respondents had taken a loan from a microfinance institution or a moneylender (high interest debt) amount to 40% of total outstanding debt. The usage of this high interest debt, however, is not particularly high as the mean is Rs. 1,482 (USD 32.15) at the time of the survey. In addition, 68% of the households receive income from the Mahatma Gandhi National Rural Employment Guarantee Scheme (MGNREGS), a ubiquitous work program for the poor in which households have the right to work for a wage for up to 100 days per year. In contrast, only 11% earned income from migration in a year when the monsoon was of poor quality. Both the MGNREGS and seasonal migration can be seen as strategies to cope with weather shocks, and thus the evidence suggest that individuals make use of such strategies (Munshi and Rosenzweig, 2016). B. Self-reported uses of insurance payouts Panel A of Table OA16 presents results of the second follow-up survey, in which farmers who were treated and received an insurance payout were simply asked to report how the cash payout was allocated among different uses such as saving, immediate consumption, gifts, and so on. Forty-five percent reported purchasing at least some inputs for Rabi following the summer season covered by the insurance policies. Since little or no rain falls during the winter, only farmers with access to a well can plant. In the data, about half of the farmers own a well, implying that nearly all farmers with well access report using part of the payout funds to purchase inputs for the winter season. Purchases of goods and services, mainly for immediate consumption, accounted for 39% of funds received, with 84% of farmers reporting using at least some funds for immediate consumption. Thirty-six percent of funds received were saved or used to pay down debt, while about one-tenth was given away. These responses, taken at face value, represent a rejection of either a full risk-sharing benchmark or a permanent income hypothesis benchmark, since more than two-fifths of funds received were used for immediate consumption or for physical agricultural investments. The survey responses are however consistent with a broad range of empirical evidence from emerging and developed countries that individuals consume or invest a significant fraction of cash windfalls (e.g., Aaronson, Agarwal and French 2012; de Mel, McKenzie and Woodruff 2008; Souleles 1999). Such evidence is generally interpreted as evidence that financial constraints, behavioral factors (e.g., rule-of-thumb consumers) and/or other frictions play an important role in consumption decisions. 6 Panel B of Table OA16 summarizes what information other parties (e.g., family, friends) had about the insurance coverage of treated farmers, the size of the payout, and the extent to which payouts were shared within and outside the immediate household. This information is important because farmers in our study areas engage in significant informal risk-sharing, which may crowd out formal insurance; social pressure to provide assistance to families, neighbors, or friends could reduce the incentive to purchase insurance in the first place, or to change investment decisions once insured. Our main results presented in Table 3 show that insurance coverage does change production decisions, implying that insurance payouts are not entirely socialized. Panel B confirms this result. While family and/or friends of treated farmers who received a payout often knew that the farmer had insurance and had received a payout, the sharing of insurance payouts was much less common. The payout was shared within the immediate household in about half of cases (48%), but with extended family in only 8% of cases, and with friends or others in only 1% of cases. This low rate of sharing outside the household occurs despite the fact that in 72% of cases, the extended family knew that a payout had been received, while friends were aware about half the time. C. Regression analysis Next we conduct regression analysis of the effect of insurance payouts on savings and debt, real outcomes such as agricultural investments, consumption and migration, as well as attitudes towards the insurance product. These outcome variables were also collected in the second follow-up survey and summarized in Table OA14. We estimate regressions of the form: outcome = f(a + b. insurance + c. (insurance x payout amount) + controls + e), where “insurance” is a dummy for whether the individual was assigned to the insurance treatment group, and “payout amount” is a continuous variable, bounded between zero and one, indicating what fraction of the maximum possible payout was received (i.e., equal to 0 for weather stations for which the contract did not pay out, and equal to1 if the contract paid out the maximum of Rs. 10,000). We include village dummies and the fertilizer treatment dummy as 7 controls, as in our earlier analysis. We therefore do not separately control for payout amount, which varies only by village as it is thus absorbed by the village dummies. Interpreting the evidence on the ex post effects of payouts requires a more nuanced view than our earlier ex ante evidence for at least two reasons. First, ex post effects measured in coefficients b and c in the above equation reflect both differential ex ante behavior (e.g., greater investment in cash crops) and ex post outcomes (weather realization and insurance payouts). Conceptually, the effects are different from the effects of an unexpected “cash drop” received after the harvest. Our experiment cannot identify what the effect of a post-harvest “cash drop” after harvest would be. Second, and more importantly, we observe only a single year’s realization of rainfall. We do not know how well the insurance performed with respect to basis risk: e.g., were payouts particularly well suited to local loss conditions, or were payouts not well matched to local loss? This limits the value of this ex post analysis. Results are presented in Table OA17. We first examine how product experience affects attitudes towards insurance, measured by asking farmers to react to interviewer questions on a 1- 10 Likert scale (1=strongly disagree; 10=strongly agree). The first column of Panel A shows that treated farmers report 0.371 higher trust in the insurance company on this ten-point scale, compared to an average response of 5.33. This effect on trust was larger for farmers who received payouts, though not statistically significantly so. The dependent variable in column 2 measures the farmer’s perceptions about basis risk, perhaps the most significant drawback of index insurance. Farmers report (on the same 1-10 scale) how strongly they agree with the statement “product pays out during times of drought”, in