WPS8120 Policy Research Working Paper 8120 Unheard Voices The Challenge of Inducing Women’s Civic Speech Ramya Parthasarathy Vijayendra Rao Nethra Palaniswamy Development Research Group Poverty and Inequality Team June 2017 Policy Research Working Paper 8120 Abstract Deliberative institutions have gained popularity in the devel- on women’s civic participation in rural Tamil Nadu. Using oping world as a means by which to make governance more text-as-data methods on a matched sample of transcripts inclusive and responsive to local needs. However, a growing from village assembly meetings, the analysis finds that the body of evidence suggests that persistent gender inequality Pudhu Vaazhvu Project significantly increases women’s may limit women’s ability to participate actively and influ- participation in the gram sabha along several dimensions ence outcomes in these forums. In response, policy makers —meeting attendance, propensity to speak, and the length have tried to induce women’s participation by leveraging of floor time they enjoy. Although women in the Pudhu the group-based format of self-help groups, which can build Vaazhvu Project villages enjoy greater voice, the study finds women’s social capital and develop their sense of political no evidence that they are more likely than women in con- efficacy and identity. This paper evaluates the impact of one trol villages to drive the broader conversational agenda such intervention, known as the Pudhu Vaazhvu Project, or elicit a relevant response from government officials. This paper is a product of the Poverty and Inequality Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at vrao@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Unheard Voices: e Challenge of Inducing Women’s Civic Speech∗ Ramya Parthasarathy† Vijayendra Rao‡ Nethra Palaniswamy§ JEL Codes O12, C49, D02, D70, J16 Key Words gender, deliberation. village democracy, India, text-as-data, participation ∗ is paper is a product of the World Bank’s Social Observatory. Financial support from the contributions of (1) UK Aid from the UK government, (2) the Australian Department’s of Foreign A airs and Trade, and (3) the European Commission (EC) through the South Asia Food and Nutrition Security Initiative (SAFANSI), which is administered by the World Bank, is gratefully acknowledged. e authors are indebted to R.V. Shajeevana, the former Additional Project Director of the Pudhu Vaazhu Project, for her advice and assistance; Kevin Crockford and Samik Sundar Das for their support; as well as Madhulika Khanna, Nishtha Kochhar, Smriti Sakhamuri, G. Manivannan, and GFK-Mode for their help with the eldwork. e authors also thank Avidit Acharya, Lisa Blaydes, Nick Eubank, Adriane Fresh, Justin Grimmer, David Laitin, Jeremy Weinstein, and participants of the Indian Political Economy working group in Washington, D.C. for comments and suggestions. e views expressed here do not necessarily re ect the UK, EC, or Australian government’s o cial policies or the policies of the World Bank and its Board of Executive Directors. † Dept. of Political Science, Stanford University. ramyap1@stanford.edu ‡ Development Research Group, World Bank. vrao@worldbank.org § Poverty Global Practice, World Bank. npalaniswamy@worldbank.org 1 I Despite formal guarantees of political equality, women across the globe are systematically under- represented in politics — whether that be elected o ce, bureaucratic posts, or everyday political participation. Women still constitute only 23.3 percent of parliamentarians (Inter-Parliamentary Union, 2017), even with the growing use of gender quotas (Krook, 2010). While women’s voter turnout rates have improved substantially across richer democracies, women are still less likely to make demands of government o cials (Karpowitz and Mendelberg, 2014) or to participate in costlier forms of political activity, like rallies, campaigns, and protests (Paxton et al., 2007). at women lack a voice in their governance is normatively problematic in its own right (Mansbridge, 1999; Sen, 2001); just as troubling, women’s absence from political life may have substantive con- sequences for policy and development outcomes, given their di ering policy preferences (Edlund and Pande, 2002; Inglehart and Norris, 2003; Miller, 2008). e dearth of women’s participation has been especially acute in developing nations like India, where the recent resurgence of deliberative democratic institutions has made the exercise of political voice that much more important (Mansuri and Rao, 2012). ese deliberative institutions, largely formed via decentralization e orts and community-driven development programs, are premised on the idea that development can be made more inclusive and be er tailored to local needs by moving decision-making from government o ces to the village itself. ese calls for participation, however, can be especially problematic for women, who o en face social costs for speaking in public, are usually less informed, and lack a sense of political e cacy (Dreze and Sen, 2002). Indeed, the extant evidence from Indian local government, or panchayati raj, shows that women are less likely to a end local village meetings, or gram sabhas (Ban and Rao, 2008b; Cha opadhyay and Du o, 2004), to participate in community resource management (Agarwal, 2001), and to run for local o ce. Recognition of these deep gendered inequalities has prompted Indian policy makers to ac- tively design deliberative institutions with social inequalities in mind (Parthasarathy and Rao, 2 2017), most notably through the use of quotas for women on village councils. Reservations, as they are known in the Indian context, have had promising results for a number of outcomes, including the delivery of women-preferred public goods (Cha opadhyay and Du o, 2004), the aspirations of young girls (Beaman et al., 2012), and gender bias among voters (Beaman et al., 2009). However, evidence that the mere presence of a female incumbent is su cient to achieve parity in participation, let alone deliberation, between citizens of both sexes is much weaker (Ban and Rao, 2008b; Cha opadhyay and Du o, 2004; Beaman et al., 2010; Parthasarathy et al., 2017).1 Indeed, Parthasarathy et al. (2017), examining deliberation in the control villages of this study, document that while female incumbents may be more likely to respond to women constituents, their presence has no discernible e ect on women’s a endance, frequency of speech, or length of oor time. As an alternative approach, the Government of India and various states have instead tried to induce women’s participation from the “bo om-up” — by building women’s organizations via a system of self-help groups (SHGs). ough the central aim of these groups has been to provide rural women with greater access to credit and livelihoods, it is also hoped that the group-based format of SHGs builds social capital, with implications for women’s sense of political e cacy and identity (Sanyal et al., 2015; Sanyal, 2014; Prillaman, 2016). is paper evaluates the e ect of one such bo om-up intervention, known as the Pudhu Vaazhvu Project (PVP), on women’s civic participation in rural Tamil Nadu. PVP is a participatory, community-driven development project implemented by the Government of Tamil Nadu that works in the poorest regions of the state. Like other SHG programs, the core economic interventions of PVP are centered on credit and livelihoods support for women that belong to project-facilitated self-help groups. In addition, however, PVP creates explicit linkages among SHGs within the village and by partnering with local government to implement credit access and job-training activities in an e ort to create social capital and improve women’s capacity to address public expenditures. 1 (Beaman et al., 2010) provides evidence for the e ect of reservation on women’s a endance and participation in gram sabhas from ve states. ey nd that women’s a endance is una ected by reservations, but do nd a positive e ect on whether women speak. While the la er results are encouraging, they are focused on the incidence of women’s speech, rather than the volume of speech or even parity in oor time with men. 3 is paper examines whether Tamil Nadu’s PVP program induces women’s participation within the gram sabha. We use text-as-data methods on a matched sample of transcripts from village assembly meetings to examine the e ects of the program on political speech at these meetings. We nd that PVP signi cantly increases women’s participation in the gram sabha along numerous dimensions — meeting a endance, propensity to speak, and the length of oor time they enjoy. Our estimates show that the PVP program nearly doubles the number women who come to the gram sabha, and boosts their frequency of speech by nearly 45 percent. is is not to say the results are all positive; we nd that women in PVP villages are no more likely than women in control villages to drive the broader conversational agenda or elicit a relevant response from government o cials. Nevertheless, these substantively signi cant gains suggest that policy interventions can have a positive impact on what has o en been thought of as something beyond the reach of small-scale interventions: shaping social norms around gender. is analysis represents one of the rst quantitative analyses of self-help groups that mea- sures objective outcomes rather than self-reports. In doing so, it not only contributes to the small but growing body of work on the political impact of self-help groups, which has qualitatively shown