Political Organizations and Political Scope Research Part 2

13 3.1 The State Ruling Party and Elections in West Bengal The Indian state of West Bengal is in eastern India and in 2011 had a population of about 90 million. India is a federation that holds national, state, and local elections. At all levels one full term lasts 5 years, but elections take place on different cycles. During our analysis period from 2013 to 2018, the All India Trinamool Congress (AITC) was in power at the state level in West Bengal. The political party is headed by its founder Mamata Banerjee, who has been the chief minister (the head of the state government) since the 2011 assembly elections when the AITC defeated the incumbent Left Front government that had held power for 34 years. This provided the AITC with access to state financial resources for the first time. In the subsequent local elections of 2013 the AITC won an absolute majority in about 55 percent of panchayats, giving it control over two levels of government in much of the state. With these electoral successes under her belt, Mamata Banerjee and the AITC were widely expected to play the key kingmaker role in the 2014 national elections. The election was expected to yield a hung parliament. Since the AITC was projected to win most of the seats in West Bengal this would give the AITC substantial bargaining power to decide on the next prime minister.4 She made repeated statements that regional parties should join forces to create a Federal Front to provide a third alternative to the two big national parties. Coming on the heels of her election campaign kickoff speech ‘Dilli chalo’ (Let’s go to Delhi), these statements were widely interpreted as indications of her ambitions to play a key role in national politics and to potentially even become the Prime Minister herself (Mitra, 2014; Paul, 2018). The AITC won 34 of West Bengal’s 42 seats in India’s Lower House in the 2014 election, a massive gain relative to the 18 seats won in the previous election. To Mamata’s big disappointment, she was sidelined due to the landslide victory of the Hindu nationalist BJP outside of West Bengal.5 The AITC’s success has little to do with vision, ideology, or appeals to caste. It is primarily a strategic vehicle for political power sustained by centralized coordination.6 Mamata Baner4 “Predictions are that the next Lok Sabha will be another hung one. . . And with twenty-nine MPs, Mamata’s bar- 5 gaining power will not only shoot up, she may even call the shots.” (Mitra, 2014) The Economic Times, ‘Election Results 2014: Despite Winning in West Bengal, Mamata Banerjee Maintains 6 Belligerence, May 17, 2014. Mitra (2014), who has followed Mamata Banerjee as a journalist, writes: ‘However, there is little question that most of [AITC] is made up of political lightweights with little or no vision of their own or, as pointed out above, with little depth of ideological commitment to Mamata’s vision of a Bengal renaissance.’ 14 jee takes a strong interest in decisions about which politicians receive party tickets and is well known to keep a close eye on activities of party members at all levels of the party hierarchy. Multiple elected state politicians have reported that they were unable to make even small changes in their constituencies, such as installing a few urgently needed handpumps, without the knowledge and permission of the party leader (Mitra, 2014). The consequence of this meticulous coordination is that the party has systematically expanded its power. As early as 2014, the journalist Mitra (2014) wrote: For the foreseeable future, there seems to be no escaping the slippery slope to quasi-autocracy in West Bengal. . . it is not voter intimidation or any other coercive power-maintenance strategy we normally associate with autocracies. Instead, Mamata’s regime will maintain power because it will continue winning election after election, just as the Left had done. 3.2 Local Government Structure in West Bengal The lowest elected tier of politicians are those elected to the panchayat council. The panchayat council makes policy decisions that apply within its jurisdiction. Somewhat confusingly, this jurisdiction can contain multiple villages, which is especially common in large states like West Bengal. This is an important feature for our empirical analysis, since we have data on the allocation of welfare benefits both between the jurisdictions of different panchayat councils, as well as between different villages within the jurisdiction of the same panchayat council. We refer to the jurisdiction of a council as panchayat to distinguish it from the villages it contains. In panchayat council elections, the panchayat area is divided into wards. Voters in West Bengal then elect the political candidate in the ward they live in. The council therefore consists of one elected member from each ward.7 The president of the council is not elected directly by voters, but is chosen indirectly in a vote by the council members. Partly because of this Westminster-style arrangement, political candidates in West Bengal campaign with explicit party affiliations (since no one candidate or election will set the direction of the government). While this makes the institutional setup somewhat unusual for local governments within India, it is a common setup in many other countries and is a direct parallel to India’s state and national government. 