other words, that the policy has low basis risk. Here, insured farmers receiving no payout feel no differently than the control group, but farmers that received a payout are statistically significantly more likely to agree that the product pays out in times of drought. The coefficient of 0.395 indicates that those receiving a Rs. 10,000 payout agree with the statement by 0.395 points more on the 1-10 Likert scale, relative to insured farmers that received zero payout. The mean reported response value is only 3.69, suggesting that the sample as a whole does view basis risk as a significant drawback of the insurance product. (Note: This impact of treatment on trust in the insurance provider and on basis risk is not correlated with education. Table OA18 replicates columns 1 and 2 of Table OA17 including the education of the head in years and its interaction 8 with the treatment dummy. We find that only the direct effect of treatment is significant. The interaction with education is not statistically significant.) Turning to financial outcomes in Table OA17, we find that treated farmers report higher levels of financial savings ex post (column 3). While the individual coefficients are not significant, a test of the hypothesis b + c = 0 can be rejected at the 5 percent level. Quantitatively, farmers who received the maximum insurance payout of Rs. 10,000 report higher savings of Rs. 1,561 in 2010 compared to untreated farmers. While we do not find any evidence that the treatment affected total indebtedness (column 4), we do find that it affected the probability that households hold expensive debt, defined as debt from money lenders, microfinance institutions (MFIs) and other sources (column 5). These three sources charge an average interest rate of 31%, compared to debt from family, typically given at zero interest rate, or debt from commercial banks at 15%. Treated households that did not receive a payout report being 28.1 percentage points more likely to hold expensive debt than the control group. In contrast, treated households that received the maximum payout were 15.7 percentage points less likely to hold expensive sources of debt than the control group (0.281 - 0.438 = -0.157). This use of insurance payouts to reduce reliance on expensive sources of debt is consistent with “pecking-order” theories of financing choice, such as Myers and Majluf (1984) -- farmers borrows from moneylenders and other expensive sources only if their liquid assets or access to cheaper sources of funding is insufficient. The finding that those insured farmers not receiving a payout actually used more expensive debt than the control group is likely to be a reflection of basis risk. Treated farmers took riskier decisions which ex post did not produce high returns because of the poor quality of the monsoon, but not all insured farmers received payouts. The remainder may have been more likely to resort to expensive sources of credit to invest or smooth consumption. Regressions in Panel B study whether insurance cash payouts affected ex post real decisions and investments in the period after payouts were received. Such effects would be expected if farmers were financially constrained. Overall, we find little statistically significant evidence of such real ex post effects, although our statistical power to detect these effects is quite 9 low given the magnitude of the measured standard errors. The low power may reflect the timing of the second follow-up survey, as already mentioned. Turning to specifics, Column 7 finds no effect of payouts on the area planted during the Rabi winter growing season. As noted, only farmers with a well can cultivate during Rabi; well owners tend to be wealthier and may be less financially constrained. Similar to our results for high interest debt, we find a positive effect of assignment to the insurance treatment group on the labor supply of children (two hours more per week relative to a mean of 12.4 hours), although this increase is not present for farmers receiving large payouts. 3 Insurance treatment status has no effect on the probability a household engages in MGNREGS or on the probability that a household reports earning income from migration. We also find no effects on the change in value of livestock and durable goods, though the standard errors are quite large. 4 The final column of Panel B reports estimates of the effect of payouts on self-reported consumption (measured per day). As in the other columns, we find no statistically significant effect of assignment to the treatment group on daily consumption. However, for farmers receiving the maximum payout, the combined effect of being treated and receiving the maximum payout is actually negative and statistically significant at the five percent level, as discussed in the main text. D. Summing up Consistent with credit constraints or hyperbolic discounting, farmers self-report using about one quarter of insurance cash payouts for immediate consumption, and about one quarter for agricultural investments or durable consumption purchases; the remainder is saved or used to pay down debt. Our formal regression evidence uncovers no consumption or investment effects, however, perhaps reflecting the late survey timing and low power of our estimates, or perhaps due to bias in farmers’ self-reports. Matching self-reports, our regressions do find that part of the payout is used to pay down more expensive forms of debt, consistent with a pecking-order model. Finally, only a small fraction of the payout is given away, and gifts are generally restricted to the recipient’s immediate family, indicating that payouts are not socialized. 3 For a farmer receiving the maximum payout of Rs. 10,000, the net effect on child labor is 2.028 – 2.665 = -0.637 hours, not statistically different from the control group. 4 We focus here on change, rather than levels, because we have pre-period data, and because there may be significant individual-level variation in how respondents report the estimated value of these goods. 10 References Aaronson, Daniel, Sumit Agarwal and Eric French. 2012. “Consumption and Debt Response to Minimum Wage Increases.” American Economic Review 102: 3111-39. de Mel, Suresh, David McKenzie and Christopher Woodruff. 2008. “Returns to Capital: Results from a Randomized Experiment.” Quarterly Journal of Economics, 123: 1329-72. Munshi, Kaivan and Mark Rosenzweig. 2016. “Networks and Misallocation: Insurance, Migration, and the Rural-Urban Wage Gap.” American Economic Review 106: 46-98. Myers, Stewart and Nicholas Majluf. 1984. “Corporate Financing and Investment Decisions when Firms Have Information Investors Do Not Have.” Journal of Financial Economics 13: 187-221. Souleles, Nicholas S. 1999. “The Response of Household Consumption to Income Tax Refunds.” American Economic Review 89: 947–58. 