largely positive results (Sanyal et al., 2015; Sanyal, 2014; Desai and Joshi, 2014), but also provides a more rigorous foundation for conclusions drawn from studies based on self-reports of empowerment like Prillaman (2016) in Madhya Pradesh. In Tamil Nadu, Khanna et al. (2015) use household survey data from the same villages as this paper and nd very similar results: women’s participation in SHGs enhances their intra-household bargaining power and their ca- pacity to participate in the public sphere; but, just as in Prillaman (2016), these results are based not on direct eld observations, but on outcomes reported by respondents. As such, we might be concerned that responses are biased by project rhetoric that have imbibed, rather than actual political behavior. To overcome this challenge, we directly examine the e ect of PVP on women’s participation using our village assembly rosters and transcripts. Indeed, in Khanna et al. (2015), which uses survey evidence from the same villages studied here, women’s self-reported a en- dance at the gram sabha is higher than the direct measures collected here, both in control and 4 treatment villages. Our direct measurement approach not only saves us from overreliance on these self reports, but also allows us to measure whether their political speech has deliberative in uence on fellow citizens and state o cials. In focusing explicitly on the speech pa erns of citizens, we also contribute to a growing lit- achtiger et al., 2005; Karpowitz and Mendelberg, erature on the empirical study of deliberation (B¨ 2014; Heller and Rao, 2015). While deliberative democracy has traditionally been the domain of normative political theorists (Habermas, 1990; Elster, 1998; Mansbridge, 1980; Gu man and ompson, 2004; Fung, 2004), scholars have increasingly tried to examine whether deliberative institutions deliver on the hopes of normative theorists. To that end, our study draws on norma- tively grounded measures of good deliberation (Mansbridge, 2015) to unpack not only the ways in which gender may a ect citizen participation, but also the types of policies that may be able to ameliorate such inequality. More speci cally, we use the methods and measures developed in (Parthasarathy et al., 2017) and focus on the political and ethical functions of deliberation. Under this conception, deliberation allows all participants to have an equal opportunity to in uence the outcome; embodies the ideal of mutual respect, whereby citizens listen a entively to one an- other; and allows citizens to be agents who participate in the governance of their society. While Parthasarathy et al. (2017) validates these measures to describe deliberative inequality in Tamil Nadu, here, we use these measures to evaluate the impact of a policy intervention on both sides of the deliberative coin — that is, not only whether citizens are able to speak, but the extent to which they are heard. Finally, this study speaks to policy makers keen on understanding the unintended conse- quences of external interventions on local governance. With the proli c growth of aid institutions and non-governmental institutions in the developing world, practitioners and policy makers alike have grown acutely aware of the ways in which external interventions may alter local commu- nity dynamics in unforeseeable ways (Gugerty and Kremer, 2008; Mansuri and Rao, 2012; Bano, 2012). In this paper, we document the ways in which inducing participation may help to amplify the voices of women in rural governance, but also shi s discourse away from the organic topics 5 raised by citizens and towards project-speci c activities. Given the nite amount of time to con- duct local assemblies, this may have the perverse e ect of crowding out discussion of issues that are broadly relevant to the community. e remainder of this paper is organized as follows: In Section 2, we describe the institutional and cultural context in which we study women’s deliberation, as well as the intervention aimed at inducing their participation. In Section 3, we describe our research design, data, and measures. In Section 4, we present our results for how PVP a ects women’s deliberation; Section 5 discusses the implications of these ndings, and Section 6 concludes. 2 I C C 2.1 Local Governance and Deliberation e institutional context in which we study deliberation is the gram sabha, or village assembly, which serves as the primary forum for citizens in rural India to demand accountability and ac- cess to public goods from local government o cials. It was created by the 73rd Amendment, which transferred responsibility for the delivery of local public goods and services to a three-tier local government. Under the constitutional mandate, all Indian villages are to be governed by an elected council, composed of ward members (representing roughly 500 people each), and a president. In recognition of historical disadvantage for women and low castes, the amendment also mandated that 33 percent of seats in village councils would be reserved for women, and a number proportionate to their population in the village reserved for disadvantaged castes. Lastly, the amendment mandated that all citizens would have the opportunity to deliberate and advise the elected council on relevant development decisions at least two times a year via a village-wide assembly, or gram sabha. ese two features — reservations for historically disadvantage castes and women, as well as the gram sabha — aim to provide an institutional check on elite domination by ensuring that all citizens have the ability to in uence development decisions. Reservations do so by explicitly 6 mandating that citizens from these underrepresented groups occupy elected o ce, and the gram sabha opens up governing decisions to public scrutiny via a deliberative forum for all citizens to a end. While there has been considerable scholarship on the distributive consequences of reser- vations, relatively li le work has examined the impact of these policies on political voice within the gram sabha itself. e evidence we do have suggests that men tend to dominate in terms of participation, and that the issue priorities of large landowners tend to take up more time within the assembly (Ban and Rao, 2008a). Despite these inequalities in participation, evidence suggests that gram sabhas tend to be democratically e cient, in the sense of re ecting the preferences of the median household (Ban et al., 2012); however, there can be a large degree of inequality of voice within households, so household preferences may simply re ect the preference of males. Indeed, women are much less likely to be aware of gram sabhas and less likely to a end (Chhib- ber, 1999; Besley et al., 2005). Since the gram sabha is an important site for citizens to demand accountability in public service delivery, these inequalities in participation may have profound consequences for citizen welfare and access to basic goods. 2.2 Women’s Status in Tamil Nadu at women are less likely to be aware of, present for, or active in the gram sabha is not sur- prising in the larger global context. Indeed, the realm of politics has been a particularly “sticky domain” for the gender gap (World Bank, 2011). e dearth of women’s politiacl activity re ects the complex and inter-related set of constraints that have limited women’s agency — from social norms about women’s roles and abilities to their limited social networks and paucity of resources both inside and outside the household. ese barriers have been well documented in the Indian context (Du o, 2012; Chhibber, 1999), where women have been largely absent from high tiers of elected o ce (they constitute only 7.8 percent of parliamentary candidates and 11.23 percent elected Members of Parliament, for example) to local, participatory institutions for ordinary cit- izens (Cha opadhyay and Du o, 2004; Beaman et al., 2010; Ban and Rao, 2008b). In Tamil Nadu, where this study is located, women have been shown to have relatively more 7 autonomy than women in other parts of rural India (Dyson and Moore, 1983; Kishor and Gupta, 2009); yet even in Tamil Nadu, women’s standing is far more complex than this optimistic account would suggest.2 It is not that patriarchy is less acute than elsewhere, but it is di erently expressed and reinforced. For example, Mines’s (1994) ethnography of private and public identity in Tamil Nadu shows that, while men in this state value and nurture a distinct civic individuality, this is not observed among women, whose sense of self derives from their role as wives, mothers, and daughters-in-law. Similarly, Kapadia et al. (1995), in her classic ethnography of Tamil women, demonstrates that among low-caste women (who are the vast majority of the female participants in the meetings we study), the seemingly high degree of female autonomy is deceptive, as it is con ned to extended family, rather than in interactions outside the family. Kapadia a ributes this to practices of marital endogamy (the practice of marrying close-kin), which create an on- going relationship between women and their birth kin a er marriage for two reasons: (a) their physical proximity, and (b) because extended families tend to have marriage ties over several generations and thus have very strong bonds. Moreover, recent improvements in education and labor market opportunities have bene ted men much more than women; this change in class has caused kinship ties to break and women to be even more restricted within the home. ough these ethnographic accounts were wri en two