WPS8119 Policy Research Working Paper 8119 Deliberative Inequality A Text-As-Data Study of Tamil Nadu’s Village Assemblies Ramya Parthasarathy Vijayendra Rao Nethra Palaniswamy Development Research Group Poverty and Inequality Team June 2017 Policy Research Working Paper 8119 Abstract The resurgence of deliberative institutions in the developing officials to bluster and read banal announcements, but world has prompted a renewed interest in the dynamics of rather, provide opportunities for citizens to challenge their citizen engagement. Using text-as-data methods on an orig- elected officials, demand transparency, and provide infor- inal corpus of village assembly transcripts from rural Tamil mation about authentic local development needs. Second, Nadu, India, this paper opens the “black box” of delibera- the study finds that across multiple measures of deliberative tion to examine the gendered and status-based patterns of influence, women are at a disadvantage relative to men; influence. Drawing on normative theories of deliberation, women are less likely to speak, set the agenda, and receive this analysis identifies a set of clear empirical standards for a relevant response from state officials. Finally, the paper “good” deliberation, based on an individual’s ability both to shows that although quotas for women on village councils speak and be heard, and uses natural language processing have little impact on the likelihood that they speak, they methods to generate these measures. The study first shows do improve the likelihood that female citizens are heard. that these assemblies are not mere “talking shop” for state 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 Deliberative Inequality: A Text-As-Data Study of Tamil Nadu’s Village Assemblies∗ 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 Introduction In the face of concerns that aid programs have not beem serving those who need them most, community-led development has moved to center stage in the aid industry (Mansuri and Rao, 2012). Participatory development programs draw on theories of deliberative democracy, which emphasize the role of citizens in coming to reasoned compromises with one another over ma ers of public interest. By moving decision-making power from government o ce to the village itself, these programs have been viewed as a way to wrest power from elite capture and improve the equity of allocations across local communities. Whether in village meetings or neighborhood as- sociations (Auerbach, 2017), citizen engagement of this sort is thought to result in more e ciently implemented and inclusively distributed development outcomes. ese instrumental aims, how- ever, are only part of the reason that international organizations and national governments have embraced community-led development; the other key reason is that we increasingly think that citizen’s voice has inherent normative value (Elster, 1998; Dryzek, 1994; Gu man and ompson, 2004). As such, this shi towards deliberation has accompanied a shi in our understanding of development itself — from narrow economic transformation to a more holistic view that includes human capabilities, social development, and justice (Sen, 2001). Despite their promise, however, there are strong reasons to think that deliberative institutions may be ill-equipped to deliver on their instrumental or normative goals (Heller and Rao, 2015). Of particular concern are the ways in which inequalities — across gender, class, caste, and position — may be reinforced or even exacerbated in deliberative fora. First, deliberative forums may perpetuate existing inequalities among citizens. Unlike aggregative forms of democracy, where standing among voters is leveled by the equal weighting of ballots, and institutional safeguards like the secret ballot protect against coercion, deliberation requires public, o en costly, exercise of voice. It takes place in highly localized se ings, where social norms shape the actions of individuals within the group. ese issues may be particularly acute for women, who tend to be perceived as less in uential than men, who are less likely to set the agenda, and who are less likely to impact outcomes (Karpowitz and Mendelberg, 2014). 2 Second, inequality between voters and state o cials can undermine the promise of deliber- ative institutions. When participation is induced by the state, as in decentralization e orts or community-driven development programs, agents of the state o en are stuck in the ironic situ- ation of having to act against their self-interest by promoting institutions whose purpose is to undermine their power (Mansuri and Rao, 2012). erefore, local bureaucrats and politicians may try to undermine these institutions by canceling them (Besley et al., 2005), crowding out mean- ingful deliberation with bureaucratic announcements, and ignoring voters’ claims and evading their requests (Bha acharjee and Cha opadhyay, 2011). Whether and to what extent these forms of dominance — of men over women, of the elite over the poor, of state o cials over citizens — a ect deliberative institutions in practice is an empirical question, but one that has been challenging to study systematically on a large scale, particularly in real-world, citizen-centered deliberative forums. In this paper, we overcome these challenges by applying Natural Language Processing (NLP, or text-as-data) methods to an original corpus of village assembly transcripts from rural India to systematically examine variation in the quality of deliberation. In particular, we examine the relationship between deliberative in uence and the gender or position (citizen versus o cial) of a speaker. By using NLP methods, we are able to quantitatively examine not only the relative oor time enjoyed by di erent types of speakers, but also their ability to in uence the topic of conversation (agenda-se ing power) and to make claims on state o cials (responsiveness of the state). We nd that, despite women’s high rates of a endance in Tamil Nadu’s village assemblies, they are indeed the “silent sex.” Women make up 58 percent of a endees on average, but are responsible for one-third of the available oor time. Moreover, when women do speak on a particular topic, they are signi cantly less likely than men to elicit a topical or relevant response from state o cials — suggesting a meaningful inequality in deliberative in uence across the sexes. Importantly, these results hold even if we control for the particular topic that is being raised; that is, for any given topic, a man is more likely to get a response from an o cial than a woman. In contrast to our ndings on gender inequality, we do not nd evidence that assemblies are 3 dominated by elected o cials. A majority of oor time is taken up by citizens, who are more likely than o cials to set the agenda. Moreover, while o cials o en read a set of announcements at the beginning of meetings, these statements are generally in response to issues raised by citizens, not e orts to direct conversation. is is consistent with previous work (Ban et al., 2012), which nds that India’s gram sabhas are more than mere “talking shops,” and that conversation within the sabhas actually re ects median household preferences within the village. Our work also speaks to the literature on the impact of descriptive representation on social norms. In particular, we explore whether and how gender quotas for the village council presi- dents a ect deliberative equality. Advocates of quotas have long argued that this policy not only improves representation of women and minorities via the election of policy makers who may share their preferences, but also creates a precedent for women voicing their own preferences (Mansbridge, 1999). We nd that the presence of a female president has a meaningful and sig- ni cant impact on the ability of women to be heard and responded to. We show that women are not only more likely to drive conversation under female presidents, but that female presidents themselves are signi cantly more responsive to women constituents, consistent with the argu- ment that “descriptive representation facilitates vertical communication between representatives and constituents” (Mansbridge, 1999, p. 641) in conditions where women have been historically marginalized. However, the mere presence of female incumbents has no e ect on the frequency or volume of women’s speech — suggesting that reservations are not a panacea for gendered inequality in these deliberative forums. is paper contributes to the growing empirical literature on deliberation, which began with rich and careful ethnographies of deliberation in Western se ings,1 and has since expanded to study developing country contexts as well. While the bulk of this literature has been limited to “successful” examples of deliberative resource allocation, such as participatory budgeting in Brazil (Baiocchi et al., 2011) and the People’s Campaign in Kerala (Heller et al., 2007), scholars 1 See Mansbridge’s (1980) study of town meetings in New England; Fung’s (2004) study of neighborhood gover- nance in Chicago’s South Side; Polle a’s (2004) and Polle a and Lee’s (2006) analyses of a variety of deliberative spaces in the US. 4 are now turning to deliberation in more challenging contexts. rough detailed qualitative and ethnographic work, scholars have shown how deliberative forums can be used as a space to make dignity claims for underprivileged groups (Rao and Sanyal, 2010), and even as a tool to solve social problems such as female genital mutilation (Mackie, 2015). Other work has expanded the scope of deliberation to include everyday communication outside the context of formal forums (Swidler and Watkins, 2015). Our work builds upon this scholarship by examining deliberative outcomes in a challenging rural context, but departs methodologically from this earlier literature by quantifying inequalities in participation. Beyond these observational studies, there is a growing literature that tests hypotheses de- rived from deliberative theory in the context of lab experiments (Fishkin and Luskin, 2005; List et al., 2013; Goeree and Yariv, 2011; Karpowitz et al., 2012; Karpowitz and Mendelberg, 2014). While such work has helped inform our understanding of how institutional contexts (e.g. deci- sion rules, moderators, etc.) a ect deliberative quality, studying systematic variation in the real world has been much more challenging. A notable exception has been the study of parliamen- tary debates, where the availability of data has enabled scholars to leverage temporal variation to study pa erns in deliberative quality (Clayton et al., 2016); unfortunately, this strand of work has o en been limited in its ability to test the e ects of institutional variation due to the focus on any single deliberative body. By contrast, our study provides a new source of transcript data from numerous local deliberative bodies — enabling us to correlate deliberative outcomes with local institutional variation, including the gender of the local politicians moderating discussion. Moreover, the focus on local, rather than national-level institutions, allows us to examine citizen voice rather than o cial debate; in doing so, we are able to address normative questions about whether and how citizens are able to participate in their own governance. is paper also provides an important bridge between empirical work and normative theories of deliberation by generating a clear set of metrics that can be coded using automated methods. Given the considerable debate within the normative literature both about deliberative standards as well as the more basic question of what constitutes deliberation, there has been meaningful 5 disagreement within the empirical literature as to how to