11 Section OA3. Additional Tables of Results Table OA1: Household Summary Statistics Table OA2: Test of Balance Table OA3: Investment in 2009 Monsoon Table OA4: OLS estimator instead of probit and tobit Table OA5: Cash crop treatment effects disaggregated by investment type Table OA6: Takeup of Credit during the Monsoon Table OA7: Heterogeneous Effects of Insurance Treatment on Land Cultivated for Cash Crops Table OA8: Heterogeneous Effects of Insurance Treatment on Investment in Cash Crops Table OA9: Heterogeneous Effects of Insurance Treatment: Additional baseline variables Table OA10: Education and the Effects of Insurance on Land Cultivated with Cash Crops Table OA11: Education and the Effects of Insurance on Investment in Cash Crops Table OA12: Different functional forms of investment dependent variable Table OA13: Treatment Effects, Winsorized Two Percent Table OA14: Summary Statistics for 2010 Follow-Up Survey Table OA15: Selected variable definitions Table OA16: Who Knew About the Payouts, and How Were They Spent? Table OA17: Ex Post Effects of Insurance Payouts Table OA18: Ex Post Effects of Insurance Payouts and Education 12 Table OA1: Additional Household Summary Statistics and Variable Descriptions Mean St. dev. p10 p50 p90 Ex Ante Variables Group Savings (1="Yes") 0.60 0.49 0 1 1 Family and Friends Credit (1="Yes") 0.79 0.41 0 1 1 Purchased rainfall insurance previously (1="Yes") 0.24 0.43 0 0 1 Instrumented probability of purchasing insurance in 2006 > 0.49 0.50 0 0 1 median probability Wealth Index (Principal Component) Squared 2.90 4.75 0.04 1.32 7.20 Ex Post Variables Informal Credit (1="Yes") 0.20 0.30 0 0 0.5 Log Bank Credit Amount, Rs. 8.61 2.89 0 9.90 10.86 Log Any Credit Amount, Rs. 10.42 2.78 9.21 11.13 12.21 Bank Credit Applied (1="Yes") 0.35 0.48 0 0 1 Any Credit Applied (1="Yes") 0.54 0.50 0 1 1 Bank Credit Approved (1="Yes") 0.34 0.47 0 0 1 Any Credit Approved (1="Yes") 0.51 0.50 0 1 1 Notes: Ex ante summary statistics for the sample of 1479 farmers that participated in both the baseline and follow- up surveys and ex post variables' summary statistics for the sample of 736 farmers in the control group. See Table OA15 for variable definitions. Table OA2: Test of Balance Treatment Control Robust p- Significance Variable Mean Mean Difference value Level A. Demographic Characteristics Household size 5.13 5.18 -0.05 0.62 Age of household head (Years) 49.84 49.35 0.49 0.45 Gender of household head (1="Male") 0.92 0.91 0.01 0.38 Number of children 6 years old or younger 0.19 0.20 -0.01 0.72 Number of children 18 years old or younger 1.67 1.72 -0.05 0.49 B. Education of Household Head Highest level of schooling completed by head (Years) 3.65 3.80 -0.15 0.54 Household head is unschooled (1 = "Yes") 0.56 0.54 0.02 0.32 Household head has 1 to 5 years of schooling (1="Yes") 0.14 0.15 -0.01 0.46 Household head has more than 6 years of schooling (1="Yes") 0.30 0.32 -0.02 0.36 Household head able to read (1 = "Yes") 0.42 0.44 -0.02 0.37 Household head able to write (1 = "Yes") 0.38 0.41 -0.03 0.25 Raven's test of analytic intelligence, household head 1.58 1.57 0.01 0.84 C. Knowledge, Trust and Expectations of Household Head Insurance knowledge index (0-5) 1.80 1.66 0.14 0.21 Head has heard of rainfall insurance (1 = "Yes") 0.43 0.40 0.03 0.22 Trust in Basix (1 if trust BASIX > 4/10, 0 otherwise) 0.39 0.41 0.02 0.50 St. dev. of expected cash crop yield (kg/acre) 47.42 45.01 2.41 0.22 D. Credit and Assets Bank credit (1= "Yes") 0.72 0.69 0.02 0.30 Any Credit (1="Yes") 0.91 0.91 0.00 0.97 Total area of agricultural land (Acres) 5.44 5.29 0.15 0.59 Wealth Index (Principal Component) -0.04 0.04 -0.08 0.39 Total amount of savings, all sources (Rs.) 22093 20607 1486 0.33 Amount of savings with bank or post office (Rs.) 1735 1413 322.1 0.27 Amount of savings in cash at home (Rs.) 1832 1597 235.1 0.16 Amount of savings in jewelry (Rs.) 13275 13396 -121.8 0.91 Amount of savings with SHG or other group (Rs.) 2186 2117 68.39 0.79 Amount of other savings (Rs.) 3065 2083 982.0 0.10 * Total amount of credit owed, all sources (Rs.) 41320 41972 -652.1 0.80 Amount of credit from bank (Rs.) 21168 19652 1516 0.36 Amount of credit from family and friends (Rs.) 6810 5998 812.2 0.42 Amount of credit from MFIs (Rs.) 557.9 827.5 -269.6 0.16 Amount of credit from moneylenders (Rs.) 11505 14000 -2496 0.12 Amount of credit from other sources of credit (Rs.) 1279 1494 -215.3 0.42 E. Livestock and other Assets (as of June 2009) Number of large animals owned 2.22 2.29 -0.07 0.64 Number of small animals owned 4.70 5.77 -1.07 0.36 Total market value of livestock owned (Rs.) 31922 36626 -4704 0.12 House type: strong structure (1 = "Yes") 0.54 0.55 -0.01 0.65 House type: semi-strong structure (1 = "Yes") 0.33 0.32 0.01 0.67 House type: weak structure (1 = "Yes") 0.13 0.12 0.01 0.81 Number or rooms in the house 2.62 2.65 -0.03 0.72 Estimated value of the house if sold (Rs.) 117097 117346 -248.7 0.98 Est. value of agricultural land (Rs.) 558434 457887 100547 0.19 Est. value of non-agri. land and other houses (Rs.) 6677 10615 -3938 0.13 F. Agricultural Investments during 2008 Monsoon Total cultivated land (Acres) - all crops 4.38 4.24 0.14 0.50 Total cultivated land (Acres) - cash crops 3.48 3.32 0.16 0.27 Any land planted - cash crops (1 = Yes) 0.92 0.92 0.00 0.66 Total amount spent on inputs - all crops 20036 20115 -78.51 0.94 Table OA2: Test of Balance (Continued) Treatment Control Robust p- Significance Variable Mean Mean Difference value Level Amount spent on hybrid seeds - all crops 853.3 844.3 9.00 0.94 Amount spent on improved seeds - all crops 4,374 4,356 18.58 0.96 Amount spent on fertilizer - all crops 3,287 3,267 20.47 0.92 Amount spent on manure - all crops 2,073 2,339 -266.1 0.10 Amount spent on irrigation - all crops 119.6 181.7 -62.02 0.18 Amount spent on hiring tractor/other impl - all crops 2,723 2,635 88.07 0.64 Amount spent on manual labor - all crops 5,028 4,896 132.4 0.63 Amount spent on bullock labor - all crops 1,578 1,597 -18.90 0.88 Total amount spent on inputs - cash crops 15,868 15,923 -54.52 0.94 Amount spent on hybrid seeds - cash crops 455.0 427.1 27.91 0.61 Amount spent on improved seeds - cash crops 4,012 3,969 42.30 0.89 Amount spent on fertilizer - cash crops 2,284 2,302 -17.67 0.88 Amount spent on manure - cash crops 1,754 1,928 -174.6 0.24 Amount spent on irrigation - cash crops 22.53 40.23 -17.70 0.29 Amount spent on hiring tractor/other impl - cash crops 2,083 1,960 123.1 0.28 Amount spent on manual labor - cash crops 3,866 3,884 -18.55 0.93 Amount spent on bullock labor - cash crops 1,392 1,412 -19.27 0.86 Notes: The table reports a randomization test run on the baseline sample, where each variable is tested for treatment assignment. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. An F-test confirms that the variables are jointly insignificant: the F-statistic is 0.91, and the corresponding p-value is 0.6656. Table OA3: Investment in 2009 Monsoon All Crops Cash Crops Amount Amount >0 Mean St. dev. p10 p50 p90 >0 Mean St. dev. p10 p50 p90 Total amount used on agricultural inputs (Rs.): Hybrid seeds 0.64 1,927 4,423 0 800 4000 Improved seeds 0.57 3,977 6,725 0 1270 10200 Manure 0.74 3,379 4,643 0 2000 8000 Pesticide 0.64 1,490 2,913 0 625 4000 Irrigation 0.27 1,134 2,843 0 0 4000 Hiring tractors or other implements 0.91 3,517 3,619 600 3000 6700 Manual labor 0.94 3,030 3,492 500 2000 6000 Bullock labor 0.68 1,253 2,004 0 1000 3000 Total amount used, all inputs 0.98 22,909 21,955 6000 16500 45050 Market value of agricultural inputs used (Rs.) Hybrid seeds 0.64 1,992 4,760 0 800 4000 0.16 506 3,209 0 0 1000 Improved seeds 0.55 3,875 6,537 0 1000 11000 0.3 2,468 5,873 0 0 8000 Fertilizer 0.93 3,364 4,425 500 2000 7500 0.43 1,080 2,051 0 0 3000 Manure 0.73 3,339 4,592 0 2000 8000 0.34 1,348 3,071 0 0 4000 Pesticide 0.63 1,490 2,908 0 600 4000 0.28 545 2,186 0 0 1500 Irrigation 0.27 1,114 2,758 0 0 4000 Hiring tractors or other implements 0.91 3,509 3,688 500 3000 6500 Manual labor 0.94 3,000 3,520 500 2000 6000 Bullock labor 0.68 1,252 2,007 0 1000 3000 Total market value used, inputs 1-5 0.96 14,060 15,531 2500 9500 29500 0.47 6,195 12,174 0 0 17700 Total market value used, all inputs 0.97 22,934 22,169 5500 16550 46000 Notes: Summary statistics for agricultural investments during the 2009 monsoon, measured for the control group (the sample of 736 farmers who did not receive the insurance treatment). Table OA4: OLS estimator instead of probit and tobit Household covariates Dependent variable: No Yes A. Investments in all crops Any ag. inputs used (1 = Yes) 0.009 0.010 (0.008) (0.008) ln(1+total cultivated land, acres) 0.029 (0.040 (0.033) (0.030) 0.079 0.108 ln(1+market value of ag. inputs used, Rs.) (0.086) (0.082) B. Investments in cash crops Any ag. inputs used (1 = Yes) 0.045** 0.047** (0.022) (0.022) ln(1+total cultivated land, acres) 0.080** 0.087** (0.035) (0.034) 0.380* 0.408** ln(1+market value of ag. inputs used, Rs.) (0.195) (0.193) Cash crop shares Share of total cultivated land planted with cash crops 0.043** 0.043* (0.020) (0.020) Share of market val. of ag. inputs devoted to cash crops 0.028 0.028 (0.019) (0.019) Notes: This table reports the marginal effect of the insurance treatment dummy on various measures of monsoon agricultural investments; each row presents a different dependent variable. The first three dependent variables relate to investments in all crops. Dependent variables in the next three regressions relate to investments in cash crops only. The final two specifications consider the share of total agricultural inputs used for growing cash crops. Two versions of each model are presented, one without additional household covariates (Column 1) and one with (Column 2). These household covariates are Age of Head, Education of Head, and the Wealth Index. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Linear estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA5: Cash crop treatment effects disaggregated by investment type Household covariates included? No Yes Dependent variable: (1) (2) A. Hybrid Seeds Positive Investment/Any ag. inputs used (1 = Yes) 0.024 0.026 (0.024) (0.024) ln(1+market value of ag. inputs used, Rs.) 0.129 0.139 (0.128) (0.127) B. Improved Seeds Positive Investment/Any ag. inputs used (1 = Yes) 0.025 0.029 (0.029) (0.029) ln(1+market value of ag. inputs used, Rs.) 0.134 0.170 (0.166) (0.165) C. Fertilizer Positive Investment/Any ag. inputs used (1 = Yes) 0.045 0.050* (0.029) (0.029) ln(1+market value of ag. inputs used, Rs.) 0.292 0.321* (0.185) (0.183) D. Manure Positive Investment/Any ag. inputs used (1 = Yes) 0.020 0.023 (0.027) (0.027) ln(1+market value of ag. inputs used, Rs.) 0.139 0.162 (0.180) (0.179) E. Pesticide Positive Investment/Any ag. inputs used (1 = Yes) 0.054** 0.056** (0.026) (0.026) ln(1+market value of ag. inputs used, Rs.) 0.310** 0.328** (0.149) (0.148) Notes: This table reports the marginal effect of the insurance treatment dummy on various measures of monsoon agricultural investments; each row presents a different dependent variable. Each of the sections A through E has a different disaggregated category of investment as the dependent variable, such as investent in hybrid seeds. Two versions of each model are presented, one without additional household covariates (Column 1) and one with (Column 2). These household covariates are Age of Head, Education of Head, and the Wealth Index. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Probit and tobit estimators. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA6: Takeup of Credit during the Monsoon Dependent Variable: Takeup of Credit during the Monsoon Log Bank Log Any Bank Bank Informal Credit Credit Credit Any Credit Credit Any Credit Credit Amount Amount Applied Applied Approved Approved (1) (2) (3) (4) (5) (6) (7) A. No Interaction Effects Treatment dummy: 0.020 -0.075 0.145 -0.016 -0.017 -0.030 -0.034 (0.026) (0.211) (0.147) (0.025) (0.028) (0.025) (0.028) B. Includes Interaction Effects Treatment dummy: -0.081 -1.049 -0.263 -0.121 -0.055 -0.166 -0.067 (0.129) (1.017) (0.728) (0.123) (0.141) (0.120) (0.140) Interaction effects: treat x Wealth Index -0.002 0.174 0.058 -0.00 0.004 -0.010 0.003 (0.016) (0.131) (0.097) (0.015) (0.018) (0.015) (0.018) treat x Age of Head 0.00 0.014 0.011 0.001 -0.00 0.002 0.00 (0.002) (0.018) (0.013) (0.002) (0.003) (0.002) (0.002) treat x Education of Head 0.008 0.062 0.037 0.003 0.006 0.003 0.004 (years) (0.006) (0.044) (0.030) (0.006) (0.006) (0.006) (0.006) treat x Insurance Knowledge -0.022 0.237 0.131 -0.010 0.031 -0.008 0.021 Index (0.031) (0.190) (0.132) (0.030) (0.039) (0.030) (0.037) treat x Head has heard of 0.211 -0.725 -0.608 0.002 -0.151 0.007 -0.118 rainfall insurance (0.141) (0.834) (0.572) (0.130) (0.169) (0.129) (0.159) treat x Trust BASIX (1=yes) -0.035 -0.034 -0.105 0.077 0.049 0.066 0.051 (0.058) (0.457) (0.326) (0.061) (0.064) (0.060) (0.064) treat x Payout (1,000 Rs.) 0.064 -0.041 -0.451 0.046 0.023 0.040 0.010 (0.066) (0.549) (0.382) (0.067) (0.073) (0.066) (0.073) Notes: This table reports the marginal effects of the treatment dummy on various dummies that represent takeup of credit during the monsoon. Panel A does not include any interaction variables; Panel B includes interactions between treatment and other household characteristics. Coefficients on uninteracted covariates (for Panel B) are not shown. Each column represents a different dependent variable. Probit and tobit estimators. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 individuals that participated in both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA7: Heterogeneous Effects of Insurance Treatment on Land Cultivated for Cash Crops Dependent variable: ln(1+land cultivated for cash crops) Baseline model Single interactions of treatment with: Multivariate [as Table 3] Wealth Age Education Knowledge Trust Payouts (1) (2) (3) (4) (5) (6) (7) (8) Insurance treatment dummy 0.086** 0.096*** 0.119 -0.012 0.079* 0.067 0.040 -0.156 (0.037) (0.036) (0.148) (0.047) (0.046) (0.047) (0.052) (0.172) Interaction effects: treat x Wealth Index -0.018 -0.031 (0.022) (0.023) treat x Age of Head 0.000 0.002 (0.002) (0.003) treat x Education of Head (years) 0.026*** 0.029*** (0.007) (0.008) treat x Insurance Knowledge Index (0-5) 0.018 0.005 (0.047) (0.046) treat x Head has heard of -0.055 -0.031 rainfall insurance (0.199) (0.194) treat x Trust BASIX (1= yes) 0.047 0.032 (0.077) (0.080) treat x Payout (1,000 Rs.) 0.109 0.100 (0.102) (0.102) Uninteracted household characteristics: Wealth Index 0.091*** 0.096*** (0.016) (0.016) Age of Head -0.001 -0.001 (0.003) (0.002) Education of Head (years) -0.005 -0.012** (0.006) (0.006) Insurance Knowledge Index -0.068** -0.066** (0.034) (0.033) Head has heard of rainfall 0.283* 0.259* insurance (1=yes) (0.154) (0.148) Trust BASIX (1=yes) 0.010 -0.018 (0.056) (0.057) Payout (1,000 Rs.) na na Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy as well as interactions between the treatment dummy and various household characteristics. Dependent variable is ln(1+ land cultivated for cash crops). Since the "knowledge of insurance" questions were only asked of farmers that were aware of insurance, the specification including the knowledge index also includes a dummy for whether the farmer had heard of insurance. No direct effect of insurance contract payout on investment reported since payout only varies at the village level. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Tobit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA8: Heterogeneous Effects of Insurance Treatment on Investment in Cash Crops Dependent variable: ln(1+investment in cash crops, Rs.) Baseline model Single interactions of treatment with: Multivariate [as Table 3] Wealth Age Education Knowledge Trust Payouts (1) (2) (3) (4) (5) (6) (7) (8) Insurance treatment dummy 0.451** 0.491** 0.156 -0.032 0.446* 0.456 0.188 -1.432 (0.218) (0.219) (0.876) (0.282) (0.271) (0.281) (0.307) (1.033) Interaction effects: treat x Wealth Index -0.124 -0.181 (0.133) (0.137) treat x Age of Head 0.006 0.023 (0.018) (0.019) treat x Education of Head (years) 0.127*** 0.166*** (0.045) (0.048) treat x Insurance Knowledge Index (0-5) 0.137 0.078 (0.284) (0.284) treat x Head has heard of -0.565 -0.361 rainfall insurance (1.204) (1.205) treat x Trust BASIX (1= yes) -0.009 -0.066 (0.456) (0.481) treat x Payout (1,000 Rs.) 0.627 0.513 (0.611) (0.622) Uninteracted household characteristics: Wealth Index 0.362*** 0.389*** (0.095) (0.096) Age of Head -0.015 -0.025* (0.013) (0.014) Education of Head (years) -0.032 -0.086** (0.034) (0.037) Insurance Knowledge Index -0.519** -0.506** (0.214) (0.211) Head has heard of rainfall 2.356** 2.234** insurance (1=yes) (0.958) (0.949) Trust BASIX (1=yes) 0.184 0.040 (0.334) (0.346) Payout (1,000 Rs.) na na Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy as well as interactions between the treatment dummy and various household characteristics. Dependent variable is equal to ln(1+investment in cash crops, Rs.). Since the "knowledge of insurance" questions were only asked of farmers that were aware of insurance, the specification including the knowledge index also includes a dummy for whether the farmer had heard of insurance. No direct effect of insurance contract payout on investment reported since payout only varies at the village level. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Tobit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA9: Heterogeneous Effects of Insurance Treatment: Additional baseline variables Dependent variable: Positive investment in cash crops (1=Yes) Single interactions of treatment with: Family Group Land Wealth Multi- Liter- Any Bank Friends Saving Land holdings Index Prev. 2006 Ins. St. dev. of Raven's variate acy Credit Credit Credit s holdings >2 Sq. ins. Instmnt. E(Yield) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) Treatment dummy: -0.001 0.064 0.035 0.051 0.062 0.066 0.081 0.077 0.079** 0.063* 0.037 0.090* -0.239 (0.039) (0.056) (0.194) (0.055) (0.065) (0.046) (0.050) (0.069) (0.035) (0.034) (0.071) (0.047) (0.321) Interaction effects: treat x Literacy 0.147** -0.006 (0.059) (0.124) treat x Raven's -0.003 0.016 (0.031) (0.033) treat x Access to any 0.026 0.045 credit (0.197) (0.255) treat x Access to bank 0.012 0.044 credit (0.065) (0.109) treat x Access to -0.004 -0.061 Family Credit (0.073) (0.125) -0.010 -0.025 treat x Group Savings (0.061) (0.065) treat x Landholdings -0.005 0.003 (0.008) (0.010) treat x (Landholdings -0.033 -0.072 >2) (0.076) (0.088) treat x Wealth Index -0.012 -0.025 (0.018) (0.025) treat x Wealth Index -0.006 -0.011 Squared (0.007) (0.010) treat x Prev. -0.006 0.072 Insurance (0.068) (0.097) treat x 2006 0.009 0.041 Insurance Instrument (0.082) (0.087) treat x SD of exp. -0.001 -0.001 (0.001) (0.001) treat x Age of Head 0.004 (0.003) treat x Educ. of Head 0.026** (years) (0.013) treat x Insurance 0.001 Knowledge Index (0.036) treat x Head has heard -0.042 of rainfall insurance (0.167) treat x Trust BASIX -0.013 (1=yes) (0.069) treat x Payout (1,000 0.038 (0.083) Hypothesis tests on interaction variables Joint significance: wealth and wealth squared 0.533 Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy as well as interactions between the treatment dummy and various household characteristics. Dependent variable is equal to 1 if the farmer invested resources in planting cash crops. