Today: Two (Closely Related) Experiments How Does Risk Management Influence Production Decisions? Evidence from a Field Experiment Shawn Cole Xavier Gine James Vickery (HBS and NBER) (World Bank) (NY Fed) Dynamics of Demand: Evidence from an Eight-Year Panel in Gujarat Shawn Cole Dan Stein Jeremy Tobacman (HBS and NBER) (World Bank) (Wharton) Shawn Cole 3ie December 17, 2013
Motivation Agriculture is the primary activity of 1/3 of the world, and 2/3 of India Rainfall is an important determinant of yield and revenue Lasting, successful, unsubsidized crop insurance was practically non-existent as recently as five years ago Financial innovation, namely introduction of index-based insurance, promises to create an important new insurance market 2
Three Questions Does the ability to hedge yield risk affect ex-ante production decisions? Could a private market in retail weather derivatives succeed? How will the market evolve? 3
Today s Talk Introduction and Motivation Field experiment I (2009): Impact of Weather Insurance Provide index insurance policies to 750 of 1500 farmers Ex-Ante: Measure subsequent investment decisions Ex-Post: Attitudes towards insurance, assets, etc. Field Experiment II (2010): The Price is Right? Elicit willingness to pay for policies Vary contract terms: does WTP differ? Field Experiment III (2005-2013): Dynamics of Demand Conclusion 4
Underinsurance and Underinvestment Estimates of marginal rates of return on investment in developing countries are generally very high (Duflo, Kremer and Robinson, 2008; Suri, 2010; McKenzie et al. 2009) Yet adoption of high return technologies is not universal. One possible explanation: high expected returns are compensation for uninsured production risk. Some evidence of costly income smoothing (Morduch, 1995, Chetty and Looney, 2006; Binswanger and Rosenzweig, 1992). 5
Related literature Cole et al. (2013) First experimental study of index insurance (2006) Demonstrates low adoption, in general (ca 5-20%) Even when price is better than actuarially fair Trust and liquidity key barriers to adoption Karlan et al. (2013) Provide large capital and insurance grants to Ghanaian farmers Insurance increases amount invested, but not type of investment Mubarak et al. (2013) Study how informal risk-sharing interacts with demand for insurance 6
The Experiment Sample: ca. 1,500 households from 2 districts of drought-prone areas of Andhra Pradesh. Two-thirds are part of earlier 2004 & 2006 surveys. Remainder from study villages + nearby villages. Randomization design: Half of the farmers (chosen randomly) were given 10 Phase-I weather insurance policies that would cover all inputs cost (seed, FYM, fertilizer and labor) for a hectare of main cash crop in the district. 7
Insurance Design (Example contract) (2000Rs) (900Rs) payout for phase Retail-level rainfall derivative Underwritten by a large insurer (ICICI Lombard) and marketed by local MFIs Insurance splits monsoon into three phases: (i) Sowing (Phase I) (ii) Podding / flowering (iii)harvest Payouts in each phase based on cumulative rainfall in the phase (each is 35-45 days) 2 nd trigger [corresponds to crop failure] (40mm) 1 st trigger (100mm) rainfall during phase 8
Key Benefits of Product No adverse selection (except possibly temporal) No moral hazard Historical rainfall data can be used to set prices Insurable in international risk markets Divisible (policies as cheap as $1.50) and easy to purchase Fast settlement and payment Little discretion, relatively difficult to politicize Identifies ex-ante who needs disaster relief 9
