WPS6215 Policy Research Working Paper 6215 Impact Evaluation Series No. 66 Conditional Cash Transfers, Political Participation, and Voting Behavior Javier E. Baez Adriana Camacho Emily Conover Román A. Zárate The World Bank Poverty Reduction and Economic Management Network Poverty, Gender and Equity Group October 2012 Policy Research Working Paper 6215 Abstract This paper estimates the effect of enrollment in a large choice, the authors find that program participants scale anti-poverty program in Colombia, Familias en expressed a stronger preference for the official party Acción, on intent to vote, turnout and electoral choice. that implemented and expanded the program. Overall, For identification the analysis uses discontinuities in the findings show that voters respond to targeted program eligibility and variation in program enrollment transfers and that these transfers can foster support for across voting booths. It finds that Familias en Acción incumbents, thus making the case for designing political had a positive effect on political participation in the and legislative mechanisms, as the laws recently passed 2010 presidential elections by increasing the probability by the Colombian government, that avoid successful that program beneficiaries registered to vote and cast anti-poverty schemes from being captured by political a ballot, particularly among women. Regarding voter’s patronage. This paper is a product of the Poverty, Gender and Equity Group, Poverty Reduction and Economic Management Network. 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 author may be contacted at jbaez@worldbank.org. The Impact Evaluation Series has been established in recognition of the importance of impact evaluation studies for World Bank operations and for development in general. The series serves as a vehicle for the dissemination of findings of those studies. Papers in this series are part of the Bank’s Policy Research Working Paper Series. 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 Conditional Cash Transfers, Political Participation, and Voting Behavior∗ Javier E. Baez Adriana Camacho Emily Conover an A. Z´ Rom´ arate October 1, 2012 Abstract This paper estimates the effect of enrollment in a large scale anti-poverty program in Colombia, Familias en Acci´ on (FA), on intent to vote, turnout and electoral choice. For iden- ti�cation we use discontinuities in program eligibility and variation in program enrollment across voting booths. We �nd that in the 2010 presidential elections FA bene�ciaries relative to non-bene�ciaries are more likely to register to vote (1.6-2.5 pp). A standard deviation increase in the proportion of FA bene�ciaries at the voting booth results in approximately a 2 pp increase in the probability of casting a ballot and of voting for the incumbent party under which the program was expanded. These results are driven by women, the direct recipients of the transfer. The magnitude of our results does not explain the �nal outcome of the presidential elections, but their direction shows that voters respond to targeted transfers and that these transfers can foster support for incumbents. JEL Classi�cation: O10, D72, P16. Keywords: Conditional Cash Transfers, voting behavior, Colombia. ∗ Contact information: Javier E. Baez, World Bank and IZA, email:jbaez@worldbank.org. Adriana Camacho, Department of Economics and CEDE, Universidad de Los Andes, email: adcamach@uniandes.edu.co. Emily Conover, Department of Economics at Hamilton College, email: econover@hamilton.edu. Rom´ an A. Z´arate, Universidad de Los Andes, email: ra.zarate22@uniandes.edu.co. Comments welcome. Thanks to Roberto Angulo, Hernando Sanchez and Staff at Acci´ on Social for granting access to the data used in the paper. We also thank comments from Paloma Anos Casero, Louise Cord, Aline Coudouel, Marcela Eslava, Francisco Gallego, Theresa Jones, Luis Felipe Lopez Calva, James Robinson, Carlos Rodriguez, Hernando Sanchez, Dean Scrimgeour, Renos Vakis and participants at the 2011 AEA Meetings, Brown University micro lunch, seminars at Lehigh University, Universidad de los Andes, Vassar College, and the World Bank. All errors are ours. The �ndings, 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. c by Javier B´ aez, Adriana Camacho, Emily Conover and Rom´ an A. Z´ arate. All rights reserved. 1 1 Introduction In the last 15 years Conditional Cash Transfer (CCT) programs have become a major component of the poverty reduction and social protection strategies of many developing countries. While most of these programs started in Latin America, currently CCTs are used in over 40 coun- tries spanning several regions of the world (World Bank, 2011). Even though certain program characteristics vary from country to country, in general standard CCTs provide money to poor families, through women, contingent upon investments in the human capital of their children, such as regular school attendance, basic preventive health care and better nutrition. A large number of CCTs have been subject to rigorous evaluations, most of which show that they have ful�lled the twin primary objectives of alleviating poverty in the short term and building the human capital of poor children. These studies have also uncovered a number of indirect