Political Organizations and Political Scope Research Part 4

i A Online Appendix: Theoretical Arguments A.1 A.1.1 Misaligned Incentives Reduce Political Targeting α > 0 and βk = 0 Let θ be the Lagrange multiplier on the state politician’s constraint (2), and µk the local politician’s constraint (4). The state politician chooses Bk to satisfy ∂Wk ∂Vk +α ∂Bk ∂Bk # " ∗ X 0 ∗ ∂bi = µk + α vi (bi ) ∂Bk (14) θ= (15) i where the second equality follows from the Envelope theorem. Define gi = ∂b∗i ∂Bk Let Ik be the number of households in locale k and let v̄i0 and ḡi denote the within-locale mean of vi0 and gi . Simple algebra yields X i vi0 gi = X (vi0 − v̄i0 )(gi − ḡi ) + v̄i0 ḡi (16) i Note that P ḡi = ∂b∗i i ∂Bk Ik = 1 Ik where the second equality follows from differentiating both sides of (4) with respect to Bk . Combining this expression with (16) and (15) yields " v̄ 0 θ = µk (Bk , Ik ) + α (vi0 − v̄i0 )(gi − ḡi ) + i Ik i     ∂b∗ v̄ 0 = µk (Bk , Ik ) + α Ik Cov vi0 , i + i ∂Bk Ik X # (17) (18) The extent to which (18) diverges from the welfare-maximizing allocation—the case where α = 0 and the Lagrange multipliers are equalized—depends largely on the size of the covariance term. ii This term captures the extent to which a marginal increase in Bk is distributed across households in a way that covaries with their marginal political importance. When vi differs substantially from ui , the covariance is small. In the limit where the covariance approaches zero and the number of households is large enough that v̄i0 /Ik ≈ 0 the first-order condition (18) converges to the condition characterizing a welfare-maximizing allocation with no political misallocation. A.1.2 α = 0 and βk > 0 We sketch this argument in two steps. First we show that the state politician will allocate less funds to a local politician who engages in political targeting. Let b̂0i (Bk ; βk ) be the targeting function of the local politician. Given that α = 0 the state politician chooses Bk to satisfy ∂Wk ∂Bk X = u0i (b̂0i (Bk ; βk ))b̂0i (Bk ; βk ) θ= (19) (20) i = X h i b̂0i (Bk ; βk ) u0i (b̂0i (Bk ; βk )) − ū0i (Bk ; βk ) + ū0i (Bk ; βk ) (21) i where ū0i (Bk ; βk ) is the average marginal utility among households in k and the third line follows P because i b̂0i (Bk ; βk ) = 1 by the local politician’s budget constraint. The first term is a measure of how well the local politician is directing each marginal dollar received to households with above-average marginal utility from benefits. If βk = 0 this term is zero because the optimal allocation will equalize marginal utility across households. To the extent that the local politician targets benefits away from the neediest households, the term will be negative. As the first term becomes more negative the equality will hold only if the second term, the average marginal utility, becomes more positive. Since ū0i (Bk ; βk ) is decreasing in Bk , a higher average marginal utility implies a lower level of Bk . In other words, the local politician’s budget is cut as his targeting becomes more political. The second step is to show that the local politician takes the state politician’s behavior into iii account. The local politician’s Lagrangian is # " L = X ui (bi ) + βk i X vi (bi ) + µk Bk ({bi }) − i X bi (22) i where the crucial feature is that the local politician understands that Bk is a function of his targeting {bi }. His optimal allocation will satisfy u0i (bi ) The term ∂Bk ∂bi + βk vi0 (bi )  = µk ∂Bk 1− ∂bi  (23) measures how much the state politician will increase or decrease the local politi- cian’s budget in response to a marginal increase in benefits targeted to i. The term will be negative if increasing benefits to i will reduce aggregate welfare, as it would if i is being targeted for its politician importance rather than its objective need. Now define the “efficiency weight” of i as ei = 1 1− ∂Bk ∂bi (24) This term is large when targeting benefits to i will increase total welfare and thus increase the number of projects received by k from the state politician. We can rewrite the local politician’s first-order condition as   ei u0i (bi ) + βk vi0 (bi ) = µk (25) This condition shows that marginal utility from targeting i will be magnified if doing so increases total welfare, and diminished otherwise. In other words, the likely impact of targeting on the local politician’s budget disciplines him from diverging too far from the choices that maximize household welfare. B B.1 Appendix: Additional Empirical Results Tests for Manipulation Figure B1 shows a McCrary test for whether there is a discontinuity in the density of the onenorm at the margin where the AITC wins an absolute majority (McCrary, 2008). Any