Key limitations Basis Risk (rainfall at farm, and consumption, imperfectly correlated with rainfall at the rain gauge) Expensive, in part due to small scale. Payout 30-60% of unsubsidized premium Complicated to understand and evaluate May crowd out informal insurance (or have unanticipated general equilibrium effects) 10
The Experiment Sample: ca. 1,500 households from 2 districts of drought-prone areas of Andhra Pradesh. Two-thirds are part of earlier 2004 & 2006 surveys. Remainder from study villages + nearby villages. Randomization design: Half of the farmers (chosen randomly) were given 10 Phase-I weather insurance policies that would cover all inputs cost (seed, FYM, fertilizer and labor) for a hectare of main cash crop in the district. Other half of farmers receive coupon for approximate expected value of the policy (Rs 350) to be redeemable after harvest, when payouts (if any) are due. We do this to control for any wealth effect. 11
Andhra Pradesh Sample and Setting Sample 1,000 households from 37 villages in two districts of Andhra Pradesh, first identified in 2004. Representative sample 500 additional households added in 2009, from same sampling frame BASIX (MFI) sells ICICI/Lombard insurance Experiments conducted by ICRISAT staff Relatively wealthy, groundnut and castor farmers 12
Hypotheses Farmers under-invest in inputs due to rainfall risk Randomly assign insurance at the start of the monsoon Effects on total investment? Substitution between cash crops and subsistence crops? Cash crops are more profitable and more risky as per Government of India estimates: Crop Estimated Profit / Hectare Rainfall Requirement Castor 2,791 625mm Groundnut 2,951 533 mm Sorghum -212 376 mm 13
Investment in Kharif 2009 All Crops Amount >0 Mean SD Cash Crops Amount >0 Mean SD Land use: Total cultivated land 0.93 4.00 3.59 0.48 1.92 2.98 In which Kartis did farmer plant? 15.68 2.78 15.26 2.29 Did farmer replant crop this Kharif? (1="Yes" ) 0.05 Did farmer abandon crop this Kharif? (1="Yes") 0.18 Market value used for all crops: Hybrid seeds 0.63 1,852 3,961 0.17 476 2,480 Improved seeds 0.56 3,779 6,012 0.31 2,402 5,537 Fertilizer 0.93 3,208 3,873 0.45 1,098 1,973 Manure 0.73 3,182 4,213 0.35 1,360 3,112 Pesticide 0.64 1,447 2,659 0.30 544 1,823 Total market value used 0.96 13,467 13,758 0.49 6,123 11,096 14
Main Results: Effect of insurance on agricultural investments Crop types: All crop types Cash crops only Insurance dummy Insurance dummy Estimator (marginal effect) (marginal effect) A. Without household covariates Any ag. inputs used (=1 if yes) 0.016 0.060** probit (0.012) (0.029) ln(1+land under cultivation, acres) 0.029 0.163** tobit (0.034) (0.070) ln(1+ag. inputs used, Rs.) 0.082 0.800** tobit (0.087) (0.387) ln(1+ag. inputs purchased, Rs.) 0.05 tobit (0.079) N 1479 1479 Village dummies yes yes =>Roughly 26% increase in land devoted to cash crops from a base of two acres 15
Timing Figure: Fraction of farmers who had planted cash crops by different points during 2009 monsoon season: difference between treatment and control group. Figure note: Vertical lines show period during which field experiment was implemented. 16
Interaction effects Production response may depend on wealth, or experience with insurance product. (If unfamiliar, may adopt a wait and see approach.) We use previous field experiments (Cole et al., 2011) as instruments for whether these farmers purchased insurance in 2006. Two stage approach: 1 st stage: P[bought ins] 06 = f( 2006 treatments) 2 nd stage: investment 09 a.insurance 09 b.insurance09 P[boughtins] 06... Other interaction variables: (i) Farmer s wealth (landholdings, PCA-based wealth measure). (ii) Education (iii) Subjective measure of variation in yields (household self-report) (iv) Self-reported knowledge of insurance 17
