The Politics of Drought Relief - Evidence from Southern India Pt 5

Appendix B B.1 Data Block and District Names The spelling of the names of Indian districts and blocks varies across the sources of data used in this paper. In order to match geographic units across data sources, I use exact name matches between any two data sources. For blocks, I use the state name, district name, and block name in the matching process as block names may be repeated across districts within a single state. Likewise, in the matching for district names I also use the state name. I then determine which blocks (or districts) in each data source are not paired and review each of the individual cases. In most cases, the lack of a match is due to a simple variation in spelling that can be adjusted by comparing the unmatched units. In a few cases, the name of a block’s headquarters is listed instead of the name of the block itself and so I identify the correct match in these instances through web searches on the administrative units. B.2 Official Drought Declarations Data I construct a block-level panel dataset of official drought declarations for the period 2008 to 2019 for Karnataka and 2009 to 2019 for Andhra Pradesh and Telangana. The Karnataka State Natural Disaster Monitoring Centre (KSNDMC) provided a list of declarations for the full period except for the 2019 season, for which data comes from an official drought declaration publication. The Telangana State Development Planning Society (TSDPS) provided a list of official drought declarations for the full period at the block level as well. Annual lists of drought declarations for the blocks in Andhra Pradesh are collected from official drought declaration documents, except for the year 2013. Data for 2013 is based on a state report listing the number of blocks per district that received a declaration. For 7 districts no blocks received a declaration and in 1 district all blocks received a declaration; for these 8 districts, I am able to code the declaration indicator. However, for 5 districts a subset of blocks received declarations but I do not observe which blocks and so the declaration indicator is set to missing for 2013. B.3 Political Data The Election Commission of India (2019) and Bhavnani (2014) provide data on state Legislative Assembly elections in Andhra Pradesh/Telangana and Karnataka. From this data, I construct the two measures described in Section 2.2 as well as voter turnout and the HerfindahlHirschman Index (HHI) of vote shares. In order to connect this constituency-level political data to the block-level declaration and environmental data, block boundaries must be mapped to the political boundaries. I map blocks to 2008-boundary assembly constituencies using information provided in the delimitation documentation for each state (Election Commission of India, 2008).55 In my sample of matched blocks-constituencies, about 92 percent of blocks map to a single constituency while the remaining blocks map to two or more constituencies. Therefore, to construct block-level political variables, for each block b in year t, I take the average of political variables over the n mapped constituencies c(b) in the prior election e(t), or variablec(b)e(t) = 1X variablece(t) . n c(b) Note, the aligned indicator at the block level is then an indicator that the average ruling party’s vote margin was greater than 0. Instrument Construction In addition to the mapping described above, a second mapping is needed for instrument construction between 2008-boundary assembly constituencies and the 2004-boundary constituencies (the latter set of boundaries were actually designated in 1976). This constituency-toconstituency mapping is made using the delimitation documents Election Commission of India (1976) and Election Commission of India (2008) as well as a mapping provided by Iyer and Reddy (2013).56 Using this second mapping allows for the construction of an instrument based on information from the pre-delimitation, 2004 election. The goal of this instrument is to estimate the counterfactual vote margin, or what voters’ support for each party would be if unaffected by local unobserved factors that also influence future declarations. However, an issue with using information from the 2004 election to estimate support for the ruling and opposition parties is that some of the parties from the 2008/9, 2013/14, and 2018/19 elections either did not exist in 2004 or did not run in all constituencies that they would later contest. Assigning a value of 0 for such a party-constituency observation would lead to a predicted vote share of 0 and so the vote margin would simply reflect the vote share of the remaining party, assuming that the party ran in 2004. Even in the case where a party did exist in 2004, it may have 55 Based on available data, I was unable to match blocks in Hyderabad, Telangana to constituencies and so these blocks are excluded from the analysis; however, these are blocks that are not subject to drought declarations during the period of analysis. 