FIGURE S2A: Z-STATISTICS OF LIKING BJP*MEMBERSHIP INTERACTION EFFECTS (NON-ELITES) FIGURE S2B: Z-STATISTICS OF LIKING BJP*MEMBERSHIP INTERACTION EFFECTS (ELITES) FIGURE S2C: Z-STATISTICS OF DISLIKING CONGRESS*MEMBERSHIP INTERACTION EFFECTS (NON-ELITES) FIGURE S2D: Z-STATISTICS OF DISLIKING CONGRESS*MEMBERSHIP INTERACTION EFFECTS (ELITES)

SECTION 1: DESCRIPTIVE STATISTICS AND SUBSTANTIVE EFFECTS  TABLE S.1: DESCRIPTIVE STATISTICS FOR VOTER ANALYSIS (UPPER CASTE SAMPLE) TABLE S.2: DESCRIPTIVE STATISTICS FOR VOTER ANALYSIS (DALIT AND ADIVASI SAMPLE) TABLE S.3: VARIABLES USED IN VOTER-LEVEL ANALYSIS TABLE S.4: SHIFTS IN SIMULATED PROBABILITY OF BJP SUPPORT AMONG DALITS AND ADIVASIS (ESTIMATED FROM MODELS REPORTED IN TABLE 1) TABLE S.5: SHIFTS IN SIMULATED PROBABILITY OF SUPPORTING BJP AMONG UPPER CASTES (ESTIMATED FROM MODELS REPORTED IN TABLE 1)


PART I: INDIVIDUAL-LEVEL ANALYSIS SECTION 1: DESCRIPTIVE STATISTICS AND SUBSTANTIVE EFFECTS
This first section contains some basic information about the voter-level analysis. First, Tables S.1 and S.2 present standard descriptive statistics for the variables employed in the paper's analysis, for both non-elite and upper caste voter samples respectively. Second, Tables S.3 present detailed information on the measures used in the analysis, the concepts they seek to represent, and the sources from which they were constructed (including the wordings of the relevant survey questions). Finally, Tables S.4 and S.5 report the full list of substantive effects for variables registering significant impacts on the electoral preferences of non-elite and elite voters from the paper's analyses (computed from the analyses presented in Table 1 of the main text).

Communalism
To what extent do you agree with these options-fully agree, somewhat agree, somewhat disagree or fully disagree: On the site of the Babri Masjud, only Ram temple should be built Q34a.
There should be a legal ban on conversions

Q36h
Income What is your total monthly household income?

Influence of Caste Identity
In deciding whom to vote for, whose opinion mattered to you most? (coded 1 only if response was 'caste/community leader' Q9

Support for Economic Liberalization
Now I will read out a few statements regarding the economic policy of the country. You tell me, do you fully agree, somewhat agree, somewhat disagree, or fully disagree with these statements?
The number of government employees should be reduced as paying for their salaries is costly for the country.

Q30b
The government factories and businesses should be sold/handed over to private companies.

Q30c
Foreign companies should not be allowed free trade in India (coding was reversed for this question)

Q30d
People are responsible for their poverty and not their government.

Q30e
Opposition to castebased employment quotas There should not be caste-based reservations in jobs.

SECTION 2: BASIC ROBUSTNESS CHECKS
In this section I present the results of some preliminary checks on the robustness of the results of the voter-level analysis. First, I present the results of some diagnostics used to examine the survey data. Table S.6 shows that collinearity between the explanatory variables is not an issue within either elite or non-elite voter samples, or among the statelevel variables. Next, I examined whether specific observations within each data set were biasing the results. To do so, I examined the Pregibon's delta beta (the equivalent of Cook's distance for logistic regression analyses) across all observations, and present the scatterplots of those values ( Figure S.1). This measure of influence revealed no observation exerted what is considered a standard 'high' degree of leverage (>1) on the overall sample. Still, I re-ran the analysis while excluding observations whose influence statistic was twice the mean of the entire sample. These results are presented in Tables S.7 and S.8 and show the key results of the analysis hold and are even improved within this edited sample.
Second, I ran trimmed down specifications that individually assess the impact of each potential confounder included in Table 1 on the relationship between membership and voting. As noted by Ray (2005: 288), potentially confounding effects are often best assessed by models with only the key independent variable and one alleged confounder. The models below suggest the key results are unchanged by such verification. Among the Dalit and Adivasi sample (Table S.9), Non-Party Member remains highly significant, while ideology (Communalism) and policy preferences for liberalization (Liberalization) do not exert a significant influence. Among upper castes (Table S.10), Non-Party Member remains insignificant across the different specifications, while Liberalization and support for the signature Hindu nationalist demand (Communalism) remain significant determinants of BJP support.
Thirdly, I include alternative measures of economic and ideological policy positions on support for the BJP. In addition to its support of economic liberalization, the BJP has voiced consistent opposition to caste-based reservations for non-elite castes, particularly for employment. This position makes sense given the party's elite profile, and distinguishes the party from its chief rivals. Could it be that the party attracts support from non-elite individuals who (albeit somewhat counter-intuitively) might oppose such a system of reservations? Alternatively could this position-rather than positions on liberalization or Hindu nationalism be explaining upper caste support for the party? To assess this possibility, Tables S.11 includes a measure of opposition to caste-based quotas (No Quotas), which does not have a significant impact on either sample. Nor does including this measure alter the impact of other variables. Additionally, I test the influence of a different agenda item of Hindu nationalism-opposition to religious conversions-on support for the BJP (Table S.12). The banning of conversions has come to occupy a central place within the Hindutva platform. Briefly, this preoccupation stems from concerns about the proselytizing efforts of minority religions, specifically attempts to induce conversions among Dalit and Adviasi communities. Once again, this measure does not distinguish non-elite supporters of the party. Interestingly, the measure also does not register a significant impact among upper castes, indicating that ideological support for the BJP among elites is specific to particular demands within the Hindu nationalist ideological agenda.

