Empirical Strategy: A Variant of the Regression-discontinuity Design

In Karnataka, Rajasthan, and Bihar, as in other Indian states, council presidencies are reserved for SCs and STs through a procedure governed by state electoral regulations and implemented by district-level bureaucrats for each subdistrict under their jurisdiction.Footnote ¹¹ First, bureaucrats use census data on group populations or population proportions to determine the total number of council presidencies that must be allocated quotas in any electoral term. For example, if 25% of the citizens in a given subdistrict are from the Scheduled Castes, then 25% of the councils in that subdistrict must have their presidencies reserved for SCs.

The following procedure is then used to allocate quotas to particular councils, across different electoral terms. First, the bureaucrat lists the council constituencies in each subdistrict in descending order, typically by the size of the SC population as measured in the previous census. In the first council elections after the passage of the 73rd amendment in 1993, the relevant bureaucrats then reserve the presidencies of the required proportion of councils appearing at the top of the list. Thus, in this example, the presidencies of the top 25% of councils on the list would be reserved for SCs in the inaugural elections (say, in 1995). The bureaucrats then work down this list in the next elections (say, in 2000), rotating reservation of the presidencies to the next 25% of councils on the list. This rotation continues until the bottom of the list has been reached and all councils have been assigned SC quotas for the presidency in some election since 1993. The assignment of quotas then rotates back up to the top of the list.Footnote ¹²

Close variants of this procedure are used across different Indian states. In Karnataka, for instance, bureaucrats rank village councils in descending order by the number of council members’ seats reserved for SCs or STs (which is in turn a proxy for the SC or ST population proportion). At the time we constructed our initial study design in Karnataka, in December 2008, we lacked data on SC members’ seats but had data on quotas for council presidencies in 2007 and census data on group proportions on which the number of SC members’ seats is based. By sorting councils in each subdistrict in descending order by proportion of the population that is SC (or ST), and using our data on reservation of the presidency, we could therefore find the lower population proportion bound between councils with reserved and unreserved presidencies.Footnote ¹³ In Rajasthan and Bihar, we also used data on SC and ST proportions, which allowed us to mimic directly the procedure used by bureaucrats.

Our empirical strategy took advantage of the fact that, in any given electoral term, village councils at the bottom of the set that receive a quota (say, the councils with the lowest SC population proportions, among the first 25% on the ranked list) are on average plausibly indistinguishable from councils at the top of the next 25%—save for the presence or absence of a quota. We thus constructed our study group by selecting pairs of councils at the bottom and the top of these respective sets, in each of our selected subdistricts. This idea is similar to standard regression-discontinuity designs, in which a pretreatment covariate such as an exam score is used to sort students into treatment and control groups (Thistlewaite and Campbell 1960), with the difference here being that the relevant threshold value of the assignment covariate (the SC population proportion) is specific to each subdistrict and varies across elections, due to the rotation of quotas.Footnote ¹⁴ Because our study group consists of pairs of councils assigned to treatment or control groups within each subdistrict, the formal properties of the design are also akin to block-randomized experiments with matched pairs (Imai, King, and Nall Reference Imai, King and Nall2009).

In Karnataka, one final detail is helpful for our strategy: If the number of councils at the threshold number of members’ seats exceeds the number of council presidencies that must be reserved for SCs (or STs) in a given subdistrict, the bureaucrat allocates quotas among these councils by drawing lots.Footnote ¹⁵ Such true randomization of quotas ensures that in expectation, there are no differences between reserved and unreserved councils near the key threshold. For about one-half of our study group of councils in Karnataka, quotas were apparently assigned through such true randomization.Footnote ¹⁶ In Rajasthan and Bihar, as for the other portion of the Karnataka study group, we relied instead on the fact that at the key subdistrict-specific thresholds—at which the SC population proportions are virtually indistinguishable but the assignment of quotas differs—the assignment of reservation is plausibly as-if random (Dunning Reference Dunning2008; Reference Dunning2012; Sekhon Reference Sekhon2009). Whether this design really produces as good-as-random assignment is an important topic we discuss in the next subsection.

