WPS8375 Policy Research Working Paper 8375 Safety Nets and Natural Disaster Mitigation Evidence from Cyclone Phailin in Odisha Paul Christian Eeshani Kandpal Nethra Palaniswamy Vijayendra Rao Development Research Group Poverty and Inequality Team March 2018 Policy Research Working Paper 8375 Abstract To what degree can vulnerability to extreme weather events expenditure, and that these reductions were indeed the larg- be mitigated by access to a rural livelihoods program, par- est for women, adding to the emerging evidence for the ticularly with regard to the impacts on women? This paper frequently-posed hypothesis that women bear the brunt addresses this question through a natural experiment arising of the effects of disasters on overall household consump- from two independent but overlapping sources of variation: tion. Participation in the livelihoods program mitigated exposure to a devastating cyclone that occurred in the Bay some of the reductions in household nonfood expenditure of Bengal region of India and the staggered rollout of a rural and women’s consumption, but not on food expendi- livelihoods intervention. Comparisons from household sur- ture. These results from a densely populated region whose veys across communities more or less exposed to the storm topography makes it particularly vulnerable to storms can before and after the introduction of the program reveal that inform future policy approaches and aid in modeling the the storm led to significant reductions in overall household impact of these policies on the effects of climate change. This paper is a product of the Poverty and Inequality 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 ekandpal@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 Safety Nets and Natural Disaster Mitigation: Evidence from Cyclone Phailin in Odisha Paul Christian Eeshani Kandpal∗ Nethra Palaniswamy Vijayendra Rao Keywords: Climate change, safety nets, gender, India JEL Codes: Q54, Q58, O13, I38 ∗ All authors are with The World Bank Group. Corresponding author email: ekandpal@worldbank.org. We thank Madhulika Khanna and Aditya Shrinivas for research assistance; and SUTRA consultancy and the Social Observatory for data collection. We gratefully acknowledge suggestions and support from Saurabh Dani, Samik Sundar Das and DV Swamy, and financial support from the South Asia Disaster Management and Climate Change team at the World Bank, and from the contributions of (1) UK Aid from the UK government; (2) the Australian Department’s of Foreign Affairs and Trade (DFAT); and (3) the European Commission (EC) through the South Asia Food and Nutrition Security Initiative (SAFANSI), which is administered by the World Bank. 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. The views expressed also do not necessarily reflect the UK, EC or Australian government’s official policies or the policies of the World Bank and its Board of Executive Directors. 1 Introduction This paper examines the effectiveness of a livelihoods program to mitigate the impact of climate-induced extreme weather events. We particularly focus on heteregeneity of reponses on the dimension of gender. Climate change is expected to increase the incidence and intensity of extreme weather events like cyclonic storms (Mendelsohn et al., 2012). While the capacity to avoid and offset the worst effects of such storms, particularly human mortality, is improving, large swaths of the developing world remain vulnerable to sizeable income losses (Intergovernmental Panel on Climate Change, 2014). In addition, much of the policy literature describes women as being particularly vulnerable to the worst insults of climate change (UN Population Fund, 2009; World Health Organization, 2011), but evidence for this hypothesis is nascent (Anttila-Hughes and Hsiang, 2013; Neumayer and umper, 2007; Hallegatte et al., 2016), especially with regard to the question of whether Pl¨ safety net programs can lessen the gender gaps resulting from these shocks. The densely populated Bay of Bengal in the North Indian Ocean is uniquely vulnerable to devastation by storms because geographical features translate strong winds into damaging, sometimes catastrophic, storm surge. The region is also home to many of the world’s poor (Alam, Hossain and Shafee, 2003), and has long been recognized has having low levels of women’s welfare (Hashemi, Schuler and Riley, 1996; Kishor and Gupta, 2004). Modeling the impact of climate change requires understanding the incidence of these storms on household welfare as well as the potential for various policy responses to offset these consequences. Microcredit interventions are a popular poverty alleviation policy tool, particularly in poor countries (Banerjee, Karlan and Zinman, 2015). The most common model of microcredit in India is the Self-Help Group (SHG), which is implemented by the National Rural Livelihoods Mission, a program reaching 600,000 villages across all of the nation’s 29 states that has leveraged more than $6 billion dollars over five years (Joshi, Palaniswamy and Rao, 2016). SHGs have a special focus on women’s empowerment, strengthening community ties, and increasing collective action (Sanyal, Rao and Majumdar, 2013). They have been found to provide cheap credit and foster women’s empowerment, but their impact impact on poverty has been shown to be minimal despite some speculation that 2 they help mitigate risk (Brody et al., 2015). Notwithstanding the lack of evidence on poverty reduction, policymakers believe that SHGs can be used to effectively attenuate the effects of severe weather events (Government of India, 2011). Cyclone Phailin made landfall in the east Indian state of Odisha on October 11, 2013 (Singh and Jeffries, 2013). It was the strongest tropical storm to hit India in 14 years, indeed stronger than Hurricane Katrina on landfall, with wind velocities of 205-220 kilometers per hour. Phailin led to one of the largest emergency evacuations on record. Over one million people were moved to shelters. 256,000 households experienced partial or severe loss, and 1.3 million hectares of agricultural land (approximately 30% of estimated agricultural land in affected areas) (Singh and Jeffries, 2013). Reconstruction and rehabilitation costs were estimated to have cost $1.45 billion (Singh and Jeffries, 2013). By coincidence, Phailin-hit areas overlapped with the study districts of a World Bank (WB)-led impact evaluation of a government-run SHG intervention, called Targeted Rural Initiatives for Poverty Termination and Infrastructure (TRIPTI). Using spatial variation in the intensity of the rainfall shock and the staggered rollout of TRIPTI, we conduct a large-scale statistical examination of (1) the negative consumption impacts of Cyclone Phailin, and (2) whether the prior presence of TRIPTI SHGs mitigated covariate risk by improving access to credit and providing a platform for government response. We construct a Geographic Information Systems (GIS)-coded dataset from two sources: the Indian Meteorological Department’s (IMD) 1◦ x1◦ gridded daily data on rainfall imputed from 6,327 weather stations across India from 1951 to 2013, and the household-level baseline and endline surveys for the impact evaluation of TRIPTI. We exploit spatial variation in the intensity of the rainfall shock experienced by a household in October 2013 as measured by the absolute deviation from the historical average of millimeters of rainfall at the nearest weather