such discon- iv Figure B1 4 4 5 5 McCrary Density Test Log Difference: -0.09 2 3 3 1 2 0 0 1 Density (0.13) -.5 -.25 0 .25 1-Norm (Taxi) Distance .5 -.5 -.25 0 .25 1-Norm (Taxi) Distance .5 tinuity would imply that the AITC or one of its competitors is able to manipulate the outcomes of elections to ensure it wins barely enough votes in barely enough seats to win a majority. Such precise manipulation is implausible because the Election Commission of India is a non-partisan bureau that is widely respected and considered free from corruption. Figure B1 confirms that there is no such discontinuity. Tables B1—B3 report a number of balance tests by estimating Equation 9 on outcomes that were determined before the 2013 panchayat election (in mid-2013). Table B1 uses data from the 2011 Census to test for whether the AITC was systematically better at winning a narrow absolute majority in areas with particular characteristics than in others. Columns 1–4 of Table B1 show, however, that there is no evidence that the AITC was more likely to hold a narrow absolute majority in areas with a larger number of households, a higher population, or more lower-caste inhabitants (Scheduled Castes and Scheduled Tribes). Columns 5–8 find no evidence of a discontinuity in the availability of infrastructure (roads), schools of different types or the distance to an internet cafe. Table B2 runs a similar set of balance tests on election outcomes from the 2011 state and 2009 national election. We find no systematic difference in the vote share of any of the four major parties in either election. Table B3 tests for differences in aggregate job allocations prior to the 2013 local election. If the AITC were able to precisely manipulate the election v Table B1 Balance Tests: Census Outcomes (1) Households (2) Population (3) SC Pop (4) ST Pop (5) Road (6) Primary Schools (7) Private Schools (8) Internet Cafe RD Estimate 61.419 (129.376) 225.536 (542.426) 111.423 (331.370) 73.257 (185.139) -0.000 (0.020) 0.285 (0.568) -0.053 (0.095) -0.001 (0.013) Obs in BW Bandwidth Robust p-val Mean Left of Cutoff Metric 1264 0.594 0.583 3984.278 1-Norm 1299 0.687 0.582 17821.926 1-Norm 1187 0.460 0.660 5068.562 1-Norm 1231 0.527 0.762 1246.921 1-Norm 1295 0.676 0.988 0.148 1-Norm 1305 0.704 0.649 16.674 1-Norm 1299 0.692 0.543 0.529 1-Norm 1214 0.498 0.819 0.065 1-Norm Note: “Robust p-val” gives the p-value after adjusting for bandwidth uncertainty (see Calonico et al., 2014). “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Bandwidths are MSE-optimal. “Mean Left of Cutoff” in Panel C gives the mean of the outcome for observations within one-tenth of the bandwidth on the left of the cutoff. Standard errors are calculated using 3 nearest neighbors. See text for description of each specification. Outcome data comes from the 2011 Indian Census. *p=0.10 **p=0.05 ***p=0.01 share by targeting resources, we might expect significant imbalance on pre-election job allocations. We test for differences in the 2013 and 2012 allocations during the dry season (the first three months of the year). We also shift back the calendar date of the 2014 national election to 2013 and 2012 to test for differences in the 4 weeks before that election (both periods were before the 2013 local election). We find no significant differences. These three tables of balance tests further support the internal validity of the multi-dimensional RD design. vi Table B2 Balance Tests: Prior Elections 2009 National Election 2011 State Election (1) AITC (2) BJP (3) CPM (4) INC (5) AITC (6) BJP (7) CPM (8) INC RD Estimate 0.000 (0.019) -0.005 (0.005) -0.004 (0.018) 0.010 (0.015) 0.018 (0.014) -0.002 (0.003) 0.002 (0.017) 0.012 (0.011) Obs in BW Bandwidth Robust p-val Mean Left of Cutoff Metric 1211 0.470 0.925 0.362 1-Norm 1219 0.477 0.456 0.054 1-Norm 1240 0.508 0.787 0.317 1-Norm 1231 0.495 0.510 0.089 1-Norm 1211 0.467 0.305 0.425 1-Norm 1151 0.380 0.646 0.040 1-Norm 1286 0.597 0.973 0.314 1-Norm 1166 0.402 0.268 0.029 1-Norm Note: “Robust p-val” gives the p-value after adjusting for bandwidth uncertainty (see Calonico et al., 2014). “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Bandwidths are MSE-optimal. “Mean Left of Cutoff” in Panel C gives the mean of the outcome for observations within one-tenth of the bandwidth on the left of the cutoff. Standard errors are calculated using 3 nearest neighbors. See text for description of each specification. Outcome data comes from aggregating polling-station-level data from the 2009 and 2011 elections to the panchayat-level and calculating the vote share for the given party. (AITC=All-India Trinimool Congress, BJP=Bharatya Janata Party, CPM=Communist