7 The number of wards depends on the population size. 15 As in other states the council president makes executive decisions, including decisions about which individuals receive NREGA benefits (see next section). But she serves at the pleasure of the majority, which has the formal authority to remove her from office through a resolution by the majority of the councilors.8 A political party that controls the majority of seats in the gram panchayat is therefore insulated from the opposition in a way in which presidents in areas where their own party falls short of the absolute majority are not. Mamata Banerjee has systematically reduced the independence of local politicians in panchayats controlled by the AITC: “‘Earlier, in the CPI(M) regime, the local party leader at the grassroots level was very important. He would decide everything . . . Mamata has reduced the role of the local politician. Today, it is all . . . controlled by her.’ (Paul, 2018) 3.3 The Welfare Program NREGA One of the largest responsibilities of the local government is to implement the welfare programs of the central and state governments in their area. The biggest anti-poverty program with some of the most sought after benefits is the National Rural Employment Guarantee Scheme, typically referred to as NREGA based on the accompanying Act. According to Dey and Sen (2016), NREGA accounts for 80 to 90 percent of local total annual expenditures, even though panchayat councils in West Bengal implement about 25 anti-poverty programs. On paper, NREGA guarantees every rural household up to 100 days of employment on publicworks projects. Households can request work whenever needed throughout the year and are paid the minimum wage. There are no further means tests (Dey et al., 2006; Government of India, 2018; Zimmermann, 2021). The central government intended NREGA to be a flexible safety net for rural households dealing with underemployment, seasonality and unexpected income shocks. In practice, the actual employment provided to households falls substantially short of the demand from households in most states, including in West Bengal. This leads to rationing, and many households report having to wait passively for work to become available (Dutta et al., 2012; Mukhopadhyay et al., 2015).9 8 9 See Section 12, West Bengal Panchayati Raj Act Nevertheless, previous research has shown that despite the shortcomings NREGA helps households better deal 16 This setup gives politicians extensive power in deciding how NREGA employment is allocated between and within panchayats. While the national government pays for most of the scheme, state governments make important decisions on how to allocate NREGA funds within the state. Within a panchayat, the panchayat council and especially the panchayat council president determine how NREGA is implemented. They register households, propose local projects to sub-district and district officials, and assign individuals to work projects. A worker who wants NREGA labor must apply at the council office. These allocations are then submitted to higherlevel officials, who approve the wage payments. To increase transparency about NREGA allocations, panchayat councils are required to keep physical records of muster rolls and to enter all NREGA-related information into a software application called NREGASoft (Government of India, 2013). The administrative data from the application is published in close to real time on a publicly available website dedicated to NREGA.1011 We use information scraped from this website for our analysis. Is the data credible enough to study the potential misallocation of NREGA resources? By 2013, any payment for NREGA labor would only be approved if registered in the online system. We can thus reasonably assume that any diversion of funds or self-dealing recorded on the website captures an actual payment made to the bank account or post office account of the named beneficiary. Though it is possible that NREGA funds may be diverted through other aspects of the program (notably through payments for materials), any payment for labor during our sample period must go through the online system. By law, expenditures on labor must account for the majority of total expenditures, ensuring that we observe the bulk of any corruption or diversion of resources. That said, it is possible that what we observe is an underestimate of favoritism in the allocation of benefits. By contrast, it is unlikely that we overestimate it, as the ruling party is unlikely to publish self-dealing in the online system while hiding legitimate payments in the with shocks, and NREGA employment is typically much higher during periods such as the agricultural off-season when households in rural areas have few alternative job opportunities. See e.g. Imbert and Papp (2015) and 10 11 Zimmermann (2021). https://nrega.nic.in To cut down on corruption, NREGA profiles are linked to biometric markers through a national identification (‘Aadhar’) number, and the central government directly transfers wages for completed work into beneficiaries’ bank accounts. Muralidharan et al. (2016) find that these features have substantially improved targeting and overall household benefits, plausibly by making it much more difficult to impersonate beneficiaries or intercept the benefits. 17 less transparent parts of the system. Expenditures on social welfare schemes increased once Mamata Banerjee came to power at the state level (Paul, 2018). Based on interviews and her own observations, Mitra (2014) describes Mamata Banerjee’s approach to social welfare programs as piecemeal and extremely pragmatic, with little interest in truly using programs like NREGA to reduce poverty. Mamata Banerjee has claimed credit for her implementation of NREGA as one of her achievements, and while actual work is done at NREGA worksites, this is used to maximize the employment opportunities and not the usefulness of projects for local development.12 Paul (2018) believes that rural welfare schemes are a key explanation for the AITC’s continued success despite multiple high-profile corruption scandals. 