Interaction effects: Education measures Dependent Variable: Treatment (1=yes) Hosehold covariate: Treatment x covariate Estimator Household head can read (1 = yes) Dummy: Investment in cash crops > 0-0.001-0.019 0.147** Probit (0.039) (0.043) (0.059) ln(1+investment in cash crops, Rs.) -0.042-0.077 0.469*** Tobit (0.093) (0.102) (0.141) ln(1+land cultivated for cash crops) -0.048-0.155 1.959** Tobit (0.523) (0.575) (0.779) ln(1+years of education) Dummy: Investment in cash crops > 0-0.007-0.013 0.073*** Probit (0.039) (0.019) (0.027) ln(1+investment in cash crops, Rs.) -0.053-0.04 0.218*** Tobit (0.094) (0.046) (0.063) ln(1+land cultivated for cash crops) -0.128-0.140 0.948*** Tobit (0.529) (0.259) (0.348) 18
Summary of findings Evidence that access to hedging instruments influences real investment for our sample of small farmers / firms. Main margin: substitution from less risky to more risky investments (subsistence crops to cash crops). Effect size in the eye of the beholder: Increase share planting cash crops from 45% to 51% Wealth doesn t seem to affect effect size Effect largest among more educated farmers 19
Will a rainfall insurance market succeed? In first five years, many studies of index insurance, few examples of commercial success Many reasons it should succeed: Rainfall risk uncorrelated with global capital markets (Gine et al., 2007) Many reasons it shouldn t succeed: Transaction costs: Gaurav, Cole, Tobacman (2011)=>Between $63 in marketing costs to sell $13 policy Basis risk 20
Necessary conditions for success Insurance makes purchasers better off AP: Insurance leads to more profitable investment, protects assets Karlan et al. (2013): Insurance increases total investment Mobarak and Rosenzweig (2013): Indian farmers switch to riskier varieties of rice Transaction costs can be reduced Perfect product for mobile payments (pending regulatory approval) Consumers will actually want to buy product Part III of this talk: measuring demand (Cf. Brown, Kapteyn, Luttmer, Mitchell, 2012) 21
Understanding Demand for Insurance Rainfall index insurance policies are complicated Contracts written on mm rainfall Farmers think about soil moisture Goal of limiting basis risk makes policies more complex If rainfall in a day < 3 mm, it doesn t count towards index If rainfall in a day >60mm, only 60 mm count towards index Payout is a non-linear function of rainfall index Much of value of policy derives from exit or tail event Question: can farmers effectively evaluate products Farmers may otherwise make mistakes, over- or under-insure Important condition for competitive, welfare-enhancing markets to develop 22
Experimental Design Evaluate willingness to pay for 2,000 farmers (1,500 old farmers and 500 farmers added to sample) for four policies (1) Actual policy designed for their geographical area E.g., Anantapur Phase II, premium 110. Pays Rs. 1,000 on exit. Gauge Strike (mm) Exit (mm) Per mm Exp Payout Anantapur 30 0 10 44 (2) mm deviation. Reduce the amount paid out per mm deficit from 10 to 5 =>Reduces expected value from 44 to 22 (3) Higher Exit. Pay Rs. 1,000 if rainfall between 0 and 5 mm =>Raises expected value from 44 to 110 (4) Basis Risk. Real policy, but written on distant rainfall station 23
Experimental Design Evaluate willingness to pay for 2,000 farmers (1,500 old farmers and 500 farmers added to sample) for four policies (1) Actual policy designed for their geographical area E.g., Anantapur Phase II, premium 110. Pays Rs. 1,000 on exit. Gauge Strike (mm) Exit (mm) Per mm Exp Payout Anantapur 30 0 10 44 (2) mm deviation. Reduce the amount paid out per mm deficit from 10 to 5 =>Reduces expected value from 44 to 22 (3) Higher Exit. Pay Rs. 1,000 if rainfall between 0 and 5 mm =>Raises expected value from 44 to 110 (4) Basis Risk. Real policy, but written on distant rainfall station 24