station. To these rainfall data, we add data on household expenditures, consumption, credit-seeking, and political engagement from the TRIPTI surveys. The baseline survey for the TRIPTI evaluation was completed in July 2011, two years before Phailin struck, while the endline was completed in August 2014. TRIPTI was assigned to the four least developed blocks (sub-districts) in each of the ten coastal districts in Odisha, which were identified using a development index produced by the Government of India in 3 1993. The blocks with the next four lowest development scores within each district were chosen as comparison (non-TRIPTI) blocks. We use regression analysis to measure the differential effects of Phailin across TRIPTI and non-TRIPTI areas. Phailin led to sizeable decreases in total household consumption, as well as a substitution away from purchased food toward the consumption of home-grown stocks. Expenditure on women’s goods, a category that includes clothing, shoes, hygiene products, and toiletries purchased for women, sees one of the largest reductions across categories of household expenditure; this finding is among the first concrete pieces of empirical evidence on the disproportionate vulnerability of women to climate change. In contrast, Phailin-hit households spend more on festivals after the storm, and the increase is greatest in magnitude in TRIPTI areas, suggesting that strengthened community ties might be used to leverage social capital in disaster-affected areas. Across the sample, households expand the number of loans taken after Phailin, but this expansion is significantly higher in TRIPTI areas. In non-TRIPTI areas, state-level aid crowds out interactions with village governments, while in TRIPTI areas Phailin-affected populations remain engaged in village politics. In particular, across the sample, households in Phailin-hit areas were more likely to know the name of the Chief Minister, who is the highest ranking political figure in the state. In non-TRIPTI areas, this awareness of the Chief Minister came at the expense of knowledge of the village governance structure, while in TRIPTI areas we see no reduction in engagement with local governance due to Phailin. The econometric analysis accounts for time invariant characteristics of the households and their location. In addition, the analysis of the effects of exposure to Phailin accounts for spatial autocorrelation common to observations sampled within the same latitude-longitude grid cell. The estimated impact of TRIPTI controls for correlation at the village level, while the estimates of TRIPTI on Phailin exposure account for spatial autocorrelation in the interactions between the latitude-longitude grid cells and villages. The approach used here thus addresses initial differences in the levels of expenditure, credit-seeking behavior or access that coincide with either exposure to Phailin or the presence of TRIPTI; however, it cannot rule out that there were differential trends in these variables arising from sources other than TRIPTI or Phailin. For example, if another weather event between baseline and 4 follow-up surveys affected only Phailin-hit TRIPTI areas but not areas affected by Phailin that were not covered by TRIPTI (or vice-versa), this would lead to spurious findings. We do not have evidence of such differential trends, but cannot fully rule them out. The analysis also only models average impacts of both Phailin and TRIPTI at a relatively coarse spatial scale. This study thus provides suggestive evidence on the impacts of both extreme weather and the safety nets designed to mitigate them. 2 Cyclone Phailin and Rainfall Data We measure the intensity of Phailin using data from the Indian Meteorological Department on daily rainfall for all 1◦ x1◦ latitude-longitude grids in Odisha from 1951 to 2013. The Indian Meteorological Department (IMD) operates 537 weather stations that measure rainfall over the past 24 hours (ending 8:30 am Indian Standard Time) (Rajeevan et al., 2006). In addition, state governments maintain rain gauges for real-time rainfall monitoring. The IMD collates, cleans and archives these data following protocols established in Rajeevan et al. (2006). The authors note that while 6,327 stations report rainfall data, only 1,803 of them did so at least 90 percent of the time between 1951 and 2003; therefore, the authors interpolated data for the remaining station-year combinations. This interpolation was subsequently updated by the IMD to include later years. Using these data, we first construct historical medians of millimeters of rainfall for each month of the year, which in turn allows us to construct the absolute deviation of rainfall from the historical median for each month from 2009 through 2013. This variable is a measure of departure from the expected volume of rainfall for each month, allowing us to understand the extent to which October 2013, the month of Phailin’s impact, represented an adverse weather shock in our sample.1 Figure 1 shows the deviation of monthly rainfall from the monthly median taken over a five-year time series, averaged over all grid squares covered by the weather station data. This plot confirms that October 2013 was a major rainfall event, and that the only other obvious anomaly picked out by this measure, July 2009, 1 The interpolated data is provided to us at a monthy scale. Summing rainfall over the month ensures that we capture the full effect of the storm. Taking deviations from historical levels ensures that any seasonally typical rainfall not associated with Phailin is not captured by our measure of shock. 5 corresponds to another known flooding event (Government of Orissa, 2010).2 To assess the impact of flooding and storm damage, we will use the spatial variation in the intensity of October 2013 rainfall.3 3 Self-Help Groups and the TRIPTI Program The Odisha Rural Livelihoods Program, TRIPTI, was launched in 2009. The program was funded by the World Bank and implemented by the Government of Odisha. TRIPTI was formed to address high rates of rural poverty in Odisha, with a particular focus on increasing diversification of livelihoods and the reduction of debt. TRIPTI operated in three steps: first, it created village-level SHGs by providing training on the management of group-based lending. Second, these federations were linked with services provided by the public and private sectors4 . And third, using a participatory identification process and a village census, TRIPTI classified households into one of four categories: (1) extremely poor and vulnerable, (2) poor, (3) manageable, and (4) well-off, and provided grants to the first two categories of households. The formalized rollout of TRIPTI was designed in conjunction with an impact evaluation in 2011. While this was nominally two years after the initiation of TRIPTI, the project had only started activities in 40 villages (out of an intended 1,020) across all project blocks. As part of the 2011 rollout, TRIPTI aimed to create 30,000 SHGs in 1,010 Gram Panchayats (GPs)5 By 2015, TRIPTI had been implemented in 38 sub-districts in 10 coastal districts of Orissa. The treatment rule stipulated that the four least developed blocks in each of the 2 The correlation between the July, 2009 shock and the October, 2013 shock across grid squares is low (-.08). 