4 4.1 Research Design Data We scrape roughly 11 million West Bengali NREGA job card profiles from the official government portal (http://nrega.nic.in). Each profile contains the name of the household head and all household members registered under the job card, as well as the election photo ID card number of the registrant; panchayat and village of the job card holder; and the project name, start date, days of labor, and total payment for each job spell. The full sample amounts to roughly 300 million job spells. We merge these job cards to data on the outcome of each local council ward race from the 2013 local elections. These data are scraped from the website of the State Election Commission of West Bengal. Each record gives the party, ward, and vote returns of each candidate (as well as the candidate’s name, caste group, and gender). We supplement these data with information collected from district offices on the names of panchayat council officers, which let us identify which elected panchayat council member is the council president. We digitize PDFs of polling station-level returns from the 2014 national election downloaded 12 Mitra (2014) quotes an AITC activist saying “. . . hundreds of people are being employed to do a job that perhaps requires about half the number. . . you will find countless men and women at the work site shovelling soil and carrying it from one spot to another. The next day you will find them shovelling it up from that spot and carrying it back to the original spot.’ 18 from the Chief Election Office of West Bengal. The station identifiers are merged to station names and geocoordinates using data from Susewind (2016). We identify the polling stations within each panchayat by querying election photo ID cards from the NREGA data against the Chief Election Office’s website, which we use to construct a station-to-panchayat crosswalk. Finally, we merge data from the 2011 Census to the panchayat-level election and NREGA data. We build a crosswalk between Census villages and panchayats using data scraped from West Bengal’s Panchayat Raj Department. We then aggregate the village-level census data by panchayat. We describe the precise steps for merging and aggregation, as well as the sources of the underlying data, in greater detail in Appendix C. 4.2 Defining the Running Variable Unlike most Indian states, West Bengal uses a Westminster system to govern its panchayats. The panchayat is divided into wards that each elect a council member, and the council then chooses a president. This system is a strength of our context because it lets us distinguish the impact of holding an absolute majority from that of leading a coalition as the largest party. But it also raises a challenge because there is no single vote count that determines control of the council. Since we cannot use a simple univariate regression discontinuity design, we instead define a multidimensional running variable (Feigenbaum et al., 2017). This approach is most easily understood through a simple example. Suppose a panchayat has 5 wards, where the AITC candidate wins in Ward 1 by a margin of 10 percent while losing in Wards 2, 3, 4, and 5 by margins of 5, 10, 15, and 20. Figure 1 illustrates this scenario. The “closest” counterfactual outcome where the AITC would have won an absolute majority is one where, holding the results in all other wards unchanged, it barely won in Wards 2 and 3. Since it lost those wards by 5 and 10 percent, one measure of the distance would be |5| + |10| = 15, which is the 1-norm. Since the AITC lost, the actual value of the running variable would be −15 (putting p it to the left of the cutoff). An alternative distance measure would be − |5|2 + |10|2 ≈ −11.18, the Euclidean or 2-norm. There is a continuum of alternative measures (|5|k + |10|K )1/K for each choice of K, but all shrink to zero as the outcome “approaches” an AITC absolute majority. More generally, in any panchayat of N wards where the AITC did not win an absolute ma- 19 Figure 1 An Example of Calculating Distance Metrics Generalized distance: 𝑣2 One-Norm / Taxi Metric: 𝑣2 + 𝑣3 Two-Norm / Euclidean: 𝑣22 +𝑣32 1 + 𝑣3 𝐾 𝐾 max(|𝑣2 |, 𝑣3 ) = |𝑣3 | 𝑣2 Inf-Norm / Minimal Swing: 5 𝐾 4 3 2 1 𝑣3 AITC Margin of victory/defeat in each ward 0 Note: The 1-norm, 2-norm, and infinity-norm in a case where the ruling party (the AITC) wins 1 seat and loses 4. Counterfactually the AITC would have had to have won seats 2 and 3 to win the majority. Each measure of “distance” to the cutoff of an AITC majority plugs the losing margins of those two seats into the expression for the norm. jority, we identify the M additional seats the AITC would have needed to have won to gain an absolute majority. Let v1 , v2 , . . . , vM be the vote margins of the AITC candidate in the M seats where it came closest to winning. Define the K-norm as Dk = M hX |vm |K i1/K m=1 Our main specification sets K = 1, but we also show results for K = 2 and K = ∞ (all of which are depicted in Figure 1). The running variable would be signed as negative, since negative values imply that the AITC does not control the panchayat. We follow an analogous procedure for panchayats where the AITC did win an absolute majority, except we identify the M seats it would have had to have lost to not have the absolute majority.13 These outcomes where the AITC won an absolute majority are signed to be positive. Our preferred metric throughout the paper is the 1-norm, which can be interpreted as the total number of percentage points of votes the AITC would need to ‘buy’ across all wards to get control of the panchayat council. This metric is more intuitive than the Euclidean norm and less 13 If the panchayat council has an even number of seats we define an absolute majority as 50% + 1 seats. 