Various institutional safeguards help protect the integrity of the process of assigning quotas. After each election, a bureaucrat appointed by the District Commissioner explains the reservation rules to council members in subdistrict assemblies; we were able to verify that at least some of these meetings did take place. Most importantly, we obtained data on the history of reservation in Karnataka, Rajasthan, and Bihar from the respective State Election Commissions, which allowed us to verify the extent to which the procedure was followed.

Table 1 shows an example of the reservation process, using data on the history of SC reservation in the subdistrict of Magadi (district of Bangalore Rural) in the state of Karnataka. Recall that in Karnataka, bureaucrats use the number of members’ seats reserved for SC (rather than SC population proportions) to sort village councils in descending order. Thus, the first column of Table 1 lists all the village councils in the subdistrict, sorted in descending order by the number of seats reserved for SC members; the next two columns show the total number of members’ seats in each council and the number of SC members’ seats. The final five columns indicate whether the presidency of the council was reserved for SCs in 1994, 2000, 2002, 2005, and 2007, respectively, with a “1” indicating presence of reservation and a blank cell indicating its absence. (In Karnataka, the identity of the reserved group and thus the presidency rotated every thirty months after 2000; in Rajasthan and Bihar, it rotates every five years, whenever there is a new village council election). For ease of presentation, here the councils are sorted by reservation status within each stratum defined by the number of SC members’ seats, so that councils that had their presidencies reserved appear first in each stratum. (In fact, however, councils with the same number of SC members’ seats located at the key cutoff value were allocated quotas for the presidency at random in Karnataka).

TABLE 1.

History of Scheduled Caste Reservation (Magadi Subdistrict, Bangalore Rural District, 1994–2007)

In the final five columns, 1 = Council presidency is reserved for Scheduled Caste. See text for explanatory notes.

The history of reservation depicted in Table 1 closely follows the expected diagonal pattern, in which the 1's move from the top left of the table to the bottom right. Where village councils that share the same number of SC seats differ in reservation status, in any electoral term, it is because some of those councils have been selected at random, through the drawing of lots, for reservation of the presidency (with one exception).Footnote ¹⁷ For example, at the bottom of the list of 1's in the final column of Table 1, the village councils of Sathanur and Shankighatta both have two SC members’ seats—and thus both could have had their presidencies reserved for SCs in 2007. Yet, Sathanur was selected at random for a quota mandating an SC president, whereas Shankighatta was not.

A similar procedure is used in Rajasthan and Bihar, with the difference that SC population totals are used to rank councils in descending order.Footnote ¹⁸ This results in quite fine-grained differences in the assignment variable (SC population) between councils assigned quotas and those not assigned quotas, at the cutoff value in any electoral term.Footnote ¹⁹ The process used to assign quotas for SC presidencies is also repeated for STs, using exactly the same procedure: Councils are sorted in descending order by the ST population or the number of members’ seats reserved for STs, and the presidencies of the required number of councils are selected for reservation. If a single presidency should in principle be reserved for both the SC and ST categories in any electoral term, due to placement on the respective lists, the presidency is reserved first for one group and then the other in a subsequent electoral term.Footnote ²⁰ In most subdistricts, however, the number of presidencies reserved for STs is relatively small (typically just one or two councils), because STs comprise only a small proportion of subdistrict populations outside of so-called tribal areas. Thus, reservation for ST presidencies has only a small impact on the process of rotation of SC reservation.Footnote ²¹

It is useful to highlight two features of this process: different lists are used in different subdistricts, and the threshold value of the assignment covariate at which councils are assigned to quotas varies across subdistricts. Thus, in some subdistricts, such as Magadi in Table 1, bureaucrats had only worked down to the middle or bottom of the descending list of councils by the election prior to our surveys. In others, such as Karnataka's Chamarajanagar subdistrict (Online Appendix Table A1), bureaucrats had cycled through the list and gone back up to the top.Footnote ²² In consequence, there is substantial variance in our study group in the SC and ST population proportions—which may mitigate in some ways the standard concern that units at the RD threshold are not representative of an interesting population (Deaton Reference Deaton2009). In fact, while the RD study group is clearly not a representative sample of village councils in our three states, the sample means of our selected councils in Rajasthan and Karnataka are statistically indistinguishable from population averages in those states on many key census variables (Online Appendix Tables A7–A8).Footnote ²³