3 To the extent that rainfall deviations are correlated with other factors creating damage such as wind or storm surge, this measure captures the effects of these factors as well. If rainfall is uncorrleated with wind or storm surge, these estimates understate the full effect of the storm on well-being. If effects of the storm spill over from areas that received high rainfall to neighboring areas, the impacts will similarly be understated. 4 SHGs were linked with the Odisha State Seed & Organic Product Certification Agency for certification and the Odisha State Seed Corporation and the Odisha Agro Industries Corporation for seed production and subsidy, while the Animal Husbandry Department subsidized poultry purchases and related inputs. District Industries Centers helped farmers’ groups establish linkages with agricultural markets. The state- run Mo-Badi program helped SHGs start kitchen gardens in member households, while the Directorate of Horticulture provided compost facilities to treated households. Many SHGs were also engaged in managing the mid-day meal program. 5 Gram Panchayats are local government units encompassing an average of eight villages each. 6 10 targeted districts would receive the intervention and that an objective assignment score would be used to rank and choose these blocks. These assignment scores were calculated as a weighted average of the following variables: (i) the number of households belonging to the two lowest caste categories, Scheduled Castes and Tribes (SCs and STs), (ii) the total number of households, (iii) the number of SHGs that were deemed credit worthy6 and (iv) a Composite Development Index developed and implemented nationally by the Indian government in 1993. In each district, all blocks were ranked by this score, and the four blocks with the highest score received TRIPTI. The SHGs were rolled out in a staggered manner across GPs, with all villages in the same GP entering the program simultaneously. In each district, the non-program blocks with the four next highest assignment scores were chosen as the control group for the evaluation. The TRIPTI impact evaluation exploits this assignment rule to evaluate TRIPTI by creating groups of communities that are treated by TRIPTI during the study period and groups that were not. To the extent that four blocks with the highest assessment scores not treated have scores that are very similar to those treated, the design approximates a regression discontinuity design. Since we cannot verify the methods of the assessment score, we prefer to interpret this design as a differences-in-differences approach, which accounts for fixed differences between treated and untreated blocks (Joshi, Palaniswamy and Rao, 2016). 3.1 Household Data from Evaluation of TRIPTI To assess the impact of both the cyclone and TRIPTI on households in Odisha, we use household data from two surveys conducted by the World Bank in support of the impact evaluation of the TRIPTI intervention. In each of the 10 block-pairs selected for the impact evaluation, one treatment block and one control block were randomly sampled for the evaluation, yielding a total sample of 80 GPs. Since all households in treated areas are eligible for treatment, we present Intent-to-Treat estimates for the impact of the program. The baseline survey, conducted between July and October of 2011 included data on 2,875 6 The rating of SHGs is an exercise that helps establish the SHG’s credit-worthiness by comparing re- payment rates and attendance to area averages. SHGs that “pass” the rating were deemed credit-worthy for TRIPTI. There are typically three grades, with the highest grades allowing SHGs to access credit directly from banks etc. The rating process is conducted by the state government federation of SHGs. 7 households from 160 randomly-selected villages, two in each of the 80 selected GPs. The endline, conducted between August and November 2014, revisited 2,874 of the same households from the baseline survey. Both survey rounds collected information on household consumption expenditure following the Indian National Sample Survey consumption module, livelihoods and debt profiles, as well as proxy measures of women’s civic engagement and inputs to household decision-making. In the face of a negative income shock, as likely caused by events like Phailin, households use multiple strategies to cope with shocks to well-being, including relying on government assistance, taking out additional loans, reallocating resources within the household, and drawing down previously-accumulated stocks of food and savings (Gallagher, 2014; Gallagher and Hartley, n.d.; Deryugina, Kawano and Levitt, 2014). Therefore, we focus on borrowing from SHGs and other lenders, home production and consumption of food and non-food items as our key outcomes of interest. These data have several useful features that allow us to understand the effect of both TRIPTI and Cyclone Phailin. The first is the timing. The baseline survey was conducted in July 2011, two years before Phailin made landfall in October 2013. The endline was collected one year later in August 2014. The second key feature is the spatial overlap between villages that received TRIPTI and Phailin’s impact zone. As shown in Figure 2, the surveyed communities overlapped significantly with Phailin-hit areas. This allows us to use spatial variation in the rainfall shock as well as the staggered roll-out of TRIPTI to estimate the impact of Phailin. The third advantage is the topical relevance of the survey data, which yields information on precisely the above-mentioned coping strategies used by households to cope with a natural disaster. TRIPTI was not randomly assigned, so one concern in identifying its effect is that households offered the chance to participate in SHGs formed by TRIPTI may differ from those who do not have this opportunity. The project’s impact evaluation report studies balance at baseline for key outcomes of interest, a few of which are significantly different (Joshi, Palaniswamy and Rao, 2016). They find that at baseline, 67.8% of households in project areas reported being SHG members compared to 74.3% in non-project areas. In addition, while 35% of households in project areas relied on SHGs for 8 savings, only 31.4% did so in treatment areas. In appendix table A1, we replicate the balance checks comparing TRIPTI and non-TRIPTI communities. Out of the 24 outcomes studied, we find baseline imbalance in three. Households in communities where TRIPTI set up SHGs consumed a higher value of food from home production, were more likely to have a current loan, and were less likely to be famililar with the village government structure before the TRIPTI intervention. This level of imbalance is consistent with what would be expected from randomized treatment assignment. The baseline values of these variables also do not vary systematically in a pattern that would suggest strategic selection of treatment villages. Previously reported results (Joshi, Palaniswamy and Rao, 2016) show that households in TRIPTI areas were 22 percent more likely to participate in self-help groups, and 7.7 percent more likely to borrow from formal or institutional sources of credit. However, the evidence on household welfare is more mixed: there was no measurable impact on household consumption, but expenditures on healthcare and women’s and children’s goods increased significantly. While there were few impacts on livelihood strategies in program areas, households in treated areas reported providing two more days labor to the National Rural Employment Guarantee Scheme (NREGS) than