20 noisy than the infinity norm, although our results are not sensitive to the choice of norm. We estimate changes at the cutoff with several specifications. 4.2.1 Tests for Manipulation of the Running Variable In Appendix B.1 we report several tests for manipulation of the running variable. We first test for whether there is a discontinuity in the density of the one-norm at the margin where the AITC wins an absolute majority. Any such discontinuity 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. Manipulation of the vote count is implausible because the Election Commission of India is a non-partisan bureau that is widely respected and considered free from corruption. Applying the test of McCrary (2008) shows no evidence of any discontinuity. Appendix B.1 also tests for discontinuities in panchayat-level outcomes measured in the 2011 Census. Since the census was taken before the 2013 election, a discontinuity would suggest there was manipulation. We find no evidence of discontinuities in population, caste composition, or the presence of various public goods. We also find no discontinuities in political outcomes in prior elections, or pre-election NREGA allocations. 4.3 4.3.1 Specifications Basic RD Our main specification estimates the discontinuity in outcomes using a local linear regression of the form Ykt = φ0 + φ1 dk + φ2 dk Mk + βMk + Xk γ + εkt for k such that |dk | < h (9) where Ykt is the outcome for a panchayat k in year t, dk is the running variable (the 1-norm in most specifications), and Mk a dummy for whether the AITC holds the absolute majority on the panchayat council. The coefficient β gives the regression discontinuity estimate. We estimate the bandwidth h using the optimal bandwidth proposed by Calonico et al. (2014). We weight observations using a triangular kernel. For specifications that pool observations across years 21 we cluster standard errors by panchayat. In those where we have only a single observation per panchayat (e.g. vote share in the 2014 election) we use the 3-nearest-neighbor estimator for standard errors (which is more conservative than the usual heteroskedasticity-robust standard error). In some specifications we control for additional variables X, typically fixed effects for the revenue district and parliamentary constituency. The Calonico et al. (2014) estimator has trouble calculating an optimal bandwidth while controlling for these fixed effects. We instead use the optimal bandwidth calculated for the analogous regression with no fixed effects and use that bandwidth in the other specifications. 4.3.2 Difference-in-Discontinuities When testing for mechanisms we also estimate a “difference-in-discontinuity” specification that measures how the difference in allocations across areas within a panchayat changes when the AITC gains control of the panchayat. Panchayats are divided into administrative units called “villages.” Let v index a village within panchayat k. We estimate a difference-in-discontinuity specification Ykvt = αk + φ1 dk skv + φ2 dk Mk skv + βMk skv + εkvt for k such that |dk | < h (10) where is αk is a panchayat fixed-effect. The main effect of the running variable and the dummy for AITC control are absorbed into the panchayat fixed-effect. What remains is the interaction of these terms with skv , which is either a proxy for a village’s historical support for the AITC or the actual vote share of the AITC candidate in the 2014 national election. In some specifications we will test for whether skv is a proxy for some other variable by controlling for the differencein-discontinuity induced by some control variable Xk : (11) Ykvt = αk + φ1 dk skv + φ2 dk Mk skv + βMk skv + φ3 dk Xkv + φ4 dk Mk Xkv + φ5 Mk Xkv + εkvt for k such that |dk | < h (12) We cluster standard errors by panchayat. As we are unaware of any method to estimate the optimal bandwidth in a difference-in-discontinuities specification, we instead apply the method 22 of Calonico et al. (2014) to the village-level analog of Equation 9.14 4.3.3 Biggest Party We disentangle the impact of being the biggest party versus holding the absolute majority by constructing a measure of “distance” to the cutoff where the AITC becomes the biggest party. There are many complications to defining this measure. The number of seats needed for an absolute majority is fixed given the total seats on the council. But the number of seats needed to be the “biggest party” depends on the number of seats won by the other parties, and the distance depends which party is counterfactually losing votes to the AITC.15 Given the ambiguities we define only the infinity-norm. We incrementally switch seats from the AITC to the runner-up in the race until the AITC is no longer the biggest party. We then measure the largest margin among the set of seats that were hypothetically switched. Even given this running variable we cannot estimate a regression discontinuity because in most cases when the AITC switches to being the biggest party it also gains the absolute majority. We cannot simply drop the cases where the AITC has the absolute majority because that would create sample selection at the discontinuity. To disentangle the effects we instead estimate ordinary least squares