A. VIEWS OF BJP PERFORMANCE
In this section, I examine whether the key results of my analysis are confounded by a voter's preferences regarding specific political parties. Most importantly, do the results hold when we control for a respondent's satisfaction with the BJP's performance as a political party? The NES allows us to derive a number of indicators of such an opinion. Tables S.13A (non-elites) and S.13B (elites) includes measures of respondent satisfaction with the BJP-led coalition's tenure in office (1999)(2000)(2001)(2002)(2003)(2004), and with improvement in a voter's pocketbook during that period. Not surprisingly, both measures of satisfaction register positive and significant impacts within both samples. More importantly, the key results within each sample remained robust to the inclusion of these variables. Associational membership remains a significant predictor of non-elite support, while preferences for economic liberalization and Hindu nationalist ideology do not. Among elite supporters, we continue to obtain the inverse of these results.

B. PARTY RATINGS: BJP VS. CONGRESS
In this section, I examine the impact of voters holding partisan preferences for the BJP. In the first set of tests, I identify voters who rate the BJP as superior to its main national rival (the Indian National Congress) across a range of measures (governance, corruption, leadership, and terrorism, and an index of all four responses constructed using principal component analysis). Not surprisingly, respondents rating the BJP favorably were more likely to vote for it (the reverse was also true, with voters rating the Congress higher emerging as more likely to vote for the Congress).
More pertinently, Table S.14A demonstrates that the impact of non-party associations remain robust to including each of these measures. In Table S.14B, we see that the main results for associational Associational Membership, Liberalization, and Communalism among elite voters also remain untouched by controlling for voter preferences for the BJP as a party across a range of dimensions. As before, associational membership remains an insignificant predictor, while preferences for liberalization and Hindu nationalist ideology are positively correlated with voting for the BJP.
However, I should note I am somewhat skeptical about whether these measures actually capture a voter's view of the relevant policy item, or just a general partisan preference for the party. For example, positive ratings of the BJP's performance on terrorism exerts a highly significant and substantial effect across the sample, yet only 3.2% of all respondents on the NES rate this as the issue most important to them.
This analysis is even more complicated among the elite sample, by the fact that the main explanatory factors for elite support of the BJP (Liberalization and Communalism) are arguably closely related to voter opinions on the BJP's quality of governance and leadership. Hindu nationalist ideological views might easily drive voters to hold more favorable views of the BJP's leaders or anti-terror strategy, and therefore be highly correlated with the latter. This appears to be what happens in Model 5, where Communalism retains a positive effect, but loses statistical significance. Similarly, proliberalization voters might prefer the BJP's record on economic governance.
All of that said, the robustness of the results in 14 of the 15 cases of these 3 key independent variables across 5 models increase confidence in this paper's argument. Further, on Tables S.15A and B I replicate the variables used in Tables S.23A and B, but code them to identify respondents who prefer the Congress on a range of issues. Once again, these variables are significant (and negative in this case), but do not confound any of the key results among elite or non-elite voters. Table S.16A shows the impact of membership is also robust to including a binary measure identifying respondents who 'like' the BJP as a party (Q15 and 15A on the NES: "Is there any party you particularly like" (If yes) which one?).. While this result helps confirm the robustness of the analysis, I have reservations about using measures of 'liking' or 'preferring Party A' as a predictor of 'voting for Party A'. First, the act of voting for a party has repeatedly been found to increase voter preferences for that party, and so concerns of endogeneity are substantial. Second, the conceptual overlap between this predictor and the dependent variable is extreme. I believe the massive coefficient on 'Liking BJP" is more a function of this conceptual proximity, not because it is a genuinely valuable explanatory factor.

C. 'LIKING' THE BJP
Indeed, Table S.16B shows the impact of including similar predictor variables for other parties examined in Table 5 of the main text (the leftist Communist parties, and lower caste 'ethnic party'). In both instances, the "liking' variables register massive coefficients that crowd out or substantially attenuate other impacts found in the analysis (antipathy to economic liberalization for leftists, and the influence of co-ethnic leaders for the Dalit ethnic party). In the face of these results, the robustness of associational membership on BJP support is even more impressive. Yet in all of these models, including such conceptually proximate predictors crowds out more analytically powerful results, leaving us with the simple fact that 'voters vote for parties they like'. The goal of my analysis is to explore why voters like the parties they like.
Finally, Table S.16C shows that among elites, the results remain robust for associational membership (not significant), liberalization, and communalism (both significant). The substantive coefficient on communalism does decrease with this specification. However, this could again be viewed as a product of the concerns mentioned above. Among the BJP's elite core especially, a measure of 'liking the BJP' might even be better conceptualized as an alternative measure of the dependent variable, rather than an autonomous independent variable. Indeed, communalism and preferences for liberalization prove to be strong predictors of an upper caste 'liking the BJP' in the first place (Model 3).
In sum, while it is heartening that my key results mostly hold even with the inclusion of such powerful measures, these predictors are highly analytically problematic. The concerns raised here are common criticisms of the analytic value of 'party ID' variables in explaining vote choice, not simply within India. Perhaps due to similar concerns, past analyses of political preferences in India (including those using the NES data, and those studying the BJP) have largely not included these variables (e.g. Chhibber 1997, Verma 2012.

D. VIEWS OF THE OPPOSITION (CONGRESS)
A slightly different question is whether the BJP is simply benefitting from voters dislike of its rivals, especially the Indian National Congress. Thus, voter support for the BJP might be motivated by disaffection with the Congress. To examine this possibility, I created a binary variable identifying respondents who specified the Congress as a party they didn't like (Q16 and 16A on the NES: "Is there any party you particularly dislike" (If yes) which one?). The results (Table S.26A and S.26B) show that the inclusion of this variable does not change any of the key results among elites or non-elites. More interestingly however, Table S.17A (Model 2) does report a significant interaction effect between non-party membership and disliking Congress. Note the coefficient is tricky to interpret since this is a logistic regression model, but subsequent analysis reveals that this interaction term exerts a positive marginal substantive effect (Table S.17A2) that is significant for most observations in the sample ( Figure S2C). Interestingly this interaction term is not significant among elite voters (Table S.17B, Figure S2D) Further, an interaction term between membership and actively 'liking' the BJP does not register consistently significant impacts among elites or non-elites (Tables S.25A-B, Figures  S2A-B).
Thus non-party networks appear particularly effective in recruiting non-elite voters who are displeased with the Congress towards the BJP, rather than those who already like the latter, which is very much in line with my argument.