effects ranging from reductions in child labor to increases in savings and productive assets, changes in sexual behavior and marriage decisions, and higher reliance on better risk-coping strategies (Fiszbein et al. 2009 for an overview). There is also a political economy aspect to social transfers. In theory, anti-poverty pro- grams such as CCTs may play a role in influencing individual political participation – in the form of voting – and preferences, strengthening democratic representation but also producing electoral rewards. For instance, by partly changing the economic circumstances of households, transfer receipts could persuade participant households to exercise their right to vote (Gleason, 2001). Politicians could also strategically allocate bene�ts to certain groups of people to raise political support or persuade recipients to cast a ballot in favor of the incumbent (Robinson and Verdier 2002, Camacho and Conover 2011, Drazen and Eslava 2012). Moreover, citizens may get signals from social policy choices and use them to infer the competence of politicians and their preferences for redistribution (Rogoff 1990; Drazen and Eslava, 2010; Banerjee et al., 2 2011). Alternative hypothesis include the possibility of voting as a result of reciprocity towards politicians, fear of losing coverage and increased access to more and better sources of information on political rights and processes (Finan and Schechter 2010, Manacorda et al., 2011, Gine and Mansuri 2011). The possibility of reaping electoral returns by strategically allocating targeted transfers to strengthen political prospects is not only a theoretical prediction but also an issue that has caught the attention in current public debates. Conjectures on possible political rewards linked to participation in CCTs have been reported in the media following presidential elections in Ecuador, Peru, Mexico and Brazil (De la O, 2009). In the speci�c case of Colombia, different media outlets speculated right before the 2010 presidential election that the official government on (FA), a large scale CCT program had used the expansion and allocation of Familias en Acci´ examined here, to systematically increase its votes.1 More recently, and perhaps due to current debates on the possible misallocation of program bene�ts by local politicians, the Government on a formal national poverty reduc- of Colombia has passed laws to make of Familias en Acci´ tion program. To further avoid political capture of the program, the law bans enrolling new bene�ciaries three months before major elections (El Tiempo, May 24, 2012).2 Despite the potential interactions between government policies and voter decision making, little evidence is available to assess whether conditional social transfers encourage people to vote and influence their political choices. Rigorous evidence on the subject however, is starting to emerge. Using individual level self-reported data, Manacorda et al. (2011) �nd that bene�ciaries of PANES - a large and temporary unconditional Cash Transfer Program in Uruguay - express larger support for the incumbent that implemented the program. The authors attribute that extra support to the inference of bene�ciaries on the politicians’ redistributive preferences as 1 See http://www.lasillavacia.com/historia/16024 and http://www.lasillavacia.com/node/30017. 2 See article http://www.eltiempo.com/archivo/documento/MAM-5444542. 3 well as from reciprocity. Similarly, evidence for Romania shows that incumbents gained political support through a program aimed at helping poor families purchase a computer (Pop-Eleches and Pop-Eleches, 2009). Using administrative data at the municipality level, analysis of Mexico’s Oportunidades CCT suggests that turnout and incumbent vote share increased in villages that were randomly assigned on (Nupia to receive the program (De La O, 2009). Previous work for Colombia’s Familias en Acci´ 2012) point to results in the same direction. The larger bene�ciary rates in regions where Uribe won by more than 50% of the vote in 2002 are associated with an increase in the proportion of votes that go to the incumbent governing coalition; there is no statistically signi�cant effect for swing regions.3 Conditional and unconditional correlation analysis for Brazil suggests that the coverage of Bolsa Familia, a large scale CCT program, is associated with Lula’s vote share when he sought a second term (Zucco, 2008). Addressing causality between social policy and political participation and preferences is empirically challenging as it requires an exogenous source of variation in social transfer receipt and rich data on voting behavior. This paper exploits discontinuities in program eligibility and variation in participation across voting booths to estimate the effect of enrollment in Familias en on (FA), a Colombian CCT program, on the intent to vote, turn out and on electoral choice Acci´ during the 2010 Colombian presidential election. The use of detailed data at the individual and voting booth levels together with the identi�cation strategies (discussed in more detailed below) allow us to control for possible confounders such as policy endogeneity, individual selection in the take-up and reverse causality. Additionally, unlike other studies we use actual voting outcome data, as opposed to self-reported data which could suffer from reporting bias. We provide evidence that relative to non-participants, FA bene�ciaries of voting age are 1.6-2.5 percentage 3 This paper relies on variation at the municipality level. However, there are concerns that identi�cation strategy (concentration of swing voters and loyal voters) may violate the necessary exclusion restriction. For instance, loyal voters’ municipalities could have been strategically targeted by politicians when the program was expanded. 