systematically assess deliberative qual- ity (Myers and Mendelberg, 2013). Here, we build upon the minimalist approach to deliberation outlined by Mansbridge (2015), and focus on a set of context-relevant standards that relate to the political and ethical functions of deliberation. In doing so, we outline qualities of good delibera- tion that are both applicable to development contexts and that can be operationalized by future scholars. Furthermore, this paper contributes methodologically to the study of deliberation. We show that the use of unsupervised but validated measures can enable scholars to evaluate large bodies of transcript data, even among ordinary citizens in developing country contexts. While this ap- proach has been used in more literate contexts — be they elite speakers in parliaments or citizens from rich countries — the bulk of scholarship in contexts similar to ours has engaged in time- intensive manual coding of data, o en to capture nuanced aspects of debate and argumentation; this can be prohibitively costly, limiting the volume of scholarship in this area. By contrast, our approach uses text as data methods to analyze a large body of transcripts and generate consistent measures for speakers’ oor time, agenda-se ing power, and ability to generate a response from the state. By validating these measures against survey data, we hope to encourage scholars that textual analysis for the study of deliberation is not prohibitively costly. e remainder of the paper is organized as follows: In Section 2, we provide the institutional context behind the Indian gram sabha, or village assembly, in which we study deliberative in- equality. In Section 3, we identify the standards of good deliberation relevant to this and other developing country contexts. In Section 4, we describe data and measures we use to evaluate the quality of deliberation. Here, we also describe the topic modeling methodology used to evalu- ate speech content. In Section 5, we descriptively document pa erns of deliberative quality and examine how design of the gram sabha, speci cally the use of gender quotas, may alter those pa erns. Finally, we discuss the implications of these results and conclude in Section 6. 6 2 Institutional Context: Panchayati Raj e institutional se ing in which we examine deliberation is the gram sabha, or village assembly. Gram sabha were created in 1993 as part of a transfer of responsibility for the delivery of local public goods and services to a three-tier local government, with the village panchayat (VP) at the bo om level. 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 pres- ident. 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. Relevant for this study, the process for assigning gender quotas is as-if-random, allowing us to interpret any observed di erences between male and female incumbents as causal.2 In doing so, we follow a signi cant body of literature that leverages this assignment process to study the e ect of female incumbents on Indian local government (Cha opadhyay and Du o, 2004; Bhavnani, 2009; Ban and Rao, 2008b; Besley et al., 2005). Lastly, in addition to this executive council, the legislature of the village would be the gram sabha, to which every citizen of the village would be a member, with gram sabha meetings held at least two times a year. While these mandates represented the minimal requirements for the village panchayat sys- tem, every Indian state was given a wide degree of leeway in how the VPs would function — leading to considerable variation in the VPs’ budgets, functions, and implementation of the gram sabha (Besley et al., 2005). In Tamil Nadu, where this study is located, the speci c functions and requirements of the VPs were de ned by the Tamil Nadu Panchayats Act (1994). Formally, the functions devolved to the VP have been to identify target populations for federal and state poverty alleviation programs; the construction and maintenance of basic public goods (village roads, streetlights, drinking water, drains); and the provision of sanitation services. 2 e speci c process for assigning women’s reservations is described in the Tamil Nadu Panchayats Rules of 1995, Section 7.3, Rule 7, which mandates the creation of a “list of Wards or Panchayats arranged in descending order of the percentage of…Women,” and then details the rotation of reservations every 10 years by proceeding down the list. Since the percentage of women in the population is roughly the same across the state’s Panchayats, assignment is as-if-random. 7 Today, these assemblies constitute the most widely used deliberative institution in human history, a ecting over 840 million people living in approximately one million villages in rural India. Deliberative democracy has deep historical roots in India where, for centuries, deliber- ative bodies were central to systems of local governance, and religious discourse and dialogue (Parthasarathy and Rao, 2017). In the period of colonial rule in the 19th century the interplay of ideas between Western liberal philosophers and Indian intellectuals led to India becoming a fertile ground for experiments in governance. e idea of self-sustaining village democracy, in particular, appealed greatly to Mohandas Gandhi, who made it a central tenet of his philosophy. In 1993, 45 years a er independence, the Gandhian push for deliberative village democracy was given constitutional sanction with the passing of the 73rd amendment to the Indian constitution. In general, Tamil Nadu has not been a front-runner in devolving much power to VPs, nor have recent improvements signi cantly improved policy devolution.3 ough Tamil Nadu VPs are not su ciently well- nanced to actually deliver public goods and services on their own, they do play a vital role in (a) implementing the last mile of various functions and programs, and (b) relaying information about local needs to the higher block tier of government, which has nal authority on the provision of key services. For example, the VPs identify areas that need more drinking water; keep track of repair and construction needs; collect census data on household toilet access; provide information on local infrastructure needs (such as roads and drainage); and identify bene ciaries from the target population for several other federal and state anti-poverty programs. e VP also provides information to higher levels of government on public service problems that range from the functioning of the public food distribution systems to glitches in the new electronic payments system for public works. Lastly, the VP is fully responsible for the full implementation (including payment of salaries) of the federal rural employment scheme (NREGA), which guarantees 100 days of work on public works for any individual who wants this work.4 3 http://www.iipa.org.in/upload/panchayat devolution index report 2012-13. pdf 4 Recently, payment of salaries has begun transitioning to an electronic system; as such, it is not directly controlled by the VP. 8 Much of this information is collected via the village-wide assembly, or gram sabha, which serves a key venue for citizens to engage with local o cials to discuss the administration of gov- ernment programs. In 1998, in response to the widely acknowledged problem of infrequent gram sabhas, the State Government of Tamil Nadu mandated that all VPs hold a minimum of four gram sabhas each year: January 26th (Republic Day), May 1 (May Day), August 15th (Independence day) and October 2nd (the birthday of Mahatma Gandhi). Since passage, this mandate has had near universal compliance; today, panchayat elections and the quarterly ritual of the gram sabha have become ingrained into the political culture of rural Tamil Nadu. 3 What Counts as (Good) Deliberation? Deliberation is a process of “mutual communication that involves weighing and re ecting on preferences, values and interests regarding ma ers of common concern,” (Mansbridge, 2015, 27). In contrast to forms of democracy that emphasize aggregate preferences via the ballot, for exam- ple, deliberation ideally fosters agreement by persuading people of a di erent way of thinking (e.g. by providing new information or changing their preferences), or by a process of reasoned compromise. When it is e ective, deliberation can be transformative; it can empower poor com- munities, enhance the capacity for collective action, and harness the capacity of communities to manage their own a airs (Heller and Rao, 2015). De nitions of deliberation, and the normative standards underlying them, have evolved con- siderably over the last decade, partly as a consequence of empirical work from eld and lab set- tings. While more traditional de nitions of deliberation (Habermas, 1990; Elster, 1998; Dryzek, 1994; Gu man and ompson, 2004; Goodin, 2005) o en presume equality among actors and limit what counts as deliberation to claims rooted in rationality and impartiality, these standards have been challenged by the rapid revival of deliberative institutions in the developing world. In this section, we de ne the metrics by which we evaluate deliberative quality in such a se ing — that is, one in which inequality and illiteracy may shape pa erns of discussion and debate. 9 We begin with Mansbridge’s (2015) minimalist de nition, which explicitly acknowledges that deliberation, particularly among the less educated, may depart from purely “rational” speech; rather, deliberation may involve story-telling and emotional claims that are meant to build em- pathy, trigger a sense of injustice, and establish credibility. Indeed prior studies show that low literacy may contribute to limited “oratory competency” (Sanyal et al., 2015), where speech may engage in identity claims and declarations rather than rational re ection focused on communi- cating, and weighing between, competing interests (Rao and Sanyal, 2010). ough such speech would be excluded by a more traditional de nition of deliberation, it still constitutes “mutual communication regarding ma ers of common concern.” Moreover, even this type of speech can still provide functional bene ts, such as improving the transmission of information, coordinating collective action, and bolstering the legitimacy of decisions (Fearon, 1998). In using this more minimalist standard for deliberation, we also depart from previous a empts achtiger et al.’s (2005) Discourse to measure deliberative quality, including, for example, B¨ al- ity Index (DQI), which derives largely from a Habermasian vision and include measures for the “level” and “content” of justi cations used in arguments — components which value “rational” arguments over persuasive story telling or identity claims. Instead, we focus on measures that relate explicitly to the political and ethical functions of deliberation (Mansbridge, 2015).5 at is, we conceive of good deliberation as that which (1) gives all participants an equal opportunity to in uence the outcome by promoting “an inclusive and egalitarian political process,”6 ; (2) embod- ies the ideal of mutual respect, whereby citizens listen a entively to one another, and (3) allows citizens to be agents who participate in the governance of their society (Mansbridge, 2015, p. 43). We address each of these in turn. 5 Deliberation also includes an epistemic function — to “generate opinions, preferences, and decisions that are appropriately informed by facts and logic and derive from substantive and meaningful consideration of relevant reasons” (Mansbridge, 2015, p. 42). Since we do not collect information on the subsequent outcomes from these assemblies, or the welfare consequences of the decisions made, we do not include measures of the epistemic quality of deliberation in this paper. 