households in control areas relative to a baseline level of slightly under two days. This increase in participation in NREGS was driven by women in treated areas. In addition, women’s autonomy over their mobility improved in TRIPTI areas, although this effect was driven by women in treated areas being more likely to go to SHGs meetings (17.8 percent) and to the bank (5.3 percent) alone. There were, however, no effects on women’s self-reported influence in household decision-making as measured by an index of women’s say in (1) the household purchase of durables, (2) goods for her personal consumption, (3) tuition expenses, (4) the household’s livelihood strategy, and (5) household decisions regarding political participation. On the other hand, public action seems to have been higher in TRIPTI areas, with women in treated blocks being 5 percent more likely to use Gram Panchayat meetings to raise problems with domestic violence and alcoholism as well as to address any issues with the Public Distribution and Mid-Day Meal Schemes. Finally, an overall index of willingness to act on these problems was higher by 8.1 percentage points while an index of willingness to 9 pursue institutional response to these community problems was 12.8 percentage points higher in treated areas (Joshi, Palaniswamy and Rao, 2016). 4 Methodology The analytical objectives of this paper are to (1) establish the effects of Phailin on household welfare, and (2) examine the extent to which TRIPTI mitigates the effect of Phailin on household well-being. To this end, we aim to measure the effect of exposure to Phailin, but our primary parameter of interest is the interacted effect of Phailin and TRIPTI. We begin by showing that Phailin represented a significant negative welfare shock to the households in the TRIPTI impact evaluation data. 4.1 Phailin Exposure in TRIPTI Communities In the endline survey for the TRIPTI evaluation, one year following Phailin’s landfall, the household survey module included questions about the household’s experience with Phailin. This allows us to document exposure to the storm in our sample and to verify that our measure of exposure to the cyclone is relevant for households in our data. The upper left panel of figure 3 shows the distribution of the rainfall shock in our sample and the proportion of households who reported that their residence was flooded during the week of Phailin. This figure confirms two key facts. First, there is a wide degree of variation in exposure to more or less intense rainfall shocks as measured in the IMD data. The households living in the areas with the biggest deviations from their grid square’s median rainfall in October experience rainfall shocks approximately three times greater than the households living in grid squares with the smallest deviations from typical rainfall in October. Second, the darker line indicates that the rainfall shock is tightly related to households’ own reports of damaging flooding. In the grid squares experiencing the smallest rainfall shocks, less than 10 percent of households reported that water had entered their house. In the grid squares that experienced the biggest rainfall shocks, this proportion was approximately 25 percent. Flooding of households is only indicative of damage, and households in our sample may 10 have been affected by consequences of the cyclone other than flooding. To assess whether the rainfall shock identifies these alternative experiences, Figure 3 shows the proportion of households who experienced the following measures according to the rainfall shock in their grid square: received an evacuation notice, left their house for shelter, received any government aid, and received aid for home repairs. These plots confirm that the rainfall shock is a strong predictor of a wide array of self-reported impacts of the storm, and gives us confidence that rainfall shock is a useful summary of the intensity with which households were affected by the storm. Figure 3 confirms that the households in the TRIPTI evaluation data were affected by Phailin and that there is substantial variation in this sample in the intensity with which these households were affected. But Phailin exposure is not randomly assigned. For example, cyclones tend to be most powerful on landfall and exhaust their strength as they move inland. So we would expect that households in coastal grid squares to be more exposed than those living in grid squares that are far inland. To understand how rainfall shocks may be correlated with characteristics of households, we perform balance tests on Phailin exposure where we test characteristics of households at baseline, prior to Phailin’s landfall, between households where the rainfall shock was above or below the median shock. Results are reported in Appendix Table A2. In general, household characteristics were relatively balanced according to Phailin intensity. Important areas where there are differences include that more exposed households consumed more food from home production, were more likely to have taken a loan, and where less likely to have heard of Gram Sabhas, which are open village meetings that are part of the Indian system of village governance. 4.2 Identifying the Effects of Phailin and TRIPTI Having checked visually that the households in our dataset were indeed affected by the storm, we next proceed to econometrically estimating the impacts of Phailin on household welfare and any mitigating effects of TRIPTI. First, to quantify the effect of Phailin on households’ well-being, we estimate the reduced form effect of Phailin on household 11 well-being using the following difference-in-differences specification: Ytigj = α + βlog (Rain)g + γP ostt + δP ostt ∗ log (Rain)g + ψj + tigj (1) where Ytij is a measure of household well-being in year t for household i in district j (like assets, income, or household expenditure). The variable “log (Rain)g ” is the natural log of deviations from monthly historical median rainfall for grid square g in October 2013. The variable P ostt indicates that an observation was taken from the endline (2014) survey rather than the baseline (2011) survey. ψj is a vector of indicator variables for the administrative block in which household i resides. This is like a difference-in-difference regression of the effect of cyclone intensity on household well-being except that the “cyclone treatment” measure is continuous rather than discrete. The coefficient α is the baseline average value of the outcome Ytij for households in locations where rainfall in October 2013 was exactly the historical average for the month of October for that location. The coefficient β is the difference in the baseline value of Ytij correlated with each additional millimeter of rainfall received in October 2013; this coefficient thus captures starting differences across locations where the Phailin shock was especially intense. If, at baseline, places that were later especially vulnerable to Phailin were different than places that were less vulnerable, the coefficient β would capture these differences. β is thus a measure of the average starting differences between households that were more exposed to the cyclone and those that were less so. The coefficient γ is the change in Ytij in the locations that were not affected by Phailin. The coefficient δ is the primary effect of interest in this specification, and captures the differences in Ytij caused by exposure to greater rainfall from