regressions of each outcome on a dummy for whether the AITC is the biggest party while controlling for whether it holds the absolute majority. To avoid comparing cases that are too dissimilar we restrict attention to a window of observations where the distance to “biggest party” is small. To the extent that observations are dissimilar it will bias estimates upwards (especially those on vote returns), which cuts against our hypothesis. Let d˜k be the distance to the number of seats where the AITC becomes the biggest party, and M̃k be a dummy for whether the AITC is the biggest party in the panchayat council. As before, Mk 14 15 Likewise, we are unaware of any method to estimate bandwidth-robust p-values for a difference-indiscontinuities. Take a simple example where the AITC wins 5 seats, the CPM wins 4 seats, and the BJP wins 3. It may seem that the “distance” to minor party status is just the vote share of the 1 seat where the AITC won the fewest votes. But that assumes the runner-up in that contest was the CPM. If the runner-up was the BJP the assumption does not hold. 23 is a dummy for whether the AITC also holds the absolute majority. We estimate the regression: Yk = φ0 + β 0 M̃k + βMk + εk for k such that |d˜k | < h (13) We estimate this regression for several choices of h to confirm that the results are robust. 5 5.1 Main Results State Government Allocates Disproportionate Funds to Co-Partisan Panchayats The left-hand panel of Figure 2 shows the regression discontinuity with the 1-norm as the running variable and the NREGA job allocation of the average household as the outcome, pooling outcomes across years t = 2014, 2015, 2016. The figure shows allocations during the dry season, the first three months of the year, when the demand for NREGA jobs is at its peak. When the 1norm switches from negative to positive the AITC switches to holding an absolute majority in the panchayat council after the 2013 local election. The figure implies that the average household in an AITC-controlled panchayat receives roughly 1 more day of NREGA labor (9 days versus 8 days—see Column 1 of Table 1). Table 1 confirms that this result holds across several specifications. Column 1, which is the same specification as Figure 2, estimates (9) without controls. Column 2 shows that the coefficient is largely unchanged by adding district and constituency fixed effects. Columns 3–4 show that the estimates are almost identical when the running variable is the 2-norm (Euclidean distance) and the infinity-norm. While columns 1–4 of Table 1 pool the post-election years 2014 to 2016, columns 5–7 report the specification from Column 2 separately by year. The premium for AITC-controlled panchayats is biggest in 2014, the year of the national election, but still systematically higher in 2015 (when there is no election). The effect is noisier and only marginally significant in 2016, possibly for the reasons we explore in Section 7. The dry season results are especially informative both because the need for poverty relief is high and because the 2014 dry season was immediately before the national election. But Column 8 shows similar impacts on average allocations for the full year. 24 Figure 2 AITC-Controlled Panchayats Get More NREGA Labor and Subsequently Return More Votes for the AITC in the 2014 National Election Relative AITC Vote Share, 2014 Election -.04 7 -.02 Relative AITC Vote Share 0 .02 Average NREGA Days of Labor 8 9 10 .04 11 .06 Average NREGA Days in Dry Season, 2014-2016 -.75 -.5 -.25 0 .25 .5 Distance to Absolute Majority (1-Norm Metric) .75 -.5 -.25 0 .25 Distance to Absolute Majority (1-Norm Metric) .5 Note: Outcome is the panchayat-level average NREGA household allocation, pooling data from 2014-2016 (analogous to Column 1 of Table 1). Each dot shows the average of the outcome within a bin of width 6.25 percentage points in the taxi metric. The observations are restricted to the optimal bandwidth (rounded to the nearest bin). The linear predictions are generated using a triangular kernel. 5.2 Panchyats Controlled by Ruling Party Return More Votes for its National Candidates We estimate Equation 9 to test whether AITC-controlled panchayats also returned more votes for the party’s candidate in the 2014 national election. We calculate the average vote share of the AITC candidate within each panchayat and define the “AITC Lean” of a panchayat as the share of AITC votes within the panchayat minus the overall share received by the candidate in the entire parliamentary constituency. We estimate Equation 9 on the vote lean using several specifications. The right-hand panel of Figure 2 visualizes the impact of AITC control of the local council on the AITC lean. Control of the council yields roughly 2 percentage points more votes for the AITC’s parliamentary candidate during the 2014 election. Table 2 confirms these results using several specifications. Column 1 is the same specification as the figure. Column 2 shows that this estimate is largely unchanged when we control for district and constituency fixed effects. 