TABLE S13B. RESULTS ARE ROBUST TO CONTROLLING FOR PERSONAL SATISFACTION WITH BJP-LED CENTRAL GOVERNMENT AND POCKETBOOK INCREASES DURING BJP TENURE (ELITE SAMPLE)
Voted for BJP (2004)  VARIABLES (1)      Υ The average coefficient of interaction terms calculated separately for each observation (using the protocol developed by Norton et. al) is slightly positive (.005) and statistically not significant (indeed none of the individual interaction effects were found to be significant). See Figure S2.0A.

SECTION 4: MEASUREMENT VALIDITY CONCERNS
Section A: Is the impact of Membership specific to non-party organizations, or does it also apply to political parties? Table S.18 checks my argument's emphasis on voter incorporation within nonparty organizations by including a measure of incorporation within party organizations. Interestingly this latter variable has no impact on poor voter support for the BJP, and does not confound the effect of non-party associations (Column 2). Further, inclusion within party networks does identify upper caste supporters of the BJP, consistent with my argument that the party arm focuses on retaining elite support.

Section B. Is the Impact of Membership greater in states where it more likely reflects participation in Hindu nationalist organizations?
These tests show membership matters precisely in those states where it is most likely to measure participation in Hindu nationalist welfare organizations.
One of the very reasonable concerns with the Associational Membership (Membership) variable is that the survey instrument provides a noisy measure of inclusion within Hindu nationalist organizations. To partially address such concerns, this section combines information from the citizen survey and Hindu nationalist records on welfare provision. For the Membership variable to capture membership within Hindu nationalist welfare organizations, it is plausible to assume those organizations have to have a local presence. If they do not, Membership is more likely to capture participation in other associations (such as those based on caste or unions). Accordingly, my theory anticipates membership to correlate with BJP support more strongly in states with dense Hindu nationalist networks. Conversely, if membership correlates more strongly with BJP support in states with weak Hindu nationalist networks, the plausibility of my interpretation of the results is weakened.
Statistically, I test this argument by examining how membership's impact on BJP support is conditioned by the state-level welfare index. 2 I do so by interacting the Hindu nationalist welfare index value for the state in which a respondent resides (Sangh Service) with their membership status in non-party associations (Member) to create the interaction variable (Sangh*Member).
The results, presented in Table S.19, are intriguing. The coefficients for both associational membership and the service index are positive and strongly significant. More importantly for this analysis is the impact of their interaction. However, the substantive and statistical significance of the interaction term within a non-linear framework cannot be interpreted in the same manner as in OLS. 3 I therefore use the inteff command in Stata, to calculate the interaction effect for each observation separately. According to this test, the mean interaction effect is positive (.488 in the full specification) and significant (p<.01). Figure S3A shows the interaction effect is positive for every voter (represented by blue dots) in the sample. Figure S3B shows this positive effect is significant for voters with a predicted probability of supporting the BJP >.2 (which includes most voters in the sample).
Of course the inclusion of a variable measured at the state-level within an individual-level model requires careful consideration and interpretation (see fn). 4 A second issue is that, since the interaction term uses a variable observed at the group-level, it essentially serves to compare the equality of coefficients across the different groups partitioned by that variable. In doing so, we assume no group differences in residual variations, which is problematic. Further most statistical efforts to address this issue involve their own problematic assumptions. 5 However, one simple way to address this issue is to examine the average marginal effect of membership on the predicted probability of supporting the BJP, and to do so for different values of the welfare index (Long 2009). 6 Figure S3C plots the difference in predicted probabilities of members and non-members supporting the BJP across the range of observed values for the Hindu nationalist welfare index. The average marginal impact of membership is negative for low values of the welfare index and is statistically not significant (the 95% confidence intervals cross zero). Yet as we move to higher values of the welfare index, the average marginal impact of membership becomes positive and steadily increases, and becomes statistically significant after the welfare index crosses a value of 0.1.

In line with my expectations, membership's positive impact on BJP support gets substantively stronger as Sangh service networks grow denser.
Of course, this test is not foolproof, as the interaction term does not uniquely identify members of Hindu nationalist welfare associations. It therefore remains technically possible that membership in non-religious associations, say unions or caste associations, still drive the results. However in the wake of these results, such an interpretation becomes seems increasingly theoretically implausible. For example, such an argument would now have to explain why membership in non-Hindu nationalist 4 Most importantly, since there is no variation in individual values on this variable within the state, the assumption of independence of individual level observations is violated. Without accounting for such clustering, our estimates of the standard errors for the state-level variable will likely be biased downwards, leading to deflated p-values. To help correct for this, the model includes standard errors corrected for clustering by state. While not a panacea, this method both presented less biased estimates than naïve errors, and is also preferable to alternatives such as hierarchical linear models, which are problematic when dealing with a small number of higher-level units. I prefer this approach to hierarchical multi-level model, because it requires fewer assumptions and data requirements. Because HLM models estimate each of the component levels using MLE it is unadvisable to use it for data with small numbers of higher-level units (Steenbergen and Jones 2002). Thus HLM is not appropriate for a dataset with only 17 state-level units. In such instances, using clustered standard errors has proven to provide more reliable estimates than using naïve standard errors, or using HLM. Further, the technique I use is not appreciably different from the widespread practice of including state dummies within the individual level regressions, which also essentially assign all respondents within the same state a score of the same value. 5 Allison (1999) proposes a test that removes the effect of residual variation by assuming that the coefficients for at least one independent variable are the same in both groups. Unfortunately, it is difficult to provide sufficient theoretical or empirical information to justify such an assumption for most analyses. 6 First, predicted probabilities can be used to compute marginal effects of variables in the model (rather than multiplicative effects indicated by the coefficients). Second, Long (2009) notes that predicted probabilities are unaffected by residual variations, and therefore can be used to provide more accurate tests of the significance of differences across groups than examining coefficients. associations increases the likelihood of BJP support more strongly in states with dense Hindu nationalist associations.