6 Mansbridge (2015) describes such a process as one that includes “multiple and plural voices, interests, concerns, and claims on the basis of feasible equality” (p. 43). 10 3.1 Equality of Participation First, good deliberation must give participants equal opportunity to in uence the outcome — at its most basic, this can be captured with a measure of oor time. While the frequency or volume of speech alone may not be a measure of equality, the ability or willingness to speak does re ect one’s authority or standing in the community. By viewing speech as a social act, we follow Karpowitz and Mendelberg (2014), who de ne speech is “a form of symbolic political or civic participation that may re ect and contribute to the sense of political e cacy and authority — in short, as a political act that creates civic standing” (Karpowitz and Mendelberg, 2014, 5- 6). Understood as a political act, then, speakers’ relative amount of oor time can be a useful indicator of social equality. Equality, of course, may be de ned across multiple axes of di erence, including class, race, caste, and gender. While each of these merits consideration, in this study, we focus on gender for three reasons: rst, there is a signi cant body of scholarship that suggests that di erences in communication styles may limit women’s ability to be heard, to exercise authority, and to shape outcomes in deliberative se ings (Karpowitz and Mendelberg, 2014). In other words, deliberation as a method of collective decision-making may have a gendered component — and it is of nor- mative importance to understand the extent of such di erences, and how they can be overcome. Second, concerns of gender equality are perhaps more acute in contexts like rural Tamil Nadu, where this study is located. In such se ings, women are o en deeply disadvantaged across key welfare metrics — from health outcomes to education and labor force participation. For example, female signature literacy in Tamil Nadu is at a mere 64.5 percent in rural areas according to the 2011 census, with male literacy at 82.4 percent.7 Gaps in labor force participation are even more acute, with rural women employed half as o en as rural men (31.8 percent versus 59.3 percent).8 Given that women enter deliberative fora at a disadvantage, then, it is important to understand the ways in which gender — as a description of a person’s social identity, as a dimension of style 7 “Signature” literacy is de ned as the ability to sign one’s own name — another minimalist standard. 8 Directorate of Census Operations, Government of Tamil Nadu, http://www.tn.gov.in/dear/ Employment.pdf 11 of interaction, as a characteristic of the se ing — a ects pa erns and content of speech. ird, from a practical perspective, gender is a relatively easy marker of social identity to observe and code in deliberative se ings; in contrast to class or caste, which may be hard to identify visu- ally, gender di erences are immediately perceptible, allowing data collection on whether men or women are speaking at any given moment. 3.2 Agenda-Setting Power Second, good deliberation is characterized by citizens “listening a entively” to one another out of mutual respect (Mansbridge, 2015, p. 43). at is, participants should acknowledge what is said by others — not merely push their own agenda forward. To capture this concept, we examine whether a given citizen is as likely as another to have his issue addressed by the speakers that follow. Consider the following example from Neganur village, in which citizens are complaining about various public goods and infrastructure needs. Female 1: ere are many wells in our village, but the wells are without a pulley wheel. Moreover, since the water is not used for any purpose, it gets wasted. So if you can de-silt the wells, we can not only use the water for drinking purposes but for other purposes also… Male 1: e kitchen has been constructed in the balwadi [pre-school] in our village. It is not used. Please arrange for the construction of a toilet for women. We also need a play ground for games. e canals are muddy. We have to de-silt the canals. We need a library. All our children are going to school with a dream of becoming IAS and IPS o cers. But to get general knowledge, they need books in the library. Our President has not say ‘no’ for any of our requests. With the hope that he will de nitely do whatever we have asked, I take leave. Male (O cial): We have a library in our panchayat. We have arranged for 5 magazines — an English paper, e Hindu and 4 Tamil magazines. All the elderly persons and children are reading. I am also asking the o cers to improve the library and have passed resolution in this regard. We have already de-silted the canal and cleaned it under Mahatma Gandhi Rural Employment guarantee scheme. Viluppuram District Vallam Block Neganur Panchayat 12 Here, a woman raises a particular issue about well water, but before she is able to get a resolution, a man interrupts to raise a separate set of issues, which then generate a response and resolution from the village o cial. at a speaker is so obviously ignored by other participants represents a marked departure from good deliberation. More generally, by examining pa erns in the topic of discussion across whole assemblies, we can identify the speakers who are most likely to drive conversation. As the example above highlights, we ought to be particularly concerned about the way in which gender my in uence agenda se ing power — a disparity that has been well documented in other contexts (Karpowitz and Mendelberg, 2014), and that may be present here as well. 3.3 Responsiveness of the State Finally, good deliberation enables citizens to be active participants in their own governance. is is particularly relevant given the se ing studied here, which was explicitly designed so that citi- zens could play a greater in local development. As described above, the rural Indian gram sabha was formalized to give communities greater voice in the development process and to improve governmental transparency and accountability. Indeed, most gram sabhas begin with an explicit call for citizen participation; for example, the opening remarks from the village secretary in Ma. Kolukkudi begin as follows: Male (O cial): is gram sabha meeting takes place on the occasion of the 65th Re- public Day. is is a special gram sabha. I greet the panchayat president, women’s self help group members and higher o cials who have come to participate in this [sabha], and I extend a warm welcome. In this gram sabha many action plans are adopted. If you, the people, nd merit and demerit, you can discuss frankly and settle [the issue]… You can ask any question; we are duty bound to reply to them. You can nd out mistakes; you can make us feel what is wrong; you can say this is wrong. We are ready to correct our mistakes. If you do not ask [anything], you will not get [anything]… Cuddalore District Komaratchi Block Ma. Kolakkudi Panchayat 13 While most assemblies begin with such a call, the extent to which o cials actually respond to citizen requests varies tremendously. In Mullangudi, for example, the village president actively engages with a citizen who requests the construction of new infrastructure — not only exchanging information about potential sources of land for the requested projects, but also identifying the a ected parties and determining who needs to approve of the proposed solution before making a nal decision. Male 1: My name is Veerapandiayan… A marriage hall is needed for our village, crematorium is needed. Drainage is needed near the tank. Also, pathway is needed for crematorium. Cement road is needed for both streets…. we place these demands before you [the president]. Importantly, community hall is necessary. President, you have to respond. Male (President): You said that community hall is needed for the village. A er se- lecting the place for this, you should ask the village administrative o cer. If you give a memorandum to him he will consider the place needed for that and give consent for the place where it can be built. I will get it built without any hesitation… You have also asked for a marriage hall. ere is a plot for it. But there is no poramboke [government] land. ere is a poramboke [government] land near the temple. In that place there is a public toilet. We do not need that. All the public are ready to give in writing that [the public toilet] is not needed? Can you get it built there? Male 1: at is women’s sanitary complex. So I cannot do as you say… e women’s self help groups should say that it is not needed… Male 2 : We will get consent from the women’s association. A toilet facility will come in the marriage hall. Let them use that. ere is no problem. at sanitary complex is only lying waste. Male (President): Your demand is, of course, correct. But to a build marriage hall, that place is not su cient… Cuddalore District Komaratchi Block Mullangudi Panchayat Here, the president acknowledges the male citizen’s request for a community hall and a marriage hall, o ers a potential solution, and solicits feedback from the community about whether that solution is feasible. In many ways, this back-and-forth re ects the ideal form of deliberation, in 14 which participants are communicating to reach a mutually agreeable decision, and where citizens are able to actively participate in their governance. By contrast, the citizens of Veeranam receive no response to their concerns about corruption within their local government. Not only does the president fail to respond to citizen’s speci c accusation, but the panchayat secretary swi ly punts the issue to the end of the meeting, and redirects the conversation to another issue. Perhaps not surprisingly, the meeting ends before the corruption charge has been addressed. Male 1: So far, no work has been without bribing anybody. (Crowd murmurs.) Male 2 : Wait. You answer his question. Male 1: So far, has our President done any work without ge ing a bribe? Male 3 (Secretary): e answer for this question will be given at the end of the meet- ing. Discussion before gram sabha now regards unused open bore wells in public lands and individual lands… Tiruvannamalai District, andarampet Block, Veeranam Panchayat ese starkly di erent excerpts suggest the meaningful variation in responsiveness by the state. As such, our last measure of deliberative quality examines how likely citizens are to receive a relevant response from o cials, and how that varies by the gender of the speaker, the content of the speech, or the characteristics of the o cials who are present. 4 Data & Measures To evaluate the quality of deliberation in Tamil Nadu’s gram sahbas, we recorded, transcribed, and translated the proceedings of assemblies conducted on Republic Day 2014, one of the four mandated days for all villages in the state to hold a gram sabha. e full sample, which consisted of 100 such assemblies, was collected as part of a broader impact evaluation of the Pudhu Vaazhvu 15 Project, a woman-centered poverty alleviation program funded by the World Bank.9 For this paper, we focus only on villages in the control group to describe what deliberation looks like, absent any additional policy interventions. ese 50 villages are spread across 9 districts, chosen to ensure geographic representation.10 From these 50 villages, 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. In order to implement both the assembly recording and collection of surveys, two eld enu- merators were assigned to each village — one from outside the village to record data on the is- sues raised, and another, who was local, to collect a endance data and help identify the speaker.11 Given that the average a endance across our gram sabhas was nearly 120 people, the introduction of a single enumerator from a neighboring village was unlikely to a ect local citizens’ behavior. e local enumerator, who was necessary to correctly provide information on the participants’ positions, only recorded a endance data and assisted the outside enumerator with information on the identities of speakers. O cial a endance data in such meetings are typically recorded 9 Village selection for the impact evaluation leveraged our knowledge of program implementation to reconstruct the selection process, thereby creating a matched sample of comparable treatment and control villages. More specif- ically, within the set of eligible districts (chosen for geograhpic representativeness, blocks were selected for assign- ment based on two sets of criterion: (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 in- frastructural 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 active project districts on these two factors. is process allowed us to nearly replicate the original assignment process for PVP. 10 Districts include: Cuddalore, Kancheepuram, Nagapa inam, Namakkal, urvallur, Tirunelveli, Tiruppur, Tiruvan- namalai, and Vuluppuram. 