25 Table 1 Panchayats Under AITC Control Receive Larger Per-Household NREGA Allocations Dry Season Full Year (1) All (2) All (3) All (4) All (5) 2014 (6) 2015 (7) 2016 (8) All RD Estimate 0.979∗∗∗ (0.354) 1.068∗∗∗ (0.292) 1.111∗∗∗ (0.348) 1.181∗∗∗ (0.387) 1.482∗∗∗ (0.464) 0.735∗∗∗ (0.278) 0.842∗ (0.461) 1.983∗∗∗ (0.604) Obs in BW Clusters in BW Control Mean Bandwidth Robust p-val Metric District FEs Constituency FEs 4200 1400 7.83 0.775 0.008 1-Norm 4200 1400 7.83 0.775 0.015 1-Norm X X 3963 1321 7.56 0.282 0.031 2-Norm X X 3843 1281 7.73 0.169 0.033 Inf-Norm X X 1326 1326 10.61 0.583 0.103 1-Norm X X 1307 1307 2.73 0.549 0.011 1-Norm X X 1294 1294 10.55 0.521 0.173 1-Norm X X 4323 1441 20.05 0.898 0.028 1-Norm X X Note: The table shows RD estimates of Equation 9. The outcome in all columns is the average per-household days of labor, where the average is over all job cards in the panchayat. Columns 1—7 measure the average over the dry season (first three months of the year) while Column 8 averages over the entire year. Columns 1–4 and Column 8 pool observations across the years 2014–2016, while columns 5—7 restrict to a single year. Bandwidths are calculated using the method of Calonico et al. (2014) on the equivalent specification without fixed-effects (see text for details). “Robust p-val” gives the p-value after adjusting for bandwidth uncertainty. “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Standard errors are clustered within panchayat. See text for description of each specification. *p=0.10 **p=0.05 ***p=0.01 Column 3 shows that using the raw AITC share (without netting out the party’s overall share in the constituency) yields nearly identical coefficients. Column 4 shows the results are unchanged when the vote share is calculated after discarding polling stations within the panchayat where some job card holders are registered to vote in a different parliamentary constituency. Columns 5 and 6 show that the estimates are almost identical when the running variable is the 2-norm and the infinity-norm. 5.3 The Results are Driven by Co-Partisan Control, Not Just Co-Partisan Officials Co-partisan areas could receive higher NREGA allocations for any number of reasons. There could be efficiencies that arise from partisan alignment because there is less asymmetric information, or because state officials can more easily force co-partisan council presidents to put in the effort of proposing NREGA projects. It is also possible that there is a desire to elevate intrafactional allies, as proposed by Persico et al. (2011). Separately, it is possible that co-partisan areas return votes at a higher rate solely because the council president endorses her party’s na- 26 Table 2 Impact on 2014 National Election (1) AITC Lean (2) AITC Lean (3) Raw AITC Share (4) Hom. AITC Lean (5) AITC Lean (6) AITC Lean RD Estimate 0.017∗∗∗ (0.006) 0.021∗∗∗ (0.005) 0.020∗∗∗ (0.005) 0.021∗∗∗ (0.005) 0.018∗∗∗ (0.005) 0.018∗∗∗ (0.006) Obs in BW Bandwidth Robust p-val Metric District FEs Constituency FEs 1263 0.480 0.010 1-Norm 1263 0.480 0.002 1-Norm X X 1129 0.329 0.002 1-Norm X X 1265 0.483 0.003 1-Norm X X 1282 0.261 0.007 2-Norm X X 1128 0.127 0.031 Inf-Norm X X 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. Standard errors are calculated using 3 nearest neighbors (as each panchayat is observed only once, clustering is unnecessary). See text for description of each specification. *p=0.10 **p=0.05 ***p=0.01 tional candidate. What distinguishes these models from ours is that the other models predict impacts driven solely by whether the council president is from the ruling party. Our model requires that the party have unchecked control over the targeting of NREGA benefits. The local institutions in West Bengal provide a crucial test between the competing models because it is possible for the AITC to hold the presidency without having unchecked control. As in many parliamentary systems, any would-be president must be backed by a majority of elected council members. But since most panchayats have multiple competing parties, in some panchayat councils the AITC wins the most seats but falls short of an absolute majority, forcing it to form a coalition. By convention the largest party almost always has the first chance to form a coalition. Hence the party of the president will generally be from the AITC regardless of whether it has a majority or is merely the largest party. The difference is that when the AITC holds an absolute majority there is no institutional check, whereas when it is only the largest party its coalition partner can threaten to bring down the government. This threat can prevent the AITC from targeting NREGA jobs for political gain. Table 3 estimates Equation 13 within three different windows around the cutoff where the AITC switches from being a minor party to being the biggest party (where the window is defined using the minimal swing distance explained in Section 4.3). Columns 1—3 show that, across a 27 Table 3 Impact of Absolute Majority Versus Coalition AITC President Biggest Party Majority Obs. Control Mean Window NREGA Days Vote Lean (1) (2) (3) (4) (5) (6) (7) (8) (9) 0.452∗∗∗ (0.044) 0.346∗∗∗ (0.040) 0.431∗∗∗ (0.047) 0.340∗∗∗ (0.041) 0.315∗∗∗ (0.058) 0.391∗∗∗ (0.049) -0.119 (0.564) 2.367∗∗∗ (0.569) -0.807 (0.589) 2.612∗∗∗ (0.597) -1.155 (0.703) 2.514∗∗∗ (0.738) 0.014∗∗ (0.006) 0.023∗∗∗ (0.006) 0.008 (0.007) 0.024∗∗∗ (0.006) -0.002 (0.007) 0.026∗∗∗ (0.007) 1026 0.192 0.15 845 0.216 0.10 496 0.278 0.05 1214 9.747 0.15 989 10.144 0.10 588 10.209 0.05 1214 -0.014 0.15 989 -0.010 0.10 588 -0.002 0.05 Note: Each column estimates Equation 13 on a different outcome and a different window h. “Biggest Party” is a binary variable equal to one when the AITC has strictly more seats than any other party. “Majority” likewise equals one when the AITC holds an absolute majority of seats. The variables are not mutually exclusive. If “Biggest Party” equals zero then “Majority” must equal zero. Thus the coefficient on “Largest Party” estimates the impact of being the biggest party without an absolute majority, and “Absolute Majority” estimates the additional impact of holding the absolute majority. “AITC President” is a binary