decades ago, recent data reinforces the distinct delineation between genders across the public and domestic spheres. Labor force participation rates for rural men in Tamil Nadu are 59.3 percent for men and 31.8 percent for women.3 An analysis of survey data from the same sample as the villages we study in this paper, Khanna et al. (2015) shows that 47 percent of married women reported that they were the primary decision makers in household decisions on durable good purchases, but only 12.5 percent reported that they a ended the last village assembly, or gram sabha. http://rchiips.org/nfhs/a subject report gender for website.pdf 2 Directorate of Census Operations, Government of Tamil Nadu, http://www.tn.gov.in/dear/ 3 Employment.pdf 8 2.3 Inducing Participation through the Pudhu Vaazhvu Project at the gram sabha speci cally is viewed as domain of men is not at all unique to the Tamilian context, but re ects the broad pa ern of gender norms that limit women’s agency in India. In response to this bias, the Government of Tamil Nadu has tried to empower women in part via the creation of Self-Help Groups (SHGs). e SHG movement in Tamil Nadu, which initially focused on reducing the economic vulnerability of women through credit, livelihoods linked economic resources, and training, began in the 1990s and was consolidated by the state under the Mahalir i am initiative in 1997-1998. e focus on women’s economic standing re ected global trends in women’s empowerment at the time, which saw access to economic development as a key lever to improve women’s agency rst within the home, and then within the community writ large. Despite the success in scaling up this initiative, however, the SHG movement continued to exclude the truly poor in Tamil Nadu; moreover, there remained an open question as to whether these institutions could support women’s civic action in the absence of explicit linkages both among various SHGs and between the SHGs and local government (Khanna et al., 2015). at is, while SHGs provided women with hyper local networks within their neighborhood, they pro- vided few opportunities for broader collective action, let alone the types of civic “training” that might help women gain the self-con dence and sense political e cacy necessary to participate in the gram sabha. Given these challenges, the Pudhu Vaazhvu Project (PVP) was explicitly designed to (a) make SHGs more inclusive, (b) support the institutional development of a village organization that would link them to credit and other sources, and (c) work closely with elected village-level gov- ernment. e core institution through which PVP achieves these ends is through the formation of a Village Poverty Reduction Commi ee (VPRC), which is composed of a federation of SHGs within the village. e VPRC’s central mandate includes credit and livelihoods, but it places sig- ni cant emphasis on several other activities, including: helping the poor to access various safety nets and social services provided by the state and central governments (e.g. India’s National Rural Employment Guarantee scheme, old age and widow’s pensions, and housing schemes); assisting 9 with the targeting of grants to the poor and disabled; and facilitating access to skilled employ- ment through youth training and job fairs. e membership of the VPRC typically contains 10 - 15 members, who are chosen to represent SHGs from each habitation, or neighborhood, within the village. PVP was initially launched in 2005, in 2,300 village panchayats (VPs) drawn from 70 blocks (a sub-district administrative unit that is made up of a cluster of VPs) in 16 selected districts of Tamil Nadu.4 e districts were chosen using a combination of objective poverty criteria, as well as other factors that captured the relative development of the district (e.g. infrastructure). Within each district, blocks were chosen on the basis of a poverty (or “backwardness”) score that included the number of households below the poverty line and the population of socially disadvantaged groups, the Scheduled Castes and Tribes (SC/STs). All villages within selected blocks were eligible to receive the program, and take-up was universal. Within each village, a set of households identi ed through the participatory identi cation process formed the core target population for the project, and were eligible to receive the targeted credit, livelihoods, and training services. For the purposes of this evaluation, however, our focus is on the village-level impact — that is, whether and how PVP’s focus on public action and inclusion a ects the quality and character of participation in gram sabhas. 3 R D 3.1 Village Selection In order to evaluate the e ect of PVP on the character and quality of deliberation, ideally, we should have randomized villages to receive the program. Since randomized assignment was not possible, we leverage our knowledge of program implementation to reconstruct the PVP selection process, thereby creating a matched sample of comparable treatment and control villages.5 More 4 Coimbatore, Cuddalore, Kancheepuram, Nagapa inam, Namakkal, Ramanathapuram, Salem, eni, iruvan- namalai, iruvalur, iruvarur, oothukudi, Tirrupur, Tirunelveli, Vellore and Villupuram. 5 e original evaluation design was based on a regression discontinuity design, in which ve or six blocks within each district would be chosen on the basis of a population score that re ected the level of backwardness of the block. 10 speci cally, within the set of eligible districts, blocks were selected for assignment based on two sets of criteria: (1) a population criterion that equally weighted the SC and the ST population proportions and the number of below poverty line (BPL) households from census data; (2) a set of block level infrastructural variables that measure the quality of infrastructure, public services and industrial backwardness. We generate our matched sample by matching project and non-project blocks within 9 ac- tive project districts6 on the two factors that determined assignment to treatment. Infrastructural variables included all available census data (from 2001, before the project started) that could mea- sure disadvantage — the number of villages in the block, average distance of the village to the nearest town, total population, percentage of villages in the block which had primary and mid- dle schools, commercial banks, cooperatives, agricultural and non-agricultural societies, medical facilities and drinking water facilities. is process allowed us to nearly replicate the original assignment process for PVP. We use a two-step matching procedure, summarized in Figure 1. First, we generate propensity- score matched blocks using a standard probit model that uses the variables listed above. Within each district, a PVP block was matched to the non-PVP block with the closest propensity score. is ensured that the chosen non-PVP block was as likely to receive the intervention as the ex- isting matched PVP block. Second, since the unit of analysis for this study is the village, we follow a similar process to identify speci c village panchayats (VP) within each matched pair of blocks. e variables used for this village-level matching are the same as those used for the block matching. us, the nally selected VPs from PVP and non-PVP blocks were ex-ante equally likely to receive the program. is two-step sampling strategy ensures pre-treatment similarity on observable covariates of treatment across treatment and control areas. However, in discussion with the implementing partners, it emerged that deviations from the rule occurred when the population score did not identify the most disadvantaged blocks that the project intended to target. In particular, the population criterion seemed, at times, to be leading to the selection of more developed and therefore arguably less poor blocks. While these changes ruled out using a discontinuity design, we combined the population criterion with other information capturing the reasons for deviation — namely, village-level infrastructure — to approximate the nal block selection criterion. 6 e sample districts were chosen to ensure representation from di erent geographic regions of PVP’s imple- mentation. 