Party-Marxist, INC=Indian National Congress) *p=0.10 **p=0.05 ***p=0.01 Table B3 Balance Tests: Pre-2013 Election Job Allocations (1) Dry Season 2013 (2) 4 Week Pre-Poll 2013 (3) Dry Season 2012 (4) 4 Week Pre-Poll 2012 RD Estimate 0.422 (0.918) -0.357 (0.299) 0.213 (1.005) -0.152 (0.478) Obs in BW Bandwidth Robust p-val Mean Left of Cutoff Metric 1072 0.397 0.783 6.286 1-Norm 1113 0.463 0.192 1.034 1-Norm 1187 0.633 0.878 7.216 1-Norm 1125 0.484 0.700 2.266 1-Norm Note: “Robust p-val” gives the p-value after adjusting for bandwidth uncertainty (see Calonico et al., 2014). “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Bandwidths are MSE-optimal. “Mean Left of Cutoff” in Panel C gives the mean of the outcome for observations within one-tenth of the bandwidth on the left of the cutoff. Standard errors are calculated using 3 nearest neighbors. See text for description of each specification. Outcome data comes from 2013 and 2012 election data. “4 Week Pre-Poll” refers to the 4 weeks prior to the calendar date of the 2014 election shifted back to 2013 and 2012. *p=0.10 **p=0.05 ***p=0.01 vii B.2 AITC Candidates are not Just Proxying for AITC Supporters (Table 5) An alternative interpretation of Table 5 is that AITC candidates get no specifical treatment. Candidacy is just a proxy for being an AITC supporter, and all AITC supporters get larger payments than non-supporters. To test this hypothesis we calculate the payment made to the median household within each village and test for how the median household’s payment changes when the AITC takes control. We calculate how this RD estimate changes as a function of the AITC’s 2014 vote share. As the AITC’s village-level vote share approaches and exceeds 50 percent, we would expect the median household to be an AITC supporter (because at least half of all villagers voted for the party). If the alternative interpretation is true—that is, if AITC candidates are just readily identifiable AITC supporters—we would expect the size of the RD estimates for the median household to converge to the size for AITC candidates from the same village. We run this test and plot the results in Figure B2. For each test we choose a vote share (0.3, 0.35, . . . 0.6) and restrict to villages where the 2014 AITC vote share was ±0.1 of the center point. We use the number of days of labor throughout the full year of 2014 for this test because the difference in Table 5 is largest for the full-year outcome, and thus the pattern of convergence ought to be the most stark. The figure shows no evidence that the RD estimates for the median household are converging to those of the AITC candidates, even when the village-level vote share is in the range 50—70 percent (the estimate furthest to the right). viii RD Estimate: Impact of AITC Majority on NREGA days, 2014 0 5 10 15 Figure B2 AITC Cand. -5 Median HH .3 .4 .5 2014 AITC Voteshare (Center of 10 pp window) .6 Note: For each test we choose a vote share (0.3, 0.35, . . . 0.6) and restrict to villages where the 2014 AITC vote share was ±0.1 of the center point. The outcome is the NREGA job allocation throughout the full year of 2014. We plot the RD impact on the median job allocation and the job allocation for AITC candidates within those villages. All estimates use the Calonico et al. (2014) method to estimate the bandwidth and include district and constituency fixed-effects (see Section 4.3). The standard errors are clustered by panchayat. ix B.3 Right-Hand Panel of Figure 5 Using Assembly Constituency Fixed-Effects Figure B3 The Contrast Between 2014 and 2016 is Equally Stark Using Assembly Constituency Fixed-Effects Relative AITC Vote Share, Minus FEs -.02 0 .02 .04 ...Closing the Gap in Returns in the Next Election 2016 -.04 2014 -.5 0 1-Norm (Taxi) Distance .5 Note: This figure is exactly the same as the right-hand panel of Figure 5 except vote leans in both years are stripped of assembly constituency fixed-effects rather than parliamentary constituency fixed-effects. x B.4 Right-Hand Panel of Figure 5 Using Raw Vote Shares Figure B4 The Contrast Between 2014 and 2016 is Equally Stark Using Raw Vote Shares .3 .35 Raw AITC Vote Share .4 .45 .5 .55 ...Closing the Gap in Returns in the Next Election -.4 -.2 0 1-Norm (Taxi) Distance .2 .4 Note: This figure is exactly the same as the right-hand panel of Figure 5 except it uses raw vote share as the outcome and does not strip out district and constituency fixed-effects. B.5 Are Close Elections Unusual, and How Might the NATE Differ from the LATE? While a close election design achieves causal inference, it does so only within a set of closely divided elections. This Neighborhood Average Treatment Effect (NATE) may not be