Selection of States, Districts, Councils, and Respondents

Our initial research took place in Karnataka, a state with a long history of village council governance and one in which substantial expenditure powers have been devolved to local councils. In some ways, this state represents a best-case setting for finding distributive effects of quotas, because councils have especially substantial resources to distribute. Yet, the nature of caste politics in Karnataka and southern India as a whole (for instance, its arguably less politically salient character than in parts of northern India, see Jaffrelot Reference Jaffrelot2003) could also plausibly moderate the distributive effect of quotas. This suggested the value of replication of our Karnataka study in settings with different caste politics and party systems. To probe the external validity of our initial findings and also to extend and test further our initial results, we thus extended our research to Rajasthan and Bihar—two states with different histories of council governance in which caste politics arguably plays a stronger local role.

In Rajasthan and Bihar, we selected several districts at random; in Karnataka, we purposively sampled six districts to maximize variation on factors such as the identity of particular dominant castes (see Dunning Reference Dunning2009).Footnote ²⁴ We then selected pairs of village councils from the subdistricts located in those districts, mimicking the reservation process described earlier as closely as possible. Thus, we used 2001 census data to sort the council constituencies in descending order of SC (or ST) population proportions (in Karnataka) and population totals (in Bihar and Rajasthan) and used our reservation data to select pairs of councils with very similar SC or ST populations but different reservation status at each subdistrict-specific threshold.Footnote ²⁵ This procedure generated a study group of 512 councils (200 in Karnataka, 148 in Rajasthan, and 164 in Bihar).Footnote ²⁶

To assess the claim of random or as-if random assignment to quotas, Table 2 presents a balance check, comparing reserved and unreserved councils on measured pretreatment covariates such as literacy and employment data drawn from the 2001 census. As the table shows, when pooling across the three states (and thus maximizing statistical power), constituencies with reserved and unreserved council presidencies are statistically indistinguishable on these covariates—just as they would be in expectation after true randomization. The nominal p-values in the final column of Table 2 assume independent tests; yet global tests that allow for dependence of the covariates also fail to reject the null hypothesis of balance. For example, the F-statistic for a regression of treatment assignment on these covariates is also insignificant (p-value 0.64). In the Online Appendix (Tables A2–A6), we show that balance holds on additional pretreatment covariates both individually within each state and for a larger study group drawn from throughout the state of Karnataka (this study group is discussed later).Footnote ²⁷ In Karnataka, balance also holds both for the subsample with quotas assigned at random (through the drawing of lots) and those assigned only as-if at random at an RD threshold.Footnote ²⁸ As discussed above, there are sometimes deviations from the prescribed assignment procedure, as noted in Table 1, and this occurs to greater or lesser degrees across different subdistricts; however, balance on pre-treatment covariates holds across different subsets of the data defined by the degree to which the assignment procedure was perfectly or imperfectly implemented. Moreover, the rotation of quotas may itself undermine the utility of lobbying officials to deviate from assignment procedure, since political actors understand that quotas will be assigned in some future election if not the current one (though they cannot readily predict when reservation will occur, inter alia because they lack the data we used to construct reservation histories).Footnote ²⁹ In sum, while we cannot be certain that assignment to quotas is as good as random—which is an Achilles' Heel of many natural experiments (Dunning Reference Dunning2008, Reference Dunning2012)—here our data on the history of reservation, our qualitative fieldwork on the assignment procedure, and our balance tests on pre-treatment covariates suggest the plausibility of this claim and give us confidence that our treatment and control groups provide valid counterfactual groups.

TABLE 2.

Balance Tests on Pretreatment Covariates

Notes: The unit of analysis is the village council constituency. Data are from the 2001 census. Standard errors are in parentheses. The p-values in the final column give the probability of observing a t-statistic as large in absolute value as the observed value, if Group (A) and Group (B) are drawn from the same distribution. Additional tests are presented in the Online Appendix (Tables A2–A6). N = 512 village council constituencies.