56 For a number of constituencies located in Bangalore, I am unable to construct a mapping between 2004boundary and 2008-boundary constituencies so blocks that map to these constituencies are dropped from the sample. had support in a given constituency but not contested the election because a coalition partner was running or the party did not anticipate sufficient support to win. Instead of assigning 0 baseline support, I assign the mean value of observed 2004 vote shares for ruling party and opposition party-constituency pairs; I chose the mean as it might be representative of how the party would have done had it run. For ruling and opposition party-constituency observations that can be matched to a 2004 vote share the mean vote share is .36 (.35 for opposition parties only and .38 for ruling parties only). As a robustness check, in Appendix Table C.8 I instead assign the first quartile value for the vote share of .28 (column 2) or the third quartile value of .46 (column 3). First, the result that declaration likelihood is decreasing in the absolute value of the vote margin is robust across these two alternative specifications. Second, the magnitude of the coefficient is larger than the OLS coefficient in all cases, though the magnitudes vary depending on which value is assigned for unobserved vote shares. I also create there alternative instruments to study the robustness of the main findings. For the first, I increase the leave-out in the state-level shift in voters’ support for a party to drop all constituencies located in the same district. Second, in order to construct an instrument in which all temporal change from nearby constituencies is removed from the shift factor, I utilize a publicly available map of modern constituencies. I find the centroid for each constituency polygon and identify all other centroids that lie within a 100 km radius (and the same state). Third, I reconstruct the instrument using data from the 1999 election, applying the same adjustment described above, for parties that did not run in a given constituency in 1999. B.4 Environmental Data I use the India Meteorological Department’s .25x.25 gridded rainfall data from Pai et al. (2014) and the .5◦ -by-.5◦ gridded, model-calculated average monthly soil moisture for van den Dool et al. (2003).57 In order to transition this gridded data to block-level variables, I use shapefiles for blocks in Karnataka and Andhra Pradesh/Telangana.58 To get the block-level averages, I first reproject the gridded data to Cylindrical Equal Area, then convert the raster to a polygon and intersect it with the shapefiles (also re-projected to Cylindrical Equal Area). I use the area of each intersected polygon within a block to get the area-weighted average of rainfall or mean soil moisture. Due to the fact that the IMD rainfall is available for grid points that fall on land only, some blocks located on the coast are not completely covered by a rainfall grid polygon; for 57 CPC Soil Moisture data is provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA. The shapefile used for Karnataka was provided by Professor Ashwini Chhatre while the shapefile for Andhra Pradesh/Telangana is from ML InfoMaps. 58 these blocks, I use the area-weighted average for the overlap. For four blocks for which there is no overlap, I use the average rainfall for a neighbouring block. The normalized difference vegetation index (NDVI) data is provided on a per pixel basis, with a spatial resolution of 1 kilometer, for each of nine 16-day periods that overlap the primary (Kharif) growing season (Didan, 2015). To construct the NDVI for each block and 16-day period, I first convert the pixels to a gridded (raster) dataset and re-project the gridded data into Cylindrical Equal Area. I then overlap the block-level polygons with the NDVI gridded data and take the average of all points that fall within the block. The NDVI gridded data is very dense and so taking an area-weighted average of all points within a block (as is done for rainfall and soil moisture) for each 16-week period in the 12 years in this sample would be a very data-intensive process. The non-area-weighted