11 Since all data had to be collected on a single day, this required a team of at least 200 enumerators — a number larger than any survey rm could provide. Moreover, familiarity with the gram sabha meetings was essential to our being able to collect this data accurately and in real time. In order to address both these constraints, we hired and trained local women as our eld enumerators. Using these enumerators was advantageous in being able to record the gram sabha proceedings without the disruption having an “outside” observer. To maintain independence of the data collection process, however, we ensured that eld enumerators who recorded the proceedings of the meeting were assigned collect data from a village in her neighboring, rather than home, district. Enumerators who helped identify the speakers were local residents, as local knowledge is essential in order to do this accurately. 16 only at the beginning of the meeting, if at all. Where available, our data on a endance at the time the meeting began were cross-validated with this o cial data by the external enumerator. e audio recordings of meetings were transcribed and translated into a corpus of textual data by an independent survey rm. Transcripts included verbatim transcriptions and translations of the assemblies, as well identi ers on the gender and position of each speaker.12 ese transcripts form the backbone of the following analysis. Each “document” in the corpus consists of an un- interrupted speech by a administrator, elected o cial, or citizen. From the 50 village assemblies, we have 1,736 such documents, each of which is identi ed by the position and gender. Table 1 presents descriptive information about the number and character of documents within each village. Assemblies have relatively good a endance (with 123 people a ending on average), and consist of roughly 34 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). Table 1: Village-Level Summary Statistics Mean Std. Dev. Median Min Max. Total A endance 123.51 83.77 103.00 25.00 462.00 Number of Speeches 34.72 22.27 29.50 4.00 97.00 Speech Length 109.92 158.22 71.68 25.60 1090.75 Percent Female 0.32 0.21 0.30 0.00 0.92 Percent Citizen 0.53 0.14 0.53 0.20 0.88 Percent Admin 0.31 0.17 0.28 0.00 0.75 Percent Politician 0.16 0.16 0.14 0.00 0.50 4.1 A Text-As-Data Approach to Deliberation While these descriptive statistics allow us to examine who speaks within the gram sabha, to un- derstand the agenda-se ing power of speakers and the state’s responsiveness to citizen issues, we 12 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); and (3) citizens, all people who neither hold a formal government job or elected o ce. ese may include members of social groups (e.g. SHGs) and other organizations, but are not direct employees of the state. 17 also examine what is said. More speci cally, we draw on natural language processing methods that use text as data to be er understand the content and character of speech. By treating our transcripts as textual data, we can estimate an unsupervised topic model, which is a computa- tional tool to “discover” a set of a salient topics within a document collection. While the complexity of language will never be fully captured by an automated method such as ours, this sort of analysis can help to overcome meaningful challenges in hand-coded analyses of deliberation — including biases due to the researcher’s priors and inconsistencies in coding across various se ings. Hand-coding begins with a pre-determined set of categories into which documents are classi ed — based on their content, tone, etc. By contrast, the unsupervised ap- proach allows us to learn the underlying features of the text without imposing our own assump- tions. ough this is necessarily imperfect and requires ex-post validation, it can be useful for identifying previously understudied or theoretically new aspects of speech in these se ings, as well as scaling up large volumes of textual data. Prior to estimating the topic model, we pre-process the set of 1,736 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.13 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.14 We also exclude numbers. From the original set of citizen speeches, 1,700 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 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, 13 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. 14 For example “repair,” “repairs,” “repairing” and “repaired” all stem to “repair.” 18 each document is characterized by a vector of proportions, representing the share of the document associated with each topics. Using STM, we identify a set of 15 topics15 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 Table 2, 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.16 Figure 1 presents the distribution of these topics across the full corpus. Table 2: Top Word Stems by Topic Topic Top Word Stems Water Highest Prob: water, road, tank, street, get, facil, arrang FREX: road, water, x, tank, pipe, street, drink Bene ciary & Voter Lists Highest Prob: get, give, given, card, name, person, list FREX: give, get, card, name, given, poverti, receiv Employment & Wages Highest Prob: ask, peopl, work, one, told, talk, know FREX: talk, told, ask, whatev, know, one, mistak Service Failures Highest Prob: come, tell, want, say, money, done, commot FREX: say, tell, commot, money, want, come, bus Greetings and anks Highest Prob: presid, take, meet, panchayat, request, o c, member FREX: request, thank, hospit, particip, presid, conduct, meet Ration Shop Highest Prob: day, need, time, proper, shop, ration, petit FREX: day, need, time, proper, petit, ration, shop Housing and Land Titles Highest Prob: hous, place, construct, month, pa a, make, everi FREX: pa a, said, construct, hous, gave, remain, make Allocation of Funds Highest Prob: rupe, scheme, govern, panchayat, fund, amount, provid FREX: rupe, amount, allot, govern, fund, thai, scheme Toilet Construction Highest Prob: build, toilet, land, built, govern, pay, use FREX: build, built, toilet, pay, land, hall, maintain Education Highest Prob: school, villag, children, women, panchayat, complex, pass FREX: school, children, complex, sanitari, pass, educ, villag Intro to PVP Highest Prob: group, loan, plf, regard, bank, vprc, inform FREX: loan, plf, bank, vprc, regard, certif, appoint SHGs Highest Prob: women, group, peopl, panchayat, help, list, self FREX: self, award, poor, status, help, women, survey Environmental Protection Highest Prob: scheme, hous, work, employ, subject, canal, select FREX: canal, gandhi, subject, employ, guarante, propos, set Announcements, Resolutions, and Voter’s Pledge Highest Prob: sabha, gram, approv, panchayat, inform, place, report FREX: sabha, gram, approv, audit, read, report, pledg Maintenance of Public Goods Highest Prob: panchayat, expens, discuss, use, regard, plastic, mainten FREX: plastic, mainten, expens, releas, avoid, discuss, instal A key challenge in the text as data literature, particularly with unsupervised methods, lies 15 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 = 15 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. For robustness, we also show full results for K =20 and K =30 models in Appendix B. 16 See Roberts et al. (2016) for a fuller explanation of FREX. 19 Figure 1: Distribution of Topics Across Corpus in how to interpret the topics that are produced. Here, we use highest probability and FREX words, as well as example documents associated with each topic, to generate a substantive label for each topic. Consider the top documents most associated with the two most frequent topics in the corpus: Topic: Water “I request you to repair the road in Mukkarumbur East colony. Drinking water, drink- ing water, drinking water, drinking water, water problem of colony has to be set right.” “If the tap is in regular use, water will be in good condition. You are not using the tap regularly so water is not in good condition.” “Please repair the pump in the junction of 3 roads. ere is no water. Or, the motor has to be repaired. we have to go around for water.” Topic: Bene ciary and Voter Lists “Checking the voters’ list, and adding names in voters list for 2014: ose who have completed 18 years recently may apply now for addition of their name. e corrected list a er addition and deletion of names, up to October 31st has been received… If anyone has come from outside to the village, they could also add their name in that special camp. Application was given to the eligible persons. Now we will readout the names, please listen…” “As per the scheme, priority should be given to di erently-abled persons. 2 or 3 persons have given a list in our Panchayat. It is not known who all have given their 20 names. NREGS cards have been given to 9 villages. NREGS cards should de nitely be given to di erently-abled persons. ey should be paid salary even when they simply stay at the site. is mission is mainly to identify the di erently-abled persons. All should participate in the peoples status survey. en only we will be able to di erentiate the poor and the di erently abled persons.” “I now read the newly included voters name at No.14, Seliambedu village. Amsaveni daughter of Ramakrishnan. Suganya daughter of Gnaprakasam. Gayathri wife of Kamalakannan. Kanimozhi wife of Devaraj. Aruldass. Babu son of Gnanaprakasam. Sridhar son of Ragupathi. Kalaiselvan son of Madasami. Arul son of Panneer Selvam. angamuthu son of Arumugam. e voters ID cards are with me. If anybody’s name is omi ed, you get the form from me and ll up the form.” 4.1.1 Topic Validation 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, as a test of predictive validity, we examine whether the topics that capture proforma features of the assem- bly are indeed more likely to be discussed by o cials, rather than citizens. More speci cally, the topic model identi es a set of standard remarks — such as the reading of resolutions, the formal greetings and votes of thanks, and discussion of government funding allocation — as dis- tinct topics. If these topics capture the rote features of assemblies as they are conducted, these should be primarily spoken by o cials, who are responsible for convening and adjourning the meeting, as well as sharing information about recent public expenditures. Figure 2 plots the dif- ference 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. Second, we also validate the topics against the survey data collected by enumerators sent to each village. More speci cally, as part of the data collection process, enumerators were asked to record information on the types of issues raised during the assemblies. Given this data, we can coarsely examine whether