To gather data on distributive and fiscal outcomes, our survey team interviewed citizens, two council members, the president, and local bureaucrats (secretaries) in each selected council constituency. In Karnataka, the sampling design called for a random sample of 10 citizens in the headquarter village of each of the 200 councils. In Rajasthan and Bihar, we selected two villages at random within each council constituency and interviewed eight citizens in each. We used an interval sampling method to select households and then attempted to interview the adult with the nearest upcoming birthday. This procedure generated a sample of 6,977 citizens across the three states.Footnote ³⁰ We asked citizens a range of questions about benefit receipt and perceptions of council priorities; we also used survey experiments in Rajasthan and Bihar to compare how caste and party affiliations shape voting preferences and expectations of benefit receipt, as well as to test observational findings from our initial work in Karnataka. Descriptive statistics are presented in the Online Appendix (Tables A9–A10). We conducted fieldwork in Karnataka in January–February 2009, in Rajasthan in August–September 2011, and in Bihar from January–March 2012. In each case, the surveys took place more than a year after the previous election had installed a new council president.

Partisan Targeting in Multi-caste Organizations

Although candidates for local councils in most Indian states are banned from using party symbols in their campaigns, and ballots do not show party affiliations, partisanship nonetheless plays a key role in village councils. Pooling across states, about 72% of our survey respondents can identify the party of the council president, with percentages of around 90% in Karnataka and Rajasthan (Online Appendix, Table A9).Footnote ⁵¹ Indeed, our data suggest that knowledge of the president's party is nearly as widespread as knowledge of the president's caste. On our questionnaires, most council members were also able to list the party affiliations of other members without difficulty. Party affiliation appears strongly related to electoral behavior: More than 80% of party members said they voted for their party's candidate in the most recent elections.

Why is party affiliation so salient? One reason is that parties play a crucial role in financing the increasingly high cost of council elections. In interviews in Karnataka, candidates as well as party officials at the district and subdistrict levels independently estimated the per-candidate cost of local campaigns at about $2,500 (100,000 Rupees)—a shockingly large sum that is presumably far more than even an entrepreneurial council member could make in bribes and kickbacks during a five-year term. In our surveys in Bihar, presidents estimated total campaign spending from all sources at 202,291 Rupees (whereas council members estimated 34,159 Rupees). The distribution of targeted benefits plays an important role in electoral mobilization (see Breeding Reference Breeding2008): Presidents estimated that for an average campaign for their office in a typical election, they spent a total of 120,254 Rupees alone on money and gifts for citizens (for members, 16,659 Rupees). Our formal surveys and qualitative interviews suggested that party leaders help fill the financing gap. In Bihar, 64% of surveyed presidents and members said their party had helped fund their campaigns for village council (though only 7% said parties had provided broader “support”). Parties also contribute to horse-trading at the council level, for instance, by supplying the funds necessary to buy members’ votes.Footnote ⁵² Finally, parties play an important role in structuring career advancement for politicians (say, in seeking seats on district or subdistrict councils), and party leaders at higher levels, such as state legislators, are often in contact with village council members and are sometimes present at council meetings (Wilkinson Reference Wilkinson, Kitschelt and Wilkinson2006).Footnote ⁵³

In return for the largess of their parties, council members and presidents are expected to mobilize votes for the party, especially in elections to fill positions in district councils, the state legislative assembly, and even the national parliament. In interviews, party workers at the district and subdistrict level described the way in which a single broker in each village—often a council member or the president—took responsibility for coordinating party mobilization efforts around election time. In our Bihar survey, council presidents reported spending an average of 3.4 hours per week doing work for their political party (compared to 1.2 hours for council members; the difference is significant at the 0.01 level). Moreover, 30% of presidents (and 24% of members, 42% of presidents who are party members, and 39% of council members who are party members) said they provide support during elections to other politicians from their parties. That council members and presidents are rewarded for mobilizing the vote at election time may in turn create strong incentives to allocate benefits to co-partisans. Stokes, Dunning, Nazareno, and Brusco (2013) provide one rationale for why local brokers seeking to build large networks might tend to invest in party loyalists: Party sympathizers are relatively cheap to mobilize, and party leaders cannot readily distinguish between the targeting of sympathizers and swing or indifferent voters.Footnote ⁵⁴ Whatever the reason why party sympathizers are targeted, however, the main point is that here local brokers may distribute resources along partisan lines. We test this hypothesis later.