4 points more likely to register to vote. A standard deviation increase in the proportion of FA bene�ciaries at each booth results in a 1.6-1.8 percentage point increase in the probability of casting a ballot and a 1.5 percentage point increase in the probability of voting for the incumbent party under which the program was expanded. This effect is stronger for women, who are the direct recipients of the cash transfer as established by the program rules. The elected candidate won in the runoff election with a large margin (69 percent of the votes), thus our results are unlikely to explain the �nal outcome. However, they show that voters respond to targeted transfers and that these transfers can foster support for incumbents, thus making the case for designing political and legislative mechanisms that avoid successful anti-poverty schemes from being captured by political patronage. The paper is structured as follows. The next section provides background information of the Colombian electoral system and the FA program. Section 3 describes the main data sets used in the analysis. Section 4 describes the identi�cation strategy. Section 5 presents the empirical results, including discussion on some robustness checks. Finally, section 6 concludes and discusses possible channels for our results. 2 Background 2.1 Government, Electoral System and Political Context Following the principle of separation of powers, Colombia’s system of government is divided into executive, legislative and judicial branches. The President is the head of the executive branch and in that position plays the role of head of state and head of government. At the sub- national level, the executive power is conferred to department governors and municipal mayors. Colombia has a long history of party politics. Within a multi-party representative democracy framework, the Colombian President is chosen directly by the electorate to serve a four-year 5 term. An amendment to the Constitution passed by the Congress in 2004 and approved by the Constitutional Court in 2005 allows a head of state to run for a consecutive term in office. In 1991, Colombia adopted a two-round system, where the runoff is avoided if in the �rst round the winner receives more than 50 percent of the valid votes. Elections in Colombia -including the presidential election- are organized by the National ıa Nacional in Spanish), the institution respon- Registry Office (RN–its acronym for Registradur´ sible for the civil registry and identi�cation of people. The RN is in charge of maintaining an updated official voter registry before every election. The right to vote is granted to all Colom- bians aged 18 years and older. Before 1988, any citizen eligible to vote and interested in voting had to register in an electoral office. Currently, any citizen who requests and obtains a national edula) from RN after turning 18 is automatically enrolled in the official identity card (known as c´ electoral registry (Art. 49 of Ley Estatutaria).4 Before 2003 these automatic registrations were assigned to the largest polling station in each municipality. Since 2003 people are automati- cally registered to vote near the office where they obtained their valid identi�cation card.5 In practice, however, this is not widely known. Dates of registration to vote are determined by the RN and usually span a two-week period in the months before elections. An individual can register in a new polling station each time registration opens, but the person can vote only in the last place where she was validly registered. For presidential elections, registered voters can cast their vote for a candidate in an election day in roughly 10,000 polling stations scattered in 1,119 municipalities across the country.6 In the last decade, the political landscape has been dominated by the parties that imple- mented and expanded the FA program. Even though it was originally conceived and designed in 2001 under the government of the Colombian Conservative Party, FA became the flagship 4 http://www.registraduria.gov.co/Reforma-politica-y-Elecciones-de.html 5 http://www.colombia.com/especiales/elecciones_2011/pedagogia-electoral/ 6 http://www.registraduria.gov.co/Informacion/inscrip_ced_2011_preguntas.htm, http://www.colombia.com/especiales/elecciones_2011/pedagogia-electoral/ 6 public anti-poverty strategy of the Partido de la U during the two terms of President Alvaro Uribe (2002-2006 and 2006-2010), reaching almost national coverage. Juan M. Santos, the for- mer Defense Minister in the second term of President Uribe and also affiliated to his political party, became the presidential candidate in early 2010. In the runoff election held on June 20, 2010, the electoral process examined in this paper, Santos achieved a landslide victory, with 69 percent of the votes. 