11 Figure 1: Illustration of Two-Stage Sampling Strategy PVP District Block A (PVP) Block B (PVP) Block C (non-PVP) Block D (non-PVP) p-score =0.11 p-score =0.07 p-score=0.12 p-score =0.0.13 VP 1 VP 1 p-score=0.56 p-score=0.67 VP 2 VP 2 p-score=0.39 p-score=0.55 VP 3 VP 3 p-score=0.53 p-score=0.57 Note: e gure above summarizes the two stage sampling construction. In the rst stage, within a selected project district, the existing PVP block (Block A) is paired with the closest non-PVP block (Block C). en, within each of these blocks, we identify matched pairs of villages, highlighted in blue (VP1 from Block A and VP3 from Block C) and red (VP3 from Block A and VP2 from Block C). e nal sample for this district will thus include four VPs from two blocks. A key assumption of propensity score matching (PSM) is that of conditional independence, which implies that program outcomes must be independent of treatment status prior to treatment, given a vector of observable covariates. While we cannot directly test for conditional indepen- dence, two facts provide con dence that we have met this bar. First, the covariates chosen for the matching procedure accurately re ect the true selection process for assignment to treatment. And second, we have a high degree of post-match balance on all observable covariates; Table 1 shows that, in 2001, the sample VPs were indeed similar on all relevant observables that possi- bly determined selection into the program. Given this, we can reasonably infer that the average di erence between the matched comparison units from treatment and control groups will yield a consistent estimate of the Average Treatment E ect on the Treated (ATT) (Rosenbaum and Rubin, 1983). A second key requirement for PSM is the existence of a region of common support, that is, for each value of a vector of observables X (or propensity score generated using X ), there is a positive probability of nding a comparison unit in both treatment and control groups. at is, 0 < P (D = 1 | X ) < 1 (1) 12 Table 1: Balance on Pre-Treatment Covariates Variable Non-PVP PVP Di . Norm. Di P-value No. of HH 657.871 736.042 -78.171 -0.143 0.443 Percent SC 0.378 0.343 0.034 0.173 0.569 Percent ST 0.012 0.010 0.002 0.063 0.525 Female Literacy Rate 0.592 0.573 0.019 0.229 0.591 I(Primary School) 0.980 1.000 -0.020 -0.200 0.421 I(Secondary School) 0.360 0.200 0.160 0.359 0.640 I(Health Center) 0.240 0.280 -0.040 -0.090 0.464 I(Hospital) 0.040 0.040 0.000 0.000 0.500 I(Clinic) 0.040 0.060 -0.020 -0.091 0.464 I(Medical Shop) 0.220 0.220 0.000 0.000 0.500 I(Big Gov’t Hospital) 0.040 0.020 0.020 0.116 0.546 I(Bank) 0.900 0.960 -0.060 -0.169 0.433 Note: e table presents di erences in means on relevant pre-treatment covariates between PVP and Non-PVP Villages. Following Imbens and Wooldridge (2008), normalized di erences and associated p-values are presented. e probability of being treated, which in our case is the probability of being a PVP village, lies between zero and one. Figure 2 shows that there is a good overlap in the propensity score distri- bution across project and non-project VPs. To impose common support, we limit the comparison to a sub-sample of observations where the propensity score is more than the minimum value in the treatment group and is less than the maximum value in the control group. For our data, the region of common support is given by (.074, .86). e nal village sample thus consists of 100 matched villages, 50 in control and 50 in treatment. 3.2 Data Collection From this matched sample, we collected two forms of data: (1) full audio recordings of the gram sabha, and (2) a standardized questionnaire to collect information on the a endance of citizens and local o cials, on the nature of issues raised by citizens, and demographic data on who raised these issues (gender and caste). is survey data also included a roster of state and local govern- ment o cials in a endance, how information on the timing of the gram sabha was communicated, the physical location of the assembly, and a endance at regular intervals. e audio recordings 13 Figure 2: Common Support across Selected Treatment and Control VPs Note: is graph plots the density of propensity scores for 268 villages across within the 18 matched blocks of the sample. e region of common support is given by (0.226, 0.688). A er imposing common support, we choose a matched sample of 50 treated and 50 control villages with the closest propensity score matches. of meetings were transcribed and translated into a corpus of textual data by an independent sur- vey rm. Transcripts included verbatim transcriptions and translations of the assemblies, as well identi ers on the gender and position of each speaker.7 Each “document” in the corpus consists of an uninterrupted speech by an administrator, elected o cial, or citizen. From the 100 village assemblies, we have 3,959 such documents, 2,223 in treatment and 1,736 in control, each of which is identi ed by the position and gender. Table 2 presents descriptive information about the number and character of documents within each village. Assemblies have relatively good a endance (with 163 people a ending on average), and consist of roughly 40 speeches, of which one-third are made by women. Citizens deliver just over half (54 percent) of speeches, with the remainder distributed between administrators (29 percent) and politicians (16 percent). 7 e original data contain rich information on the position of each speaker, from school headmasters and ration shop owners, to elected o cials and administrators. For the purpose of our analysis, we code the speaker into three types: (1) administrators, who include all persons employed by the state or local government (e.g. panchayat secretary, block development o cer, school headmaster, village administrative o cer, etc.); (2) elected o cials, who include all persons who are in elected o ce (e.g. president, vice president, ward member); (3) citizens, all people who neither hold a formal government job or elected o ce. Within treatment in treatment areas, we also code for “activated” citizens, who were a liated with PVP. 14 Table 2: Village-Level Summary Statistics Mean Std. Dev. Median Min Max. Total A endance 163.896 114.641 124.000 25.000 720.000 Number of Speeches 39.590 28.296 31.000 4.000 172.000 Speech Length 100.915 118.795 75.036 25.600 1090.750 Percent Female 0.351 0.179 0.344 0.000 0.920 Percent Citizen 0.542 0.138 0.539 0.190 0.879 Percent Admin 0.294 0.150 0.285 0.000 0.750 Percent Politician 0.164 0.163 0.121 0.000 0.537 3.3 A Text-as-Data Approach to Deliberation While these descriptive statistics provide an initial picture as to who speaks within the gram sabha, we examine the nature of deliberative in uence using a text-as-data approach to the doc- ument transcripts. More speci cally, we use recent a computational tool known as unsupervised topic models to inductively “discover” a set of salient topics within the document collection, as- sociate those topics with each document and speaker, and examine pa erns of speech within each assembly. ough this approach will never fully capture the nuanced and complex nature of human conversation, it can help us to uncover underlying features of our data without imposing our own assumptions about the set of categories or issues that are discussed. Prior to estimating the topic model, we pre-process the set of 3,959 documents such that infrequent words (those with fewer than 5 occurrences in the corpus) and certain proper nouns, as well as overly common “stopwords” are removed.8 Infrequent and proper nouns are o en names of bene ciaries, townships, or neighborhoods that are mentioned in meetings, but are not in common usage. e remaining terms are then “stemmed” such that various forms of the same word are counted together.9 We also exclude numbers. From the original set of citizen speeches, 3,894 documents remain a er processing. Using this processed corpus, we adopt the approach of Roberts et al. (2016) to estimate a Structural Topic Model (STM), which allows us to inductively discover topics, or clusters of words 8 Stopwords are overly common words which are ltered out before the use of natural language processing meth- ods to improve the estimation process. ey o en include functional words, including articles, prepositions, basic verbs such as “is,” and pronouns. 9 For example, “requesting,” “requested,” and “requests” will all be stemmed to their root word “request.” 15 that commonly co-occur within the data. e model outputs (1) a set of topics, which are de ned as mixtures of words, where each word has a probability of belonging to each topic, and (2) for each document analyzed, the proportion of the document associated with each topic. As such, each document is can be characterized by a vector of proportions, representing the share of the document associated with each topic. Using STM, we identify a set of 25 topics10 discussed within the gram sabhas, and explore how these topics vary with the identi able characteristics of speakers and villages — including the gender of the speaker, the position of the speaker, and the reservation status of the village council president (female and/or Scheduled Caste). e generated topics are presented in Appendix Table A.1, which lists the highest probability words in each topic, as well as the FREX words, which are both frequent and exclusive, thereby identifying the words that distinguish topics. We also validate these topics in Appendix 3A using two tests of predictive validity. Below, Figure 3 presents the distribution of these topics across the full corpus. 3.4 Measures Having interpreted and validated the topic model output, we now turn to our measures of delib- erative participation and in uence — that is, whether women are able to speak and how well they are heard. As a measure of their participation, we estimate the e ect of PVP both on women’s a endance (measured in raw numbers and as a percentage of female voters), as well as the fre- quency and volume of speech. For measures of frequency, we examine the share of all speech delivered by women, as well as the share of female speeches among only citizens (excluding politi- cians and administrators). As a measure of volume, we look at the length of speeches — in terms of the number of words — to capture the amount of oor time enjoyed by women versus men. Collectively, these measures capture the extent to which PVP encourages women to be present and active participants in the civic space. 10 Since this method assumes a xed, user-speci ed number of topics, we rst assess the relative performance of models under a range of values (K ∈ 5, 50), and choose K = 25 for the preferred speci cation. is speci cation performs relatively well on a number of empirical tests (residuals t, held-out likelihood, semantic coherence, and exclusivity of topics), and yields topic clusters consistent with our substantive understanding of village assembly discussions. We also re-ran the analysis for 15, 20, and 30 topics, and results remain largely robust to these alternative speci cations. 