informative about the average treatment effect if close elections occur in a relatively small and unusual set of places. A party might use its power differently in a highly contested battleground than in a stronghold. To some degree this problem is less salient for our mutlidimensional design because a more diverse set of outcomes will appear near the cutoff. A directly elected council president who xi barely won is clearly insecure. But a multi-constituency local council where the AITC barely won a majority could comprise only closely won constituencies, or it could comprise dozens of stronghold constituencies alongside one or two pivotal swing constituencies. That may explain why the majority of panchayats are close to the cutoff. The bars in Figure B5 show the share of our sample that lies within a widening set of windows around the cutoff (as measured with the one-norm). Even a very narrow window of 0.05, meaning the absolute majority was won or lost by a cumulative total of only 5 percentage points, accounts for nearly 25 percent of the sample. A window of 0.3 contains over half of the sample. For comparison the optimal bandwidth of the baseline specification in Table 1 is 0.775, implying our main estimates are based on well over half the sample. Though the RD estimate is weighted disproportionately towards the very closest elections, Figure B5 also shows that even a simple comparison of means between AITC and non-AITC panchayats yields similar estimates. We restrict to observations within the window and run a simple ordinary least squares regression of the outcome on a dummy for whether the AITC holds an absolute majority. The left-hand panel shows estimates when the outcome is the number of dry season NREGA days from 2014 to 2016. The estimates show no obvious pattern as the window widens. All lie within each other’s 95 percent confidence intervals, and within the 95 percent confidence interval of the RD estimate. At the widest window of 0.3 the estimate is nearly identical to the RD estimate. The right-hand panel shows a more pronounced positive relationship between the size of the window and the estimate for the 2014 election return, but that is to be expected for a political outcome. Places that elected (rejected) an AITC majority in 2013 by a wider margin will also likely vote for the AITC by a wider (narrower) margin in 2014. Taken together these results are not consistent with the fear that areas right at the cutoff are dramatically different from those within 0.3 of the cutoff. Though this window may seem narrow, it accounts for most of the sample. We can also test more directly whether there are major differences between areas decided by close versus non-close elections. Using the full sample, we estimate a simple OLS regression of several outcomes on a dummy for whether the running variable is “close” to the cutoff (meaning the outcome of the 2013 election lies within a range of 0.05 based on the one-norm metric). Table B4 shows the estimates from such regressions on panchayat-level outcomes from xii the 2011 census, and Table B5 for dry season NREGA labor for several years. Table B4 suggests close elections may occur in places that are slightly further from roads, but otherwise there is no difference of statistical or economic significance. Table B5 suggests closely decided panchayats may receive slightly smaller NREGA allocations, but the result is at best marginally significant in one year. These results cannot prove that randomly assigning non-close areas to AITC control would have impacts similar to those estimated by the RD. But they also do not give obvious reason to expect the NATE is highly unrepresentative. Figure B5 Share of Observations and Differences in Outcomes Between AITC and non-AITC Panchayats as the Window around the Cutoff Widens .75 .05 Note: .1 .15 .2 Window .25 .3 .02 Estimate 0 .25 0 Fraction .5 0 -.5 0 .25 Fraction .5 .5 Estimate 1 .75 1.5 1 .04 2014 AITC Lean 1 Dry Season NREGA Days .05 .1 .15 .2 Window .25 .3 xiii Table B4 Close Elections vs. Non-Close Elections: Census Outcomes (1) Households (2) Population (3) SC Pop (4) ST Pop (5) Road (6) Primary Schools (7) Private Schools (8) Internet Cafe Close GP -44.465 (87.782) -216.295 (379.533) 17.246 (210.839) -71.002 (116.401) -0.031∗∗ (0.013) 0.390 (0.413) -0.076 (0.065) -0.009 (0.008) N Mean Outcome 1852 4080.553 1852 18210.514 1852 4984.143 1852 1344.313 1848 0.172 1852 16.885 1852 0.534 1848 0.076 Note: We estimate a simple OLS regression of the outcome on a dummy for whether the running variable is “close” to the cutoff (meaning the outcome of the 2013 election lies within a range of 0.05 based on the one-norm metric). Outcome