Section C. Do Members Primarily Belong to Non-Religious Organizations?
These tests show members are distinguished by attributes and attitudes more consistent with participation in religious organizations than those organized around labor or caste.
In the first set of tests, I specifically examine whether members are distinguished by traits we would expect of personnel within a) trade and labor unions, b) caste associations, or c) religious organizations. Table S.20A shows that members are not distinguished by employment in occupations that enjoy higher levels of unionization within India (these include public sector employees, and workers in manufacturing). The variable is statistically insignificant across all four specifications. Further, many union members depend directly on the public sector, which provides most formal employment opportunities in India. Thus if union members dominated the Membership measure, we would expect to see members oppose moves to reduce or privatize the public sector. Yet the tests in Table S.20A (especially Columns 1 and 2) show no such opposition. Table S.20B similarly examines whether members are distinguished by traits we would expect of personnel within caste associations. Here I examine three variables: the degree to which respondents agree they should vote in the same way as their caste community members, whether respondents identified their caste leaders as the most important influence on their vote choice, and the degree to which they agree that boys and girls from different castes should not marry. Since caste associations function as socio-political organizations that both regulate the marriage market within caste groups, and function as political lobbies for their members, we would expect members within such associations to respond positively to all three of these measures. Yet these expectations are met in none of the three cases. Further, in two instances, we find significant effects in the opposite direction. Members are significantly less likely to agree they should vote with their co-ethnics, and significantly less likely to disapprove of intercaste marriage. These results cut strongly against the idea of that voters identified by the Membership variable are largely participants within caste associations.
Finally, Table S.20C examines whether members are distinguished by traits of those participating in religious organizations. Specifically, I examine if members are marked by higher levels of religious activity, which may make them more likely to affiliate with Hindu nationalist associations. I test whether members are more likely to pray frequently, attend temple, and participate in ritual religious occasions. I anticipate that if Nonparty membership is driven by religious associations, we would expect to see members positively distinguished across these three criteria. The results show that the coefficients for all three measures are positive, and in two cases statistically significant. Religious participation also has a substantial impact. The predicted probability of a respondent being an associational member increases from 11.15 percent for someone who never participates in religious rituals, to 19.18 percent if they do so frequently (an increase of 72%).     Table S.21 show that while associational membership significantly distinguishes voters switching to support the BJP between 1999 and 2004 (Column 1), it does not distinguish those already supporting the party in 1999 (Column 2). This increases confidence that non-party associations are drawing voters towards the party, rather than supporters of the party being drawn into these associations.
However, this analysis does not preclude a voter deciding to vote for the BJP at some point in the five years between 1999 and 2004, and only then joining a non-party organization, again during the same five-year span. One possible way to address such concerns is to test the impact of associational membership on the subsample of respondents who reported making their electoral decision within a few days of the 2004 election. Table S.21 shows that membership increased the likelihood of these 'late deciders' voting for the BJP (Column 3), and late deciders switching to vote for the BJP (Column 4). In this final specification membership precedes vote choice, except in the unlikely event of the respondent joining a non-party association on election day or just before.
The data is unable to deal with a related concern, that people may have joined associations before 1999, voted for a party other than the BJP, and then changed their mind in 2004. However, this sequencing is less problematic for my argument, which emphasizes that organizational incorporation precedes vote choice, not necessarily that these shifts must be immediate. If a voter joined a Sangh affiliated organization in 1998, and took until 2003 to decide to switch their political allegiance to the BJP, that would not necessarily contradict the logic implied by my analysis.

SECTION 6: FURTHER DETAILS ON MATCHING ANALYSIS
This section contains a couple of additional tests that help confirm the validity of the propensity matching protocol used in Table 3. First, since the propensity matching technique uses replacement, it is possible that only a few control units are being matched with multiple treatment units. As a consequence, we may have members compared with relatively few non-members. Table S.37A shows this is not the case, as 85% of control units within the matched sample are only used once, and 97% are used once or twice.
A second concern is whether 'nearest' neighbors are actually closely matched (since the match did not specify calipers around the propensity score differential). Table  S.37B shows this is not the case, as most matched pairs had a minimal propensity score difference of less than .001 (on a 0-1 probability scale), and no pair had a difference of greater than .011.

TABLE S.22A: IS MATCHING HIGHLY DEPENDENT ON A FEW CONTROL UNITS?
No-the vast number of control units for the matching are used only once.

SECTION 7: ADDITIONAL TESTS A. OTHER MARGINALIZED COMMUNITIES: HINDU NATIONALISM AND RELIGIOUS MINORITIES IN INDIA (SURVEY EVIDENCE)
Does the BJP's success extend to other poor Indians, specifically poor religious minorities? Many minority communities, especially Muslims, make up a significant part of India's poor. Yet there is also greater income heterogeneity within these communities than within Dalit and Adivasi populations. Only 54.98% of Muslims lie within the bottom two income categories on the 2004 NES (earning less than $22 and $44 per month). By contrast 71.33% of Dalits and 76.17% of Adivasis are within these two categories. Thus, I replicated the results for the sample of poor Muslims (defined as those in the bottom two income categories, which was roughly half the Muslim sample). Table  S.23 shows that non-party associations do not exert the same positive effect on support for the BJP among poor Muslims. The second column shows a similar non-impact among low-income Christians (again those within the bottom two income categories).
The evidence also suggests that relatively few poor Muslims or Christians are incorporated within these organizations. However, it is difficult to tell if these low numbers are due to supply or demand-side constraints. Are the low numbers of religious minorities within these associations due to a Hindu nationalist aversion to incorporating religious minorities? 7 Or is it due to the fact that religious minorities are unlikely to join such organizations, despite the potential material benefits of doing so, because of the BJP's Hindu nationalist? 8 These are important questions worthy of further inquiry.

B. STATE AND CONSTITUENCY--LEVEL EFFECTS:
I examined if the key results shifted if I did not include state-level fixed effects. Tables S.24 and S.25 present the results for random effects models for non-elite and elite samples respectively. Once again the results were practically identical among the lower caste sample. Membership distinguished non-elite supporters of the BJP not simply within states, as the fixed effects specification measured, but across the national sample. Membership also continued to mark those poor voters shifting to support the party, while economic and cultural preferences once again did not. Among upper castes as well, the results were highly similar to those in the main text. The only difference here was regarding the influence of caste leaders, which was previously significantly negatively related to BJP support among elites, and was not significant in these specifications. Overall, the consistence of these findings across fixed and random effects specifications is encouraging. 7 Observations from qualitative fieldwork suggests that service activists in central India didn't prevent anyone from coming to their chapters (and there were occasional Christian and Muslim beneficiaries). This suggests the low rates of religious minority incorporation might remain due to the concerns among such communities with Hindu nationalist ideology 8 The average level of support for Hindu nationalism, measured on a 4 point scale, was one full point lower among Christians (1.94) and Muslims (1.78) than among Hindus (2.91), a difference that was significant at the .001 level. Tables S.26A and S.26B repeat the analysis using constituency-level fixed effects and robust standard errors corrected for clustering at the constituency level. The results are robust to this specification: non-party membership significantly distinguishes nonelite supporters of the BJP, while support for economic liberalization and Hindu nationalism significantly distinguish elite supporters. The robustness of these results offer strong confirmation of my argument, showing that associational membership marks supporters of the BJP from co-ethnic non-supporters, even within the same electoral constituency. However, including constituency fixed effects does attenuate the sample, as the dummies for constituencies in which the BJP won no votes perfectly predict the outcome variable, and so are dropped from the analysis.