2.2 on Program The Familias en Acci´ Colombia suffered a severe economic downturn during the late 1990s that led to a serious in- crease in poverty and a deterioration of several social indicators. One of the responses of the Government of Colombia to mitigate the effects of the crisis on the poor was to launch a social transfer program in 2001 inspired in a successful scheme applied in Mexico, the widely known Progresa/Oportunidades. The Colombian FA program offers cash transfers -ranging from ap- proximately US $8 to US $16 per school age child and nutrition packages of approximately $28 dollars for children below 7- contingent upon school attendance of the bene�ciary children and regular participation in growth monitoring sessions and nutrition workshops, respectively. In addition to having children in the relevant ages, households are offered the program based on their poverty score index in a proxy means test system known as SISBEN. The SISBEN poverty index score is constructed with information from a registry of the poor. This register, through household level interviews, collects information on households’ demographics, structure, durable goods, housing characteristics, human capital, labor force participation, income, and access to basic services. The poverty index score is calculated with an algorithm that weighs several variables to predict household welfare. The score takes values from 0 to 100 for the poorest and less poor households, respectively. At the time of this study, the distribution of the score is divided into 6 brackets and households assigned to the lowest 7 bracket (SISBEN level 1), corresponding nearly to the population in extreme poverty, are deemed eligible to bene�t from the FA program. The program was initially piloted in 22 municipalities in 2001 but it has been continuously expanded since then. The �rst major expansion took place in the second half of 2002, targeting initially 400 municipalities with fewer than 100,000 inhabitants amongst other conditions.7 In 2005, the program was extended to include displaced families and municipalities which either be- came able to offer the required education and health services or with services accessible in nearby towns. In 2007, the program was further expanded to large urban centers and municipalities not covered before. By 2010 the program reached almost national coverage, bene�ting nearly 2.5 on Social, 2010; Attanasio et al, 2010). In 2012 million households in 1,093 municipalities (Acci´ the program became a law, so that all people in SISBEN level 1 are eligible for the transfer, regardless of the place of residence.8 Figure 1 depicts the cumulative enrollment in the program over time, where the vertical lines indicate the timing of the presidential elections. 3 Data This paper uses three administrative data sets to identify the effects of FA on voting behavior: (1) an electoral census, (2) the FA’s information system of bene�ciaries (SIFA) and, (3) the SISBEN. The �rst one has information on the most recent date of registration and voting place (municipality, polling station and voting booth) for all adults who were registered to vote in the 2010 Presidential election.9 Once registered in a speci�c polling station, citizens are assigned to a voting booth within the polling station at the time of the elections. 7 Municipalities could not be departmental capitals, had to have at least one bank branch in the municipality (to deliver the transfers), and had access to education and health facilities that allowed for the implementation of the program. 8 http://www.eltiempo.com/politica/ARTICULO-WEB-NEW_NOTA_INTERIOR-11929202.html 9 In Colombia citizens become eligible to vote at age 18. Colombia has very high registration rates. According to LAPOP, a Latin American political survey, registration rates between 2004 and 2010 were approximately 90%. Yet, actual turnout rates are much lower, usually in the order of 50% percent 8 The second dataset, SIFA, is used by the national agency that runs FA for administration and monitoring purposes. The data set is a longitudinal census of program bene�ciaries from 2001 to the present. There is information on nearly 2.5 million families who have participated or are currently participating in the program. We limit this dataset to all adults from bene�ciary households with a valid identity card (needed to vote) and match it with the electoral census (using each person’s national identi�cation number). Overall, most people in SIFA (96.5%, corresponding to 3,608,733 program participants) are registered to vote. Finally, we use the SISBEN to identify non-participants that are comparable to FA ben- e�ciaries, namely that they live in households with similar socio-economic and demographic characteristics as determined by their corresponding poverty index score. These data are also matched with the electoral census. Then we append Sisben-electoral census with SIFA-electoral census to produce a sample that is comprised of individuals below and above the cutoff score that establishes eligibility to FA. The matched sample shows that 86% of the FA bene�ciaries are registered in the SISBEN. The rest is mostly comprised