data comes from the 2011 Indian Census. “Road” refers to the share of villages within the panchayat that are connected by a “major district road.” The columns that mention schools use as the outcome the number of that type of school in the panchayat. “Internet Cafe” is the share of villages with an internet cafe. *p=0.10 **p=0.05 ***p=0.01 Table B5 Close Elections vs. Non-Close Elections: Dry Season NREGA Days of Labor Close GP N Mean Outcome (1) 2012 (2) 2013 (3) 2014 (4) 2015 -0.548∗ (0.301) -0.049 (0.263) -0.560 (0.389) 0.030 (0.230) 1959 7.919 1959 5.753 1959 11.560 1959 2.949 Note: We estimate a simple OLS regression of the outcome on a dummy for whether the running variable is “close” to the cutoff (meaning the outcome of the 2013 election lies within a range of 0.05 based on the one-norm metric). The outcome in each column is the average per-household NREGA allocation during the dry season of that year. *p=0.10 **p=0.05 ***p=0.01 xiv B.6 Is the Sample of Candidates who Ran Again in 2018 Unusual? Section 7 infers party-switching using a sample of 2013 local candidates who ran again in the 2018 local election. One concern is that this sample is selected in ways that might produce misleading results. This appendix studies how the sample differs from the full set of candidates. We restrict to the subset who were registered in 2013 with one of the four major parties: the AITC, the BJP, the CPIM, and the INC. Within this subset roughly 8 percent are in the sample used in Section 7. Table B6 shows estimates from regressions of a dummy for inclusion in the sample on several political characteristics of the candidate. Candidates who won their election and received a higher share of votes in 2013 are more likely to be in the sample. That is not surprising, as unsuccessful candidates have lower chances of success in the next election. Constituencies that were contested and with more candidates are also more likely to feature repeat candidates. The incumbency advantage of a winner is likely more valuable in a contested seat, as these are less likely to be safe party strongholds. After controlling for the success of the candidate, AITC candidates are less likely to run for re-election (compared to the BJP, the excluded category in the regression). That may be because these candidates are more likely to seek higher office, or because the ruling party’s brand and political machine makes it less reliant on incumbency effects. Candidates from the opposition parties are even less likely to field repeat candidates, perhaps because holding office is less valuable when they control so few local councils. These patterns by themselves are not necessarily a problem for our analysis. Since we argue that the AITC seeks recruits to gain majorities on councils where it didn’t win enough seats, it helps our argument that the sample is mostly candidates who actually won in 2013. What could potentially be problematic is the selection by party, as much of our analysis studies whether average payments differ by party. It would be a problem if, for example, the AITC candidates who don’t run for re-election are paid systematically less than the ones who do. We can test directly for this pattern by regressing 2015 NREGA days of labor on the dummy for whether the candidate is in the sample, and its interaction with the party dummies. Columns 1 and 2 of Table B7 show that although candidates in the sample get more days of labor than those that are not, the effect vanishes after we control for whether the candidate wins xv Table B6 Correlates of Being in the Sample of Candidates Who Run in 2018 In Sample Winner (1) (2) (3) (4) 0.080∗∗∗ (0.002) 0.070∗∗∗ (0.003) 0.031∗∗∗ (0.005) 0.066∗∗∗ (0.003) 0.057∗∗∗ (0.006) 0.008∗∗∗ (0.001) 0.002 (0.006) 0.063∗∗∗ (0.003) 0.071∗∗∗ (0.007) 0.007∗∗∗ (0.001) 0.018∗∗∗ (0.007) -0.008∗∗ (0.004) -0.031∗∗∗ (0.004) -0.017∗∗∗ (0.003) 84082 84079 84079 84079 Vote Share Numer of Candidates Contested Election AITC CPIM INC N Note: We estimate a simple candidate-level OLS regression of a dummy for whether the candidate is in the sample used for Section 7. We restrict to the subset of candidates from one of the four major parties. The BJP is the excluded category (among the four party dummies). We cluster standard errors by race. *p=0.10 **p=0.05 ***p=0.01 xvi Table B7 Candidates in the 2013-2018 Sample Are Not Unusual in their 2015 NREGA payments 2015 NREGA Days In Sample (1) (2) (3) 2.177∗∗∗ -0.440 (0.701) 9.861∗∗∗ (0.368) -2.023 (2.234) 8.820∗∗∗ (0.382) 3.390∗∗∗ (0.684) -1.133∗ (0.686) -0.902 (0.726) 1.605 (2.436) 1.620 (2.621) 0.568 (2.775) 27715 27715 0.365 (0.707) Winner AITC CPIM INC In Sample × AITC In Sample × CPIM In Sample × INC N P-val on In-Sample Vars 27715 Note: We estimate a simple candidate-level OLS regression of days of NREGA labor in 2015 on several predictors including a dummy for whether the candidate is included in the sample used in Section 7. We restrict to the subset of candidates from one of the four major parties. The BJP is the excluded category (among the four party dummies). We cluster standard errors by race. The row “P-val on InSample Vars” shows the p-value from an F-test on the joint significance of “In Sample” and the terms interacted with it. *p=0.10 **p=0.05 ***p=0.01 their election. Column 3 shows that the interactions between the dummy for being in the sample and the dummies for party affiliation are not significant. An F-test for the joint significance of the main effect and the interactions cannot reject the null that all are zero. This suggests there is no systematic difference in payments to candidates in the sample, regardless of the 2013 party of the candidate. xvii C Data Appendix C.1 Raw Sources of Data We rely most heavily on 3 datasets that we constructed by converting unstructured administrative data into structured data files. As described in later sections, we supplement these data with several other sources that were scraped, digitized, or obtained directly from government officials. The most important original dataset is the NREGA job card dataset, which was scraped in late 2018 through early 2019 from the government’s public web portal (https://nrega.nic.in). Figure B6 shows an example of a job card (the names and identification numbers have been replaced in this figure to protect the identity of the household). The parts of the record we use are the job card details, the family details, and the employment given. We also scrape outcomes of the 2013 and 2018 gram panchayat elections from the website of the State Election Commission of West Bengal ( http://www.wbsec.gov.in ). Figure B7 shows an example of the results for a single panchayat. Some panchayats did not report results through this portal and are excluded from our study. Our third major source of data is the official ”Form 20” tally sheet of booth-level vote counts for each parliamentary constituency and assembly constituency, drawn from the website of the Chief Election Officer of West Bengal (http://ceowestbengal.nic.in/). Figure B8 shows an example of a tally sheet from the 2014 national election. We hired a contractor to apply optimal character recognition to convert these results to structured data, and validated the totals using basic consistency checks. The Form 20 sheets for the 2016 state election were too blurry to be read by machine. Instead we hired four contractors to manually enter the data, which was then validated and corrected by an undergraduate research assistant. C.2 Constructing the Dataset Used to Show the Impact of AITC Control on National and State Election Outcomes We construct 3 distance metrics to the frontier of an AITC absolute majority using the 2013 gram panchayat election outcomes. These panchayat-level distance metrics are the running xviii variable for the regression discontinuity design. The next step is to merge the running variable to panchayat-level aggregates of the vote count in the 2014 and 2016 elections. As we were unable to find any record specifying which gram panchayat contained each polling station, we constructed our own crosswalk through the NREGA job card data. Some fraction of households registered for their job cards using their election photo ID card (EPIC). As shown in Figure B6, those households have the EPIC number listed on the job card. We constructed a random sample of 10 epic numbers from each village in each gram panchayat and queried those numbers against an online portal created by the Chief Election Officer of West Bengal to let voters find their polling station for the 2019 election (which, at this stage of the data construction, was in progress). Since most locations used for polling in 2014 were retained for 2019, this dataset gave us a mapping between job card numbers and the names of buildings used by the job card holder. We assigned a polling location to a gram panchayat if the plurality of job cards linked to EPICs registered to vote at that location were also registered under the gram panchayat.25 This gave us a crosswalk between the names of gram panchayats and the names of polling locations. Since the Form 20 tally sheets record only the numerical ID of a polling station, we had to link the station-level vote counts by ID number to the data constructed by Susewind (2016), which contains the ID number and name of each 2014 polling booth. We cleaned these names and