variable equal to one when the council president is a member of the AITC. “NREGA Days” gives the average per household NREGA days during the 2014 dry season. “Vote Lean” is as defined in Table 2. *p=0.10 **p=0.05 ***p=0.01 range of windows, becoming the largest party raises the probability of an AITC council president by 30 to 45 percent. The point estimate implies a doubling or tripling over the baseline where the AITC is a minor party. The additional impact of also having an absolute majority is of a similar magnitude. Together these coefficients suggest that even when the AITC does not hold the absolute majority, simply being the biggest party more than doubles the chance of an AITC president. But Columns 4—6 show no evidence of an accompanying increase in the aggregate NREGA allocation. Only when the AITC has the majority does the panchayat receive more days of NREGA labor. Columns 7—9 show a similar pattern for the AITC lean of the 2014 election returns. Only at the widest window is there any evidence that being the biggest party increases votes. The effect vanishes as the window narrows, suggesting it is caused by selection bias. The effect of holding the majority is by contrast consistently positive and significant. In summary, panchayats where the AITC is the biggest party are substantially more likely to have an AITC president but receive no more NREGA jobs and return no more votes for the AITC. Only panchayats where the party holds the absolute majority, and thus faces no external check on its control of the council, receive extra NREGA jobs and return additional votes. These results 28 are inconsistent with models of co-partisan efficiency and simple partisan favoritism, but they are consistent with our model of political complementarities. 6 6.1 Mechanism There is Political Targeting Within the Panchayat Section 5 shows that the state targets NREGA allocations to co-partisan panchayats, but not in the way predicted by simple models of partisan favoritism. A further distinction of our model is that the state over-allocates funds to co-partisan local councils because it knows they will use the funds for political targeting within the panchayat. Is there evidence of within-panchayat political targeting? For obvious reasons we cannot observe individual votes, but we can measure votes by ‘village,’ a sub-panchayat administrative unit with a median size of 200 households. We test whether within a panchayat, villages that have historically voted for the AITC are disproportionately rewarded when it controls the panchayat council. We estimate Equation 10, the difference-indiscontinuities. We interact the dummy for AITC control with three measures of past political support: the average 2013 share of AITC candidates in the village, the AITC vote share in the 2011 election, and the average of these two measures. Columns 1—5 of Table 4.a show the difference-in-discontinuity estimates for allocations during the dry season. The first three columns imply that when the AITC controls the local council, pro-AITC villages (based on all three measures of historical support) get systematically higher NREGA allocations in the 2014 dry season. Column 4 shows similar albeit weaker patterns for the full 2014-to-2016 stretch. Targeting in later years is thus no longer as strongly linked to pre2014 support (either because there is less targeting in non-election years, or because they have the 2014 vote share as a more current guide to political support). Column 5 shows that the result holds even after controlling for whether there is an elected AITC official living in the village, suggesting historical support is not just a proxy for electing an AITC candidate. Column 6 shows similar effects for the full year of allocations. Could these measures of past support simply be proxying for demographic characteristics 29 that are correlated with objective need for poverty relief? We test this alternative interpretation by calculating for each panchayat the average days of labor per household within each of several demographic characteristics: Scheduled Caste/Tribe versus Other Caste, Muslim versus Non-Muslim, and Below versus Above Poverty Line. We take the difference between the average within each category and the overall average within the panchayat. This difference represents the excess allocation to this group. If AITC-controlled panchayats are simply targeting low income households, we would expect the excess allocation for Scheduled Castes and Tribes, Muslims, and Below-Poverty-Line households to increase at the cutoff. Table 4.b shows that this is not the case. There is no evidence of targeting to disadvantaged demographics. 6.2 AITC Candidates Receive Excess NREGA Payments Trilateral coordination requires local officials to exert effort on behalf of a national candidate who has no official authority over them. Even a successful candidate has relatively little personal power to reward that effort. The organization must somehow sustain local cooperation. One possible means is through direct payments through NREGA itself. Since we observe beneficiary names in the NREGA administrative dataset, we can match the candidates standing for election in the 2013 election to their NREGA job card profiles. We use this measure to test whether AITC local candidates receive more generous payouts when the AITC controls the council. We estimate Equation 9 on the average days of labor within the subset of households of AITC candidates. For comparison we estimate the same equation for all other households in the panchayat. Columns 1—4 of Table 5 show that while both sets of households