Phailin. A negative sign on δ means that locations exposed to higher rainfall in October 2013 had greater declines (or slower growth) in Ytij than places that had lower rainfall. Because we allow Ytij to vary at baseline according to vulnerability to Phailin (as captured in the coefficient β ), this specification also prevents any correlation between the fixed characteristics of households or villages that may be correlated with both Phailin exposure and the outcomes Ytij from contaminating δ . Finally, as in the usual differences-in-differences approach, the causal effect of Phailin is identified by δ if 12 differences in outcomes between more and less affected areas are constant over time for reasons other than the cyclone. In this specification, we cluster standard errors at the rainfall grid level and include block fixed effects. Clustering at the rainfall grid level accounts for the fact that two observations from the same grid are not statistically independent draws from the population of interest. To explore the question of whether TRIPTI mitigates consequences of Phailin exposure for households, we estimate the effect of TRIPTI on Phailin-affected households using the following triple difference specification: Ytighj = α + βlog (Rain)g + γP ostt + δP ostt ∗ log (Rain)g +θT RIP T Ih + λT RIP T Ih ∗ P ostt + πT RIP T Ih ∗ log (Rain)g (2) +ρT RIP T Ih ∗ P ostt ∗ log (Rain)g + ψj + tighj As in the double difference specifications above, the coefficients α, β , γ , δ , θ, λ and π capture differences in starting values or time trends that are not associated with the combined impact of TRIPTI and Phailin. The vector ψj captures fixed differences between administrative blocks. Additionally, in this specification, the effect of living in a TRIPTI-treated village h in a Phailin-hit area is estimated by a triple difference. Specifically, the coefficient ρ is the effect of TRIPTI on the year-to-year change in household welfare, between baseline and endline, across areas with different exposure to absolute deviations in rainfall due to Phailin. For measures where higher (lower) values of Y indicate greater welfare, when ρ is positive (negative), we conclude that the project had a buffering effect on households that were worst-hit by Phailin. In this specification, we cluster standard errors at the grid-village level and include block fixed effects. Grid-village clustering accounts for non-independence of observations from the same village or the same grid square, while allowing for a different correlation structure for different grids in the same village or different villages in the same grid. 13 5 Results 5.1 Direct Impacts of Phailin The double difference of Phailin exposure as shown in equation (1) yields evidence of a significant reduction in total household expenditure per capita. Results shown in the upper panel of Table 1 suggest that a doubling in the size of the rainfall shock reduced per capita consumption by a third. This reduction was primarily driven by lower per capita food expenditure (Anttila-Hughes and Hsiang, 2013). Households appeared to switch away from purchased food to home produced food, suggesting that a coping strategy might have included drawing down their stocks of food. While health and education expenditures did not change after Phailin, per capita festival expenditures increased by a magnitude of 8 percent, significant at the 99 percent level of confidence. The increase in festival expenditures is similar in magnitude and significance to a decrease in women’s expenditures, suggesting that women are buffering the households after Phailin. The timing of Phailin relative to our fieldwork allows us to estimate expenditures on festivals both before and after the cyclone. To confirm that this is indeed being driven by Phailin and not an artifact of the data, we can exploit the timing of festivals relative to the survey to conduct a placebo test of sorts by comparing expenditures on festivals prior to Phailin to expenditures on festivals following Phailin. Dussehra was coincident with the start of Phailin, while Diwali was right after. Other festivals followed, including Christmas, then New Year’s, then Raja eight months after Phailin’s landfall. In results presented in Appendix Table A3, we show that the most-affected areas spent more on Raja and other festivals, suggesting that the increase in per capita festival expenses stems from expenditures on festivals that occurred after Phailin. At least three reasons may explain this increase: that such expenses are a demonstration of gratitude for having survived a traumatic event; households have more disposable income due to the influx of post-Phailin aid; and they are an investment in social capital that may yield future private returns such as lower prices on food or lower interest on loans, as found in previous literature on festival expenditures (Rao, 2001). The first panel of Table 3 examines the impacts of Phailin on women in further detail. The 14 National Rural Employment Guarantee Scheme is the world’s largest public works program, providing 100 days of paid work to the rural poor. We find that this program appears to have been less accessed in the worst-hit areas. Facing a twice-as-intense rainfall shock led to three fewer days of NREGS-provided work for women. These women were also half as likely to have been aware of the last village council meeting, but more likely to know the Chief Minister’s name. Although NREGS is supposed to prevent households from falling into poverty, this finding is indicative that it may fail serve this protective function in the case of consumption shocks arising from extreme weather events. 5.2 Additional Contribution of TRIPTI on Buffering Effects of Phailin: Total Effect of Phailin in Triple Differences To examine the impact of TRIPTI on Phailin-affected households, we estimate the triple-difference specification presented in equation (3). Table 1 shows that TRIPTI offset the decline in total non-food expenditures after Phailin but had no effect on food expenditure. Non-food items may be easier to finance through credit, thus these markets may have been less affected by Phailin than food markets (Carleton and Hsiang, 2016). Tellingly, the decrease in expenditure on women’s goods was primarily observed in non-TRIPTI areas, which may reflect the female empowerment effects of SHG participation (Brody et al., 2015; Sanyal, Rao and Majumdar, 2013; Datta, 2015). While the triple difference estimate is not significant, the estimated effect is approximately half of the double difference estimate for Phailin, suggesting that TRIPTI may have buffered the impacts of Phailin for some women. The effect of Phailin on children’s goods is proportionally similar to the effect on women’s goods, but we cannot reject the hypothesis of a zero effect on children’s expenses. This pair of results is consistent with the finding by Anttila-Hughes and Hsiang (2013) that households reduce investments in human capital and luxury foods, but not non-food expenditure. Results presented in the lower panel of Table 2 show that the government’s response to Phailin suggests that the government used TRIPTI’s infrastructure to distribute aid. Estimates show that TRIPTI expanded access to loans; in addition, Phailin and 15 non-Phailin affected areas had similar access to credit. In particular, in Phailin affected areas, TRIPTI households took out a greater number of loans than in non-Phailin areas. The double