The reason that partisan distribution may then mitigate the distributive effects of quotas is twofold. First, our research suggests that parties are often to a greater or lesser extent comprised of multiple castes and caste groups at the local level. Many candidates for village councils draw support from both dominant and marginalized castes; correspondingly, voters of the same caste, even in the same village, frequently support different parties. In Karnataka, where our within-village sample of SC citizens is largest, we interviewed up to four SC citizens in each village (out of 10 sampled respondents); in 63% of the villages in which at least two SC respondents identified the party for which they voted in the most recent election, they had voted for at least two different parties. Moreover, in village councils in which more than one SC or ST council member answered our party affiliation question, we found that they were affiliated with the same party only 56% of the time.

Party leaders also stressed the multi-caste character of their local organizations in our field interviews—even those from traditionally upper caste parties such as the BJP, whose leadership in Karnataka's Malavalli subdistrict went to pains to point out the presence of local SC party leaders.Footnote ⁵⁵ In Bihar, about 50% of SC or ST party members among surveyed council members were allied with the Janata Dal (United), whereas the rest were split among other parties, especially the Rasthriya Janata Dal and the BJP; similar percentages held for SC/ST citizens as well. In Rajasthan, SC and ST members tilted more heavily toward the Congress party, but still about one-quarter of SCs and STs in our sample were members of or felt closest to the BJP.Footnote ⁵⁶ Such party splits within caste groups and within villages may reflect the alignment of “factions” within villages with different parties.Footnote ⁵⁷ Yet, the crucial point is that those alignments cut across caste lines. Of course, as we discuss later, the extent to which parties are locally multi-caste may vary across contexts, with possible implications for the mitigating effect of partisanship on quotas. Here, the key point is that if parties are locally multi-caste—and local party organizations have incentives to target voters along partisan lines—then benefits may be distributed within parties to multiple castes.

A second, related point concerns the dynamic incentives engendered by the rotation of quotas, which may itself also heighten the tendency of party affiliations to trump caste ties in shaping distribution. In each electoral term, the caste or tribe identity of the president varies exogenously, due to the quota policy. Consistent with standard theories of dynamic bargaining, regular exogenous alternations in power may moderate how much policy changes whenever the identity of the group in power shifts. For example, Dixit, Grossman, and Gul (Reference Dixit, Grossman and Gul2000), building on Alesina (Reference Alesina1988), construct an infinite-horizon model in which two groups rotate in power according to some fixed exogenous probability, and they characterize the set of efficient allocation rules that arise in equilibrium. The key insight of their dynamic model is that each group alters policy less dramatically when in power than it would in a one-shot interaction. Such dynamics might reduce variation in caste- or tribe-based targeting across electoral terms in which quotas are present or absent—because the randomized application of reservation corresponds well to the exogenous process described in Dixit, Grossman, and Gul's (Reference Dixit, Grossman and Gul2000) model.Footnote ⁵⁸

Yet, the key point here is that multi-caste parties are the actors within which such dynamic bargains can be struck. The identity of the party in power is not assigned exogenously by the quota policy but instead depends in part on distributive targeting. Parties therefore have stronger incentives to use resources endogenously to achieve and maintain power, and they may do this by distributing some resources to both upper and lower caste affiliates. Thus, when council leadership rotates exogenously among co-partisans from different castes—due to the quota policy—we may observe endogenous intra-party, cross-caste targeting. In sum, models such as Dixit, Grossman, and Gul (Reference Dixit, Grossman and Gul2000) may nicely characterize the process of intra-party bargaining between dominant and marginalized castes within the council's governing party; exogenous alternation of council leadership may create incentives for inter-temporal smoothing of targeting, but such dynamics occur within multi-caste parties.Footnote ⁵⁹ Note that here, too, the cross-cutting nature of party and caste ties at the local level is again central: If party and caste or tribe were isomorphic, the presence of a caste- or tribe-based quota might lead to bigger shifts in the targeting of benefits, because the quota would then also engender a shift in the party of the council president as well as in his or her ethnic identity.