the type and frequency of issues counted in the survey-collected data correspond to their counterparts in transcript data. is comparison, while helpful, is necessarily 21 Figure 2: 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. imperfect for two reasons: First, while the survey-collected data merely count whether an issue was raised within a village assembly, the transcript data shares are calculated based on the propor- tion of documents associated with that topic. As such, the transcript data will overweight topics that are discussed at length or by many speakers, relative to those that are brie y mentioned. Second, while many topics have clear analogues across the datasets, others are coded di erently across the two sources. For example, whereas the survey data identify a single topic for envi- ronment and sanitation issues, in the transcript data, the inductive process of topic modeling distinguishes between environmental protection and the maintenance of public goods, including sanitation and recycling issues. Given these discrepancies, we nd the closest possible analogues, or aggregate where necessary. ere are also a handful of topics for which clear analogues are not available. For example, while the unsupervised topic model identi es “Voter and Bene ciary Selection” this does not come up in the survey data as an explicit issue — likely because the pro- cess of identifying the target poor is a regular procedure at most gram sabhas, and therefore was not picked by enumerators as an explicit issue. Despite these di erences in measurement, however, we can still evaluate whether the relative 22 Table 3: Validation of Topical Prevalence Using Survey Data Transcript Data Survey Data Water 0.1487 0.1743 Wages and Employment 0.0990 0.0647 Housing 0.0668 0.0540 Ration Shop 0.0735 0.0625 Toilets 0.0606 0.0625 Environment and Sanitation 0.0617 0.0511 Education 0.0446 0.0945 Funding 0.0612 0.0260 Women’s Issues 0.0810 0.1261 Note: is table presents the relative frequency of topics across both our survey and transcript data. Categories collected in the survey data were post-coded by issue area. For transcript data, documents were coded as a mixture of topics. As such, we take the share of all documents associated with that topic. Direct comparisons across the dataset was not possible for all topics, as there were only a limited set of clear analogues. frequency of speci c topics (water, housing, etc.) are roughly similar across the two datasets (Ta- ble 3). e similar proportions (both in levels and rank) for topics with ready analogues suggests that our unsupervised methods re ect substantively what hand-coded results would yield. 4.2 Measures of Deliberative Equality Having validated the output of the topic model, we can generate a set of quantitative measures to capture deliberative quality across our sample of villages. Deliberative quality here is assessed based on the three metrics identi ed above — namely, the equality of participation, agenda-se ing power, and responsiveness by the state. To evaluate the equality of participation, we look at both the frequency and volume of speech by gender and position. at is, we can examine counts for the number of speakers with each demographic category of interest (men versus women, citizens versus o cials). We also examine the length of speech as a proxy for the amount of oor time that speakers occupy. To be er understand who drives the topic of conversation, we examine the sequence of speech topics to estimate the likelihood that a given speech is followed a speech that addresses the same topic. Since any given speech is modeled as a mixture across multiple documents, we focus on the 23 primary and secondary topics that are associated with each topic. More speci cally, we generate three measures for agenda se ing power: (1) an indicator if either the the primary or secondary topic of speech i is the same as the primary or secondary topic of speech i + 1 (nextSame); (2) the share of the next ve speeches that address either the primary or secondary topic of speech i (prop5same); and (3) the length of speeches for which the primary or secondary topic of speech i is continues to be addressed (lengthTopic ). Given the frequency of topic changes, we only measure this for a maximum of 5 subsequent speeches. Based on these measures, we can then examine whether features of the speaker or assembly are associated with greater agenda-se ing power within the gram sabha. Lastly, since a key objective of the gram sabha is to provide citizens with the opportunity to speak directly to the state — 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. e la er consideration ensures that o cials are not merely co-opting the conversation by switching topics, but actually engaging with the concerns raised by citizens. 5 Patterns of Deliberative ality Using these measures — for equality of participation, agenda-se ing power, and ability to address the state — we can now examine pa erns of deliberative quality within Tamil Nadu’s gram sabhas. 5.1 Equality of Participation e most basic measure of equality relates to whether everyone has relatively equitable access to the oor. To examine this, we rst look at the share of speeches within each sabha that are made by citizens versus o cials, as well as for men versus women. Given that a key aim of 24 the assembly is to give citizens a chance to voice needs to o cials, and for o cials to respond, we would expect a healthy sabha to have roughly equal shares of speeches from both groups. Indeed, we nd that on average, citizens deliver 55.41 percent of speeches, while o cials deliver the remaining 44.59 percent. ese raw speech shares support the notion that the gram sabha is not merely a state-dominated space, in which o cials disseminate info or overtake the space; rather, citizens are able to speak up and engage others in a deliberative fashion. In terms of gender equity, however, we focus on speeches made by citizens and nd that di erences in speaking frequency are quite stark — a full 65 percent of speeches are made by men, while women speak only 35 percent of the time (Row 1, Table 4). Of course, such a disparity may simply re ect the shares of men and women in a endance; as such Table 4 also presents di erences in speaking frequency normalized by percent of men / women in a endance (Row 2), and normalized by percent of men / women among voters (Row 3). For these measures, a value of 1 indicates that women (or men) are speaking as frequently as their population share would suggest, while values greater than 1 indicate they women (or men) are speaking more frequently than their population share would warrant. Even with these normalizations, however, we see that the gender gap remains wide and signi cant. In part, this is because the a endance gap between men and women is not pronounced in Tamil Nadu, with women o en a ending the sabha in greater numbers than men. Table 4: Frequency of Citizen Speeches, by Gender Mean, Male Speeches Mean, Female Speeches t-statistic p value Raw Di erences 0.6623 0.3377 7.1362 0.0000 Normalized by A endance Share 2.5208 0.5979 3.7940 0.0004 Normlized by Population Share 1.3202 0.6801 6.8730 0.0000 To understand what might be driving the relative infrequency of female speech, we perform a series of multivariate regressions, which allow us to correlate village-level factors with the likeli- hood of female speech. Here, we look not just at citizens, but also at administrators and politicians to examine what role formal status may have in improving the women’s voice. In particular, we examine three factors that theoretically should improve the frequency of women’s speech: the 25 presence of a female president, the level of female a endance, and village-level female literacy. ough female a endance and literacy are likely endogenous, we can intepret the coe cient on the female incumbent causally. In doing so, we follow a signi cant body of literature that lever- ages the as-if-random assignment of gender quotas in Indian local government (Cha opadhyay and Du o, 2004; Bhavnani, 2009; Ban and Rao, 2008b; Besley et al., 2005) to determine the ef- fect of the incumbent’s gender on local governance outcomes. e speci c process for assigning women’s reservations is described in the Tamil Nadu Panchayats Rules of 1995, Section 7.3, Rule 7, which mandates the creation of a “list of Wards or Panchayats arranged in descending order of the percentage of…Women,” and then details the rotation of reservations every 10 years by proceeding down the list. Since the percentage of women in the population is roughly the same across the state’s Panchayats, assignment is as-if-random. Results are presented in Table 5. Models 1, 3, and 5 present the basic results: While female citizens speak slightly more o en when women a end the gram sabha in greater numbers, they are not more vocal in the presence of a female president or in more educated villages. Among politicians and administrators, the presence of a female president does positively correlate with female politician speech, likely due to the actions of the president herself. ese results hold even when we control for the overall “backwardness” of the district, using an indexed score that includes demographic and infrastructural variables (Models 2, 4, and 6).17 In addition to looking at the frequency of speech, we can also examine whether the total oor time occupied by men and women is roughly equal. Given that women speak signi cantly less o en than men, they would have to speak longer per speech to equalize oor time — but perhaps consistent with our expectations, they do not. Women on average speak a mere 55 words per speech, where as men average roughly 77 words per speech (Table 6). In other words, these per speech disparities only exascerbate the overall gender gap in oor time within each village. If 17 Variables used for the indexed score include: the number of villages in the block, average distance of the village to the nearest town, total population,the population shares of the Scheduled Caste and Scheduled Tribe communities, the number of households below the poverty line, the percentage of villages in the block which had primary and middle schools, commercial banks, cooperatives, agricultural and non-agricultural societies, medical facilities and drinking water facilities. 