average is instead used because it is calculable within a reasonable timeframe; however, because of the density of data points for NDVI, taking an area-weighted average might not be as important as for gridded datasets with more sparsely available data. The guidelines offer specific variables and cutoffs that can be used for the soil moisture and NDVI data in order to generate indicators needed for meeting the sufficient condition for a declaration, particularly in the more recent drought manual. However, instead of generating indicator variables, I use the block-level average of NDVI for the nine 16-day periods and monthly mean soil moisture for the 4 months of the main agricultural season discussed above to flexibly control for soil moisture and vegetation. To account for variation in hydrological indices, I focus on groundwater levels and utilize monthly, district-level data (NWIC, 2020). The groundwater level data is not available at the block level or for all months, but is available for most districts in May and August. These months are also of particular relevance because they occur immediately before the growing season and at the end of the season, giving a sense of the variation in groundwater over the full season. For crop area sown, the drought manual suggests that an indicator can be formed for area sown being less than 50 percent of total cultivable area (DAC, 2009). Again, data for crop area sown is only available at the district level and for the years 2008 to 2018 (DAC&FW, 2020). Because district boundaries in Karnataka and Telangana changed between 2008 and 2018, I combine districts that split post 2008 into the original 2008 district.59 Instead of generating the measure suggested in the manual, which requires information on cultivable land, I generate the 59 For three of Telangana’s newly formed districts (Jangaon, Vikarabad, and Siddipet) the district was formed from two or three 2008-districts and so I assign the crop area statistics to the primary district from which it was separated. ratio of the total area sown in the main agricultural season (the Kharif season) for a given year to the average total area sown during the season for the years 2008 through 2018; this measure is meant to capture how area sown in a given year varies from that of a normal year, measured as the average over the period.60 I also use this district-level data in the analysis of the impacts of drought and declarations on agricultural production. To do so, I also create measures of the share of cropped area sown that is devoted to each crop or crop group for each district-year.61 One concern with utilizing the difference between early season NDVI and maximum NDVI as a proxy for agricultural production is that it does not take into account changes in cropping patterns. If drought or declarations impact the cropping decisions of households and this NDVI difference is crop specific, this would impact the interpretation of the results reported in Section 6. In order to ascertain whether this might be an issue, I re-estimate Equation 6 on the reduced sample in column 1 of Table C.5 and add the crop share controls in column 2. Importantly, the addition of controls for the crop shares does not impact the coefficients on prior-year drought or declarations. Because rainfall indicators are the primary measure for determining declarations, I take the block-level area-weighted averages and construct indicators for drought-affected areas. According to the 2009 guidelines, the primary criteria for a drought declaration in the primary agricultural season is met if (1) the total rainfall for June and July is below 50 percent of normal (with or without a dry spell) or (2) the total Kharif-season rainfall (June to September) is below 75 percent of normal. To construct these indices, I start by constructing the ratio of current rainfall to the long-term-average rainfall, or ratioperiod = current total rainfall in given period long-term-average rainfall for given period for the relevant period of months or weeks.62 Here a ratio of 1 indicates that the block received normal or average rainfall while a ratio less than (greater than) 1 indicates below-average 60 For the six districts in Telangana from which Jangaon, Vikarabad, and Siddipet districts were separated, my measure of total crop area necessarily changes in 2016 because I only assign the newly-formed district to the primary 2008-district. However, when running the regression in column 7 of Table 5, where this crop area sown ratio is used as a control, I can also include an indicator for being in one of these six districts post 2016 and the interaction between the area ratio and the indicator and it does not significantly impact the results. The coefficient on the absolute value of the ruling party’s vote margin becomes -.772 (s.e. .242) as opposed to -.766 (s.e. .244). 