In sum, distribution may take place along partisan rather than caste lines, with important implications for the impact of quotas. That is, linkages between upper and lower caste council members within multi-caste party organizations may create incentives for allocating benefits to both upper and lower caste party members, regardless of the presence or absence of quotas. If these hypotheses are correct, we should expect the party of the council president to influence the distribution of benefits, as presidents use their discretion to target co-partisans. Moreover, if multi-caste parties shape distribution in the context of dynamic incentives created by rotating quotas, we may also expect linkages across castes, within parties, to be especially important in shaping benefit receipt. We turn in the next two subsections to testing these observable implications of the theory.

Testing the Impact of Partisanship

First, does the distribution of council benefits in fact follow a partisan logic? Our surveys asked citizens and council presidents to which political party they belong; a follow-up question asked citizens (including those who professed no party membership) to which party they feel closest.Footnote ⁶⁰ We use these responses to code two indicator variables.Footnote ⁶¹ The first equals one if the respondent is a member of the party of the council president and is zero otherwise; the second equals one if the respondent feels closest to the party of the president and is zero otherwise. Because the probability of belonging to the council president's party varies across states—as does the average level of benefits received—we weight within-state estimates by the shares of respondents from each state.Footnote ⁶²

We find that citizens who are members of the party of the council president are significantly more likely than others to have received a job or benefit from the council in the previous year (Table 5, first row). Indeed, the estimated effect represents an increase of 30% in the probability of benefit receipt over those who do not share the president's party. This also holds for SC and ST citizens alone (second row), though the effect is estimated more imprecisely. The relationship is pronounced for the MGNREGA job scheme (third row)—in which, as discussed earlier, the president appears to play a particularly active role—but co-partisanship is also statistically related to the probability of receiving a benefit from any government scheme. In other analyses, we found that citizens who share the party of the council president are 13 percentage points more likely than other citizens to say they had received a gift from a party or candidate before an election, in return for turning out to vote (p-value < 0.001). The effects of party appear somewhat stronger in Karnataka and Rajasthan than in Bihar: For example, co-partisanship is associated with a 12 percentage point increase in the probability of receiving benefits in Karnataka, whereas it is associated with a significant 9 percentage point increase in the probability of employment through the MGNREGA scheme in Rajasthan (Online Appendix Tables A17).Footnote ⁶³ We discuss such heterogeneity in more detail later.

TABLE 5.

Party Affiliation and Benefit Receipt (Difference of Percentages, Members of President's Party Minus Nonmembers)

Notes: The table presents the percentage point difference in benefit receipt by co-partisanship with the village council president. Standard errors are in parentheses. Members of the council president's party are citizens who identify the party of the president, in response to the questions, “Are you a member of any political party? If so, which one?” The party of the council president is coded from survey responses of presidents and council members (see text). Respondents were also asked the party to which they feel closest (final row). The cells present weighted averages, where the weights are the share of respondents in each state. N = 1,966 citizens in Karnataka, 2,370 citizens in Rajasthan, and 2,641 citizens in Bihar. Questions about the MGNREGA scheme and all government schemes were asked only in Rajasthan and Bihar.

+ p < 0.1, * p < 0.05, *** p < 0.001.

Of course, the association between co-partisanship and benefit receipt does not conclusively indicate a causal effect of party affiliation: Party membership is not randomly assigned, and the effect of co-partisanship could be confounded by several variables, such as the partisan composition of the council or individual attributes associated with both co-partisanship and benefit receipt. Some concerns about reverse causality may be somewhat allayed by our finding that merely feeling closest to the party of the council president is not statistically related to benefit receipt (final row of Table 5): Integration into party networks through membership may thus be most important in causing citizens to be rewarded by the party in power with material benefits. Nonetheless, citizens who receive benefits may affiliate with the party of the person who gave them the benefit, rather than the other way around. Moreover, we have greater power to detect party effects than for those of quotas because of the structure of our data: Co-partisanship is an individual-level variable that varies within villages, whereas assignment of quotas is clustered at the village council level. Thus, our observational analysis might not provide a fair basis for comparing the effects of party affiliation to those of caste- and tribe-based quotas.

Party versus Caste: A Survey Experiment

To overcome these limitations, we conducted a survey experiment in Bihar and Rajasthan in which we experimentally manipulated the party and caste of a hypothetical candidate for village council president. This design obviates several important concerns about confounding and reverse causality, while also generating tests with equivalent power for party and caste effects. The survey experiment also allows us to test several implications of our argument about cross-cutting affiliations and partisan targeting within multi-caste parties.