Experimental Design Evaluate willingness to pay for 2,000 farmers (1,500 old farmers and 500 farmers added to sample) for four policies (1) Actual policy designed for their geographical area E.g., Anantapur Phase II, premium 110. Pays Rs. 1,000 on exit. Gauge Strike (mm) Exit (mm) Per mm Exp Payout Anantapur 30 0 10 44 (2) mm deviation. Reduce the amount paid out per mm deficit from 10 to 5 =>Reduces expected value from 44 to 22 (3) Higher Exit. Pay Rs. 1,000 if rainfall between 0 and 5 mm =>Raises expected value from 44 to 110 (4) Basis Risk. Real policy, but written on distant rainfall station => No effect on expected value (in expectation) 25
Measuring Willingness to Pay Becker-Degroot-Marschak Mechanism Scratch-card with price hidden Respondent writes down willingness-to-pay (WTP) Respondent scratches off price If WTP>price, farmer purchases policy at price If WTP<price, no sale Practice with chocolate bar first 26
Summary Statistics Bid Type Ordering 1 Ordering 2 N Mean Median Real Policy 68.97 67.90 1978 68.43 70.00 Policy (Exit) 79.79 78.81 1978 79.30 80.00 Policy (mm Deviation) 56.66 56.40 1978 56.53 55.00 Basis Risk 38.90 39.07 1978 38.98 35.00 Bid is scaled as percent of policy premium Average and median willingness to pay generally exceed actuarial value: there is scope for insurance 27
Regressing Willingness to Pay on Policy Characteristics Phase I Phase II All All (1) (2) (3) (4) VARIABLES Bid Bid Bid Bid Change in payout per mm of deviation -13.23*** -11.48*** -11.90*** -11.90*** (0.555) (0.356) (0.302) (0.302) Change in exit level 11.33*** 10.71*** 10.86*** 10.86*** (0.196) (0.173) (0.140) (0.140) Change in basis risk -30.69*** -29.06*** -29.45*** -29.45*** (0.712) (0.487) (0.408) (0.408) Phase 2 Policy -5.001*** -5.050*** (0.881) (0.862) New Participants 6.328*** (1.032) Policy Ordering -0.484 (0.792) Income distribution (>25% & <55%) 2.021* (1.152) Income distribution (>50% & <75%) 2.842** (1.263) Income distribution (>75%) 6.392*** (1.249) 28
Taking Stock Farmers get direction right, but magnitudes wrong Change in mm deviation Reduces expected value by Rs. 22 Affects payouts in bad, but not catastrophic, states of world Reduces willingness to pay by Rs. 13 Change in exit level Triples expected value, from 44 to 110 Payout occurs in exactly the worst state of the world Increases willingness to pay by 11 Private market may not work well (Anagol et al., 2012) Sales agents may not recommend appropriate products Government ownership of products doesn t solve problem 29
Policy Deviation Change in mm Deviation Increase in exit threshold Basis Risk Change in E[Value] Low Soph. (Q1) 44 to 22=>-22-9 -14 44 to 110=>+66 +9 +12 E[Value] unchanged, real value much lower -22-34 High Soph (Q4) 30
Dynamics of Demand Can demand sustain over long periods of time? How important are payouts in repurchase decisions? 31
Experimental Setting SEWA markets rainfall insurance to residents of up to 60 villages, from 2006-2013 Each year, each household in the study was randomly assigned a marketing package, which induced exogenous variation in takeup. From 2009 through 2013, we elicited households willingness to pay for insurance using a Becker-deGroot-Marschak (BDM) mechanism Final sample of people assigned to receive marketing was 1160. Occasional attrition. Analyze demand for balanced sample of 989 households. 32