16 Figure 3: Distribution of Topics Across Corpus Second, as a measure of whether women are more likely to be heard, we use pa erns in the topics discussed to identify who drives the topic of conversation, and which speakers are most likely to receive a response from the state. More speci cally, we examine whether women who speak are as likely as men to steer the conversation towards the issues they raised (agenda-se ing power). To operationalize this concept, we rst identify the topic of each speech using the STM, and then examine whether the speeches that follow continue to address the same issue. Given that each speech is modeled as a mixture of topics, we focus on the primary and secondary topic associated with each document. We also examine the share of the following ve speeches that continue to address the same topic, and the length that a topic persists. Finally, we examine whether the state (i.e. administrators or elected o cials) is more likely 17 to respond to certain speakers. Given that a key goal of the gram sabha is to provide ordinary citizens with an avenue to speak directly to their elected representatives — to ask questions, to demand accountability, to voice complaints — one measure of deliberative in uence is whether state o cials directly address citizen concerns. To measure this, we generate a series of indicator variables to capture (a) whether a citizen’s speech is followed by an o cial, either elected or administrative, and (b) whether that response addresses the topics raised by the citizen. 4 E PVP D E 4.1 Equality of Participation We rst examine whether PVP boosts a endance and frequency of speech among women. While a endance levels among women are already quite high in Tamil Nadu, the presence of PVP still aims to foster collective action among women and explicitly link SHG activities to local govern- ment. Table 3 presents the results. Models (1) through (4) present the e ect of PVP on women’s a endance, measured in raw numbers, while Models (5) through (8) present the e ect of PVP on women’s a endance, measured as a percentage of female voters in the village. e baseline spec- i cations suggest that PVP leads to roughly 70 more women in a endance, or an 8 percentage point increase (from a baseline of 8.5 percent). is represents a doubling of female a endance at the gram sabha. ese results are robust to the inclusion of a variety of demographic and infras- tructural controls, and are consistent with those of Khanna et al. (2015), in which women from the same villages are asked about their a endance at the the most recent asssembly. ey nd that PVP boosts women’s a endance by 65 percent, from a baseline of 11 percent in control villages to nearly 20 percent in treatment areas; our ndings are substantively similar, though smaller in both level and magnitude — lending support to the concern that self-reported measures of women’s political activity may overestimate actual behavior. Second, we look at whether this boost in a endance is accompanied by a greater frequency of women’s speech (Table 4). Here, the unit of analysis is the document, and we examine whether 18 Table 3: E ect of PVP on Women’s A endance Dependent variable: Female A endance (Raw) Female A endance (% of Voters) (1) (2) (3) (4) (5) (6) (7) (8) I(PVP) 68.63∗∗∗ 70.03∗∗∗ 79.61∗∗∗ 58.16∗∗∗ 0.08∗∗∗ 0.09∗∗∗ 0.10∗∗∗ 0.07∗∗∗ (24.34) (24.45) (24.12) (21.59) (0.03) (0.03) (0.04) (0.02) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 96 96 96 96 95 95 95 95 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01.Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages. Demographic controls include: number of households, percentage Scheduled Caste, percentage Scheduled Tribe. Infrastructure controls include indicators for the presence of a primary school, secondary school, health center, hospital, clinic, medicla shop, government hospital, and bank. Table 4: E ect of PVP on Frequency of Women’s Speech Dependent variable: P(Female), All Speeches P(Female), Citizen Speeches (1) (2) (3) (4) (5) (6) (7) (8) I(PVP) 0.07∗∗ 0.06∗∗∗ 0.05∗∗∗ 0.07∗∗ 0.22∗∗∗ 0.21∗∗∗ 0.22∗∗∗ 0.22∗∗∗ (0.03) (0.03) (0.02) (0.03) (0.05) (0.04) (0.04) (0.05) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 3,894 3,894 3,894 3,894 2,130 2,130 2,130 2,130 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages. Demographic controls include: number of households, percentage Scheduled Caste, per- centage Scheduled Tribe. Infrastructure controls include indicators for the presence of a primary school, secondary school, health center, hospital, clinic, medicla shop, government hospital, and bank. the likelihood of having a female speaker is greater in treatment rather than control villages. Models (1) through (4) of Table 4 present results for all speakers (o cials and citizens), while Models (5) through (8) focus on speeches only by citizens. Once again, we see that PVP has a substantial impact on the frequency of women’s speech. We see a roughly 6 to 7 percentage point increase in the incident of any women’s speech. Given a baseline frequency of 35 percent, this increase represents an 18 percent change. e e ect is even more pronounced when looking at citizen speeches alone. From a baseline rate of 38 percent, PVP increases female speech by 17 percentage points, which represents a 57 percent increase. ese results hold to a variety of speci cations, including those that control for demographic and infrastructural characteristics. 19 Table 5: E ect of PVP on Length of Women’s Speech Dependent variable: Speech Length (All Speakers) Speech Length (Citizens Only) (1) (2) (3) (4) (5) (6) (7) (8) I(Female) −16.97∗ −13.38∗ −16.95∗ −17.32∗ −4.33 −3.72 −7.48∗∗ −4.54∗ (8.83) (7.74) (9.13) (8.86) (2.68) (2.84) (3.49) (2.55) I(PVP) −10.02 −9.33 −7.63 −8.73 −5.46 −4.88 −5.74 −5.04 (9.96) (7.98) (8.75) (9.07) (3.77) (3.33) (4.11) (3.70) I(Female) x I(PVP) 21.91∗ 19.02 23.82∗ 22.43∗ 51.92∗∗∗ 51.44∗∗∗ 56.04∗∗∗ 52.13∗∗∗ (12.68) (12.13) (13.54) (12.97) (9.36) (8.93) (10.25) (9.49) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 3,894 3,894 3,894 3,894 2,130 2,130 2,130 2,130 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages. Demographic controls include: number of households, percentage Scheduled Caste, percentage Scheduled Tribe. Infrastructure controls include indicators for the presence of a primary school, secondary school, health center, hospital, clinic, medicla shop, government hospital, and bank. ird, we examine the e ect of PVP on the length of oor time enjoyed by women (Table 5). Here, we proxy for oor time using the word count of each speech. Consistent with previous work, we nd that women generally occupy less oor time then men, about 16 fewer words per speech (compared to an average of 78 words per speech for men in control villages); given that fewer women speak overall, this leads to a massive disparity in oor time. PVP, however, has a substantial impact on women’s length of speaking, increasing the average speech length by over 20 words for the full sample, and by over 50 words per speech among citizen speeches. is disparity not only closes the gender gap in oor time, but actually enables women to take up a majority of the conversation. 4.2 Deliberative In uence While women are speaking signi cantly more in our treatment villages, their voices may still go ignored. Previous empirical work has shown that women are signi cantly less likely than men to drive conversation or set the agenda (Karpowitz and Mendelberg, 2014; Parthasarathy et al., 2017). To examine whether PVP improves women’s ability to in uence discussion, Table 6 regresses two measures of agenda-se ing power — the likelihood that the following speech is on 20 the same topic (Models 1 - 4), and the length of subsequent speeches that are on the same topic (Models 5 - 8) — on an interaction between the speaker’s gender and the village treatment status. Table 6: E ect of PVP on Deliberative In uence Dependent variable: Next Same Length Same (1) (2) (3) (4) (5) (6) (7) (8) I(PVP) −0.05∗ −0.05∗∗ −0.05 −0.06∗ −0.09 −0.15 −0.15 −0.16 (0.03) (0.03) (0.03) (0.03) (0.10) (0.11) (0.10) (0.11) I(Female Speaker) −0.01 −0.01 −0.0001 −0.004 0.02 −0.02 0.02 −0.01 (0.03) (0.03) (0.03) (0.03) (0.11) (0.11) (0.12) (0.11) I(PVP)xI(Female Speaker) 0.003 −0.001 −0.01 −0.001 0.03 0.04 0.003 0.04 (0.05) (0.05) (0.05) (0.05) (0.17) (0.17) (0.18) (0.17) Matched Pair FE Topic FE Demographic Controls Infrastructure Controls p-Score Control Observations 2,099 2,099 2,099 2,099 2,061 2,061 2,061 2,061 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. Demographic controls include: number of households, percentage Scheduled Caste, percentage Scheduled Tribe. Infrastructure controls include indicators for the presence of a primary school, secondary school, health center, hospital, clinic, medicla shop, government hospital, and bank. Across all speci cations, we nd no evidence that PVP improves the agenda-se ing power of women; point estimates are small and statistically insigni cant — suggesting that the presence of this intervention does not increase the likelihood that women are able to drive conversation. Moreover, we nd no evidence that PVP improves women’s ability to elicit a response from the state (Table 7), let alone from elected o cials (Table 8). To ensure that these results are robust to alternative speci cations of the topic