C. COALITIONAL EFFECTS:
Could the BJP's success be the product of the party's alliances with parties who appeal to lower caste and tribal voters? Some of the concerns about this being a 'coalitional effect' is addressed by having the DV measure votes for the BJP specifically, and not a coalitional partner. However, it is possible the BJP itself does better with the poor in states where it has coalition partners, specifically those that are seen as low-caste parties or headed by low-caste chief ministers. To assess this possibility, I created an indicator variable that identified those states in which the BJP had coalition partners in the 2004 national elections, and a second variable that specifically identified states in which the BJP had joined a coalition or offered external support to parties seen as having a lower caste base. The results of this analysis are presented in Table S.27 (see fn for a list of parties and states in each case). 9 Neither coalition variable has a significant impact, and each actually registers a negative coefficient.

D. IS THIS CLIENTELISM?
Given that welfare is privately provided out of electoral considerations, can we not think of it as a form of clientelist exchange? In the main text (p. 34), I argue that the benefits provided by the BJP's non-party affiliates are not part of a clientelist exchange because a) they are not provided with exclusion locally, and b) quid pro quo is not enforced among recipients. In support of the latter point, I present evidence that the BJP does not appear to attempt to monitor the reciprocity of voters incorporated within its non-party networks (a key feature of clientelism). 10 Finding evidence of the presence or absence of micro-level monitoring is of course extremely difficult, and has been the subject of significant debate within studies of distributive politics (e.g. the debate between Stokes 2005 and Nichter 2008). However, the NES 2004 does provide measures of party contact with individual voters. Specifically respondents were asked if they had been visited by party personnel during the election campaign (NES 2004, Q8). 11 If members who were visited by such personnel were more likely to vote for the BJP than members who were not visited, then this would suggest the importance of monitoring efforts in translating non-party activities into votes. A second variable related to monitoring is whether a respondent attended a campaign rally prior to the election. Prior studies have noted that parties often use attendance at these rallies as a signal of a voter's intention to reciprocate at the polls (e.g. Auyero 2001). Table S.28 show that both measures of monitoring effort are actually negatively correlated with the likelihood of a member voting for the BJP, and statistically insignificant. This runs against the expectation of non-party associations forging clientelist ties with their members.  Membership remains significant, and in fact the coefficient increases in a random-effects specification. Support for Hindu nationalism remains insignificant. The only change is that in this specification support for liberalization has a significant effect, but one that is substantively and statistically far less significant than that of membership (and one that is eliminated when excluding prior supporters of the BJP).  Observations 3,452 Note: *** p<0.01, ** p<0.05, * p<0.1 Robust standard errors clustered by constituency in parentheses. Sample is smaller since this specification drops constituencies in which BJP did not win any votes.    (N=25,433) available at the time this study was conducted. In this section I provide additional details about the survey, and show it compares favorably with other widely used surveys, including the Afrobarometer and World Values Survey. Not surprisingly, the NES studies have been the most widely utilized electoral surveys within the study of Indian politics (e.g. Chhibber 1999, Yadav 1999, Chandra 2004, Heath 2005, Ray and Wallace 2007. Since the Election Commission of India does not keep individual-level voting data, nor does it provide aggregate voting statistics for different caste groups, the NES provides invaluable empirical opportunities to examine the sources of BJP support from different caste constituencies. The CSDS has regularly conducted surveys of voter opinion since 1996, and has done so for both the Lok Sabha (National Parliament) and Vidhan Sabha (State Assembly) elections. The CSDS studies are the largest, most systematic and widely utilized electoral surveys available in the country, and are particularly invaluable for this project as they contain information on both the electoral choices and social profile of a large, representative sample of voters.
For this particular study I am employing data collected during the 2004 Lok Sabha [National Assembly] elections. I was given access to this data as a visiting researcher at CSDS in 2007. Specifically, I was given the raw data files for Dalit, Adivasi, and upper caste Hindu voters in the 17 major states in India: Andhra Pradesh, Assam, Bihar, Chhattisgarh, Gujarat, Haryana, Jharkhand, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Orissa, Punjab, Rajasthan, Tamil Nadu, Uttar Pradesh, and West Bengal. The sample covers 3411 Dalit respondents, 2048 Adivasi, and 2711 upper caste respondents.
In the words of its executors, the NES is not an exit poll, but a post-poll survey, whose primary purpose is not 'predicting the number of seats that a…coalition is likely to get. Rather, it is a survey that takes the results as given, and then tries to understand the reasons why voters chose the parties they did.' 12 As the survey team was particularly interested in understanding state-level politics even within national elections, they use a stratified random sampling technique that makes sure the samples within each state are reflective of the state's share of the country's total. The interviews were conducted after the votes had been cast in each constituency, but before the results were known. This had the advantage of having respondents report their actual (rather than anticipated) vote choice, but without knowledge of the outcome influencing their answers to the questions.