26 Table 5: Frequency of Female Speech Dependent variable: Female Speech Citizens Citizens Admin. Admin. Politicians Politicians (1) (2) (3) (4) (5) (6) Female President 0.10 0.11 0.21∗∗∗ 0.19∗∗∗ 0.90∗∗∗ 0.88∗∗∗ (0.08) (0.09) (0.09) (0.08) (0.07) (0.08) Female A endance 0.001∗∗ 0.001∗∗ 0.0003 0.0000 −0.0002 −0.0002 (0.0005) (0.001) (0.001) (0.001) (0.0004) (0.0004) Female Literacy 0.34 0.42 0.06 −0.14 −0.11 −0.34 (0.40) (0.41) (0.66) (0.81) (0.43) (0.46) District FE Backwardness Score Control Observations 913 913 473 473 322 322 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess Score is an measure of village level development, calculating using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. we compare average male to female oor time (measured by total number of words spoken by men versus women within a village), we see that men almost always occupy signi cantly more oor time than women (Table 7). e one exception is female politicians, who still speak less on average than male politicians, though the di erences are not statistically signi cant. at female politicians are able to approach parity with men in terms of oor time suggests that providing women with a formal role or position may be important to closing the gender gap in deliberation. Table 6: Length of Speeches, by Gender Mean, Male Speeches Mean, Female Speeches t-statistic p value All Speakers 77.3307 55.0601 2.7035 0.0069 Citizens Only 34.1925 32.0133 0.6526 0.5142 Administrators Only 152.8220 184.6585 -0.9009 0.3690 Politicians Only 70.1786 41.0845 2.2511 0.0251 Table 7: Assembly Floortime, by Gender Mean, Male Floortime Mean, Female Floortime t-statistic p value All Speakers 1758.5000 659.6200 6.1285 0.0000 Citizens Only 399.4600 233.0698 3.1440 0.0023 Administrators Only 1204.3778 590.5172 3.4689 0.0009 Politicians Only 529.0385 343.1765 1.1740 0.2478 27 5.2 Agenda-Setting Power While the mere amount of speech — in terms of frequency or volume — can be a useful indicator of deliberative equality, neither provides a full picture of a speakers’ ability to in uence discussion. A er all, a long-winded speech may be ignored just as easily as a short one. To that end, we examine the pa erns in agenda-se ing power. Here, we are speci cally concerned with whether there is a disparity between men and women in their ability to re-direct conversation toward their own ends. For this, we examine three measures of agenda se ing power — whether a speech is followed by one on the same topic (nextSame), the share of the following ve speeches that are on that same topic (prop5same), and the number of uninterrupted speeches that continue to discuss that topic (lengthTopic ). Table 8: Agenda Power by Position (All Speeches) Mean, O cials Mean, Citizens t-statistic p value Next Topic Same 0.5309 0.6006 -2.8386 0.0046 Perc. Same (Next 5 speeches) 0.4473 0.5152 -4.3642 0.0000 Length Topic 1.1620 1.3709 -2.6726 0.0076 Table 9: Agenda Power by Position (New Topics Only) Mean, O cials Mean, Citizens t-statistic p value Next Topic Same 0.4698 0.5287 -1.5152 0.1303 Perc. Same (Next 5 speeches) 0.3870 0.4773 -3.8086 0.0002 Length Topic 0.9457 1.1205 -1.5720 0.1165 Tables 8 through 10 present an initial look at the results. Strikingly, across all measures of agenda se ing power, citizens seem to have a much greater in uence on the direction of conver- sation than do o cials (Table 8). When a citizen raises a topic, the probability that the following speech will continue that topic is nearly 7 percent higher than when o cials raise a topic; sim- ilarly, citizen speeches are likely to generate conversation for a greater share of the following speeches and for longer uninterrupted stretches. Of course, this may simply be a function of o - cials’ resolution power, or ability to de nitely end a subject on a particular ma er, thus providing an open avenue for a new subject to be raised. To address this concern, we do two things: rst, we include topic xed e ects to make sure that it is not the speci c content that is driving the re- 28 sults (Table 11), and second, limit our sample to only those speeches in which a speaker is raising a new topic, and even then, the pa erns generally hold (Table 9).18 is suggests that the gram sabha is not merely a state-dominated space, in which o cials disseminate info or overtake the space. Rather, citizens are able to raise coherent issues and have others engage in a deliberative fashion. Table 10: Agenda Power by Gender (Citizen Speeches) Mean, Male Speakers Mean, Female Speakers t-statistic p value Next Topic Same 0.6150 0.5785 1.1011 0.2712 Perc. Same (Next 5 speeches) 0.5234 0.5025 1.0094 0.3132 Length Topic 1.4444 1.2592 1.7064 0.0883 With respect to gender, Table 10 initially suggests that di erences across the sexes are not striking — this, despite the fact that women speak signi cantly less o en than men. However, when we break down the results by position to see if these pa erns in agenda-se ing power holds across both citizens and o cials, the pa erns suggest important gendered di erences. In Table 11, we regress our measures of agenda se ing power on the interaction between an indicator for female speakers, and an indicator for a citizen speaker. Here, we nd that male citizens are the most likely to set the agenda; they are 10 percentage points more likely than the male politicians (the omi ed category) to have the speech following theirs stay on the same topic; given that only 56 percent of male politician speeches drive the conversation, this is an 18 percent increase in the agenda se ing power of male citizens — suggesting that the common man is incredibly powerful within the gram sabha. Notably, the dominance of male citizens persists to the inclusion of topic xed e ects, suggesting it is not that men are merely raising particular issues that others care about. e dynamic for women, however, is markedly di erent. ough village citizenship confers a relative advantage on men, it tends to disadvantage women. While male citizens are more likely to drive the agenda than male politicians, the same does not hold true for female citizens relative to female politicians. To be er understand the ways in which one’s position may condition the e ect 18 Here, a new topic is de ned simply as a deviation from the previous speech; the issue may have been raised at a much earlier point within the assembly. 29 Table 11: Agenda Se ing Power, by Gender and Position Dependent variable: Next Same % Next 5 Same Length Topic (1) (2) (3) (4) (5) (6) Female 0.07 0.06 0.02 0.02 0.13 0.10 (0.05) (0.04) (0.04) (0.03) (0.17) (0.14) Citizen 0.11∗∗∗ 0.08∗∗∗ 0.08∗∗∗ 0.06∗∗∗ 0.32∗∗∗ 0.24∗∗∗ (0.03) (0.03) (0.02) (0.02) (0.07) (0.08) Female x Citizen −0.11∗ −0.10∗ −0.05 −0.04 −0.33 −0.31 (0.06) (0.05) (0.05) (0.04) (0.23) (0.20) District FE Backwardness Score Control Topic FE Female President Control Observations 1,651 1,651 1,456 1,456 1,605 1,605 Note:∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the district, in parenthesis. e Back- warndess Score is an measure of village level development, calculated using demographic and infrastructal vari- ables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. of gender, we plot the interaction between gender and position in Figure 3. Among politicians, women are slightly more likely to shape the agenda than men (Figure 3a); by contrast, among citizens, women are consistently less likely than men to drive the agenda, and for the length of the topic discussed, these di erences are statistically signi cant at the 0.05 level (Figure 3b). Finally, to ensure that these results are robust to alternative speci cations of the topic model itself, we re-run the analysis with varying number of topics (K =∈ {20, 30}) and nd largely consistent results (presented in Appendix 2B). To be fair, disparities in agenda-se ing power may be inconsequential from a development perspective if men and women tend to discuss the same issues; however, if there are issues that are disproportionately addressed by women, who are also more likely to get ignored, then we may be particularly worried about development outcomes. To examine whether men and women do in fact overlap or di er in the issues they discuss, we plot the expected di erence in topic proportions between male and female citizens, along with the 95 percent con dence interval, for all non-proforma topics (Figure 4). While we see no signi cant di erences between men and women for the bulk of issues (sanitation, employment, service failures, housing, etc.), we do see stark di erences on particularly gendered issues, including self-help groups (women) and the 30 Figure 3: Agenda-Se ing Power by Gender and Position (a) Next Same (b) Length Topic Note: e gures above plot the interaction between gender and position on agenda se ing power within the gram sabha. e x-axis charts the whether speakers are citizens, and the y -axis graphs the coe cient for the e ect of being a woman and the 95 percent con dence interval. e model speci cation includes controls for village level demographics and infras- tructure, district xed e ects, and topic xed e ects. introduction of PVP (which is a women-centered poverty alleviation project). To the extent that female citizens are more likely to be ignored in the gram sabha, then, we ought to be concerned that issues that uniquely impact women will be the least likely to be resolved. 5.3 Responsiveness of the State While the ability to drive conversation is a meaningful indicator of one’s in uence in a deliber- ative se ing, perhaps of even more relevance for the gram sabha is whether citizens are able to elicit a meaningful response from state actors. at is, when citizens raise an issue to administra- tors or politicians, how likely are they to get an on-topic response, and does this responsiveness vary by gender? To examine state responsiveness, we generate an indicator variable which takes on a value of 1 if a citizen’s speech is followed by an administrator or politician and addresses either the primary or secondary topic of that speech. Table 12 examines citizen speeches and presents basic 31 Figure 4: Topical Prevalence of Issues, by Gender (Citizens Only) Note: e gure above plots the expected topic proportion and 95% con dence interval for each issue area, by the speaker’s gender. Data include only citizens speeches. Coe cients greater than zero indicate topics that are more frequently raised by women, while those less than zero indicate topics that are more frequently raised by men. di erences in means across the genders both on whether an o cial responded, and whether that response was on topic. Results are further broken down by the o cial’s position: administrator or politician. While men and women are equally likely to get a response from o cials, men are signi cantly more likely to get an on-topic response. Interestingly, this di erence is driven pri- marily by politicians; while politicians respond in a relevant manner to male speakers 70 percent of the time, they only respond to women 49 percent of the time. By contrast, administrators respond to all citizens on topic about 60 percent of the time. Table 12: Likelihood of O cial Response, by Gender Mean, Male Citizens Mean, Female Citizens t-statistic p value Any O cial Response 0.5657 0.5541 0.3503 0.7262 On Topic O cial Response (All) 0.6316 0.5415 2.0461 0.0414 On Topic Politician Response 0.7034 0.4860 3.5253 0.0005 On Topic Administrator Response 0.5730 0.6020 -0.4674 0.6408 To be fair, these di erences may be driven simply by whether the topic raised is new to the discussion — that is, if women are bringing up issues that few other people care about, politicians may be less likely to respond than if the issue were more popular. To address this, we not only control for whether a topic is “new” to the discussion (Table 13, Model 1), but also include topic 32 xed e ects (Table 13, Model 3); unsurprisingly, new topics are 20 percentage points less likely to elicit a response from politicians; however, even when we control for this, women are 18 percentage points less likely than men to receive a response from their elected o cial. Table 13: O cial Responsiveness, by Gender Dependent variable: On Topic Politician Response On Topic Admin. Response (1) (2) (3) (4) (5) (6) Female −0.18∗∗∗ −0.27∗∗∗ −0.28∗∗∗ −0.003 −0.17∗∗∗ −0.20∗∗∗ (0.05) (0.03) (0.04) (0.06) (0.05) (0.04) Female President −0.10 −0.12 −0.11 −0.08 (0.07) (0.08) (0.10) (0.10) New Topic −0.20∗∗∗ −0.20∗∗∗ −0.20∗∗∗ −0.12 −0.12 −0.09 (0.04) (0.04) (0.04) (0.09) (0.10) (0.08) Female x Female President 0.18∗∗∗ 0.20∗∗∗ 0.43∗∗∗ 0.42∗∗∗ (0.08) (0.09) (0.07) (0.08) District FE Backwardness Score Control Topic FE Observations 251 251 251 259 259 259 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess Score is an measure of village level development, calculating using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. 