We first read citizens a short statement by a hypothetical candidate for president of a local village council:

Our goal was to stimulate identification of the candidate with a specific caste (jati) and caste category, as well as a political party. Thus, in Rajasthan, we experimentally varied [caste name] to read either “Rajesh Singh” or “Rajesh Dhobi”—names associated with the Rajput (forward) and Dhobi (SC) castes, respectively; in Bihar, [caste name] was either “Rajesh Yadav” or “Rajesh Chamar”—again common names associated with a locally dominant caste and an SC caste.Footnote ⁶⁵ Finally, we varied [party] to indicate one of the two major parties currently vying for power in each respective state—in Rajasthan, the Bharatiya Janata Party (BJP) or Congress, and in Bihar the Rashtriya Janata Dal (RJD) or Janata Dal (United). Thus, each respondent was exposed to a hypothetical candidate from a dominant or a subordinate caste from one of the two major parties in the state.

After being read the candidate's speech, voters were asked how much they agreed with a series of statements, on a scale of 1 to 7. We focus here on three:

1. 1. “You would vote for a candidate such as [caste name] for president of the village council.”

2. 2. “If [caste name] were elected, people like you would receive more jobs from the village council.”

3. 3. “If [caste name] were elected, people like you would receive more benefits from welfare schemes through the village council.”

The first question is intended to tap vote intention, whereas the second and third tap expectations of benefit receipt—where the phrase “people like you” is left intentionally vague to allow respondents to think of either their party or caste identities (or both or neither) in responding. We combined the second and third questions to create a “jobs/benefits” scale running from 2 to 14. Although the caste name is repeated in each question, party ties were stimulated only by the candidate's initial statement—which might conceivably weaken party effects in this experiment. Note that membership in or closeness to a specific party is a fixed attribute of each respondent, as is his or her caste, but the caste and party of the candidate are manipulated at random. Thus, finding that the partisan relationship between respondents and candidates influences respondents’ evaluations or expectations of benefit receipt cannot plausibly be due to omitted variables or reverse causality.

One way to analyze these data is to assess how co-partisan and co-caste relationships shape evaluations of the candidates. Thus, respondents who share the party of the candidate can be compared to respondents who do not, and respondents who share the caste of the candidate can be compared to those who do not. This approach allows us to compare directly the effects of partisan ties to the effects of caste ties in shaping voting preferences and expectations of benefit receipt.Footnote ⁶⁶ Here, we can understand “sharing” the candidate's party in broader or narrower ways: We can require that the respondent identify him- or herself as a member of some party—as recorded by responses to the questions, “Are you a member of any political party? If so, which one?” Or, we can code the party to which the respondent feels closest, a broader criterion. In the case of caste, we can slice the data three ways, ordered by increasing narrowness in the understanding of the concept:

1. (1) Broad Caste Group: here, we include all respondents and consider an SC or ST respondent to share the group of the SC candidate, whereas forward-caste and Other Backward Classes (OBC) respondents share that of the dominant-caste candidate.

2. (2) Narrow Caste Group: here, we consider SC respondents to share the caste group of the SC candidate, whereas forward castes share the group of the Rajput candidate (in Rajasthan) and OBC respondents share the group of the Yadav candidate (in Bihar) (and other respondents are omitted).

3. (3) Caste (jati): here we include only respondents from the Dhobi and Rajput castes (in Rajasthan) and the Chamar and Yadav castes (in Bihar); we consider respondents to share the caste identity of the candidate only if they are from the same caste (Dhobi, Rajput, Chamar, or Yadav, respectively).

Obviously, we have the most data and thus statistical power with option (1), but this likely introduces heterogeneity and measurement error in the caste variable. Option (2) corresponds most closely to the legal category of caste used to allocate quotas, which is desirable. Finally, (3) taps the narrowest notion of shared caste (jati), and so caste might plausibly be expected to boost evaluations of candidates the most here. Note that whatever definition of shared caste and party membership we use, we estimate causal effects using about the same number of respondents per treatment condition—so we have approximately equal statistical power for detecting either co-partisan and co-caste effects.