A1. Sampling Protocol:
The NES survey is a post-poll survey, not an exit poll. This means that the sampling is not sensitive to whoever is willing to answer survey questions as they walk out of a polling station. Instead, the NES looks to survey a representative random sample of Indian voters. The NES used a multi-stage weighted sampling design, similar to the Afrobarometer protocol. Within each state, the NES selected a random number of parliamentary constituencies (adjusted for the size of the state's electorate). In total, the survey sampled 420 out of 543 parliamentary constituencies. Within each parliamentary constituency, the NES selected a random number of state assembly constituencies (the units relevant for state-level politics). In total, 932 Assembly Constituencies were selected for the survey. Within each Assembly constituency the local sampling unit was individual polling station areas (typically villages or urban wards), which were again randomly selected. A total of 2380 such polling stations were selected. From each polling station, 15 people were selected randomly from the electoral register (this is nearly double the number of people interviewed at each principal local sampling unit for the Afrobarometer surveys). 13 Investigators visited each individual at their home residence. The survey was administered in the local regional language, and even accounted for variations in the local dominant language within states. The respondents were interviewed with partial replacement (i.e. only if another household member fit the same age and gender profile of the intended interviewee). While this substitution violates the principle of purely random sampling, it accounts for a small portion of the final sample (substitution was used in 8.5% of all cases). Further, replacement was only followed if the investigator was not able to meet the interviewee after 2 attempts.
The response rate on the survey is below: For a breakdown of sample sizes and response rates for each of the 17 states used in the analysis, see Table S.29 below.

A2. Representativeness:
The NES sample is more than ten times the India sample for the World Values Survey, and is less biased towards more privileged respondents (see below). The large sample size was in part designed to ensure representative samples of individual states, not simply at the all-India level. Towards this end, the NES looked to achieve at least 1000 interviews in every major state (including the 17 used in this study), which ensured a sampling error of just 3.1% (with a 95% confidence interval). The smallest state-level sample among the cases used in this study was 843 (Tamil Nadu), yielding a 3.3% error margin.
Further, the NES sample was excellent in its representation of conventionally underrepresented groups (see Table S.30). 46.5% of the respondents were women, only 1.8% below their proportion of the Indian population. Further women were under-

B. Survey Team Profiles:
The survey team was trained in two phases. In the first phase, the state coordinators of the survey, all university faculty, were trained in the survey design, sampling strategy, and interview protocol. In the next phase, each of these coordinators then recruited and trained university students over three days (comparable to the length of training for Afrobarometer) to serve as the enumerators. The students were mostly graduate and undergraduate students in political science, and therefore familiar with electoral politics (especially in their home state). The survey coordinators were told to make efforts to ensure they recruited a suitable number of lower caste and female enumerators. They had mixed success in this endeavor. A quarter of enumerators were women, which ensured they had a significant presence on the survey teams, but were clearly underrepresented.
Caste profiles of the enumerators for the 2004 survey were not available from the Lokniti research institute, but a publicly available report on their 2009 survey team does contain this information. 15 Since enumerators in 2004 and 2009 were selected through a highly comparable process, this information is instructive. The 2009 report found that 21% of the NES enumerators came from a Dalit background, and 11% came from Adivasi backgrounds. In each case this was actually higher than their respective proportions within the national electorate (roughly 16 and 9 percent respectively). This level of representation of non-elite social groups is impressive. I should also note that the enumerators identified themselves as representatives of the national survey institute in Delhi, not by their personal names (which would have identified their caste status to interviewees).
Further, the survey teams were designed to match the caste profile of enumerators with those of the majority of interviewees. This strategy does have limits, given the caste heterogeneity of the Indian electorate, even at the local level. Survey teams were assigned to specific polling areas, most of which had members of multiple caste groups. It was therefore not feasible to give each caste group a local representative on each survey team. However, this would be an issue with any survey of an ethnically diverse country (and is no different from issues on the Afrobarometer, or Asian Barometer surveys). Indeed, the NES efforts in India to match respondent and enumerator ethnicity are exemplary, given the challenges involved, and at least as impressive as these other more famous surveys.
The 2009 report found that 21% of the NES enumerators came from a Dalit background, and 11% came from Adivasi backgrounds. In each case this was actually higher than their respective proportions within the national electorate (roughly 16 and 9 percent respectively). Further, the survey teams were designed to match the caste profile of enumerators with those of the majority of interviewees. Of course, this strategy does have limits, given the caste heterogeneity of the Indian electorate, even at the local level.

C. Sensitive questions:
Of the questions asked on the survey that were actually used within the paper's analysis, the three that are the most potentially sensitive to response biases are those asking respondents about their vote choice, their income, and their views on Hindu nationalist ideology.

C1. Electoral Preferences:
In terms of ensuring that people's political preferences remain confidential, the survey team did not ask direct questions about people's vote choice. Instead, each respondent was given a blank dummy ballot that replicated the design of those used by India's Electronic Voting Machines, and placed their ballot choices in a sealed box. The ballots were matched with the rest of the survey instrument using anonymous serial numbers printed on both. This procedure helps minimize concerns about systematic response biases in political responses, and is more careful than those used on a number of global surveys. For example, both the Afrobarometer and World Values Survey ask respondents to openly report their electoral preference to the enumerator.
C2. Income: Asking respondents their income prompts concerns that respondents will be nervous about declaring their income, specifically if they are worried that such information is being collected for a government agency (for tax purposes). Poor respondents working in the informal sector might feel particularly insecure about such questions. For this reason, the NES asks about monthly household income at the very end of the survey questionnaire, so that this item does not induce non-responses for subsequent questions. However it is worth noting that only 186 respondents out of the sample of 20,204 that I worked with did not respond (<1%) to the income question. This proportion was not different amongst the Dalit and Adivasi subsample (51 out of 5460, <1%), suggesting this was not a huge concern. Of course it is still possible people merely underreport their income (rather than not responding to the question).
However, the percentage of Dalit and Adivasi respondents to the NES living in poverty did appear to match official figures. In 2004-05, the rate of poverty (according to India's low official poverty line of $10/15 per month in rural/urban India) was 36.8% (rural) and 39.8% (urban) for Dalits, and 47.7% (rural) and 33.9% (urban) for Adivasis. The rate among all other castes was 17.86%, and will likely be far lower among upper castes (Planning Commission 2011: 116). These figures matched those in the 2004 National Election Study used here, in which 41% of Dalit respondents and 40.02% of Adivasi respondents came from households earning less than $1 a day (all other castes: 18.12%).
In addition, I created an index of asset ownership, based on assets the enumerator observes the household to have, which are less susceptible to underreporting. Using an asset index (constructed via component analysis) does not affect any of the major results with either elite or non-elite subsample.
C3. Support for Hindu Nationalism: The final question of concern was the measure of Hindu nationalist ideology, which asked respondents the degree to which they agreed with the demand to exclusively build a temple at the site of a mosque demolished by Hindu nationalist activists in 1992. However, if respondents had felt uncomfortable with responding to this question, we would have expected a large number of responses within the 'don't know' category. However, a vast majority of respondents expressed a firm opinion on this question (with 77.26 responding, and only 22.74 percent in the 'don't know' category'. Further, over 70% of Dalits and Adivasis provided active responses. Further, within these active responses, each opinion category is well represented (with 36 percent of respondents disagreeing, while 64 percent agree). Nor did the rates of active responses differ substantially between those who supported the BJP (77%) and those who didn't (70%).
In addition, there may have been a specific concern that Muslims would not feel comfortable responding openly to this question, yet 84.96% of all Muslims and a similar proportion of poor Muslims (81.43%) provided an active response. One final concern might be that Muslims feel pressure to falsify their preferences and rate Hindu nationalist agendas favorably. However, the survey reports the vast majority of Muslims disagreeing with this agenda (68%), which would appear to follow conventional expectations of how this community would feel on this specific issue.
This evidence should help mitigate concerns of asymmetries in response rates among members of any particular constituency, which would be the primary problem for the present analysis. Of course, as with any survey, it is impossible to definitively determine the level of bias within response patterns, as we cannot observe the 'true' values of each response category. However, the evidence provided above does mitigate concerns that such biases are substantial enough to affect the key findings of the analysis.