consolidated the data to the building-level.26 We fuzzy-matched this dataset by the name of the polling location to the crosswalk constructed above, and hired two native Bengali speakers to independently validate the matches. We matched 2016 names back to 2014 stations based on their numerical identifiers.27 25 26 In the vast majority of cases all EPICs registered to vote at a polling station were linked to a job card from the same panchayat. Some polling stations are actually separate rooms within the same building. Since the crosswalk between panchayats and 2019 stations gives only the polling location, we consolidated 2014 and 2016 vote counts within 27 building. We have found no publicly available official correspondence between 2014 and 2016 stations, nor any description of how station identifiers were assigned. But based on a limited subset of assembly constituencies where we observe the names of the 2016 stations, and on information recovered from old versions of the West Bengal CEO website found on the Internet Archive, we have deduced that in the vast majority of cases a 2014 station reused in 2016 will retain the same ID. In the vast majority of cases, new stations were given names that contain letters or slashes to avoid having to renumber existing stations. Staff affiliated with the Election Commission of India has confirmed that this is the standard practice followed in recent elections. We estimate that this rule xix C.3 Constructing the Dataset Used to Show the Impact of AITC Control on NREGA Allocations We restrict our sample of job cards to the subset living in panchayats linked to polling stations (see previous section). Using the station to panchayat crosswalk we infer the polling date of each panchayat, which lets us calculate the the number of days of labor received by each job card within the period 4 weeks after the date of polling as well as during the dry season (which we define as the first 12 weeks of the calendar year) and each full calendar year. To be precise, we record a job card as having received some number of jobs within a period if it was given a job spell whose start date fell within the period. If a job card did not have a job spell within that period, it was recorded as receiving zero days of labor. We identify the subset of these job cards that are AITC candidates by fuzzy string-matching candidates by name to the full set of NREGA recipients (individuals registered under any job card). We discard cases where the gender or caste group of the recipient is inconsistent with the reservation status of the district being contested by the candidate (which would imply the match is incorrect). We also discard any cases where multiple candidates or NREGA recipients have the same name (as they are functionally indistinguishable). This set of machine-identified matches was then hand verified by two native Bengali speakers working independently, and cases where the two disagreed were adjudicated by the authors. We tag a jobcard as matched to an AITC candidate if any individual registered under the job card was matched. After discarding ambiguous cases, in the full sample we match just over half of AITC candidates to a job card. We calculated the mean days received by all job cards within the panchayat (our measure of held in 95% of assembly constituencies. In the remainder, a few stations are added to the middle of the list and subsequent stations are renumbered accordingly. But even in these cases, we estimate (using the cases where we know the names of 2014 and 2016 stations) that we only assign the 2016 polling station to the wrong panchayat 30% of the time. This is because numerically adjacent stations are generally in the same panchayat, minimizing the impact of the kind of transposition error caused by an insertion and renumbering. Taken together, these estimates imply that in 95 to 97% of cases the 2016 station is assigned to the correct panchayat. We confirm that when we use our matching approach, there is a strong correlation (R-squared of 0.9) between total valid votes cast in 2014 and 2016. We have also verified that Figure 5 is unchanged when we restrict the sample to the ACs that do not appear to have been renumbered. Finally, there is a small number of polling locations created for the 2016 election that were not used in the 2014 election. These stations necessarily were lost because we could not identify their names. The number of new stations was small enough—typically a handful in each assembly constituency—that the resulting measurement error is small. Most importantly, there is no reason to expect the resulting measurement error to change at the discontinuity. xx “aggregate” allocations) as