Tables 6A and 6B show estimated effects of party and caste ties on vote intentions—the first question—using the three ways of defining shared caste; for co-partisanship, 6A uses shared party membership, whereas 6B uses closeness to the party of the candidate. The bottom row of each table gives the estimates of party effects—the difference in average candidate evaluations, across those exposed to a co-partisan and those exposed to a candidate from a different party—whereas the final column gives the caste effects: the difference in average evaluations of those exposed to candidates who share their caste and those who do not. Several findings are notable.

TABLE 6A.

Survey Experiment—Party Membership, Caste, and Vote Intentions

TABLE 6B.

Survey Experiment—Party Closeness, Caste, and Vote Intentions

Notes to Tables 6A and 6B: The cells record average answers, by treatment condition, to the question, “On a scale of 1 to 7, how much do you agree with the following statement: You would vote for a candidate such as [candidate's caste name] for village council president.” Differences of means estimate the causal effect of varying the candidate's party and caste. Standard errors are in parentheses.

*p < 0.001 ** < 0.01 * p < 0.05* + p < 0.1. N = 5,011 (2,370 citizens in Rajasthan and 2,641 citizens in Bihar; for some treatment conditions, as noted in the text, the N is smaller).

First, co-partisanship sharply boosts respondents’ reported likelihood of voting for the candidate. The average estimate along the bottom row of Table 6A—which measures the effects of exposure to a candidate who is a co-partisan, among party members—is 0.49, an effect size of one-quarter of one standard deviation; most of the individual estimates are significant. Effects are somewhat weaker when we measure the party to which respondents feel closest, with an average effect of 0.33 points (last row of Table 6B), but we still find significant party effects.

Second, the impact of party consistently appears to trump the effects of caste in shaping voting intentions. As expected, caste effects are stronger as we narrow the definition of caste, although we lose statistical power because the analysis only includes respondents who can be assigned to a candidate from their own jati. Yet, the point estimates from the party treatments are always as large as, and sometimes substantially larger than, the point estimates from the caste treatments. We do not think these results indicate that caste is irrelevant for shaping voting behavior, but they do suggest the relative importance of party in shaping preferences over candidates for village council president.Footnote ⁶⁷

How do respondents assess the likelihood that people like them would receive jobs and benefits if the candidate were elected? Here, too, party effects are typically at least as strong as caste effects (Tables 7A and 7B). Interestingly, however, sharing the party of the candidate does not appear to add much when the candidate already shares the respondent's caste, or vice versa. In contrast, co-partisanship sharply boosts such expectations when the candidate and respondent are from different castes (fourth, fifth, and sixth columns of 7A and 7B)—a finding consistent with the claim that distribution is shaped by intra-party bargains between dominant and marginalized castes.

TABLE 7A.

Survey Experiment—Party Membership, Caste, and Benefit Receipt

TABLE 7B.

Survey Experiment—Party Closeness, Caste, and Benefit Receipt

Notes to Tables 7A and 7B: The cells record average responses, by treatment condition, to questions tapping expectations of benefit receipt if the hypothetical candidate were elected. The dependent variable is a 2–14 scale created by summing degree of agreement, on a 1–7 scale, with the following statement: (1) “If [candidate's caste name] were elected, people like you would receive more benefits from welfare schemes through the village council;” [and] (2) “If [candidate's caste name] were elected, people like you would receive more jobs from the village council.” The differences of means in the final columns and final rows of each portion of the table estimate the causal effect of experimentally varying the candidate's party and caste. Standard errors are in parentheses.

* p < 0.001 **p < 0.01 * p < 0.05* + p < 0.1. N = 5,011 (2,370 citizens in Rajasthan and 2,641 citizens in Bihar; for some treatment conditions, as noted in the text, the N is smaller).

In sum, our observational and survey-experimental results do suggest that party affiliations shape vote intentions and benefit receipt as much or more than caste ties. Moreover, party ties that cross-cut caste may be especially important in shaping distribution. Together with our observation that local party organizations are often comprised of multiple castes, these results imply that cross-cutting relationships between party and caste or tribe help explain why caste- and tribe-based quotas do not have more impact. Our findings suggest that benefits are distributed along partisan—not so much caste—lines, which has important implications for quotas’ distributive effects.