of displaced individuals and indigenous people, who by law are not required to have a score below the eligibility threshold to participate in FA, and thus are not included in our analysis. The sample is restricted in two ways. First, considering that the program is offered only to families with children under 18 years, the data are restricted to only include adults 25-46 years old at the moment of the presidential election. This age range is determined by the age distribution of mothers in urban Colombia as reported in vital statistics records in 2010. This is done to ensure that we focus on the group of parents whose voting behavior is more likely to be influenced by FA given the timing in which the program was rolled out throughout the country. The second restriction limits the data to only urban households. The reason for doing this is that by design, the algorithm that determines the poverty score in rural areas was de�ned to produce substantial discontinuities around the threshold of program eligibility for variables 9 that should not be affected by the program (sex and age of head of the household). This clearly invalidates the underlying assumption necessary to perform a regression discontinuity design given that individuals would be different in some observables around the threshold. Summary statistics for the �nal matched dataset (electoral census + SIFA + SISBEN) are reported in Table 1. Panel A reports statistics for the sample at the individual level, which is used for the analysis on registration to vote. Approximately 46% of the individuals in the sample are eligible to bene�t from FA but around 42% of the eligible (19.6%) actually participate in the program.10 93.2% of individuals are registered to vote in the elections, which is similar to the rate calculated using a political survey done in Colombia, LAPOP, 90%. 28.3% people in our sample registered to vote after the onset of FA on the municipality. The average person is 38 years old and has seven years of education at the moment they were interviewed for the SISBEN; more than half (58.8%) of individuals in the sample are women. Panel B, reports individual level statistics but restricting the data to females due to the fact that mothers are the direct recipients of the transfer. In total we have over 2 million women, with 25.5% eligible for FA and 95.5% registered to vote. Overall, the characteristics of this group are similar to those reported in Panel A for the complete sample. Panel C of Table 1 in turn presents descriptive statistics for the booth-level sample used to estimate the effects of FA on voter turnout and choice. People registered to vote in urban centers were assigned to 103,367 voting booths, each of them with an average of 420 individuals. Turnout rates at the booth level for both Presidential elections (�rst round and runoff) were close to 60 percent.11 Finally, the average fraction of votes for the incumbent party candidate was 38% and 51% in the �rst round and runoff respectively, with a margin of victory at the booth level of 15 (�rst round) and 28 (runoff) percentage points. 10 FA program administrators estimate that the overall take up rate in 2010 was around 65 percent. However, take up rates have been found to be much lower in urban areas. 11 These turnout rates are higher than the national level ones (46% for 2010) because they are calculated for urban areas and at the booth level. 10 4 Empirical Strategies 4.1 Regression Discontinuity We �rst rely on a Regression Discontinuity (RD) framework to identify a causal effect of FA on decisions regarding registering to vote. We exploit the discontinuity in FA eligibility for people around the threshold, since people just below the eligibility threshold, the �rst bracket of the SISBEN score, were able to enroll in FA, while people just above were not. Moreover, the cutoff for the �rst bracket of the SISBEN score does not determine eligibility to other major social programs except for a child feeding program given to children between 6 months and 5 years old, which is an in-kind daily transfer that does not coincide speci�cally with the payment of FA, therefore this is not expected to invalidate our empirical strategy. In section 5.3 we discuss the robustness of the results when potential bene�ciaries of the child feeding program are excluded from the sample of analysis. We need to validate several assumptions for the RD strategy to work in this context. Specif- ically, if the eligibility indicator affects registration to vote only through enrollment in FA, then two necessary conditions should hold. First, there should be no bunching in the distribution of the SISBEN score around the eligibility threshold. Figure 2 shows the histograms for all the individuals in the SISBEN data base (top graph), and in our matched sample for men and women (bottom left) and for women only (bottom right). By design of the proxy-means test algorithm, the distribution is bimodal and the overall distribution is very similar to the one in our matched sample. Furthermore, the density is higher above the eligibility threshold12 than below it, which is contrary to the idea of manipulation of the score in order to become eligible for FA. Neither distribution shows bunching at the eligibility