well as by the subset of job cards matched to an AITC candidate and the subset of individuals not recorded as AITC candidates. Though it is possible that some candidates are missed or mismatched, there is no reason to expect the level or sign of the resulting measurement error to systematically change at the discontinuity. We calculate most of the average allocations within demographic group using data recorded on the NREGA job card. The one exception is “Muslim” versus ”Non-Muslim,” which we identify probabilistically using the names of the head of household and the head’s closest male relative. For each job card we construct a list of name tokens (e.g. “Ajay Manohar Shenoy” comprises tokens “Ajay,” “Manohar,” and “Shenoy”). We use the listing sheets from the West Bengal subset of the 2006 Rural Economic Development Survey to create a large sample of tokens where the individual’s religion is known. For each token we calculate the share of individuals in REDS that are Muslim and the share that are non-Muslim. We then estimate a person-level linear probability model based on whether any of the name tokens in the person’s name are “positive for being Muslim,” meaning at least 80 percent of people with that name token are Muslim, or “negative for being Muslim,” which is the converse set. If the predicted probability of “Muslim” is greater than 50 percent, we predict the individual is Muslim. This person-level predictive model, when applied to a hold-out sample in REDS, is highly accurate. Within the set predicted to be Muslim, 98.7 percent are in fact Muslim. Within the set predicted not to be Muslim 94.8 percent are in fact non-Muslim. We then bring the token-level Muslim shares and the predictive procedure into the NREGA job card data to predict whether the job card holder is Muslim (based on whether the head of household’s name is predicted to be Muslim). C.4 Constructing the Village-level Datasets Used to Show Within-Panchayat Outcomes All within-panchayat analyses use a village-level dataset constructed analogously to the two panchayat-level datasets above. We match polling stations to villages using the same method, and identify AITC candidates within each village using the same job card-level matches. The one new variable is the 2009 and 2011 average vote share within the village. We digitize stationlevel data from the Chief Election Officer, and merge the counts to station names scraped from xxi an archived version of the website. We consolidate station-level data to the building-level using a similar method as described above for 2014 stations before fuzzy string-matching to 2014 locations by name (as above, the machine-generated matches are validated by two independent India-based research assistants). Since there is a nontrivial number of stations that do not match (and unlike the 2014 and 2016 vote counts, the 2009 and 2011 vote shares are used on the right-hand side of the regression) we average party-level vote shares within village rather than summing vote counts and subsequently calculating shares. This procedure reduces the risk of putting undue weight on one or two large stations, though in practice the results are almost identical to using the other method. C.5 Constructing the Datasets On Party-Switching We constructed the candidate-level dataset by fuzzy string-matching candidates by name and panchayat from the 2013 data to the 2018 data. We discard matches where the gender and caste of the 2018 candidate is inconsistent with the reservation status of the 2013 candidate.28 We discard matches with a low match probability (below 98.95 percent) or cases where multiple candidates are matched. The final sample includes only candidates who appear in 2013 and are matched to 2018. We machine-match this subset of candidates by name to the job card data using a similar procedure to that outlined above.29 We calculate the total labor allocation to the job card of each matched candidate for each year (restricting the sample only to those who did match to a job card). The final dataset is at the candidate-level (which we combine with the aggregate-level datasets by panchayat). 28 29 Unfortunately we do not know the actual gender and caste of the candidate in the 2013 data, only the reservation status of the seat being contested. For this phase we did not hire native speakers to hand-validate the matches, as we noted that the additional step made almost no difference when applied to the AITC candidates. Nevertheless, for this phase we use only machine-matched candidates (even though we have hand validated matches for the AITC candidates) to ensure there is no systematic measurement error based on political party.