difference also shows the credit expansion effect of TRIPTI. First time borrowing from SHGs increased in the wake of Phailin, but is not driven by new lending by TRIPTI suggesting that credit may have expanded to meet greater demand for loans after Phailin. In addition to these extensive margin effects on borrowing from SHGs in non-TRIPTI areas, we also find evidence of intensive margin effects (larger sums borrowed) in TRIPTI areas. The estimates also show that although Phailin reduced the amount of borrowing in non-TRIPTI areas, that TRIPTI was able to mitigate this restriction in credit after Phailin. Finally, Table 3 shows that TRIPTI increased awareness of the last village council meeting, offsetting the negative effect of Phailin. This triple interaction is consistent with TRIPTI making it easier to facilitate community meetings in order to provide information about aid programs or indeed to leverage aid. 6 Interpretation and Discussion Cyclone Phailin significantly reduced consumption expenditures for affected households in Odisha, India. The SHG intervention, TRIPTI, was able to mitigate some of these impacts. Specifically, the ability of Phailin-hit households in TRIPTI areas to increase their consumption of non-food expenditures relative to Phailin-hit non-TRIPTI households implies that SHGs can help rebuild capital assets. However, the effect on food consumption by these households is just as large in TRIPTI areas as in non-TRIPTI areas, suggesting that SHGs are not a substitute for emergency food aid programs. The worst-hit households spend less on women’s goods but more on social expenditures, providing concrete evidence that women buffer their households from negative consumption shocks and are likely to be among the most vulnerable to the effects of climate. Results also show that governments can use SHGs to channel financial aid after extreme weather events as post-Phailin borrowing increased to a greater degree in TRIPTI areas, and households stayed engaged with the local governance structure used by TRIPTI whereas in non-TRIPTI areas knowledge of local governance was replaced by increased salience of state-level politics. 16 One interpretation of our results is that credit options were limited after Phailin, which in turn induced people to turn to SHGs as a source of loans. In non-TRIPTI areas, households that were more affected by Phailin were more likely to take out their first loans from SHGs, while in TRIPTI areas, the worst-hit households took out additional SHG loans. Both results suggest that expansion of credit may have been a response to the damages inflicted by Phailin. Because extreme weather events are not randomly assigned, no single study can ever rule out that confounding trends led to spurious correlation. However, this context provides a useful example of one policy tool that may be used to improve resilience to extreme weather events. Future research should investigate the ability of social safety nets, including microcredit interventions, at mitigating the impacts of extreme weather events. Such research may include a comparison of the cost-effectiveness of SHGs to other disaster-response programs. Author Contributions: PC and EK designed and conducted the research. PC, EK, and VR wrote the paper. NP designed the trial of the livelihoods intervention. 17 References Alam, Md. Mahbub, Md. Arif Hossain, and Sultana Shafee. 2003. “Frequency of Bay of Bengal cyclonic storms and depressions crossing different coastal zones.” International Journal of Climatology, 23(9): 1119–1125. Anttila-Hughes, Jesse K., and Solomon M. Hsiang. 2013. “Destruction, Disinvestment, and Death: Economic and Human Losses Following an Environmental Disaster.” Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman. 2015. “Six randomized evaluations of microcredit: Introduction and further steps.” American Economic Journal: Applied Economics, 7(1): 1–21. Brody, Carinne, Thomas de Hoop, Martina Vojtkova, Ruby Warnock, Megan Dunbar, Padmni Murthy, and Shari L. Dworkin. 2015. “Economic Self-Help Group Programs for Improving Women’s Empowerment: A Systematic Review.” Campbell Systemic Reviews, 19. Carleton, Tamma A., and Solomon M. Hsiang. 2016. “Social and Economic Impacts of Climate.” Science, 353(6304). Datta, Upamanyu. 2015. “Socio-economic impacts of JEEViKA: A large-scale self-help group project in Bihar, India.” World Development, 68: 1–18. Deryugina, Tatyana, Laura Kawano, and Steven Levitt. 2014. “The economic impact of hurricane katrina on its victims: evidence from individual tax returns.” National Bureau of Economic Research. Gallagher, Justin. 2014. “Learning about an infrequent event: evidence from flood insurance take-up in the United States.” American Economic Journal: Applied Economics, 6(3): 206–233. 18 Gallagher, Justin, and Daniel Hartley. n.d.. “Household Finance after a Natural Disaster: The Case of Hurricane Katrina.” American Economic Journal: Economic Policy. Government of India. 2011. “National Rural Livelihoods Mission Framework for Implementation.” Government of Orissa. 2010. “Special Relief Commissioner’s Annual Report on Natural Calamities 2009-10.” Hallegatte, Stephane, Adrien Vogt-Schilb, Mook Bangalore, and Julie Rozenberg. 2016. Unbreakable: building the resilience of the poor in the face of natural disasters. World Bank Publications. Hashemi, Syed M, Sidney Ruth Schuler, and Ann P Riley. 1996. “Rural credit programs and women’s empowerment in Bangladesh.” World Development, 24(4): 635–653. Intergovernmental Panel on Climate Change. 2014. Climate Change 2014–Impacts, Adaptation and Vulnerability: Regional Aspects. Cambridge University Press. Joshi, Shareen, Nethra Palaniswamy, and Vijayendra Rao. 2016. “Impact Evaluation Framework and Results: Odisha Rural Livelihoods Project.” World Bank. Kishor, Sunita, and Kamla Gupta. 2004. “Women’s empowerment in India and its states: evidence from the NFHS.” Economic and Political Weekly, 694–712. Mendelsohn, Robert, Kerry Emanuel, Shun Chonabayashi, and Laura Bakkensen. 2012. “The impact of climate change on global tropical cyclone damage.” Nature climate change, 2(3): 205–209. umper. 2007. “The gendered nature of natural Neumayer, Eric, and Thomas Pl¨ disasters: The impact of catastrophic events on the gender gap in life expectancy, 1981–2002.” Annals of the Association of American Geographers, 97(3): 551–566. 19 Rajeevan, M., J. Bhate, J.D. Kale, and B. Lal. 2006. “A high resolution daily gridded rainfall for the Indian region : analysis of break and active monsoon spells.” Current Science, 91(3). Rao, Vijayendra. 2001. “Celebrations as Social Investments: Festival Expenditures, Unit Price Variation and Social Status in Rural India.” The Journal of Development Studies, 38: 71–97. Sanyal, Paromita, Vijayendra Rao, and Shruti Majumdar. 2013. “Recasting Culture to Undo Gender: A Sociological Analysis of Jeevika in Rural Bihar, India.” World Bank Policy Research Working Paper. Singh, Deepak, and Andrew Jeffries. 2013. “India Cyclone Phailin in Odisha: Rapid Damage and Needs Assessment.” Asian Development Bank, Government of Odisha, and World Bank. UN Population Fund. 2009. “State of the World Population: Women, Population and Climate Change.” World Health Organization. 2011. “Gender, Climate Change and Health.” 