5.4 E ect of Mandated Representation on Deliberative Equality ese pa erns suggest that women, and women citizens in particular, are at a considerable disad- vantage in the gram sabha. ey speak less, are less likely to drive conversation, and are less likely to get a response from government o cials. And these disadvantages hold even in sabhas when we control for the issues that are raised. Indeed, it was in recognition of these deeply gendered inequalities that the Government of India pro-actively designed the gram sabha with quotas for women to serve on the village council and as village president. Here, we examine whether the presence of a female incumbent — to lead the discussion, to respond to citizens, and to act as a role model for female villagers — has an impact on the measures of deliberative equality explored above. Importantly, the assignment process for women’s reservations, established by the Tamil Nadu Panchayats Act (1994), is as-if-random — allowing us to interpret these e ects in a causal 33 manner (Ban and Rao, 2008b). First, we nd that the presence of a female president has no discernible impact on the like- lihood that female citizens participate within the gram sabha (Table 4, Models 1 and 2). at is, the mere presence of a female president does not seem to encourage more women to a end or speak within the gram sabha, suggesting that the “role model” e ects of such incumbents may not be su cient to a ect deliberative participation in the short term. ough these results might seem surprising given the optimism around female quotas, they are quite consistent with evi- dence from Bengal in Cha opadhyay and Du o (2004) and South India in Ban and Rao (2008a), which nd no e ect of female reservation on the political behavior of ordinary women. By contrast, we nd that the presence of a female president does have a meaningful and sig- ni cant impact on the ability of women to be heard and responded to. Focusing only on citizen speakers, Table 14 regresses our measures of agenda-se ing power on indicators for gender of the speaker and the gender of the village council president. As expected, women speakers are at a considerable disadvantage relative to male speakers (roughly 14 percentage points less likely to drive the next issue discussed), but this disadvantage is essentially reversed under female presi- dents. For clarity, we visualize the interactions in Figure 5, which plots the coe cient estimates for the e ect of being a female speaker under male and female presidents respectively. While un- der male presidents, women are signi cantly less likely than men to set the agenda, under female presidents, di erences between the genders are not only smaller in magnitude, but statistically insigni cant. To be er understand whether and how female presidents themselves might be elevating the voices of other women, we also look at whether women citizens are more likely to generate a relevant response from state o cials. Table 13 presents the results for both politician responsive- ness (Models 1 - 3) and administrator responsiveness (Models 4 - 6). e data are striking: overall, women are 18 percentage points less likely than men to receive a relevant response from elected o cials – a meaningful decline given that men receive topical responses 70 percent of the time. Importantly, however, the presence of a female president can ameliorate the neglect that female 34 Table 14: Agenda Se ing Power, by Gender of Speaker and Gender of President Dependent variable: Next Same % Next 5 Same Length Topic (1) (2) (3) (4) (5) (6) Female −0.14∗∗∗ −0.14∗∗∗ −0.05 −0.05 −0.37∗∗∗ −0.40∗∗∗ (0.03) (0.03) (0.04) (0.04) (0.11) (0.12) Fem. Pres. −0.09∗ −0.09∗ −0.07 −0.08∗ −0.33 −0.33∗ (0.05) (0.05) (0.05) (0.05) (0.20) (0.19) Female x Fem. Pres. 0.20∗∗∗ 0.19∗∗∗ 0.02 0.02 0.30∗∗∗ 0.28∗∗∗ (0.05) (0.06) (0.05) (0.04) (0.13) (0.10) District FE Backwardness Score Control Topic FE Observations 924 924 818 818 895 895 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01.Sample include only speeches delivered by citizens (all administrator and politician speech is excluded). Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess Score is an measure of village level development, calculated using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. Figure 5: Agenda-Se ing Power by Gender of Speaker and Gender of President (a) Next Same (b) Length Topic Note: e gures above plot the interaction between speaker’s gender and president’s gender on agenda se ing power within the gram sabha. e x-axis charts the whether the president is a woman, and the y -axis graphs the coe cient for the e ect of being a woman and the 95 percent con dence interval. e model speci cation includes controls for village level demographics and infrastructure, district xed e ects, and topic xed e ects. citizens face. Basic results are reported in Models 2 and 4 of Table 13, while Models 3 and 5 also include xed e ects for the topic discussion. For clarity, we plot the interaction between an indi- 35 cator for female speakers and an indicator of a female president in Figure 6, which shows that the female incumbents signi cantly reduce the di erential treatment that men and women receive from the state. In fact, Figure 6a shows that women are slightly more likely than men to receive a topical response from elected o cials when a female president presides over the gram sabha. Interestingly, among (largely male) administrators, the gender di erentials under male and fe- male presidents follow the same general pa ern. Even though di erences are not statistically signi cant (Figure 6b), this still suggests that administrators follow the lead of female presidents. Figure 6: State Responsiveness by Gender of Speaker and Gender of President (a) On Topic Response from Politicians (b) On Topic Response from Administrators Note: e gures above plot the interaction between speaker’s gender and president’s gender on responsiveness by the state. e x-axis charts the whether the president is a woman, and the y -axis graphs the coe cient for the e ect of being a woman and the 95 percent con dence interval. e model speci cation includes controls for village level demographics and infrastructure, district xed e ects, and topic xed e ects. 6 Discussion Taken together, these pa erns suggest that we need to pay more a ention to the ways in which inequalities among citizens may a ect the ability of deliberative democratic institutions to deliver on their promise — to engage citizens in the development process and produce more inclusive development outcomes. Of course, scholarship on Indian local government has examined these 36 inequalities in development outcomes, including those along gender lines, but very li le research has been able to open the “black box” of the deliberative bodies that are at the core of India’s decentralization e ort. In this paper, we do just that. By using text-as-data methods on an original corpus of village assembly transcripts from rural Tamil Nadu, we show that these assemblies are not merely empty spaces where state o cials bluster and read banal announcements; rather, they provide meaningful forums for citizens to challenge their elected o cials, demand transparency, and provide information about very real local development needs — from water and sanitation issues, to wage payments and government service failures. We also show, however, that among citizens, inequalities in power and status meaningfully impact the citizens’ ability to be heard. More speci cally, we show that across multiple measures of deliberative in uence, women are at a considerable disadvantage. ey are less likely to be heard, less likely to drive the agenda, and less likely to receive a relevant response from state o cials. Indeed, even when we account for the particular issues raised, women still remain at a disadvantage — o en ignored while their male peers receive a direct response. In the excerpt below, for example, a woman raises a genuine concern about the lack of ration shop facility in the village, to which the o cial does not respond; a man then raises the exact same issue a er her, and receives an immediate response from the elected o cial, who promises to speak to higher-ups about what can be done to address the issue: Female 1: In Pa upalli village, so far, there is no fair price shop. ey are keeping it in the Women’s Health Building. Women are quarrelling. e village people want it built new. ere is ght in the panchayat. So people are going to the neighboring village. But the pa a [titled] land owners are preventing them from using their land for going to the next village, so resolution should be passed for construction of a ration shop here. Male 1: For so many years, there is no ration shop here. Only rental shops are here. So long, it was in rented place and now it is kept in Women’s Health Association. Now, women ask for the building and want a fair price shop built. So there is a lot of problem. Please establish for us a ration shop. Male (President): Regarding this ration shop, we should talk with MLA [Member of the Legislative Assembly] and BDO [Block Development O cer]. e request will be made… 37 iruvallur District Minjur Block Sengayam Panchayat To be fair, one might think the above excerpt is not problematic insofar as the male politi- cian eventually responded to her substantive concern about the ration shop. However, from the perspective of deliberative equality, for women to in uence conversation as democratic equals, they should not have to wait for a man to elevate their concern before an o cial responds. ese pa erns of gendered discourse are perhaps unsurprising, but they do reiterate a need to be er design deliberative institutions to elevate the voices of women. In fact, our evidence suggests that women’s voices are more likely to be ampli ed with female presidents — under whom women are more likely to be heard and more likely to receive a state response. In the excerpt below, for example, we see a female president speci cally calling out women’s needs and using that as justi cation for a proposed resolution around liquor shop and ration shop concerns. Female 1: We need a ration shop for our village. We nd it di cult to go up to Devireddikuppam. We have to walk for 7 days in a month. We can t walk such a long distance keeping the rice bag in hands. You have to nd a solution for this problem and, at least, arrange a part time ration shop in our village. You take action for removing the liquor shop. We can t use the road a er 7 o clock. Drunken people are giving much trouble and using vulgar words. Female (President): Women are talking much about the ration shop and liquor shop. We will include these subjects in the resolution… Tiruvanamalai District andarampet Block Kolamanjanur Panchayat at the president explicitly elevates the requests of the women who are talking in her village underscores the notion that descriptive representation can improve the vertical communication between citizen and politician. However, our evidence suggests that this is no panacea for the deeper problem of women’s general silence. While women in Tamil Nadu are more likely to a end the gram sabhas than their male peers, they are signi cantly less likely to speak — even when 38 a female president sits in power. is lack of participation has been documented in a variety of se ings, both in the developed and developing world, and is undoubtedly driven by a combination of factors, from information and resource constraints to cultural biases against women. But it also underscores a deeper challenge in deliberation — namely, that using one’s voice can be a costly exercise, and one that di erentially burdens those who are already least-advantaged. Whether and how the common woman can be encouraged to overcome that cost is a deeply important question facing development practitioners, and one that will undoubtedly impact the ability of local government to address women’s needs. 