model itself, we re-run the analysis with 30 topics and nd largely consistent results (presented in Appendix 3B). Given that one key function of the gram sabha is to provide a forum for citizens to make requests of and demand accountability from politicians, elected o cials’ failure to respond to women suggests that they remain unheard within the gram sabha. Interestingly, despite women’s lack of substantive in uence, Khanna et al.’s (2015) survey- based evaluation of PVP suggests that women feel more e cacious a er program implementa- tion. More speci cally, when presented with hypothetical vigne es about various village and 21 Table 7: E ect of PVP on Responsiveness of the State Dependent variable: On Topic O cial Response (1) (2) (3) (4) I(PVP) −0.05 −0.05 −0.06∗ −0.06∗ (0.03) (0.03) (0.03) (0.03) I(Female Speaker) 0.005 0.002 0.005 0.004 (0.05) (0.05) (0.05) (0.05) I(PVP)xI(Female Speaker) 0.02 0.02 0.02 0.02 (0.07) (0.07) (0.08) (0.07) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 1,141 1,141 1,141 1,141 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clus- tered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. Table 8: E ect of PVP on Politician Responsiveness Dependent variable: On Topic Politician Response (1) (2) (3) (4) I(PVP) −0.06 −0.06 −0.10 −0.06 (0.05) (0.05) (0.07) (0.05) I(Female Speaker) −0.06 −0.05 −0.03 −0.06 (0.05) (0.06) (0.06) (0.06) I(PVP)xI(Female Speaker) 0.001 −0.02 −0.03 0.001 (0.10) (0.09) (0.10) (0.10) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 485 485 485 485 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clus- tered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. household level issues (including public service delivery, infrastructure, local law and order, and family disputes), women in the same treatment villages as those studied here were nearly 25 per- cent more likely than women in control villages to state they would take some form of action — be it speaking to a village o cial or raising the issue at a gram sabha. Of course, such hypothet- icals do not capture whether their promised action will yield results, but the boost in women’s sense of self-e cacy is a positive step towards their public and political action. 22 5 D Taken together, these pa erns highlight the opportunities and challenges to improving equitable deliberation. Against a backdrop in which women are less likely to a end and participate in local governance, we show that a bo om-up intervention can signi cantly increase women’s engagement in the gram sabha. We do so using multiple measures of voice — their presence, the share of speeches made by women, and the oor time that they use. However, we nd li le evidence that this newfound voice is able to improve women’s deliberative in uence. Across multiple measures of agenda se ing power and state responsiveness, we nd that women in PVP villages are no more likely to get a relevant response from peers or from the state. at we see li le shi on these la er indicators underscores some of the challenges in trying to improve deliberative equality; it is not enough to induce women to speak, but we must also encourage others to listen. Another possible explanation for women’s lack of agenda-se ing power may be that women in treatment areas are bringing up a set of new topics, related to Pudhu Vaazhvu itself, that do not elicit responses from their fellow villagers or elected o cials. Indeed, PVP villages discuss 2.66 more topics on average than non-PVP villages — a di erence that is signi cant at the 0.001 level. To address this, we examine the variation in topics raised by gender; more speci cally, we estimate the di erence in expected topic proportion between men and women across treatment and control villages (Figure 4a). In control villages, we nd that men and women are generally likely to discuss the majority of topics with the same frequency, with a few notable exceptions that re ect the gendered nature of social life in rural Tamil Nadu. More speci cally, men are signi cantly more likely to discuss employment and expenditure-related topics (like NREGA, the rural employment guarantee, and the ration shop), while women are more likely to raise water and housing concerns, as well as education. By contrast, in treatment villages, women speak much more than men about project activities, like loans, spending audits, and job training, and do not speak signi cantly more than men about 23 Figure 4: Distribution of Topics by Gender (a) Control Villages (b) Treatment Villages Note: e Figures above plot the expected topic proportion and 95% con dence interval for each topic among female speakers, by treat- ment status. Coe cients less than zero indicate topics that are more. frequently raised by women, while those greater than zero indicate topics that are more frequently raised by men in non-PVP villages 24 those issues that they had spoken more about in control villages. In other words, it may be that PVP is fundamentally shi ing the content of conversation that women engage in — moving them from discussing primarily domestic ma ers (water collection, education) to issues related to the administration of the program. Table 9 presents the average number of speeches within a village devoted to speci c issues. When we look at canonically “women’s” issues, such as water, housing, education, etc., we see a marked decline in the frequency that these issues are discussed. e two notable exceptions are entitlement requests and animal husbandry, both of which are emphasized by PVP’s livelihoods and social safety net programs. By contrast, for the canonically “male” issues, such as employment, ration, and garbage, we see no discernible di erence. is suggests that even though women are speaking more o en in treatment villages, they are speaking speci cally about the project activities, whereas the men continue to raise their usual governance concerns. Table 9: Topic Counts, by Treatment Status Avg. Speeches, Control Avg. Speeches, PVP t-statistic p value Water 9.3200 6.5000 2.1102 0.0376 Housing 5.1400 3.1200 2.5822 0.0115 Entitlement Requests 7.1200 9.5000 -1.5262 0.1305 Education 1.8000 1.6400 0.3752 0.7083 Public Infrastructure 2.2400 2.6800 -0.8833 0.3795 NREGA 2.7200 1.8000 1.6389 0.1048 Ration Shop 3.6200 3.1000 0.5158 0.6073 Garbage 1.4200 1.3600 0.1887 0.8507 Voter Lists 2.7600 4.4000 -2.2051 0.0299 Sanitary Complex 0.8800 0.6800 0.7922 0.4303 SHGs 1.4400 1.2200 0.6521 0.5159 Whether this is normatively problematic or not remains to be seen. On the one hand, if PVP encourages its members to raise issues that are important and consequential for previously disempowered groups, it may not be worrisome that other citizens have less time to discuss ma ers relevant to them. On the other hand, if the newly vocal constituency of women created by PVP crowds out discussion that is broadly relevant to other marginalized groups, we may have reason to worry that the gram sabha is no longer dominated by men, but by the project’s participants. at is, we should be cautious about programs that so alter the organic processes of 25 the gram sabha, which has been a broadly e ective forum for local governance. e notion that outside intervention might have unintended consequences on local dynam- ics of citizen participation is not new. For example, Bano (2012) studies the consequences of externally funded NGOs and Pakistan, and nds that these organizations o en displace organic community-based groups and upset informal processes by which could monitor one another. In Kenya, Gugerty and Kremer (2008) evaluate the impact of a funding program to strengthening women’s associations and nd that the introduction of external funds has li le impact on the groups’ activities, but leads to a substantial change in the membership of groups, encouraging the entry of younger, more educated women. In our study, the consequences of external inter- vention have less to do with membership, but we do see that the introduction of this external program may meaningfully shi conversation about relevant governance issues. Of course, the growth of project-speci c conversation may be a consequence of PVP’s unique design, which speci cally encourages women to publicly administer the program’s activities within the village. is feature of PVP is just one of the many channels by which this interven- tion may boost women’s civic engagement; other channels include the provision of credit access, livelihoods training, and social networks fostered by the group-based format of the program. Evi- dence from an earlier economic evalution of PVP by Khanna et al. (2015) suggests that all of these mechanisms may be at play; in their household survey, they nd that PVP reduces the high cost debt burden of target households, improves women’s intrahousehold decision-making power, and boosts their willigness to engage with public o cials. eoretically, each of these components may individually raise women’s ability to participate in the public sphere: With greater access to credit and livelihoods, women may have more decision-making power and autonomy within the home — power which earns them more public freedoms. e social networks developed by the group-based format of the program may facilitate collective action (Sanyal, 2014). And the particular focus on women-led administration of PVP may “mechanically” boost the participation of women in the gram sabha, as they use this public space to announce programs and publicize PVP’s activities. Unfortunately, while the bundled nature of this intervention precludes us from 26 unpacking the e ects of any given channel, parsing these mechanisms and understanding their interactions is a ripe area for future research, particularly for policy makers. Additionally, future scholarship is needed to understand the heterogenous impacts that such programs can have on women from di erent caste, religious, and class backgrounds, which have been shown to mediate the impact that self-help groups and livelihoods training can have on women’s empowerment (Field et al., 2010). is study’s focus on the speech acts of women in the gram sabhas limited data collection of those individual characteristics which could not be immediately visually ascertained. Collecting such information would have interfered with the natural functioning of the meeting. As a result, we cannot identify the extent to which PVP’s focus on including Scheduled Caste and Tribe households from the poorest of the poor may be relevant to understanding the null results on deliberative in uence. 