20 7 Figures and Tables Figure 1: Rainfall shock caused by Cyclone Phailin Source: Authors’ Calculations from IMD Data 21 Figure 2: Overlap in Exposure to Cyclone Phailin and Treatment by TRIPTI Source: Authors’ Calculations from IMD Data and TRIPTI IE Survey Data 22 Figure 3: Distribution of households experiencing evacuation in our sample Source: Authors’ Calculations from IMD and TRIPTI IE Survey Data 23 Table 1: The Effects of TRIPTI and Phailin on Per Capita Expenditures Total Home Purchased Women’s Children’s Expenditure Food Production Food Non-Food Festivals Health Education Goods Goods LN(Rainfall Deviation) -1398.790 -1066.135 -131.424∗∗∗ -939.625 -556.111∗∗ -725.319∗∗∗ -417.692 -4164.835∗∗ -723.671∗∗∗ -255.020∗∗∗ (1082.373) (1072.164) (32.396) (1073.945) (246.743) (227.890) (312.340) (1654.985) (259.608) (96.842) Post 19254.891∗∗∗ 19874.685∗∗∗ -720.596∗∗∗ 20590.443∗∗∗ -675.154 -1686.978∗∗∗ -521.013 -15741.115 2541.409∗∗∗ 820.558 (3625.866) (3578.995) (121.967) (3576.005) (933.767) (416.025) (1066.659) (16452.641) (918.656) (565.589) Post X LN(Rainfall Deviation) -3439.441∗∗∗ -3575.257∗∗∗ 122.680∗∗∗ -3699.893∗∗∗ 162.702 367.069∗∗∗ 74.563 3135.084 -422.415∗∗∗ -104.219 (645.155) (636.553) (22.186) (636.023) (168.928) (74.272) (188.180) (3032.097) (162.847) (101.538) Constant 10168.961∗ 7860.538 812.590∗∗∗ 7075.055 3820.935∗∗∗ 4341.837∗∗∗ 3215.409∗ 24432.075∗∗∗ 4791.983∗∗∗ 1613.930∗∗∗ (6038.682) (5981.524) (181.233) (5991.504) (1379.807) (1280.720) (1751.285) (9211.699) (1454.364) (541.529) Total Effect of Phailin 15815.451 16299.427 -597.916 16890.549 -512.452 -1319.910 -446.450 -12606.031 2118.993 716.339 p-value 0.00 0.00 0.00 0.00 0.50 0.00 0.61 0.35 0.01 0.12 TRIPTI village 2179.408 2838.539 -144.441 2997.286 -1021.667 -1230.416∗∗∗ -2638.238 -14857.354 -879.300 -1078.935∗∗ (5728.593) (5609.671) (125.949) (5589.350) (1170.832) (395.086) (2294.249) (9742.777) (881.190) (457.417) LN(Rainfall Deviation) -1286.942 -1019.102 -115.393∗∗∗ -909.414 -448.310∗ -687.363∗∗∗ -380.352 -2984.548∗∗ -775.608∗∗∗ -327.232∗∗∗ (1119.341) (1106.253) (32.511) (1107.988) (261.477) (227.092) (363.888) (1278.726) (268.385) (102.377) Post 19741.189∗∗∗ 19182.441∗∗∗ -516.012∗∗∗ 19701.043∗∗∗ 1313.702 -1015.484∗ 283.280 1396.508 3268.860∗∗ 731.752 (4240.514) (4144.961) (169.083) (4138.629) (1195.649) (602.011) (1447.033) (5709.481) (1342.122) (556.431) TRIPTI X Post -1681.851 840.148 -413.618∗ 1231.955 -4291.630∗∗ -1160.976 -1177.214 -34224.221 -1369.706 459.801 (7501.915) (7393.084) (231.397) (7380.776) (1836.029) (791.289) (2155.450) (27408.490) (2004.950) (1109.501) 24 TRIPTI X LN(Rainfall Deviation) -384.082 -502.955 28.108 -533.541 183.066 222.889∗∗∗ 449.380 2693.801 173.212 193.229∗∗ (1033.872) (1012.242) (22.571) (1008.641) (209.504) (70.656) (404.951) (1773.382) (157.245) (81.966) Post X LN(Rainfall Deviation) -3511.005∗∗∗ -3431.508∗∗∗ 87.884∗∗∗ -3522.795∗∗∗ -202.058 246.009∗∗ -93.954 58.672 -554.753∗∗ -97.595 (761.296) (744.403) (30.795) (743.234) (213.704) (106.426) (255.454) (1046.710) (236.293) (100.365) TRIPTI X Post X LN(Rainfall 269.453 -190.457 69.953∗ -256.299 785.571∗∗ 208.975 264.793 6140.218 245.368 -64.324 Deviation) (1339.757) (1319.580) (42.069) (1317.331) (332.642) (141.754) (382.660) (5009.890) (357.278) (199.444) Constant 9528.703 7584.552 717.782∗∗∗ 6898.135 3218.046∗∗ 4123.533∗∗∗ 3057.043 17750.426∗∗ 5049.306∗∗∗ 2018.859∗∗∗ (6207.325) (6133.389) (181.344) (6143.147) (1459.518) (1278.363) (2028.791) (7076.496) (1500.646) (570.243) Total Effect of TRIPTI 2064.780 2145.127 -46.380 2207.446 -53.030 -798.553 -1924.065 -6023.335 -460.721 -950.029 p-value 0.71 0.69 0.72 0.68 0.96 0.05 0.35 0.30 0.52 0.02 Total Effect of Phailin 16499.637 15560.475 -358.175 15921.950 1897.214 -560.500 454.120 7595.398 2959.475 569.833 p-value 0.00 0.00 0.04 0.00 0.12 0.36 0.76 0.13 0.03 0.32 Control (TRIPTI) Observations 2694 2694 2694 2694 2694 2613 2663 2665 2624 2656 Treated (TRIPTI) Observations 2095 2095 2095 2095 2095 2013 2054 2048 2029 2053 Notes : ***sig. at 1%, **sig. at 5%, *sig. at 10%. All specifications include block fixed effects. The top panel includes standard errors clustered at the village; the bottom panel includes standard errors clusters by grid-village. Table 2: The Effects of TRIPTI and Phailin on Household Borrowing Total Amount Any Current Total Number of Currently Loans from an Any Current Loan Current Loans Borrowed SHG Post -0.124 -0.291 75645.547 -0.299 (0.696) (2.932) (63629.029) (0.187) LN(Rainfall Deviation) -0.038 0.077 -10845.947 -0.015 (0.092) (0.506) (11328.846) (0.039) Post X LN(Rainfall Deviation) 0.049 0.089 -10606.043 0.057∗ (0.122) (0.521) (11279.328) (0.033) Constant 0.831 1.304 81984.901 0.165 (0.519) (2.845) (63425.972) (0.221) Total Effect of Phailin -0.074 -0.202 65039.503 -0.242 p-value (0.90) (0.93) (0.22) (0.12) TRIPTI village -0.674 0.072 63209.621 0.214 (0.765) (2.838) (52984.540) (0.495) Post -0.223 4.287∗∗ 113704.287∗ -0.623∗∗∗ (0.471) (1.732) (65476.226) (0.223) LN(Rainfall Deviation) -0.059 0.297 -7253.530 -0.020 (0.096) (0.414) (11358.829) (0.041) TRIPTI X Post 0.431 -10.405∗∗∗ -112395.327 0.667∗∗ (0.952) (3.271) (110989.166) (0.309) TRIPTI X LN(Rainfall Deviation) 0.117 -0.007 -11522.355 -0.035 (0.135) (0.505) (9436.419) (0.088) Post X LN(Rainfall Deviation) 0.062 -0.737∗∗ -18299.386 0.113∗∗∗ (0.084) (0.302) (11518.075) (0.039) TRIPTI X Post x LN(Rainfall Deviation) -0.066 1.872∗∗∗ 21853.711 -0.117∗∗ (0.167) (0.586) (19678.155) (0.055) Constant 0.957∗ 0.064 62499.853 0.187 (0.540) (2.328) (63782.764) (0.234) Total Effect of TRIPTI -0.623 1.937 73540.977 0.062 p-value 0.39 0.45 0.14 0.89 Total Effect of Phailin -0.228 5.423 117258.612 -0.627 p-value 0.65 0.00 0.08 0.01 Control (TRIPTI) Observations 2736 2189 2732 2736 Treated (TRIPTI) Observations 2116 1741 2114 2116 Notes : ***sig. at 1%, **sig. at 5%, *sig. at 10%. All specifications include block fixed effects. The top panel includes standard errors clustered at the village; the bottom panel includes standard errors clusters by grid-village. 25 Table 3: The Effects of TRIPTI and Phailin on Women’s Civic Engagement Voted in the Last Village Has Heard Council Days Days of Knows the of Village Attended (Gram Worked Paid Work Name of Meeting the Last Pan- under under the Chief (Gram Village chayat) NREGA NREGA Minister Sabha) Meeting Election LN(Rainfall Deviation) 2.543 1.561 -0.159∗ 0.117 0.022 -0.012 (1.864) (2.011) (0.086) (0.082) (0.041) (0.033) Post 8.031 21.060∗∗ -0.825∗∗ 1.670∗∗∗ 0.116 -0.238 (6.635) (7.928) (0.312) (0.470) (0.176) (0.178) Post X LN(Rainfall Deviation) -1.306 -3.563∗∗ 0.171∗∗∗ -0.274∗∗∗ -0.016 0.038 (1.179) (1.391) (0.056) (0.085) (0.032) (0.032) Constant -13.385 -7.450 1.332∗∗∗ -0.527 -0.108 1.008∗∗∗ (10.431) (11.266) (0.484) (0.460) (0.228) (0.186) Total Effect of Phailin 6.725 17.497 -0.654 1.396 0.100 -0.201 p-value 0.22 0.01 0.01 0.00 0.49 0.17 TRIPTI village -0.441 3.604 -1.552∗∗ 0.020 0.064 0.183 (10.882) (13.251) (0.718) (0.492) (0.191) (0.247) LN(Rainfall Deviation) 2.626 1.756 -0.191∗∗ 0.142∗ 0.021 -0.006 (2.333) (2.570) (0.086) (0.077) (0.034) (0.038) Post 7.718 21.150∗ -0.547 2.263∗∗∗ 0.069 -0.183 (6.907) (10.991) (0.447) (0.490) (0.176) (0.183) TRIPTI X Post 0.327 -1.876 -0.062 -1.382∗∗ 0.079 -0.189 (13.537) (20.318) (0.608) (0.633) (0.311) (0.310) TRIPTI X LN(Rainfall Deviation) 0.127 -0.599 0.280∗∗ -0.008 -0.013 -0.035 (1.922) (2.345) (0.128) (0.090) (0.035) (0.045) Post X LN(Rainfall Deviation) -1.448 -3.746∗ 0.129 -0.384∗∗∗ -0.008 0.029 (1.215) (1.900) (0.078) (0.087) (0.032) (0.033) Treat X Post X LN(Rainfall Deviation) 0.300 0.636 -0.005 0.257∗∗ -0.014 0.031 (2.413) (3.562) (0.107) (0.114) (0.056) (0.056) Constant -13.908 -8.601 1.507∗∗∗ -0.657 -0.099 0.982∗∗∗ (13.090) (14.422) (0.485) (0.434) (0.192) (0.214) Total Effect of TRIPTI -0.013 3.640 -1.277 0.268 0.037 0.179 p-value 1.00 0.77 0.04 0.53 0.82 0.44 Total Effect of Phailin 6.570 18.040 -0.423 2.136 0.048 -0.122 p-value 0.37 0.12 0.35 0.00 0.80 0.52 Control (TRIPTI) Observations 2713 2713 2713 2713 2713 2713 Treated (TRIPTI) Observations 2085 2085 2085 2085 2085 2085 Notes : ***sig. at 1%, **sig. at 5%, *sig. at 10%. All specifications include block fixed effects. The top panel includes standard errors clustered at the village; the bottom panel includes standard errors clusters by grid-village. 