receive higher dry season NREGA allocations under AITC control, the point estimate for AITC candidates is twice as large. Columns 5—8 show that the difference is even bigger when we measure allocations over the entire year. All specifications show bigger RD estimates for AITC candidates than other households. We test for whether the difference is statistically significant using a nonparametric clustered bootstrap. The p-value is roughly 0.06 for dry season allocations and less than 0.01 for allocations over the entire year. One natural objection to this test is that it may reflect personal corruption by the individuals in power rather than a payment by the party. Though councilors other than the president have 30 Table 4 Targeting a) There is Targeting by Political Affiliation... Dry Season (1) 2014 Majority X GPSHARE (2) 2014 Full Year (3) 2014 (4) 2014-2016 (5) 2014 (6) 2014 5.022∗∗∗ (1.938) 1.851∗ (1.008) 4.312∗∗ (1.990) 0.484 (0.461) 9.382∗∗∗ (3.581) 11107 1323 0.588 1-Norm X 34986 1398 0.766 1-Norm X 11107 1323 0.588 1-Norm X 10929 1303 0.550 1-Norm X 4.953∗∗ (2.446) 5.594∗ (3.144) Majority X AITC2011 Majority X AVERAGE Majority X AITC REP Obs in BW Clusters Bandwidth Metric Panchayat FEs 6637 1325 0.789 1-Norm X 7696 1160 0.572 1-Norm X b) ...Not by Proxies for Poverty or Need Caste Group Religion Below/Above Poverty Line (1) SC (2) ST (3) Other (4) Muslim (5) Non-Muslim (6) Below (7) Above RD Estimate 0.156 (0.285) -0.275 (0.664) 0.096 (0.128) 0.331 (0.485) 0.099 (0.062) -0.594 (0.411) -0.023 (0.057) Obs in BW Clusters in BW Bandwidth Robust p-val Metric District FEs Constituency FEs 1347 1347 0.64 0.536 1-Norm X X 1108 1108 0.47 0.732 1-Norm X X 1308 1308 0.56 0.850 1-Norm X X 1293 1293 0.56 0.368 1-Norm X X 1297 1297 0.53 0.406 1-Norm X X 1307 1307 0.65 0.133 1-Norm X X 1424 1424 0.84 0.996 1-Norm X X Note: a) The table shows estimates of Equation 10. “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Bandwidths are calculated using the method outlined in Section 4.3. The unit of observation is a village-year, and standard errors are clustered by panchayat. “AITC2011” refers to AITC vote share in 2011 state election, “GPSHARE” to the average vote share of all resident AITC candidates in the 2013 local election, and “AVERAGE” to the average of these two measures (if one of them is not observed for a village, “AVERAGE” simply equals the other measure— hence that specification has the most observations). “AITC REP” is a dummy for whether the village is home to an elected AITC council member. b) We estimate Equation 9 on the excess days of labor allocated to each demographic group (“SC” and “ST” abbreviate Scheduled Caste and Scheduled Tribe). The outcome is the average days of labor per household within the sub-population minus the overall average for the panchayat. All averages are for the 2014 dry season. *p=0.10 **p=0.05 ***p=0.01 31 Table 5 AITC Candidates Receive Excess Payments Dry Season Full Year (1) AITC Cand. (2) Other HH (3) AITC Cand. (4) Other HH (5) AITC Cand. (6) Other HH (7) AITC Cand. (8) Other HH RD Estimate 2.046∗∗∗ (0.602) 0.994∗∗∗ (0.343) 2.094∗∗∗ (0.538) 1.086∗∗∗ (0.283) 4.917∗∗∗ (1.401) 1.792∗∗ (0.783) 5.317∗∗∗ (1.220) 2.025∗∗∗ (0.628) Obs in BW Clusters in BW Bandwidth Robust p-val Metric District FEs Constituency FEs P-val: Difference 3798 1266 0.499 0.0028 1-Norm 4293 1431 0.917 0.0053 1-Norm 3798 1266 0.499 0.0444 1-Norm X X 4293 1431 0.917 0.0059 1-Norm X X 0.0602 3990 1330 0.624 0.0011 1-Norm 4140 1380 0.751 0.0335 1-Norm 3990 1330 0.624 0.0026 1-Norm X X 4140 1380 0.751 0.0616 1-Norm X X 0.0076 0.0604 0.0054 Note: We estimate Equation 9 for average payments to households of AITC candidates and all other households. “Metric” gives the distance metric (to AITC absolute majority) used as the running variable. Bandwidths are calculated using the Calonico et al. (2014) method. Standard errors are clustered within panchayat. “P-val: Difference” tests for whether the coefficients in adjacent columns are significantly different (the value in Column 2, for example, tests for the difference between RD Estimates in Columns 1 and 2). The test is run using a clustered bootstrap that nonparametrically accounts for the uncertainty in the choice in bandwidth. *p=0.10 **p=0.05 ***p=0.01 no formal authority over NREGA allocations, one can imagine a deal where the president buys off the other members of the council to keep everyone complicit in the misdirection of funds. We test this possibility by splitting the AITC candidates into those that won their seat on the council and those that did not, and estimating the discontinuity separately for each. Table 6 shows that the increase in benefits is very similar for winners and losers, both for the dry season and the full year. Since losers have no direct control over the program, these excess payments could only happen with the complicity of the party apparatus. That is consistent with the idea that the payments are in return for maintaining, organizing, and delivering votes rather than serving in office. In unreported results we also find that the AITC council president’s payments on average are roughly equal to those of a regular AITC council member, reinforcing that the payments are not linked to their position on the council. One concern given these results is that the AITC rewards all of its voters, and being an AITC candidate is just a proxy for being an AITC voter. In Appendix B.2 we show that even in villages where more than 50 percent of people voted for the AITC, the median household does not receive a premium under AITC control anywhere close to that enjoyed by AITC candidates.