Summary Statistics 2006 2007 2008 2009 2010 2011 2012 2013 Pooled Balanced Treatment Sample No. of households 405 649 649 989 989 989 989 989 6,648 No. of households (Lagged) 405 649 649 989 989 989 989 5,659 No. of villages 32 52 52 60 60 60 60 60 60 Take-up Average market price per policy (Rs.) 214 69 190 151 75 195 200 200 161 Average price paid per policy (Rs.) (if purchased) 104 70 140 58 21 62 63 63 59 Average price paid per policy/market price (if purchased) (%) 50 100 74 37 28 32 32 32 40 No. of purchasers 74 251 131 157 556 448 468 558 2,643 No. of purchasers (Lagged) 74 251 131 157 556 448 468 2,085 No. of non-purchasers 331 398 518 832 433 541 521 431 4,005 No. of non-purchasers (Lagged) 331 398 518 832 433 541 521 3,574 Average policy units purchased (if purchased) 1.03 1.02 1.07 2.33 4.52 2.16 1.96 1.99 2.40 Re-purchase rate (%) - 43 35 41 64 56 60 68 53 New-purchase rate (%) - 87 33 66 82 30 43 43 55 Quit rate (%) - 57 65 59 36 44 40 32 47 Payouts Payout (yes/no) 0 0 38 64 353 64 341-860 Average payout (if purchased) 0 0 165 92 321 23 346-146 Average payout per policy (Rs.) (if purchased) 0 0 165 39 77 13 171-59 Average payout (if payout >= Rs. 1) 0 0 570 225 505 158 475-449 Average payout per policy (Rs.) (if payout >= Rs. 1) 0 0 570 96 121 93 234-182 Average number of people per village who received payouts (if village payout per policy >= Rs. 1 0 0 10 12 29 11 15-17 Crop Loss Experienced crop loss (yes/no) 319 146 202 496 296 223 283-1965 Average agricultural revenue lost due to crop loss (Rs.) (if payout>=rs. 1) 0 0 2892 914 2580 423 1271-1790 Average agricultural revenue lost per village due to crop loss (Rs.) (if payout>=rs. 1) 0 0 3018 3492 2882 472 1317-2326 This table gives the summary statistics for the entire (balanced treatment) sample. In 2010, although the premium for internal per policy use was only INR 150, Nabard was subsidising the policies with a 'buy one get one free' offer. 33
Insurance Purchasers Pooled Individual Fixed Effects (1) (2) (3) (4) (5) (6) Village Payout per Policy in Previous Year (Rs. '000s) 0.842 *** 0.795 *** 0.682 *** 0.493 *** 0.593 *** 0.523 ** (0.123) (0.145) (0.148) (0.145) (0.192) (0.199) Payout received Previous Year (Rs. '000s) 0.023-0.002-0.039-0.059 (0.048) (0.048) (0.047) (0.048) Number of Insurance Policies Bought Previous Year 0.007 0.009 0.014 0.015 * (0.008) (0.007) (0.009) (0.009) Number of Households in Village who received a Payout Previous Year 0.004 ** 0.003 * (0.002) (0.002) Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) 0.004 0.002 (0.012) (0.014) Mean Village Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) 0.088 *** 0.107 ** (0.030) (0.045) Constant 0.168 ** 0.161 ** 0.616 *** 0.426 *** 0.622 *** 0.544 *** (0.068) (0.069) (0.075) (0.153) (0.078) (0.078) r2 0.159 0.16 0.169 0.148 0.15 0.159 N 2085 2085 2085 2085 2085 2085 Sample restricted to insurance purchasers from 2006-2012, with households entering and exiting the sample each year based on their insurance purchase decisions. The dependent variable is a dummy for purchasing insurance in current year. The sample consists of 882 households who purchased insurance at least once. All specifications include year dummies and the complete set of same-year marketing variables as additional controls. All specifications are OLS. The Fixed Effects specifications include household fixed effects. Variation in the fixed effects specifications is provided by the 505 households who purchased insurance more than once and experienced variation in the payouts received. All standard errors are clustered at village level. Column 1: a payout per policy of Rs. 1,000 causes a 84% increase in the probability of purchasing insurance in the next season. Column 3: mean village revenue lost is significant as well 34