26 8 Appendix 27 Table A1: Baseline Balance in TRIPTI Treatment Assignment (1) (2) T-test Non-TRIPTI TRIPTI Difference Variable N Mean/SE N Mean/SE (1)-(2) Total Expenditure Per Capita (Rs.) 1547 2388.851 998 2413.096 -24.245 (325.133) (388.437) Total Food Expenditure Per Capita (Rs.) 1569 1895.222 1029 1913.821 -18.599 (308.292) (360.493) Value of Home Produced Food Per Capita (Rs.) 1569 72.357 1029 85.047 -12.689** (6.998) (10.859) Expenditure on Food from Market Per Capita (Rs.) 1569 1822.599 1029 1828.222 -5.623 (305.772) (356.660) Expenditure on Non-Food Per Capita (Rs.) 1547 466.677 998 439.827 26.850 (27.930) (26.371) Ann. Expenditure on All Festivals (Rs.) 1547 285.033 998 269.097 15.936 (21.370) (34.483) Expenditure on Raja (Rs.) 1550 69.180 1009 62.113 7.067 (7.714) (8.186) Expenditure on Dusshera (Rs.) 1558 39.780 1011 38.966 0.815 (2.708) (3.890) Expenditure on Diwali (Rs.) 1556 27.011 1011 24.117 2.894 (2.617) (2.120) Expenditure on Other Festivals (Rs.) 1571 145.960 1009 140.740 5.220 (15.417) (31.537) Expenditure on Health (Rs.) 1547 915.824 998 843.024 72.800 (77.989) (106.668) Expenditure on Education (Rs.) 1547 1188.440 998 1116.090 72.350 (100.814) (149.312) Expenditure on Women’s Goods (Rs.) 1525 726.989 993 760.123 -33.134 (34.164) (61.126) Expenditure on Children’s Goods (Rs.) 1547 199.398 998 167.150 32.248* (20.081) (14.816) =1 Any Loan 1608 0.618 1050 0.614 0.004 (0.038) (0.058) Number of Loans 1066 1.687 677 1.873 -0.186 (0.082) (0.154) Total Value of Loans (Rs.) 1608 22822.886 1050 18626.689 4196.197* (2203.956) (2585.603) =1 Any Loan from an SHG 1608 0.069 1050 0.104 -0.035* (0.010) (0.027) Days Worked Under NREGA 1586 0.793 1019 1.055 -0.262 (0.273) (0.413) Days of Paid Work Under NREGA 1586 1.255 1019 1.473 -0.218 (0.343) (0.621) =1 Knows the Name of Chief Minister 1586 0.433 1019 0.459 -0.027 (0.022) (0.028) =1 Has Heard of the Gram Sabha 1586 0.146 1019 0.102 0.044** (0.024) (0.025) =1 Attended the Last Village Meeting 1586 0.016 1019 0.018 -0.002 (0.004) (0.005) =1 Voted in Last Village Council 1586 0.945 1019 0.940 0.004 (0.006) (0.009) Notes : The value displayed for t-tests are the differences in the means across the groups. Standard errors are clustered at variable cluster. Fixed effects using variable district code are included in all estimation regressions. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A2: Baseline Balance in Phailin-affected Villages (1) (2) T-test Above Median Rainfall Below Median Rainfall Difference Variable N Mean/SE N Mean/SE (1)-(2) Total Expenditure Per Capita (Rs.) 1313 1942.842 1232 2883.823 -940.980 (343.162) (485.942) Total Food Expenditure Per Capita (Rs.) 1333 1476.075 1265 2352.030 -875.955 (342.222) (449.203) Value of Home Produced Food Per Capita (Rs.) 1333 91.315 1265 62.702 28.613** (10.437) (8.816) Expenditure on Food from Market Per Capita (Rs.) 1333 1384.527 1265 2288.794 -904.267 (336.570) (444.159) Expenditure on Non-Food Per Capita (Rs.) 1313 444.284 1232 468.792 -24.509 (21.661) (35.316) Ann. Expenditure on All Festivals (Rs.) 1313 276.914 1232 280.778 -3.864 (22.019) (32.352) Expenditure on Raja (Rs.) 1313 61.014 1246 72.062 -11.048 (9.123) (9.906) Expenditure on Dusshera (Rs.) 1313 40.965 1256 37.887 3.078 (3.563) (3.216) Expenditure on Diwali (Rs.) 1313 26.978 1254 24.713 2.265 (2.814) (2.621) Expenditure on Other Festivals (Rs.) 1313 147.866 1267 139.828 8.038 (19.639) (22.878) Expenditure on Health (Rs.) 1313 947.955 1232 822.608 125.346 (123.038) (76.363) Expenditure on Education (Rs.) 1313 1138.219 1232 1183.355 -45.136 (121.474) (119.718) Expenditure on Women’s Goods (Rs.) 1308 732.250 1210 748.494 -16.244 (60.193) (43.814) Expenditure on Children’s Goods (Rs.) 1313 195.522 1232 177.407 18.115 (25.606) (14.646) =1 Any Loan 1333 0.680 1325 0.553 0.126* (0.054) (0.048) Number of Loans 952 1.661 791 1.877 -0.217 (0.106) (0.174) Total Value of Loans (Rs.) 1333 22224.246 1325 20099.851 2124.395 (2841.853) (2384.584) =1 Any Loan from an SHG 1333 0.104 1325 0.062 0.042 (0.023) (0.014) Days Worked Under NREGA 1313 1.023 1292 0.766 0.257 (0.376) (0.321) Days of Paid Work Under NREGA 1313 1.716 1292 0.959 0.757 (0.627) (0.322) =1 Knows the Name of Chief Minister 1313 0.450 1292 0.436 0.014 (0.034) (0.025) =1 Has Heard of the Gram Sabha 1313 0.079 1292 0.179 -0.100*** (0.023) (0.025) =1 Attended the Last Village Meeting 1313 0.012 1292 0.021 -0.009 (0.004) (0.005) =1 Voted in Last Village Council 1313 0.939 1292 0.947 -0.008 (0.009) (0.007) Notes : The value displayed for t-tests are the differences in the means across the groups. Standard errors are clustered at variable cluster. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A3: The Effects of TRIPTI and Phailin on Per Capita Festival Expenses Other Festivals All Festivals Raja 2014 Dussehra 2013 Diwali 2013 in 2013-14 LN(Rainfall Deviation) -725.319∗∗∗ -174.666∗∗∗ -149.025 -88.479∗∗ -469.477∗∗ (227.890) (52.231) (95.075) (40.886) (189.413) Post X LN(Rainfall Deviation) 367.069∗∗∗ 147.423∗∗∗ 10.402 -3.745 263.843∗∗∗ (74.272) (32.513) (31.705) (15.966) (50.473) Post -1686.978∗∗∗ -633.470∗∗∗ -5.614 79.249 -1334.495∗∗∗ (416.025) (180.691) (178.844) (89.434) (283.216) Constant 4341.837∗∗∗ 1044.327∗∗∗ 873.986 521.283∗∗ 2773.214∗∗∗ (1280.720) (292.301) (532.243) (228.860) (1066.677) Total Effect of Phailin -1319.910 -486.047 4.788 75.504 -1070.651 p-value 0.00 0.00 0.97 0.30 0.00 TRIPTI village -1230.416∗∗∗ -38.981 -167.377∗∗ -51.023 -814.885∗∗∗ (395.086) (184.592) (82.718) (61.979) (261.173) LN(Rainfall Deviation) -687.363∗∗∗ -105.767∗∗ -118.101 -84.817∗∗ -523.525∗∗∗ (227.092) (52.343) (83.281) (35.893) (194.181) Post -1015.484∗ -235.836 208.497 208.199 -1255.236∗∗∗ (602.011) (249.196) (299.899) (139.753) (377.733) TRIPTI X Post -1160.976 -877.194∗∗ -425.494 -279.728 36.967 (791.289) (354.103) (331.372) (173.041) (540.113) TRIPTI X LN(Rainfall Deviation) 222.889∗∗∗ 4.837 30.406∗∗ 9.838 149.941∗∗∗ (70.656) (32.998) (14.761) (11.133) (46.805) Post X LN(Rainfall Deviation) 246.009∗∗ 79.146∗ -25.883 -27.275 244.174∗∗∗ (106.426) (44.504) (52.837) (24.789) (66.375) TRIPTI X Post X LN(Rainfall Deviation) 208.975 152.036∗∗ 72.135 50.785 2.981 (141.754) (63.760) (58.710) (30.899) (96.977) Constant 4123.533∗∗∗ 662.797∗∗ 699.604 499.338∗∗ 3067.600∗∗∗ (1278.363) (292.322) (465.941) (200.733) (1096.344) Total Effect of TRIPTI -798.553 117.892 -64.835 9.600 -661.963 p-value 0.05 0.52 0.47 0.87 0.02 Total Effect of Phailin -560.500 -4.654 254.748 231.709 -1008.081 p-value 0.36 0.99 0.40 0.10 0.01 Control (TRIPTI) Observations 2613 2673 2679 2677 2645 Treated (TRIPTI) Observations 2013 2067 2068 2071 2025 Notes : ***sig. at 1%, **sig. at 5%, *sig. at 10%. All specifications include block fixed effects. The top panel includes standard errors clustered at the village; the bottom panel includes standard errors clusters by grid-village. 30