61 I include shares for individual crops including cotton, dry chilies, black pepper, ginger, coriander, tumeric, arecanut, niger seed, tobacco, cashew nut, sugarcane, tapioca, and mesta. For the remaining crops, I include shares for crop groups: millets, grain, oilseeds, fruit, vegtables and tubers, or pulses and legumes. 62 The long-term-average should consider 30 prior years, as a minimum. For a given year t, I take the average rainfall over the period 1998 to t − 1. (above-average) rainfall. I use these ratios themselves as continuous controls for rainfall in some specifications and also combine the ratios with cutoffs specified in the guidelines to create indicator variables. The baseline indicator for drought affected blocks (used throughout the paper) is then, June/July droughtbt = 1{ratiobt < .5 ∨ ratioJune−Sept < .75}. bt While this indicator is proscribed by the national guidelines for the majority of the sample period, there are other indicators of drought that might be used by state governments during the period. For example, during part of the sample period, official drought declarations in Andhra Pradesh/Telangana cite an alternative, state-specific cutoff for the June-September rainfall ratio: .75 for areas where normal rainfall is above 1000 millimeters; .80 where normal rainfall is 750-999.9 millimeters; and .85 where normal rainfall is below 750 millimeters (GoAP, 2013). I take the baseline drought indicator, described in the previous paragraph, and adjust the full-season criteria for Andhra Pradesh/Telanga to create an alternative indicator droughtAP bt . Further, the 2016 national guidelines add an additional criteria, based on dry spells, to the two listed above to define the necessary condition for a drought declaration. According to the guidelines, dry spell is defined as 3-4 weeks of consecutive rainfall below 50% of normal following the onset of the monsoon. I do not attempt to estimate the beginning of the monsoon in each year, instead I create an indicator that takes a value of one if ratioW eek is less than .5 for any consecutive 4 weeks during June to September.63 Again, I construct an alternative indicator for drought-affected blocks that takes the baseline indicator and changes the value to one if the ratiois less than .5 for any consecutive 4-week period, droughtDS bt . Appendix Table C.9 provides the results of estimating Equation 4 using each of the alternative drought indicators. While there are some small differences in the coefficient on the absolute value of the ruling party’s vote margin, the results showing that declaration likelihood is higher in competitive areas is robust to these alternatives. Finally, to identify blocks that might be considered urban, I use the harmonized nighttime lights dataset provided by Li et al. (2020) and overlay each image with block-level polygons. I then take an area-weighted average of data points as the block-level night lights value. I consider a block to be urban if its night lights in 2008 was above the 95th percentile for all blocks in the sample. 63 Also, prior to the change in guidelines, the Karnataka state government used rainfall deficiency, soil moisture, and more than 4 weeks of dry spell to determine drought declarations (GoK, 2016). Appendix C Additional Tables and Figures Table C.1: Instrument Construction for Vote Shares and Ruling Party Vote Margin Dependent variable: votesharepce margince (1) p ^ ce voteshare 0.290∗∗∗ (0.038) ^ ce margin Observations R2 F-Statistic Constituency FE Election FE (2) 2,929 0.295 58 Yes Yes 0.252∗∗∗ (0.044) 1,470 0.403 33 Yes Yes Note: This table reports regressions of observed political variables on constructed political variables for each stage of instrument construction (see Section 4.1) at the constituency-election level. All specifications include constituency and election fixed effects. Column 1 reports results for a regression of observed vote shares on constructed vote shares for the ruling party and main opposition in each constituency. Columns 2 provides an estimation of the regression of the ruling party’s vote margin on the constructed vote margin. The F-statistic for the constructed instrument is reported. Standard errors are clustered by constituency. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.2: Drought Declarations and Political Motivations, Alternative Functional Form Declarationbt (1) droughtbt marginc(b)e(t) margin2c(b)e(t) Observations First Stage F-Stat (1st Variable) First Stage F-Stat (2nd Variable) KP First Stage F-Stat Specification Block FE State-Year FE 0.192∗∗∗ (0.007) −0.013 (0.020) −0.404∗∗∗ (0.079) 14,043 OLS Yes Yes (2) 0.186∗∗∗ (0.007) −0.098 (0.089) −1.711∗∗∗ (0.563) 14,043 58 22 10 IV Yes Yes Note: This table reports an estimation similar to Equation 4 but the indicator for block b receiving a drought declaration in year t is regressed on the a quadratic in the ruling party’s vote margin (instead of the absolute value) and an indicator for drought occurrence (OLS specification in column 1 and IV specification in column 2). All specifications include block and state-year fixed effects. The instrument construction for the ruling party’s vote margin is explained in Section 4.1 and the results of the first-stage regressions are provided in Table C.3. The Sanderson-Windmeijer first stage F-statistics are reported for each of the endogenous variables followed by the Kleibergen-Paap Wald rk F statistic for the IV specification. The Anderson-Rubin Wald test has a p-value of 0.0002 for the IV specification. Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.3: First Stage Regressions of Vote Margin and Squared Vote Margin on Instruments Column 2 \ c(b)e(t) margin 2 \ c(b)e(t) margin droughtbt Observations R2 Block FE State-Year FE marginc(b)e(t) margin2c(b)e(t) (1) (2) 0.282∗∗∗ (0.026) −0.011∗ (0.007) −0.769∗∗∗ (0.132) −0.0001 (0.003) 14,043 0.436 0.238∗∗∗ (0.054) −0.004∗∗∗ (0.001) 14,043 0.603 Yes Yes Yes Yes Note: This table displays the results of the first stage regressions for the IV regression shown in column 2 of Table C.2. Column 1 of this table provides the first stage for the the ruling party’s vote margin and column 2 shows the first stage for the square of the ruling party’s vote margin. Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.4: Robustness of Agricultural Production Results Instrumenting for Prior-Year Declaration with Prior-Year Drought IHS(∆N DV Ibt ) Declarationbt−1 (1) Declarationbt−1 0.265∗∗∗ (0.033) droughtbt−1 Observations KP First Stage F-Stat Specification Block FE State-Year FE Env. Variables (2) 14,037 IV Yes Yes Yes 0.178∗∗∗ (0.007) 14,037 584 First Stage Yes Yes Yes Note: Column 1 of this table reports an IV estimation of Equation 6 where the dependent variable is the inversehyperbolic sine (IHS) of the NDVI difference, or the difference between maximum and early-season vegetation index, for block b in year t and prior-year drought becomes an excluded instrument for the independent variable of interest, the indicator for block b having received a declaration in the prior year, t − 1. Column 2 reports the first stage regression of the declaration indicator for block b in t − 1 on an indicator for drought occurrence in block b and year t − 1. Both specifications include block and state-year fixed effects; the ratio of current June-September rainfall to its long-term average (and the squared ratio); the June/July rainfall to its long-term-average (and the squared ratio); and mean monthly soil moisture. The Kleibergen-Paap Wald rk F statistic is reported for the IV specification. Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.5: Robustness of Agricultural Production Results Controlling for Cropping Patterns IHS(∆N DV Ibt ) (1) Rainbt Rain2bt droughtbt−1 Declarationbt−1 droughtbt−1 × Declarationbt−1 Observations R2 Block FE State-Year FE Env. Variables Political Variables Crop Shares 0.263∗∗∗ (0.046) −0.040∗∗ (0.017) −0.008 (0.010) 0.032∗∗∗ (0.009) 0.035∗∗∗ (0.012) 12,750 0.769 Yes Yes Yes Yes No (2) 0.261∗∗∗ (0.045) −0.042∗∗∗ (0.016) −0.005 (0.010) 0.030∗∗∗ (0.009) 0.034∗∗∗ (0.012) 12,750 0.776 Yes Yes Yes Yes Yes Note: This table reports estimations of Equation 6 where the dependent variable is the inverse-hyperbolic sine (IHS) of the NDVI difference, or the difference between maximum and earlyseason vegetation index, for block b in year t. The independent variables are the ratio of current June-September rainfall to its long-term average (Rain) in block b and year t, an indicator for drought occurrence in year t − 1, and an indicator for receiving an official drought declaration in year t − 1. Both specifications also include block and state-year fixed effects; mean monthly soil moisture; the June/July rainfall to its long-term-average (and the squared ratio); the absolute value of the ruling party’s vote margin; the indicator for being aligned to the state ruling party; a comparable indicator for alignment to the federal ruling coalition; voter turnout; the HHI of vote shares; the number of candidates who ran; indicators for years since the last election; and indicators for the political party of the MLA(s). Additionally, column 2 also adds controls for the share of agricultural area sown devoted to each crop or crop group (for details, see Appendix B). Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.6: Planned Relief Expenditures Relative to Drought Occurrence/Severity Planned Relief Expenditures (1) droughtbt −2.530 (2.800) Rainbt Rain2bt Areabt Observations R2 District FE (2) 0.016∗∗∗ (0.004) 100 0.291 No (3) (4) 0.279 (2.896) 14.958 (23.387) −7.507 (8.449) 0.016∗∗∗ (0.004) 100 0.294 No 0.016∗∗∗ (0.005) 100 0.613 Yes −18.151 (27.575) 6.417 (9.309) 0.015∗∗∗ (0.005) 100 0.616 Yes Note: This table uses data on planned relief expenditures for agricultural, horticultural, and animal husbandry subsidies in 2018 for blocks located in Karnataka (GoK, 2018). The table reports the results of specifications where the dependent variable is the amount of planned relief expenditure (in 100,000,000 rupees) and the independent variables are the geographic area of the block (in square meters) and an indicator for drought occurrence (column 1) or a quadratic in the ratio of June-September rainfall to is long-term average (column 2). Columns 3 and 4 repeat the specifications in columns 1 and 2, respectively, but adding district fixed effects. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.7: Impact of Drought and Declarations on Night Lights Average (1) droughtbt−1 Declarationbt−1 droughtbt−1 × Decbt−1 Observations R2 Block FE State-Year FE Env. Variables Political Variables 0.018 (0.012) −0.006 (0.007) −0.024∗ (0.014) 12,268 0.859 Yes Yes Yes Yes IHS(N Lbt ) 50th Percentile 5th Percentile (2) 0.056∗∗ (0.024) −0.012 (0.013) −0.049∗ (0.027) 12,268 0.729 Yes Yes Yes Yes (3) −0.057∗ (0.030) 0.065∗∗∗ (0.022) −0.005 (0.038) 12,268 0.746 Yes Yes Yes Yes Note: This table reports estimations of Equation 6 where the dependent variable is the inverse-hyperbolic sine (IHS) of night lights for block b in year t (sample is limited to blocks with 2008 nigh lights below the 95th percentile). Column 1 considers the area-weighted average light and columns 2 and 4 consider the median and 5th percentile of night lights within a block, respectively. The independent variables are an indicator for drought occurrence in year t − 1, and an indicator for receiving an official drought declaration in year t − 1. Controls include block and state-year fixed effects: the ratio of current June-September rainfall to its long-term average (and squared ratio); mean monthly soil moisture; the June/July rainfall to its longterm-average (and the squared ratio); the absolute value of the ruling party’s vote margin; the indicator for being aligned to the state ruling party; a comparable indicator for alignment to the federal ruling coalition; voter turnout; the HHI of vote shares; the number of candidates who ran; and indicators for the political party of the MLA(s). Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.8: Robustness to Alternative Baseline Vote Share Assumptions Parties Missing a 2004 Vote Share Assigned: Mean 1st Quartile 3rd Quartile (1) Panel A: droughtbt |marginc(b)e(t) | Observations droughtbt Observations R2 KP First Stage F-Stat Block State-Year (3) Declarationbt 0.184∗∗∗ (0.008) −0.909∗∗∗ (0.302) 14,043 0.188∗∗∗ (0.008) −0.550∗∗ (0.257) 14,043 0.183∗∗∗ (0.008) −1.079∗∗∗ (0.336) 14,043 |marginc(b)e(t) | Panel B: \ c(b)e(t) | |margin (2) 0.148∗∗∗ (0.038) −0.009∗∗∗ (0.002) 14,043 0.582 15 Yes Yes 0.156∗∗∗ (0.040) −0.009∗∗∗ (0.002) 14,043 0.582 15 Yes Yes 0.132∗∗∗ (0.035) −0.009∗∗∗ (0.002) 14,043 0.582 14 Yes Yes Note: Panel A of this table reports IV estimations of Equation 4, where the indicator for block b receiving a drought declaration in year t is regressed on the absolute value of the ruling party’s vote margin and an indicator for drought occurrence. Panel B reports the first stage regression for each IV specification. Column 1 reports the specification with the baseline instrument, where parties that did not run in a constituency in 2004 are assigned the mean observed vote share. Columns 2 and 3 instead use an instrument where the first quartile and third quartile, respectively, of the observed vote shares is assigned to those parties. The Kleibergen-Paap Wald rk F statistic for the endogenous variable is reported. The Anderson-Rubin Wald test has a p-value of 0.0000 in columns 1 and 3 and 0.0136 in column 2. Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table C.9: Robustness to Alternative Definitions for Drought-Affected Blocks Declarationbt (1) droughtbt (2) 0.184∗∗∗ (0.008) 0.201∗∗∗ (0.008) droughtAP bt droughtDS bt |marginc(b)e(t) | Observations KP First Stage F-Stat Block Year Election Cycle (3) −0.909∗∗∗ (0.302) 14,043 15 Yes Yes Yes −0.882∗∗∗ (0.306) 14,043 15 Yes Yes Yes 0.119∗∗∗ (0.007) −0.997∗∗∗ (0.317) 14,043 15 Yes Yes Yes Note: This table reports IV estimations of Equation 4, where the indicator for block b receiving a drought declaration in year t is regressed on the absolute value of the ruling party’s vote margin and an indicator for drought occurrence. Column 1 utilizes the baseline indicator for drought while columns 2 and 3 use alternative indicators for droughtaffected blocks described in Appendix B. The Kleibergen-Paap Wald rk F statistic for the endogenous variable is reported. The AndersonRubin Wald test has a p-value of 0.0000 in columns 1 and 2 and .0001 in column 3. Standard errors are clustered by block. ∗ , ∗∗ , ∗∗∗ denote statistical significance at the 10%, 5%, and 1% levels, respectively. Figure C.1: Official Drought Declaration Likelihood for Aligned and Non-Aligned Blocks (a) Likelihood of Official Drought Declaration versus Ruling Party’s Vote Margin (b) Distribution of Running Variable Note: This figure provides information relevant to a regression discontinuity design testing for the affect of being aligned to the ruling party (having a positive vote margin) on the likelihood a block receives an official drought declaration. Panel A plots a regression of the indicator for a block receiving a declaration on a fourth order polynomial in the running variable, the ruling party’s vote margin, allowing for a discontinuity at a vote margin of 0; in addition, the figure includes a binscatter of the mean declaration likelihood against the mean vote margin where each point represents about 100 block-year observations. Panel B considers constituencyelection observations, showing the distribution of the running variable in a histogram (left) and a non-parametric regression of the distribution allowing for a discontinuity at a vote margin of 0 (right) as in McCrary (2008). The coefficient estimate for the discontinuity in the density is .389 (s.e. .093) suggesting that, at least in some elections in the sample, the ruling party is more likely to win a close election than an opposition party. Figure C.2: Impact of Declarations on Agricultural Production Controlling for Additional Prior-Year Environmental Measures Note: This figure plots coefficient estimates and confidence intervals for an estimation of Equation 6 where the dependent variable is the inverse-hyperbolic sine (IHS) of the NDVI difference, or the difference between maximum and early-season vegetation index, for block b in year t and the indicator for drought has been replaced by indicators for the quintile of the prior-year rainfall shock distribution. The rainfall shock is measured as the ratio of June-September rainfall in the prior-year to its long-term-average. The figure displays the coefficient estimates for the declaration indicator interacted with quintile of the prior-year rainfall distribution and the x-axis provides the range of prior-year rainfall shock values for each quintile. The lowest quintile then displays the impact of declarations in areas affected by a negative rainfall shock (less than 78% of normal rainfall) while the highest quintile contains areas with excessive rainfall (more than 28% above normal). The third quintile represents areas with rainfall around the average (about a ratio of 1). Repeating the results shown in Figure 4, the first set of coefficients (grey) are from a regression that controls for block and state-year interacted fixed effects; mean monthly soil moisture; the June-September rainfall to its long-term-average (and the squared ratio); the June/July rainfall to its long-term-average (and the squared ratio); the absolute value of the ruling party’s vote margin; the indicator for being aligned to the state ruling party; a comparable indicator for alignment to the federal ruling coalition; voter turnout; the HHI of vote shares; the number of candidates who ran; indicators for years since the last election; and indicators for the political party of the MLA(s). The second set of coefficients (black) are for a regression that additionally controls for the prior-year NDVI difference; prior-year monthly soil moisture; and the prior-year June/July rainfall to its long-term-average (and the squared ratio). Standard errors are clustered at the block level.