6 C Motivated by the concern that inequalities among citizens may limit the ability of deliberative democratic institutions to produce more inclusive development outcomes, this paper opens the “black box” of the gram sabhas at the core of India’s decentralization e ort. We use text-as-data methods on an original corpus of village assembly transcripts from rural Tamil Nadu to show that bo om-up e orts to empower and induce women’s participation can be useful in mitigating the gaping gender gap in political participation. More speci cally, we evaluate the impact of the a woman-centered poverty alleviation program, which explicitly aims to bring women into greater contact with village government and to provide them with greater agency in the administration of a government program. We nd that PVP is able to signi cantly improve gram sabha participation by women in terms of their a endance, their propensity to speak, and the oor time they enjoy. However, we also show that greater voice for women does not lead to greater agenda se ing power or responsiveness from the state. is may result from the fact that project-facilitated participation encourages women to speak up primarily about activities related to PVP itself, in ways that potentially fail to engage the broader village community. Or it may simply re ect the 27 deeper di culty in improving deliberative equality, which requires not only that citizens have an equal ability to speak, but also to be heard. 28 R B. Agarwal. Participatory exclusions, community forestry, and gender: An analysis for South Asia and a conceptual framework. World development, 29(10):1623–1648, 2001. achtiger, M. Sp¨ A. B¨ orndli, M. R. Steenbergen, and J. Steiner. e deliberative dimensions of leg- islatures. Acta Politica, 40(2):225–238, 2005. R. Ban and V. Rao. Is deliberation equitable? Evidence from transcripts of village meetings in south India. World Bank Policy Research Working Paper Series, (4928), 2008a. R. Ban and V. Rao. Tokenism or agency? e impact of women’s reservations on village democ- racies in south India. Economic Development and Cultural Change, 56(3):501–530, 2008b. R. Ban, S. Jha, and V. Rao. Who has voice in a deliberative democracy? Evidence from transcripts of village parliaments in south India. Journal of Development Economics, 99(2):428–438, 2012. M. Bano. Breakdown in Pakistan: How Aid is Eroding Institutions for Collective Action. Stanford University Press, 2012. L. Beaman, R. Cha opadhyay, E. Du o, R. Pande, and P. Topalova. Powerful Women: Does Exposure Reduce Bias? e arterly Journal of Economics, 124(4):1497–1540, 2009. L. Beaman, E. Du o, R. Pande, P. Topalova, et al. Political reservation and substantive represen- tation: Evidence from Indian village councils. India policy forum, 7:159–201, 2010. L. Beaman, E. Du o, R. Pande, and P. Topalova. Female leadership raises aspirations and educa- tional a ainment for girls: A policy experiment in India. Science, 335(6068):582–586, 2012. T. Besley, R. Pande, and V. Rao. Participatory democracy in action: Survey evidence from South India. Journal of the European Economic Association, 3(2-3):648–657, 2005. R. Cha opadhyay and E. Du o. Women as policy makers: Evidence from a randomized policy experiment in India. Econometrica, 72(5):1409–1443, 2004. P. Chhibber. Democracy without Associations: Transformation of the Party System and Social Cleav- ages in India. University of Michigan Press, 1999. R. M. Desai and S. Joshi. Collective action and community development: Evidence from self-help groups in rural India. e World Bank Economic Review, 28(3):492–524, 2014. J. Dreze and A. K. Sen. India: Development and participation. Oxford University Press, USA, 2002. E. Du o. Women empowerment and economic development. Journal of Economic Literature, 50 (4):1051–1079, 2012. T. Dyson and M. Moore. On kinship structure, female autonomy, and demographic behavior in India. Population and development review, pages 35–60, 1983. 29 L. Edlund and R. Pande. Why have women become le -wing? e political gender gap and the decline in marriage. e arterly Journal of Economics, 117(3):917–961, 2002. J. Elster. Deliberative democracy, volume 1. Cambridge University Press, 1998. E. Field, S. Jayachandran, and R. Pande. Do traditional institutions constrain female entrepreneur- ship? A eld experiment on business training in India. e American Economic Review, 100(2): 125–129, 2010. A. Fung. Deliberation’s darker side: six questions for Iris Marion Young and Jane Mansbridge. National Civic Review, 93(4):47–54, 2004. M. K. Gugerty and M. Kremer. Outside funding and the dynamics of participation in community associations. American Journal of Political Science, 52(3):585–602, 2008. A. Gu man and D. ompson. Why deliberative democracy. 2004. J. Habermas. Moral consciousness and communicative action. MIT press, 1990. P. Heller and V. Rao. Deliberation and Development: Rethinking the Role of Voice and Collective Action in Unequal Societies. World Bank Publications, 2015. G. W. Imbens and J. M. Wooldridge. Recent Developments in the Econometrics of Program Eval- uation. NBER Working Paper Series, (14251), 2008. R. Inglehart and P. Norris. Rising tide: Gender equality and cultural change around the world. Cambridge University Press, 2003. Inter-Parliamentary Union. Women in national parliaments, 2017. URL http://www.ipu. org/wmn-e/world.htm. K. Kapadia et al. Siva and her sisters: gender, caste, and class in rural South India. Westview Press, Inc., 1995. C. F. Karpowitz and T. Mendelberg. e silent sex: Gender, deliberation, and institutions. Princeton University Press, 2014. M. Khanna, N. Kochhar, and N. Palaniswamy. A Retrospective Impact Evaluation of the Tamil Nadu Empowerment and Poverty Alleviation (Pudhu Vaazhvu) Project. e Journal of Devel- opment Studies, 51(9):1210–1223, 2015. S. Kishor and K. Gupta. Gender equality and womens empowerment in India. National Family Health Survey (NFHS-3) India 2005-06. 2009. M. L. Krook. otas for women in politics: Gender and candidate selection reform worldwide. Oxford University Press, 2010. J. Mansbridge. Beyond adversarial democracy. Chicago, IL: University of ChicagoPress, 1980. J. Mansbridge. Should blacks represent blacks and women represent women? A contingent yes . e Journal of politics, 61(03):628–657, 1999. 30 J. Mansbridge. A Minimalist De nition of Deliberation. In V. Rao and P. Heller, editors, Deliber- ation and Development: Rethinking the Role of Voice and Collective Action in Unequal Societies, pages 27–50. World Bank Publications, 2015. G. Mansuri and V. Rao. Localizing development: does participation work? World Bank Publications, 2012. G. Miller. Women’s su rage, political responsiveness, and child survival in American history. e arterly Journal of Economics, 123(3):1287–1327, 2008. M. Mines. Public Faces, Private Lives: Community and Individuality in South India. Univ of Cali- fornia Press, 1994. R. Parthasarathy and V. Rao. Deliberative democracy in india. World Bank Policy Research Working Paper, (7995), 2017. R. Parthasarathy, V. Rao, and N. Palaniswamy. Deliberative Inequality: A Text-as-data Study of Tamil Nadu’s Village Assemblies. Working Paper, 2017. P. Paxton, S. Kunovich, and M. M. Hughes. Gender in politics. Annual Review of Sociology, 33: 263–284, 2007. S. A. Prillaman. Strength in Numbers: How Women’s Networks Close India’ ’s Political Gender Gap. 2016. URL http://www.soledadprillaman.com/job-market-paper. M. E. Roberts, B. M. Stewart, and E. M. Airoldi. A model of text for experimentation in the social sciences. Journal of the American Statistical Association, 111(515):988–1003, 2016. P. R. Rosenbaum and D. B. Rubin. e central role of the propensity score in observational studies for causal e ects. Biometrika, 70(1):41–55, 1983. P. Sanyal. Credit to capabilities: A sociological study of microcredit groups in India. Cambridge University Press, 2014. P. Sanyal, V. Rao, and S. Majumdar. Recasting culture to undo gender: A sociological analysis of Jeevika in Rural Bihar, India. World Bank Policy Research Working Paper, (7411), 2015. A. Sen. Development as freedom. Oxford Paperbacks, 2001. World Bank. World Development Report 2012: Gender, Equality, and Development. World Bank Publications, 2011. 