threshold. Second, as Figure 3 shows, other variables which are not affected by the program are continuous at the eligibility 12 Normalized to a value of zero and depicted with a vertical line. 11 threshold.13 We check for household size and number of children, given that these variables might be endogenous to the program eligibility criteria. We also checked that other variables such as years of schooling, cohabitation, age and unemployed that could potentially affect po- litical preferences and voting decisions do not exhibit discontinuities at the cutoff either. These two conditions of no bunching at the threshold and continuity in observables at the threshold, lend support to the use of regression discontinuity as the �rst identi�cation strategy. Program effects are estimated using a “fuzzy� RD design given that the probability of being in FA does not change from 1 to 0 at the eligibility threshold. Speci�cally the �rst stage regression is given by: F Aij = δ0 + δ1 eligij + f1 (scoreij |scoreij ≤ score) (1) + f2 (scorei |scorei > score) + γj + χ Xij + ij Where F Aij is an indicator variable equal to one if the individual i of municipality j is enrolled in FA. eligij is an indicator equal to one if the person is eligible for FA. score is the threshold score used to determined eligibility for FA, and it is normalized to a value of zero at the eligibility threshold. f1 (·) and f2 (·) are functions of the SISBEN score that are allowed to be different at each side of the threshold. Xij correspond to individual covariates such as age and years of schooling.14 γj controls for municipality �xed effects and ij is an error term. Then, using eligibility as an instrument for FA enrollment status, to determine the effect of FA on the probability of registering to vote, we estimate the second stage regression: Registeredij = π0 + π1 F Aij + f1 (scoreij |scoreij ≤ score) (2) + f2 (scoreij |scoreij > score) + γj + ν Xij + ηij 13 Even though the functions that characterize these variables are not always monotonic on the poverty score –due to the way the score algorithm was constructed, importantly there are no abrupt changes occurring at the discontinuity threshold. 14 The results are robust and very similar when omitting these covariates. 12 where Registeredij corresponds to an indicator variable of whether an individual i in munici- pality j registered to vote. The coefficient of interest is π1 . Here we assume that the eligibility indicator affects registration to vote only through enrollment in FA. This is a plausible assump- tion given the smoothness of the distribution of the SISBEN score and other covariates at the program eligibility threshold (Figure 2 and Figure 3). Since eligibility is assigned at the house- hold level, we cluster the RD results at this unit. We identify tables using this RD methodology with the word “RD� in the title. 4.2 Fixed Effects and 2SLS Regressions Individual level voting data is rarely (legally) recorded. To overcome the absence of information on the actual vote for each person, we use voting outcomes at the voting booth level, which is the most disaggregated level available. We start by using OLS polling station �xed effects regressions of the variable of interest (turnout, incumbent’s party vote share, margin of victory) on the proportion of FA participants (men and women or women) in the voting booth (b) of the form: Ypbr = β0 + β1 P rop F Apb + β2 SISBENpb + β3 controlspbr + γp + ηr + pbr (3) Where p designates a polling station, b a voting booth, and r an electoral round (�rst round or runoff). Y is the outcome of interest which is identi�ed in each table. P rop F A is the propor- tion of registered voters at the booth who are FA bene�ciaries (men and women bene�ciaries, or women bene�ciaries). SISBEN corresponds to the booth level average of the SISBEN score to proxy for poverty. controls corresponds to a vector of other booth (b) level characteristics speci�ed in each table and usually includes: female proportion, age, household size, children, cohabitation, education level, employment status, homemaker, student status, renter, pensioner, and length of time between the SISBEN interview and the elections (distance), and all of the 13 previous variables interacted with the distance variable. We use polling station (p) and elec- toral round (r) �xed effects. We also estimate similar regressions for women by substituting P rop F Apb in regression 3 for a variable that capture the proportion of female voters enrolled in FA in each booth P rop F A f emalepb . In these regressions we omit the control of female proportion due to collinearity, but keep all other controls. We estimate 2SLS regressions to further control for potential endogeneity in FA enrollment. We instrument the proportion of FA enrolled in the booth (men and women bene�ciaries, or women bene�ciaries) with the proportion of FA eligible at each booth as determined by the SISBEN score of each person assigned to cast a ballot in the corresponding voting booth. We keep the same controls as in equation 3, including polling station �xed effects, and report 2SLS results. In summary, by using the identi�cation strategies outline in this section, we are able to obtain causal estimates of the effect of FA on voting behavior. In particular, the RD strategy allow us to estimate effects for people around the threshold, and the polling station �xed effect strategy with the 2SLS allow us to avoid selection concerns due to the fact that individuals are assigned to a voting booth within a station. 