Insurance Non-Purchasers Pooled Individual Fixed Effects (1) (2) (3) (4) (5) (6) Village Payout per Policy in Previous Year (Rs. '000s) 0.408 *** 0.358 *** 0.361 *** 0.245 ** 0.201 * 0.203 * (0.077) (0.079) (0.079) (0.108) (0.105) (0.105) Number of Households in Village who received a Payout Previous Year 0.003 * 0.003 * 0.005 *** 0.005 *** (0.002) (0.002) (0.002) (0.002) Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) 0.000 0.01 (0.006) (0.012) Mean Village Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) -0.01-0.018 (0.030) (0.036) Constant -0.109 ** -0.108 ** -0.108 ** 0.578 0.507 0.508 (0.047) (0.046) (0.047) (0.068) (0.067) (0.070) r2 0.175 0.178 0.178 0.177 0.184 0.185 N 3574 3574 3574 3574 3574 3574 Sample restricted to households who did not purchase insurance from 2006-2012, with households entering and exiting the sample each year based on their insurance purchase decisions. The dependent variable is a dummy for purchasing insurance in current year. The sample consists of not 989 but 977 households, as 12 households purchased insurance in each year that it was available. All specifications include year dummies and the complete set of same-year marketing variables as additional controls. All specifications are OLS. The Fixed Effects specifications include household fixed effects. Variation in the fixed effects specifications is provided by the 515 households who did not purchase insurance more than once and experienced variation in the payouts received. All standard errors are clustered at village level. Column 1 shows that an insurance payout of Rs. 1,000 results in a 41% increased probability of purchasing insurance the following year by non-purchasers. 35
Entire Sample Pooled Individual Fixed Effects Pooled Indiv OLS OLS OLS OLS IV IV I (1) (2) (3) (4) (5) (6) ( Village Payout per Policy in Previous Year (Rs. '000s) 0.457 *** 0.389 *** 0.305 *** 0.271 *** 0.434 ** 0.382 ** 0.15 (0.079) (0.081) (0.091) (0.092) (0.201) (0.193) (0.235 Payout received Previous Year (Rs. '000s) 0.102 ** 0.081 * 0.067 * 0.047 0.103 0.022 0.342 (0.042) (0.042) (0.036) (0.035) (0.314) (0.329) (0.367 Number of Insurance Policies Bought Previous Year 0.037 *** 0.037 *** -0.01-0.01-0.001 0.004-0.019 (0.006) (0.006) (0.007) (0.007) (0.026) (0.027) (0.029 Number of Households in Village who received a Payout Previous Year 0.003 ** 0.003 ** 0.004 *** (0.001) (0.001) (0.002) Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) 0.007 0.003 0.007 (0.007) (0.008) (0.007) Mean Village Revenue Lost Due to Crop Loss Previous Year (Rs. '0000s) 0.016 0.01 0.02 (0.030) (0.031) (0.032) Constant -0.014-0.016 0.608 *** 0.559 0.654 *** 0.631 (0.041) (0.041) (0.047) (0.050) (0.046) (0.048) F Stat. (First Stage, Instrumented: Number of Insurance Policies Bought Previous Year) 2098.54 25064.2 82.00 F Stat. (First Stage, Instrumented: Payout Received Previous Year) 154.366 22751.2 60.03 r2 0.177 0.182 0.168 0.172 0.161 0.166 0.150 N 5659 5659 5659 5659 5659 5659 5659 Regressions include entire sample. All specifications include year dummies and the complete set of same-year marketing variables as additional controls. In the IV Specifications, "Payo Previous Year" and "Number of Insurance Policies Bought Previous Year" are instrumented with the full set of marketing variables lagged one year. Errors clustered at village level. IV Results: Auto-correlation between purchasing may be endogenous, not causal Policy payouts still matter (a lot) Village-level payouts still matter 36
Some Concluding Policy Remarks Many facets of the NAIS vs. mnais vs. WBCIS decisions are technically complex Some are not: Demand from poor farmers, at actuarially fair prices Transaction costs are key Should subsidies be available only to bank borrowers? Could an index-like product be made available to the poor in similar spirit to NREGA? 37
Thanks! 38