39 References Auerbach, A. M. (2017). Neighborhood Associations and the Urban Poor: India’s Slum Develop- ment Commi ees, World Development . URL: h p://www.sciencedirect.com/science/article/pii/S0305750X15311797 achtiger, A., Sp¨ B¨ orndli, M., Steenbergen, M. R. and Steiner, J. (2005). e deliberative dimensions of legislatures, Acta Politica 40(2): 225–238. Baiocchi, G., Heller, P., Silva, M. K. and Silva, M. (2011). Bootstrapping democracy: Transforming local governance and civil society in Brazil, Stanford University Press. Ban, R., Jha, S. and Rao, V. (2012). Who has voice in a deliberative democracy? Evidence from transcripts of village parliaments in south India, Journal of Development Economics 99(2): 428– 438. Ban, R. and Rao, V. (2008a). Is deliberation equitable? Evidence from transcripts of village meet- ings in south India, World Bank Policy Research Working Paper Series (4928). Ban, R. and Rao, V. (2008b). Tokenism or agency? e impact of women’s reservations on village democracies in south India, Economic Development and Cultural Change 56(3): 501–530. Besley, T., Pande, R. and Rao, V. (2005). Participatory democracy in action: Survey evidence from South India, Journal of the European Economic Association 3(2-3): 648–657. Bha acharjee, S. and Cha opadhyay, R. (2011). e Information De cit: Use of Media in Delib- erative Democracy, Economic and Political Weekly pp. 45–52. Bhavnani, R. R. (2009). Do electoral quotas work a er they are withdrawn? Evidence from a natural experiment in India, American Political Science Review 103(01): 23–35. Cha opadhyay, R. and Du o, E. (2004). Women as policy makers: Evidence from a randomized policy experiment in India, Econometrica 72(5): 1409–1443. Clayton, A., Josefsson, C. and Wang, V. (2016). otas and Women’s Substantive Representation: Evidence from a Content Analysis of Ugandan Plenary Debates, Politics & Gender pp. 1–29. Dryzek, J. S. (1994). Discursive democracy: Politics, policy, and political science, Cambridge Uni- versity Press. Elster, J. (1998). Deliberative democracy, Vol. 1, Cambridge University Press. Fearon, J. D. (1998). Deliberation as discussion, in J. Elster (ed.), Deliberative democracy, Cam- bridge University Press, Cambridge, pp. 44–68. Fishkin, J. S. and Luskin, R. C. (2005). Experimenting with a democratic ideal: Deliberative polling and public opinion, Acta Politica 40(3): 284–298. Fung, A. (2004). Deliberation’s darker side: six questions for Iris Marion Young and Jane Mans- bridge, National Civic Review 93(4): 47–54. 40 Goeree, J. K. and Yariv, L. (2011). An experimental study of collective deliberation, Econometrica 79(3): 893–921. Goodin, R. E. (2005). Sequencing deliberative moments, Acta Politica 40(2): 182–196. Gu man, A. and ompson, D. (2004). Why deliberative democracy. Habermas, J. (1990). Moral consciousness and communicative action, MIT press. Heller, P., Harilal, K. and Chaudhuri, S. (2007). Building Local Democracy: Evaluating the Impact of Decentralization in Kerala, India, World Development 35(4): 626–648. Heller, P. and Rao, V. (2015). Deliberation and Development: Rethinking the Role of Voice and Collective Action in Unequal Societies, World Bank Publications. Karpowitz, C. F. and Mendelberg, T. (2014). e silent sex: Gender, deliberation, and institutions, Princeton University Press. Karpowitz, C. F., Mendelberg, T. and Shaker, L. (2012). Gender inequality in deliberative partici- pation, American Political Science Review 106(03): 533–547. List, C., Luskin, R. C., Fishkin, J. S. and McLean, I. (2013). Deliberation, single-peakedness, and the possibility of meaningful democracy: evidence from deliberative polls, Journal of Politics 75(01): 80–95. Mackie, G. (2015). Traveling to the Village of Knowledge, in V. Rao and P. Heller (eds), Deliberation and Development: Rethinking the Role of Voice and Collective Action in Unequal Societies, World Bank Publications, pp. 85–106. Mansbridge, J. (1980). Beyond adversarial democracy, Chicago, IL: University of ChicagoPress . Mansbridge, J. (1999). Should blacks represent blacks and women represent women? A contingent yes , e Journal of politics 61(03): 628–657. Mansbridge, J. (2015). A Minimalist De nition of Deliberation, in V. Rao and P. Heller (eds), Deliberation and Development: Rethinking the Role of Voice and Collective Action in Unequal Societies, World Bank Publications, pp. 27–50. Mansuri, G. and Rao, V. (2012). Localizing development: does participation work?, World Bank Publications. Myers, C. D. and Mendelberg, T. (2013). Political deliberation, e Oxford handbook of Political Psychology pp. 699–734. Parthasarathy, R. and Rao, V. (2017). Deliberative democracy in india, World Bank Policy Research Working Paper (7995). Polle a, F. (2004). Culture is not just in your head, Rethinking social movements: Structure, mean- ing, and emotion, pp. 97–110. 41 Polle a, F. and Lee, J. (2006). Is telling stories good for democracy? Rhetoric in public deliberation a er 9/11, American Sociological Review 71(5): 699–721. Rao, V. and Sanyal, P. (2010). Dignity through discourse: poverty and the culture of deliberation in Indian village democracies, e Annals of the American Academy of Political and Social Science 629(1): 146–172. Roberts, M. E., Stewart, B. M. and Airoldi, E. M. (2016). A model of text for experimentation in the social sciences, Journal of the American Statistical Association 111(515): 988–1003. Sanyal, P., Rao, V. and Prabhakar, U. (2015). Oral Democracy and Women’s Oratory Competency in Indian Village Assemblies: A alitative Analysis, World Bank Policy Research Working Paper (7416). Sen, A. (2001). Development as freedom, Oxford Paperbacks. Swidler, A. and Watkins, S. C. (2015). Practices of Deliberation in Rural Malawi, pp. 133–166. Tamil Nadu Panchayats Act (1994). 42 Appendices A Topic Model Validation Using Survey Data In addition to the predictive validation exercises, we also validate the output of our topic model by comparing the distribution of topics generated against topics coded by survey enumerators who recorded the proceedings of each gram sabha. ough this comparison is useful, it is neces- sarily imperfect because clear analogues do not always exist across the topic model output and the enumerators pre-determined categories. As such, where possible, we aggregate topics for comparison. Table A.1 presents the topics from the topic model, along side the survey data topic used for comparison. Certain categories across both data sources had no clear comparison, and were thus excluded from the validation exercise. Table A.1: Topic Comparisons for Validation Transcript Topics Survey Data Topics Panchayat Expenses Allocation of Funds Taxes Maintenance of Public Goods Sanitation and Environment Environmental Protection Employment & Wages Employment Water Water Toilet Construction Toilets Education Education Childcare Ration Shop Ration Shop Housing and Land Titles Housing Analogues not available Resolution Announcements Greetings and anks Bene ciary and Voter List Intro to PVP SHGs Service Failures Health Roads Women’s Issues Elderly Care Animal Care Electricity Voter ID Cards Village Organizations 43 B Results Under Alternative Topic Model Speci cations To ensure that the main results for agenda se ing power and state responsiveness are not sensi- tive to a particular topic model speci cation, we re-run our topic model with K = 20 and K = 30 topics, generate new measures of deliberative in uence, and present results below. B.1 Agenda Setting Power We rst re-examine how agenda-se ing power varies with the gender and status of the speaker. Consistent with the main results presented (for K = 15 topics), we see that even under these alternative model speci cations, male citizens are more likely to drive the agenda than male politicians, and female citizens are less likely to drive the agenda than female politicians. Point estimates all follow the pa erns presented in the main results (Table 11), but lose statistical sig- ni cance for K = 30. Table B.1: Agenda Se ing Power, by Gender and Position (K = 20) Dependent variable: Next Same % Next 5 Same Length Topic (1) (2) (3) (4) (5) (6) Female 0.13∗∗∗ 0.11∗∗∗ 0.09∗∗∗ 0.07∗∗∗ 0.29∗ 0.24 (0.05) (0.05) (0.03) (0.03) (0.17) (0.16) Citizen 0.11∗∗∗ 0.10∗∗ 0.11∗∗∗ 0.09∗∗∗ 0.26∗∗∗ 0.25∗∗∗ (0.05) (0.05) (0.03) (0.02) (0.11) (0.09) Female x Citizen −0.13∗ −0.13∗ −0.13∗∗∗ −0.11∗∗∗ −0.29 −0.27 (0.07) (0.07) (0.05) (0.04) (0.21) (0.19) District FE Backwardness Score Control Topic FE Female President Control Observations 1,651 1,651 1,456 1,456 1,607 1,607 Note: Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Score is an measure of village level development, calculated using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. 44 Table B.2: Agenda Se ing Power, by Gender and Position (K = 30) Dependent variable: Next Same % Next 5 Same Length Topic (1) (2) (3) (4) (5) (6) Female 0.07∗ 0.05 0.05∗ 0.02 0.21∗∗∗ 0.13 (0.03) (0.04) (0.03) (0.03) (0.09) (0.09) Citizen 0.10∗∗∗ 0.07∗ 0.07∗∗∗ 0.04∗∗∗ 0.24∗∗∗ 0.18∗∗∗ (0.04) (0.04) (0.02) (0.02) (0.08) (0.08) Female x Citizen −0.04 −0.03 −0.04 −0.03 −0.21 −0.19 (0.07) (0.08) (0.05) (0.05) (0.18) (0.19) District FE Backwardness Score Control Topic FE Female President Control Observations 1,651 1,651 1,456 1,456 1,624 1,624 Note: Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Score is an measure of village level development, calculated using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. 45 B.2 State Responsiveness Next, we examine whether women are less likely to receive a relevant response from state o cials, as they do in the main results. Once again, point estimates are consistent with two broad pat- terns: rst, female citizens are signi cantly less likely to receive a topical response from elected male politicians; and second, they are signi cantly more likely to receive a relevant response from female incumbents. While evidence of women’s relative neglect is consistent and statistically sig- ni cant across both topic model speci cations, the coe cient on female president responsiveness to female citizens loses statistical signi cance in the K = 30 speci cation. Table B.3: O cial Responsiveness, by Gender (K = 20) Dependent variable: On Topic Politician Response On Topic Admin. Response (1) (2) (3) (4) (5) (6) Female −0.09 −0.17∗ −0.20∗∗∗ −0.06∗∗∗ −0.10 −0.16∗∗∗ (0.09) (0.10) (0.08) (0.03) (0.06) (0.07) Female President 0.01 0.02 0.13∗ 0.13 (0.10) (0.11) (0.07) (0.08) New Topic −0.19∗∗∗ −0.18∗∗∗ −0.19∗∗∗ −0.09 −0.09 −0.06 (0.05) (0.05) (0.06) (0.07) (0.07) (0.07) Female x Female President 0.23 0.21 0.12 0.16 (0.17) (0.17) (0.14) (0.15) District FE Backwardness Score Control Topic FE Observations 233 233 233 262 262 262 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess Score is an measure of village level development, calculated using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. 46 Table B.4: O cial Responsiveness, by Gender (K = 30) Dependent variable: On Topic Politician Response On Topic Admin. Response (1) (2) (3) (4) (5) (6) Female 0.03 −0.12∗∗∗ −0.15∗∗ 0.06 −0.01 −0.003 (0.07) (0.04) (0.07) (0.08) (0.07) (0.06) Female President −0.16∗ −0.15∗ 0.13∗ 0.09∗∗ (0.08) (0.09) (0.07) (0.04) New Topic −0.08 −0.09 −0.02 −0.06 −0.06 −0.03 (0.06) (0.07) (0.08) (0.07) (0.07) (0.07) Female x Female President 0.35∗∗∗ 0.33∗∗∗ 0.18 0.18 (0.10) (0.13) (0.14) (0.12) District FE Backwardness Score Control Topic FE Observations 233 233 233 262 262 262 Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01.Robust Standard Errors, clustered at the district, in parenthesis. e Backwarndess Score is an measure of village level development, calculated using demographic and infrastructal variables, including the share of population belonging to the Scheduled Castes and Tribes, as well as indicators for the presence of a primary or secondary school, hospital or medical clinic, and bank. 47