31 A A T I V A key challenge in the text as data literature, particularly with unsupervised methods, lies in how to interpret the topics that are produced. We label topics based on both a careful examination of the highest probability and FREX words presented in Table A.1, as well as a reading of the documents most associated with a given topic. While the topics identi ed by this method are largely consistent with what we would expect in a gram sabha meeting, we further validate the topics generated in two ways. First, we employ two tests of predictive validity — that is, we test whether certain topics are more prevalent based on the characteristics of the speaker and village. First, we examine whether the “proforma” topics generated by the topic model are more likely to be discussed by o cials, rather than citizens. Since the topic model identi es a set of standard, routine remarks — in particular, the reading of resolutions, the formal greetings and votes of thanks, and discussion of government funding allocation — as distinct topics; if these topics capture the rote features of assemblies as they are conducted, these should be primarily spoken by o cials, who are responsi- ble for convening and adjourning the meeting, as well as sharing information about recent public expenditures. Figure A.1 plots the di erence between the expected proportion of these proforma topics between citizens and o cials (both elected and administrative) for the documents in the corpus. As expected, these proforma speeches are all signi cantly more likely to be raised by o cials, suggesting that the topics re ect our substantive interpretation of their content. Figure A.1: Topical Prevalence of Proforma Topics, by Position of Speaker Note: e gure above plots the expected topic proportion and 95% con dence interval for each proforma topic, by the speaker’s position. Coe cients greater than zero indicate topics that are more frequently raised by o cials, while those less than zero indicate topics that are more frequently raised by citizens. As a second test of predictive validity, we examine whether topics explicitly related to the 32 intervention, PVP, are more likely to be raised in treatment villages. Since these topics are ex- plicitly related to the treatment, they should be largely absent from our control areas. Figure A.2 plots the di erence between the expected proportion of these proforma topics between treat- ment and control villages for the documents in the corpus. As expected, these PVP topics are all signi cantly more likely to be raised in treatment areas, suggesting that the topics re ect our substantive interpretation of their content. Figure A.2: Topical Prevalence of PVP Topics, by Treatment Status Note: e gure above plots the expected topic proportion and 95% con dence interval for each PVP-related topic, by the village’s treat- ment status. Coe cients greater than zero indicate topics that are more frequently raised in control villages, while those less than zero indicate topics that are more frequently raised in treatment villages. 33 Table A.1: Top Word Stems by Topic Topic Top Word Stems Water Highest Prob: water, facil, tank, problem, drink, well, come FREX: water, pipe, drink, tank, x, tap, motor Entitlement Requests Highest Prob: get, give, card, year, petit, dont, even FREX: give, get, petit, said, card, even, dont Greetings Highest Prob: come, panchayat, presid, member, request, inform, ward FREX: thank, behalf, ward, grievanc, presid, today, come Moderation of Debate Highest Prob: ask, one, want, say, commot, talk, keep FREX: ask, say, want, anyth, commot, talk, question Ration Shop Highest Prob: tell, told, ration, shop, good, month, much FREX: told, tell, ration, much, readi, good, answer Infrastructure Requests Highest Prob: villag, need, time, also, arrang, mani, pleas FREX: pleas, need, demand, time, requir, sit, speak Land Management Highest Prob: road, street, canal, land, light, pond, lake FREX: canal, pond, lake, road, tree, coloni, street Women’s Livelihood Programs Highest Prob: peopl, scheme, person, famili, bene t, start, mani FREX: peopl, may, poverti, famili, bene t, scheme, start PLF Loans Highest Prob: group, loan, plf, got, member, function, repay FREX: loan, repay, group, outstand, plf, repaid, got Actions and Resolutions Highest Prob: take, write, chang, resolut, problem, pass, action FREX: write, nd, week, bdo, take, see, solut Housing Subsidies Highest Prob: hous, build, toilet, construct, allot, built, place FREX: hous, construct, build, built, toilet, pa a, allot Voter and Bene ciary Lists Highest Prob: gram, sabha, list, name, place, read, resolut FREX: gram, sabha, name, includ, list, voter, read Animal Husbandry Highest Prob: given, money, know, cow, pay, insur, thing FREX: money, know, cow, buy, die, thing, yet NREGA Highest Prob: work, day, done, panchayat, complet, job, number FREX: work, done, day, wage, yes, complet, agricultur Youth Job Training Highest Prob: train, vprc, person, youth, abl, di er, target FREX: train, youth, mental, abl, comput, di er, vprc VPRC Audits Highest Prob: fund, bank, receiv, amount, expens, account, incom FREX: fund, interest, receiv, incom, balanc, expens, account Vote of anks Highest Prob: meet, o c, conduct, particip, district, also, rst FREX: particip, o c, collector, meet, conduct, a end, rst Women’s Sanitary Complex Highest Prob: govern, women, given, per, complex, sanitari, use FREX: govern, marriag, complex, per, sanitari, alloc, maintain Education Highest Prob: school, children, bus, educ, child, hospit, studi FREX: children, bus, studi, hospit, school, child, std Garbage and Sanitation Highest Prob: use, panchayat, plastic, remov, prevent, avoid, improv FREX: plastic, garbag, prevent, remov, avoid, vaccin, ca l Rules for Bene ciary Selection Highest Prob: scheme, panchayat, toilet, year, employ, bene ciari, discuss FREX: guarante, gandhi, employ, memori, price, bene ciari, propos Discussion of Women’s SHGs Highest Prob: panchayat, group, women, help, self, peopl, award FREX: award, elig, self, support, survey, mission, help VPRC Loans Highest Prob: rupe, lac, drive, instal, fund, given, panchayat FREX: rupe, drive, lac, driver, instal, licens, total VPRC Administration Highest Prob: provid, detail, inform, regard, appoint, certif, centr FREX: centr, certif, appoint, provid, detail, communiti, util Panchayat Expenses Highest Prob: panchayat, sabha, approv, report, inform, regard, scheme FREX: report, approv, releas, nanc, commiss, mainten, usag 34 B R A T M S To ensure that the main results for agenda se ing power and state responsiveness are not sen- sitive to a particular topic model speci cation, we re-run our topic model with K = 30 topics, generate new measures of deliberative in uence, and present results below. We rst re-examine how agenda-se ing power varies with the gender of the speaker and village treatment status. Consistent with the main results presented (for K = 25 topics in Table 6), we see that even under this alternative model speci cation, we nd no evidence that PVP improves the agenda- se ing power of women; point estimates are small and statistically insigni cant — suggesting that the presence of this intervention does not increase the likelihood that women are able to drive conversation. Table B.1: E ect of PVP on Deliberative In uence (K = 30) Dependent variable: Next Same Length Same (1) (2) (3) (4) (5) (6) (7) (8) I(PVP) −0.04 −0.04 −0.05 −0.04 −0.07 −0.11 −0.12 −0.10 (0.03) (0.03) (0.04) (0.03) (0.09) (0.09) (0.09) (0.08) I(Female Speaker) −0.03 −0.03 −0.04 −0.03 0.04 0.001 −0.01 −0.01 (0.02) (0.02) (0.02) (0.02) (0.08) (0.09) (0.09) (0.09) I(PVP)xI(Female Speaker) 0.02 0.02 0.02 0.02 −0.001 −0.02 −0.01 0.01 (0.04) (0.04) (0.04) (0.04) (0.11) (0.10) (0.10) (0.10) Matched Pair FE Topic FE Demographic Controls Infrastructure Controls p-Score Control Observations 2,099 2,099 2,099 2,099 2,061 2,061 2,061 2,061 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. Demographic controls include: number of households, percentage Scheduled Caste, percentage Scheduled Tribe. Infrastructure controls include indicators for the presence of a primary school, secondary school, health center, hospital, clinic, medicla shop, government hospital, and bank. 35 When we look at the response of state o cials, we nd evidence that under the alternative speci cation, PVP actually has a positive and signi cant e ect on the likelihood of women re- ceiving a relevant response (Table B.2); however, this e ect is largely driven by administrators. When we look speci cally at politician responsiveness (Table B.3), we nd that PVP has no pos- itive e ect on whether women are heard or addressed by the state. Table B.2: E ect of PVP on Responsiveness of the State (K = 30) Dependent variable: On Topic O cial Response (K = 30) (1) (2) (3) (4) I(PVP) −0.12∗∗∗ −0.12∗∗∗ −0.13∗∗∗ −0.12∗∗∗ (0.03) (0.04) (0.04) (0.04) I(Female Speaker) −0.06∗∗ −0.06∗∗ −0.07∗∗∗ −0.06∗∗ (0.03) (0.03) (0.03) (0.03) I(PVP)xI(Female Speaker) 0.12∗∗∗ 0.11∗∗∗ 0.13∗∗∗ 0.12∗∗∗ (0.03) (0.04) (0.04) (0.03) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 1,141 1,141 1,141 1,141 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01.Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. Table B.3: E ect of PVP on Responsiveness by Elected O cials (K = 30) Dependent variable: On Topic Politician Response (K = 30) (1) (2) (3) (4) I(PVP) −0.15∗ −0.16∗∗ −0.16∗∗∗ −0.15∗ (0.08) (0.08) (0.07) (0.08) I(Female Speaker) −0.10∗ −0.08 −0.08 −0.10∗ (0.06) (0.06) (0.06) (0.06) I(PVP)xI(Female Speaker) 0.06 0.04 0.03 0.06 (0.09) (0.09) (0.09) (0.09) Matched Pair FE Demographic Controls Infrastructure Controls p-Score Control Observations 485 485 485 485 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the block-pair, in parenthesis. Data are taken from the full sample of villages, but include only citizen speakers. 36