D. National versus State Surveys?
Using a national election dataset instead of a compilation of state assembly election data was a major decision with significant ramifications for the research design, and hence requires brief explanation. There would have been obvious advantages to using Vidhan Sabha [State Assembly] data for this study. Under the Indian federal system, state governments possess a great deal of autonomy over a variety of public policies, and for this reason state elections are often more politically charged for local populations than Lok Sabha elections. Further, given that I am specifically looking at disadvantaged populations for whom the most important public services are those controlled by state administrations, there is a strong rationale for examining their political behavior when voting in assembly elections.
Having said this, there were stronger reasons for employing national parliamentary data that ultimately led to the decision to use the 2004 Lok Sabha survey. The biggest concern was the lack of comparability between separate state assembly surveys, as the questions on each differed greatly. This variance made it impossible to aggregate responses across the state-level surveys for more than a small handful of questions. Responses to questions crucial for this analysis were thus either unavailable or unevenly measured across states. Further, since state assembly elections happen by rotation across the country, the responses were not collected synchronously, compounding problems of comparability. Finally, the methodologies used on the various surveys differed somewhat in their techniques of sampling, including their methods for replacing missing respondents. Given these concerns, the national election data appeared to offer the best opportunities for systematically examining support patterns for the BJP across India.

Conclusion:
In conclusion, it is worth mentioning that the NES 2004, like any survey, is far from perfect. In particular the proportion of female enumerators remains inadequate, and the post-survey quality checking protocol is not transparent, and requires better scrutiny. However, these flaws suggest room for improvement, not discarding the data altogether. This survey remains the most reliable electoral survey in India, and represents a tremendous effort to grapple with the unparalleled complexity of India's vast democracy in the fairest possible manner. In terms of its multistage design, sample representativeness, enumerator diversity (in terms of caste), and instrument design (including the use of dummy ballots), the NES is not only adequate, but arguably superior to many of its better known counterparts (including the Afro-and Asian Barometer, and World Values Survey).

PART II: STATE-LEVEL ANALYSIS SECTION 9: STATE-LEVEL BASIC STATISTICS AND TRENDS
This first section of Part II presents descriptive information on the state-level analysis on religious welfare in India. Table S.31 presents detailed information on the measures used in the state-level regression analysis, the concepts each measure seeks to represent, and the sources from which they were constructed (including the wordings of the relevant survey questions).
Next I include information on the spread of religious welfare across India. Table S.32 and Figures S.4-S.5 provide additional information on the cross-sectional and temporal spread of service chapters across India from the internal records of Hindu nationalists. The figures further highlight the rapid proliferation of Sangh service wings across India over the past two decades, which supports my contention that this is a relatively recent approach.
Next, Table S.33 provides a simple descriptive disaggregation of shifts in the BJP's performance between 1996 and 2004. The table shows the party made substantial gains with non-elite voters in seven major Indian states, but continued to struggle in the remaining ten. In addition, Figure S.5 shows states where service networks were most rapidly expanded between 1996 and 2004 were also those in which the BJP made the most dramatic gains with poor voters during this same period.
Finally, Figures S.6 and S.7 present descriptive evidence refuting alternative explanations of the BJP's success. Figure S.6 shows the weak relationship between the BJP's representation of Dalit and Adivasi candidates and non-elite support for the party even when removing the outlier case of West Bengal. Figures S.7A and B shows a similarly weak relationship between communal conflict and BJP vote share, especially when removing the outlier caste of Gujarat.      This section addresses some additional robustness checks for the state-level analysis presented in Table 6 of the main text in three main parts:

A. Basic checks
First Table S.34 shows that collinearity between the explanatory variables is not an issue within the state-level variables. Table S.35 then examines if the relationship between service network strength and state-level BJP performance is robust to models where the measures of these two factors are not log-transformed (as they are in the models presented in Table 6 of the main text). The results indicate the key relationship holds even with this alternative specification across both random and fixed-effects specifications, and with a lagged dependent variable.