5 Findings 5.1 Intent to Vote and Turnout We initially explore the overall effect of enrollment in the FA program on individuals’ intent to vote in the 2010 presidential election using voter registration as a proxy for intent to vote.15 This part of the analysis relies on a Regression Discontinuity (RD) framework exploiting the discontinuity in FA eligibility to the program for people around the threshold. Figure 4 presents a graphical representation of the �rst stage regression and shows a sharp 15 Our sample restrictions of keeping parents aged 25-46 addresses potential concerns due to automatic regis- edula (ID card). tration when people �rst obtain the c´ 14 discontinuity in FA enrollment at the eligibility threshold both for the men and women data set (left) and the women sample data set (right). Men and women to the left of the threshold are between 29 and 33 percentage points (panel A “Eligibility� of Table 2) more likely to be bene�ciaries, while the estimated discontinuity using the sample of women is between 35.5 and 41 percentage points as reported in panel A “Eligibility� of Table 3. An initial graphical examination of the data - shown in the left top panel of Figure 5- suggests that individuals covered by FA were more likely to register to vote relative to those who were similar but barely ineligible for FA because their SISBEN score was just above the eligibility threshold. The corresponding 2SLS results are shown in the second set of results in panel A “FA� of Table 2. In columns (1)-(3) we report parametric regressions with a different functional forms of the running variable which are allowed to be different on either side of the threshold. Each �rst column uses the whole sample and the second and third column uses a narrower bandwidth around the eligibility threshold. In column (4) we report results using an optimal bandwidth for an RD setting following Imbens and Kalyanaraman (2009) (I & K). Overall, there is strong evidence that FA encourage people to participate in the elections as voters. Table 2 shows that the probability of registering to vote is signi�cantly higher (1.6-2.5 percentage points) among FA bene�ciaries relative to comparable non-participants whose poverty scores are just above the threshold. To further explore whether the availability of FA is inducing people to register to vote, we also exploit the variation in the roll-out of the program over time and across municipalities. We condition the sample to all individuals who signaled a desire to vote (i.e. registered to vote) and rede�ne the dependent variable in equation 2 to take a value of 1 if the person registered after the onset of FA in their municipality and 0 if the person registered before the onset of FA. The results on this subsample provide some indication that people covered by FA are more likely to register to vote than non-participants after the Cash Transfer Program is introduced in their 15 municipalities (Table 2, Panel B and bottom panels in Figure 5). However, these �ndings do not hold when the models are estimated in a narrower sample around the threshold (Columns (3) and (4 )). Another possibility is that FA could have differential effects among members of the family, especially on those that are the direct recipients of the transfer and more exposed to different sources of information due to the interventions of the program. More speci�cally, in the case of CCTs, mothers of the bene�ciary child are usually designated to receive the transfer and are also required to participate in a number of activities at the community level (e.g. growth monitoring sessions and workshops on good nutrition, health and hygiene practices). This pos- sibly encourages women more than men to interact and exchange information with community leaders and other program participants. We split the sample between women and men to test whether program effects vary with the gender of the parents. Tables 3 and A1 (in the appendix) show the same speci�cations of Table 2 for these two groups. They indicate that the results for the whole sample are mainly driven by a higher probability to register to vote among women (Table 3). Women are between 1.5 and 3 percentage points more likely to register to vote (panel A “FA�) and 2.6 to 3.5 percentage points more likely to register to vote after the onset of FA in their municipality (panel B “FA�).16 In contrast, there is no evidence of systematic differences in the intent to vote between FA participant and non-participant men (Appendix Table A1). Next we look at whether the relatively higher intention to vote attributed to FA translates into actual votes. Considering that voting is not compulsory, this point is of major importance in light of the large historical differences in registration and turnout rates in Colombia. For instance, 44.3% of the population registered in the electoral census casted a ballot in the 2010 Presidential elections17 . Given the absence of voting records at the individual level, we exploit 16 The average registration rate for women in the sample of analysis is 95.5 percent. 