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Probit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA10: Education and the Effects of Insurance on Land Cultivated for Cash Crops Dependent variable: ln(1+land cultivated for cash crops) Education Interaction Variable(s): Education Additional Literacy + Raven's score + Education interacted with: (years) [as education years of years of Land- Bank credit St. dev. of OA8 Col 8] dummies education education holdings usage expected ag yield (1) (2) (3) (4) (5) (6) (7) Insurance treatment dummy -0.156 -0.160 -0.176 -0.195 -0.189 -0.098 -0.081 (0.172) (0.175) (0.174) (0.178) (0.178) (0.187) (0.176) Interaction effects: treat x Education of Head (years) 0.029*** 0.042* 0.020 0.029*** (0.008) (0.022) (0.014) (0.008) treat x Education = 1-5 years 0.051 (0.138)   treat x Education = 6 + years -0.150 (0.236) treat x Head can read 0.109 (0.134) treat x Raven's Test score 0.030 (0.038) treat x Education x landholdings > 0.036*** median (0.009) treat x Education x landholdings < 0.012 median (0.012) treat x Education x bank credit 0.031*** (0.009) treat x Education x No bank credit 0.022 (0.017) treat x Education x st dev of exp yield 0.049*** > median (0.011) treat x Education x st dev of exp yield 0.013 < median (0.010) Uninteracted education measures Education of Head (years) -0.012** -0.032** -0.014 -0.013** (0.006) (0.016) (0.010) (0.006) Education = 1-5years 0.064 (0.096) Education = 6+ years 0.230 (0.172) Head can read 0.015 (0.097) Raven's Test score -0.032 (0.027) Education x landholdings > median -0.023*** (0.007) Education x landholdings < median 0.001 (0.009) Education x bank credit -0.014*** (0.007) Education x no bank credit -0.006 (0.011) Education x st. dev. of exp yields > -0.025*** median (0.009) Education x st. dev. of exp yields < -0.003 median (0.008) Landholdings dummies no no no no yes no no Bank credit dummies no no no no no yes no St dev of exp yield dummies no no no no no no yes Other covars. from Table 4? yes yes yes yes yes yes yes Hypothesis tests on interaction variables Joint significance: 1-5 and 6+ yrs education 0.387 Joint significance: literacy and education 0.003 Joint significance: Raven's and education 0.003 Equality: high vs low landholding 0.272 Equality: Bank credit vs no credit 0.381 Equality: High SD vs Low SD of exp yields 0.004 Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy and various interactions between the insurance treatment and measures of education and cognition. All specifications include other covariates from Table 4 (wealth, age, knowledge of insurance/heard of insurance, trust, payouts) and their interactions with the treatment dummy. “Landholdings dummies” refers to dummy for (landholdings > median) and the interaction term (landholdings > median) x treatment. “Bank credit dummies” and “St dev of exp yield dummies” are similarly defined. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Tobit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA11: Education and the Effects of Insurance on Investment in Cash Crops Dependent variable: ln(1+investment in cash crops, Rs.) Education Interaction Variable(s): Education Additional Literacy + Raven's score + Education interacted with: (years) [as education years of years of Land- Bank credit St. dev. of OA9 Col 8] dummies education education holdings usage expected ag yield (1) (2) (3) (4) (5) (6) (7) Insurance treatment dummy -1.432 -1.350 -1.458 -1.627 -1.783 -1.233 -0.915 (1.033) (1.050) (1.048) (1.077) (1.106) (1.145) (1.057) Interaction effects: treat x Education of Head (years) 0.166*** 0.287** 0.151* 0.168*** (0.048) (0.132) (0.085) (0.049) treat x Education = 1-5 years -0.308 (0.840)   treat x Education = 6 + years -1.382 (1.408) treat x Head can read 0.188 (0.815) treat x Raven's Test score 0.077 (0.230) treat x Education x landholdings > 0.199*** median (0.058) treat x Education x landholdings < 0.102 median (0.078) treat x Education x bank credit 0.183*** (0.053) treat x Education x No bank credit 0.114 (0.106) treat x Education x st dev of exp yield 0.300*** > median (0.070) treat x Education x st dev of exp yield 0.058 < median (0.062) Uninteracted education measures Education of Head (years) -0.086** -0.258** -0.132** -0.087** (0.037) (0.103) (0.065) (0.037) Education = 1-5years 0.968 (0.617) Education = 6+ years 1.971* (1.066) Head can read 0.541 (0.612) Raven's Test score -0.214 (0.166) Education x landholdings > median -0.151*** (0.045) Education x landholdings < median 0.013 (0.056) Education x bank credit -0.108*** (0.040) Education x no bank credit -0.016 (0.073) Education x st. dev. of exp yields > -0.170*** median (0.055) Education x st. dev. of exp yields < -0.019 median (0.046) Landholdings dummies no no no no yes no no Bank credit dummies no no no no no yes no St dev of exp yield dummies no no no no no no yes Other covars. from Table 4? yes yes yes yes yes yes yes Hypothesis tests on interaction variables Joint significance: 1-5 and 6+ yrs education 0.387 Joint significance: literacy and education 0.003 Joint significance: Raven's and education 0.003 Equality: high vs low landholding 0.272 Equality: Bank credit vs no credit 0.381 Equality: High SD vs Low SD of exp yields 0.004 Notes: This table reports the marginal effects on cash crop investments of the insurance treatment dummy and various interactions between the insurance treatment and measures of education and cognition. All specifications include other covariates from Table 4 (wealth, age, knowledge of insurance/heard of insurance, trust, payouts) and their interactions with the treatment dummy. “Landholdings dummies” refers to dummy for (landholdings > median) and the interaction term (landholdings > median) x treatment. “Bank credit dummies” and “St dev of exp yield dummies” are similarly defined. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Tobit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA12: Different functional forms of investment dependent variable Dependent variable Functional Form of Total Cultivated Land Market Value of Ag. Inputs Used Dependent Variable Household covariates Household covariates No Yes No Yes A. All Crop Types ln (10+x) 0.007 0.011 0.059 0.085 (0.011) (0.009) (0.068) (0.064) ln(1+x) 0.028 0.039 0.082 0.110 (0.034) (0.031) (0.087) (0.083) ln(0.1+x) 0.042 0.054* 0.082 0.110 (0.036) (0.032) (0.087) (0.083) ln(0.01+x) 0.044 0.056* 0.082 0.110 (0.036) (0.032) (0.087) (0.083) Inverse hyperbolic sine 0.038 0.051 0.088 0.117 (0.042) (0.039) (0.093) (0.088) B. Cash Crops Only ln (10+x) 0.019** 0.021** 0.264* 0.285** (0.009) (0.008) (0.136) (0.134) ln(1+x) 0.086** 0.093*** 0.451** 0.485** (0.037) (0.036) (0.218) (0.216) ln(0.1+x) 0.076** 0.084*** 0.451** 0.485** (0.032) (0.030) (0.218) (0.216) ln(0.01+x) 0.075** 0.083*** 0.451** 0.485** (0.031) (0.030) (0.218) (0.216) Inverse hyperbolic sine 0.111** 0.120*** 0.487** 0.523** (0.047) (0.046) (0.235) (0.233) Notes: This table reports the marginal effect of the insurance treatment dummy on various measures of monsoon agricultural investments; each row presents a different functional form of the dependent variable. The first five dependent variables relate to investments in all crops. Dependent variables in the next five regressions relate to investments in cash crops only. The final two specifications consider the share of total agricultural inputs used for growing cash crops. Two versions of each model are presented, one without additional household covariates and one with. These household covariates are Age of Head, Education of Head, and the Wealth Index. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Tobit estimator. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA13: Treatment Effects, Winsorized Two Percent Household covariates included? Dependent variable: No Yes (1) (2) A. Investments in all crops Positive investment/Any ag. inputs used (1 = Yes) 0.012 0.009 (0.011) (0.008) ln(1+total cultivated land, acres) 0.029 0.041 (0.033) (0.030) ln(1+market value of ag. inputs used, Rs.) 0.088 0.118 (0.087) (0.083) B. Investments in cash crops Positive investment/Any ag. inputs used (1 = Yes) 0.060** 0.064** (0.029) (0.030) ln(1+total cultivated land, acres) 0.086** 0.094*** (0.036) (0.035) ln(1+market value of ag. inputs used, Rs.) 0.455** 0.490** (0.218) (0.216) Cash crop shares Share of total cultivated land planted with cash crops 0.047** 0.048** (0.021) (0.021) Share of market value of ag. inputs devoted to cash crops 0.034* 0.035* (0.019) (0.019) Notes: This table reports the marginal effect of the insurance treatment dummy on various measures of monsoon agricultural investments after 2 percent winsorization; each row presents a different dependent variable. The first three dependent variables relate to investments in all crops. Dependent variables in the next three regressions relate to investments in cash crops only. The final two specifications consider the share of total agricultural inputs used for growing cash crops. Two versions of each model are presented, one without additional household covariates (Column 1) and one with (Column 2). These household covariates are Age of Head, Education of Head, and the Wealth Index. Cash crops are defined as castor in Mahbubnagar and groundnut in Anantapur. Probit and tobit estimators. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1,479 farmers that completed both the baseline and follow-up surveys. *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA14: Summary Statistics for 2010 Follow-Up Survey Variable N Mean St. Dev. p10 p50 p90 A: Attitudes Towards Insurance and Financial Outcomes How Trustworthy: ICICI Lombard (0-10) 1,445 5.33 2.09 3 5 8 BASIX Risk Product Pays Out During Drought (1-5) 1,459 3.69 0.71 3 4 5 Savings in Bank or Cash (Winsorized) 1,459 4,918 9,821 0 2,000 12,000 Total Outstanding Debt (Winsorized) 1,459 3,708 5,404 30 100 10,080 Dummy: Any High Interest Debt 1,459 0.59 0.49 0 1 1 Sum of High Interest Debt (Winsorized) 1,459 1,482 3,345 0 20 10,000 B: Ex Post Real Outcomes Log (Acres Planted in Rabi) 1,459 0.43 0.56 0 0 1 Children Mean Hours Worked per Week 1,081 12.44 11.31 0 11 28 Dummy: HH Has NREGA Earnings 1,459 0.68 0.47 0 1 1 Dummy: HH Earned Income from Migration 1,459 0.11 0.32 0 0 1 Change in Value of Livestock Excluding Chickens and 1,459 4,989 11,670 0 0 19,000 Other Animals (Winsorized) Change in Value of Durable Goods (Winsorized) 1,459 1,267 6,151 0 0 0 Total Daily Consumption (Winsorized) 1,459 208 107 122 183 302 Notes: Summary statistics for the sample of 1459 individuals that participated in the 2010 Follow-Up survey. Table OA15: Selected variable definitions Variable Descriptive information Survey Baseline Variables [Summary statistics in Table OA1] Group Savings Dummy Variable equal to one if, at the time of the 2009 baseline 2009 Baseline survey, the household has savings from an self-help group (SHG) or other group (revolving fund). Family and Friends Credit Dummy variable equal to one if, at the time of the 2009 baseline 2009 Baseline survey, the respondent either currently has outstanding loans from family and friends, or equal to one if indicates that has applied for credit from family and friends since the end of 2008 monsoon and was approved at least once. Purchased Insurance Previously Dummy Variable equal to one if, at the time of the 2009 baseline survey, the respondent has purchased rainfall insurance in previous 2009 Baseline years. Instrumented Probability of Purchasing Insurance in 2006 Dummy variable equal to one if the respondent had a greater than Greater than Median Probability median probability of purchasing insurance in 2006. Uses as instruments the marketing treatments applied in our prior research 2006 (see Cole et al., 2013) because these randomly assigned treatments affect the probability that farmers purchased insurance in 2006. High Interest Debt Amount of credit from a microfinance institution (MFI), moneylender or other sources (other than credit from banks or from 2009 Baseline family or friends), at the time of the 2009 baseline survey. Ex-Post Investment-Related Variables [Summary statistics in Table OA14] Informal Credit Variable averaging two dummy variables, both asked at the time of the 2009 follow-up survey: 1.) Received more help from 2009 Follow-Up relatives/neighbors relative to same time last year, and 2.) Received more gifts from relatives/neighbors relative to same time last year. Log Bank Credit Amount The logarithm of the total amount of outstanding loans owed to 2009 Follow-Up banks at the time of the 2009 follow-up survey, in Rs. Log Any Credit Amount The logarithm of the total amount of outstanding loans owed to banks, family and friends, microfinance institutions (such as BASIX, 2009 Follow-Up etc.), and moneylenders at the time of the 2009 follow-up survey, in Rs. Bank Credit Applied Dummy variable equal to one if, at the time of the 2009 follow-up survey, have applied for credit from a bank since Bharani (April 27, 2009 Follow-Up 2009). Any Credit Applied Dummy variable equal to one if, at the time of the 2009 follow-up survey, have applied for credit from a bank, family and friends, 2009 Follow-Up microfinance institutions (such as BASIX, etc.), or a moneylender since Bharani (April 27, 2009). Table OA15 (continued): Selected variable definitions Variable Descriptive information Survey Bank Credit Approved Dummy variable equal to one if, at the time of the 2009 follow-up survey, have approved for credit from a bank since Bharani (April 2009 Follow-Up 27, 2009). Any Credit Approved Dummy variable equal to one if, at the time of the 2009 follow-up survey, have approved for credit from a bank, family and friends, 2009 Follow-Up microfinance institutions (such as BASIX, etc.), or a moneylender since Bharani (April 27, 2009). Children: Mean Hours Worked Per Week Mean number of hours worked by each household member between ages 6 and 20. Households that did not report any members between 2009 Follow-Up ages 6 and 20 are omitted from the Table OA14 regression. Ex-Post Living Standards Variables in 2010 Follow-Up [summary statistics in Table OA14] How Trustworthy: ICICI Response to being asked, on a scale of 0-10, how trustworthy is your Lombard (0-10) BASIX LSA? 2010 Follow-Up BASIX Risk Product Pays Out Response to being asked, how strongly do you agree or disagree with During Drought (1-5) the statement, "Times when the BASIX rainfall insurance product pays out mathc up well with times of drought (i.e. the product has low basis risk)"? Scale: 1= "strongly disagree," 2="disagree," 3= 2010 Follow-Up "neither agree nor disagree," 4="agree," and 5="strongly agree." Livestock Buffalos, cows, young calves/young stock, oxen/bullocks, goats and 2009 Follow-Up & sheep . 