B. Causal sequencing
Next, I address issues of causal sequencing: Column 1 of Table S.36 presents the results for the lagged DV model without the full battery of controls. This spare model might be preferred by some readers, given concerns about limited degrees of freedom are heightened by including a lagged dependent variable (which shrinks the sample from 47 to 32 units).
A second concern is that about the sequencing of the organizational and electoral date. The data for service organizations was collected in 1995, 1997, and 2004, while the electoral data is from 1996, 1999, and 2004. Thus the independent variable is in fact observed prior to the electoral data for the first two periods in the panel, but not for 2004. Model 2 therefore tests the results without the 2004 data and finds the results still hold in this specification.
However, omitting the 2004 data also attenuates the sample to 30 units. Another strategy is to match 2004 welfare data with electoral data from after the 2004 election. Unfortunately the NES data from India's next parliamentary election (in 2009) is still not available to outside researchers. However, while the raw data remains unavailable, some of the aggregate data on support for the BJP from Dalit and Adivasi voters was published in secondary sources. From this data, I was able to calculate the BJP's vote share from these voters in 2009 for 15 of the 17 major states used in this analysis (see footnote for the procedure used for such calculations). 17 This data allows me to test a specification where welfare data precedes electoral data in each period of the panel (see footnote). 18 Unfortunately, without the raw NES data, I am not able to calculate a number of the key control variables for 2009 (Elite support for the BJP, Ethnic voting, etc). Consequently, I estimate these models with state fixed effects (Model 3).
The results in Model 3 show that the welfare index again significantly correlates with non-elite support for the BJP in this specification (also see Figure S2.3). However I should reiterate that the results for Model 3 are forced to utilize estimates for the 2009 electoral data, and hence should be read as provisional rather than conclusive evidence.
Finally, Model 4 reverses the specification and regresses levels of Hindu nationalist welfare provision on prior levels in the BJP's vote share. If prior BJP performance correlates with subsequent concentrations of welfare, the causal logic of my argument is violated, and our concern with endogeneity would be heightened. With the service data from 1997 and 2004, and electoral data from 1996 and 1999, we can construct a panel of 30 units for which electoral data precedes service data (see fn). 19 Since we have measures of all relevant control variables for these years, we can estimate the standard specification used for this time-series data. The results (model 4) are revealing: prior BJP electoral performance does not appear to correlate with subsequent concentrations of Sangh welfare.
To summarize, the tests in Models 2-4 suggest that while the prior density of the welfare index correspond to subsequent BJP performance among non-elites, the prior level of BJP performance does not correlate with subsequent density of Hindu nationalist welfare.

C. Further Selection Effects
Finally, I further explore potential selection effects in where welfare chapters were built in the first place. Section 7A of Part I of this supplement examined whether the impact of welfare extended to poor Muslims and Christians and found evidence it did not. However, even if individual level evidence finds no electoral impact of welfare, religion might still play a role in shaping where services are provided in the first place. For example, we might ask whether religious proselytism or outreach from Christians or Muslims affect the BJP's welfare distribution. Data from the original version of this paper suggests there is some impact of the religious composition of Indian states. Model 1 in Table 6 of the main text shows that the percentage of Christians (but not of Muslims) in a state's population did register a positive impact on where welfare chapters were located, although this selection effect did not confound the impact of service wings on BJP vote share.
However, one might argue that Hindu nationalists were less concerned with the overall numbers of Muslims and Christians, and more specifically with the number of Muslims and Christians within a state's lower caste and tribal population. This is because, in the eyes of Hindu nationalists, these non-elite populations are part of a Hindu community, but have proven ripe conversion targets for Christian and Muslim proselytizers. Hindu nationalists worry that the marginalized status of Dalits and Adivasis within Hinduism has provided an opportunity for rival faiths to prey on these communities. Rightly or wrongly, the presence of Dalit and Adivasis Muslims or Christians might serve as a signal to Hindu nationalists of the proselytizing efforts of these rival faiths.
This perception, in turn, might lead Hindu nationalists to build welfare chapters in areas with large Dalit and Adivasi religious minority populations. Table S.37 therefore replaces the measure of a state's Muslim population with a specific measure of the percentage of Dalits and Adivasis within a state who are Muslim (using data from the NES survey sample). 20 Second, Hindu nationalists are particularly concerned with Christian missionary efforts among Adivasi (tribal) populations specifically. 21 I therefore replace the measure of a state's Christian population with a specific measure of the percentage of Adivasis who are Christians. However, neither measure significantly correlates with the density of welfare chapters.
A final measure examined whether welfare chapters were concentrated in states where Christian missionaries have been historically active. I obtained data on the number of Protestant missionaries within each of the 17 major Indian states as of 1925 from the dataset used by Woodberry (2012). However, this measure did not appear to significantly correlate with where Hindu nationalists subsequently built their welfare networks ( Figure  S10).

PART III: THE ARGUMENT BEYOND INDIA: THE CASE OF ISLAH IN YEMEN
The main text (pp. 36-37) reports the results of an analysis of support for Islamist politics in Yemen, which uses responses from a recent Arab Barometer survey conducted in 2006. While the survey did not report the specific electoral preferences of respondents, it did ask them whether they thought it 'suitable to have a parliamentary system in which only Islamic political parties and factions compete.' 22 While far from a perfect measure, it seems reasonable to assume that Yemeni respondents who favor a system exclusively comprised of Islamic parties are more likely to support the Islamist Islah party than those who do not.
The survey also recorded respondent participation in civil society associations, in a manner roughly similar to the Indian surveys used in this book. A little less than a third of Yemeni respondents identified as members of formal non-party organizations. Table S.38 presents results from logistic regression models testing the influence of associational activity on a voter's preference for a system comprised solely of Islamic parties. The tests include a range of control variables derived from the survey (including their piety, the influence of ethnic leaders, and demographic factors), which attempt as far as possible to mimic similar controls from the study of the BJP in India. The analysis reveals that among poorer voters (defined as those self-reporting in the bottom six income deciles), associationally active members were in fact significantly more likely to voice support for an exclusively Islamic party system than non-members. The substantive effects of organizational inclusion were also substantial, doubling the likelihood of such a preference being expressed (from 14 to 30 percent). Conversely, a respondent's religiosity (measured as the frequency of reading the Quran) did not significantly influence their opinions on this issue.
Equally interestingly, these effects were inverted within the pool of relatively privileged Yemeni respondents. Within this pool of voters, organizational inclusion is negatively and insignificantly correlated with support for Islamic parties. However measures of piety do appear to correlate very strongly with elite political preferences. A comparatively elite Yemeni who reads the Quran everyday is over 30 percentage points more likely to voice support for such a system than one who rarely does so. 23 Thus, the above analysis reveals a strikingly similar divergence in how organizational inclusion and cultural values affect elite and non-elite political preferences in India and Yemen. 22 Question 246, part 2 on the 2006 Arab Barometer Survey Instrument. The survey was conducted in Jordan, Palestine, Algeria, Morocco, Kuwait and Yemen. 23 The likelihood increases from 11 to 41 percent. These effects of piety and membership do not change if we restrict the sample to the top one, two, or three income deciles.