2010 Follow-Up Durable Goods Includes value of tractors, animal-pulled equipment, electric motor/oil engine/pipeline, sprinkler set/drip equipment set, hand tools, thresher, insecticide pump, manuals, sprayer & dusters, 2009 Follow-Up & processing units, ox/bullocks cart, furniture, refrigerator, bicycle, 2010 Follow-Up motorcycle, sewing machine, electrical goods, telephone/cell phone, others Household has NREGA Earnings Dummy variable equal to 1 if any household member has worked in the NREGA (National Rural Employment Guarantee Act) since June 2010 Follow-Up 8, 2009. Household has Earned Income from Migration Dummy variable equal to 1 if any household member has received 2010 Follow-Up income paid in cash or kind from migration since June 8, 2009. Total Daily Consumption Total daily consumption, measured by summing consumption of different items measured over different time intervals between one 2010 Follow-Up day and three months (normalized to a per-day basis in each case). Table OA16: Who Knew About the Payouts, and How Were They Spent? Panel A. Self-Reported Use of Insurance Payout Responses to "What have you done with this money you % of Payout % Farmers have received?" Received Reporting > 0 N Invested in agricultural activities for the Rabi 16.43 44.61 529 Bought goods and services to be used straight away 26.93 83.64 538 Bought goods that will last a longer time 12.04 59.40 532 Paid off debts 12.44 53.53 538 Saved for 2010 Monsoon 20.41 70.76 537 Saved for later in the future 2.66 20.93 516 Gave to family and friends 10.36 3.75 520 Panel B: Knowledge About Insurance Status Household Family Friends Others Knew you had insurance 0.99 0.88 0.66 0.26 Knew you received a payout 0.99 0.72 0.49 0.21 Knew the size of your payout 0.93 0.51 0.31 0.09 Asked for money because of payout 0.70 0.19 0.03 0.01 Received money because of payout 0.48 0.08 0.01 0.01 Notes: This table provides self-reported data on the use of insurance payounts and the knowledge about insurance status by household, family, friends, and others. Data were collected during the second follow-up survey conducted in 2010. The sample includes the 535 treated farmers that participated in both the baseline, first and second follow-up surveys and received a positive rainfall insurance payout. Table OA17: Ex Post Effects of Receipt of Insurance Payouts Panel A: Effects on Attitudes Towards Insurance and Financial Outcomes (1) (2) (3) (4) (5) (6) ICICI BASIX Risk Total Lombard is Product Pays Savings in Outstanding Dummy: Any Sum of High trustworthy Out During Bank or Cash Debt High Interest Interest Debt (0-10) Drought (1-5) (Winsorized) (Winsorized) Debt (Winsorized) Insurance Treatment 0.357** -0.00423 123.9 83.18 0.281*** 282.0* (yes = 1) (0.154) (0.0495) (558.6) (281.4) (0.100) (167.2) Insurance Treatment x Payout (fraction of max) 0.144 0.395*** 881.9 -0.423 -0.438*** -383.3 (0.212) (0.0920) (806.2) (527.1) (0.164) (328.7) N 1445 1459 1459 1459 1459 1459 Mean of dep var (full sample) 5.33 3.69 4917.8 3708.46 0.59 1482.21 P-value of Test: Max Payout = Control 0.000 0.000 0.047 0.833 0.173 0.683 Estimator Linear Linear Tobit Tobit Probit Tobit Panel B: Effects on Real Outcomes (7) (8) (9) (10) (11) (12) (13) Children Dummy: HH Change in Log (Acres Mean Hours Has Dummy: HH Value of Change in Value Total Daily Planted in Worked per MGNREGS Earned Income Livestock of Durable Goods Consumption Rabi) Week Earnings from Migration (Winsorized) (Winsorized) (Winsorized) Insurance Treatment -0.0131 1.606* -0.114 -0.115 759.2 255.4 -4.251 (yes = 1) (0.0388) (0.922) (0.109) (0.148) (922.9) (509.0) (7.995) Insurance Treatment x Payout (fraction of max) 0.0195 -2.110 0.0906 -0.00844 -1817.6 -92.29 -10.97 (0.0680) (1.400) (0.186) (0.235) (1252.1) (567.0) (9.547) N 1459 1081 1336 1270 1459 1459 1459 Mean of dep var (full sample) 0.429 12.440 0.704 0.128 4989.0 1267.2 207.6 P-value of Test: Max Payout = Control 0.897 0.600 0.857 0.406 0.177 0.564 0.012 Estimator Tobit Tobit Probit Probit Linear Linear Linear Notes: This tabel reports ex-post effects of insurance payouts; each column is a different dependent variable. The table reports marginal effects for the treatment dummy and for the interaction between the treatment dummy and amount paid out. Panel A examines how product experience affects attitudes towards insurance, measured by asking farmers to react to questions on a Likert scale. Panel B analyzes whether insurance cash payouts affected ex post real decisions. Data were collected during the second follow up survey in 2010.Children hours worked per week is for children ages 6-20 (380 households have no children). In specifications 9 and 10, the dependent variable is zero for all observations in several villages, so that the observations are dropped from the probit regression. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. Sample includes the 1459 individuals that participated in the baseline survey and first and second follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively. Table OA18: Ex Post Effects of Insurance Payouts and Education (1) (2) ICICI Lombard is trustworthy BASIX Risk Product Pays Out (0-10) During Drought (1-5) Treatment dummy 0.404*** 0.190*** (0.133) (0.046) Interaction effects: treat x Education of Head (years) 0.006 -0.006 (0.022) (0.007) Uninteracted education measures Education of Head (years) 0.036** 0.007 (0.016) (0.006) N 1445 1459 Estimator Linear Linear Notes: This tabel reports ex post effects of education on two dependent variables: trust in ICICI Lombard and belief that the risk product pays out during drought. The table reports marginal effects for the treatment dummy, the interaction between the treatment dummy and education, and for uninteracted education. The dependent variables are measured by asking farmers to react to questions on a Likert scale. Data were collected during the second follow up survey in 2010. Robust standard errors. All specifications include village dummies and a dummy for whether the household received the fertilizer discount. The sample includes the 1459 individuals that participated in the baseline survey and first and second follow-up surveys. Symbols *, **, *** denote significance at the 10, 5 and 1 percent level, respectively.