17 Source: IDEA, website:http://www.idea.int/vt/country_view.cfm?id=48#pres 16 variation in eligibility to FA across more than 100,000 voting booths in both OLS polling station �xed effect regressions and in an instrumental variable setting to look at the effect of FA on voter turnout in the 2010 presidential election. OLS and 2SLS results indicate that, in addition to encouraging people to register to vote, the receipt of FA transfers has a positive effect in the probability of casting a vote (Table 4). An increase of one standard deviation in the proportion of FA bene�ciaries at the voting booth (around 10%), results in an increase in turnout of .07 and .09 of a standard deviation respectively. This corresponds to nearly a 2 percentage point increase in the turnout rate or 600,000 additional votes in the 2010 elections.18 We also disaggregate the effects of FA on turnout by gender (Table 4, Panel B). Consistent with the �ndings on the probability of being registered to vote, we �nd that the increase in turnout rates among FA bene�ciaries is mostly explained by women voting more. 5.2 Political Support The FA program was originally conceived, designed and piloted in 2001 under the administration of the Conservative party. One year later, the independent candidate Alvaro Uribe was elected president for the period between 2002 and 2006. He then ran successfully for a second term and was president until 2010. FA became the government’s flagship anti-poverty program during his two terms, which led to a notable expansion until achieving almost national coverage by 2009. The results discussed above indicate that FA bene�ciaries signaled a greater intent to vote but actually were more likely to vote in the Presidential elections. Seeking to shed light on the possible influence of targeted transfers on voter’s choices, we examine whether FA fostered political support for the incumbent party candidate that implemented and expanded the program between 2002 and 2009. Two variables are used to measure political support in each of the electoral rounds: the per- 18 The total number of registered people to vote in the 2010 elections was 29,853,299, and of those, 13,296,924 voted in the 2010 runoff. 17 centage of votes that went to candidate of the incumbent party; and the margin of victory de�ned V otes Incumbent−Runner U p as: Incumbent+Runner U p , where a value close to 1 favors the incumbent party, whereas a value close to -1 favors the runner-up. By using polling station (p) �xed effects (each polling station has several voting booths), we exploit the variation in the proportion of FA bene�ciaries among adults registered to vote across voting booths within the same polling station, where polling station proxies for neighborhood. We also pool the two electoral rounds and include election round �xed effects. Table 5 presents OLS and 2SLS results that use variation on voting outcomes at the voting booth level to identify the effects of FA enrollment. Overall, the �ndings are indicative of stronger preferences among FA bene�ciaries for the incumbent party. A one standard deviation increase on the average FA participation rate at each voting booth (approximately a 10% increase) raises the share of votes of the elected candidate by 1.5 percentage points. Results based on the margin of victory yield analogous results, namely that the gap in the votes between the incumbent party candidate Santos, and the runner up broadened in favor of the former (by 1.7 percentage points) with an increase of a standard deviation (around a 10% increase) in the proportion of FA bene�ciaries. The two set of results are robust to empirical models that use each electoral round independently (Appendix Table A2). Along the same lines of the heterogeneity observed in program effects on intent-to-vote and voter turnout by gender, the stronger support for the incumbent party attributed to FA is driven by the preferences of women to vote relatively more for Santos (Table 5, panel B). 5.3 Robustness The validity of RD designs hinges on the underlying assumption that the running variable (the SISBEN score) generates exogenous variation in program participation. We have provided evidence that this assumption holds in the context of FA in urban areas. First, as a check for 18 manipulation and nonrandom assignment, an inspection of the SISBEN distribution reveals that there is no sorting of people around the threshold that determines eligibility to the program. Second, continuity checks for an array of socio-economic characteristics (age, household size, school attainment, etc.) that may be linked to voting behavior do not exhibit statistically signi�cant jumps at the threshold of eligibility, ruling out the possibility that other factors rather than program enrollment explain the effects on political participation and preferences. Third, our results are generally not sensitive to multiple speci�cations of the SISBEN function, including several polynomial functions, non-parametric methods and sample trimming procedures, thus minimizing possible misspeci�cation errors - a critical issue in RD design. We perform two additional tests to assess the robustness of the �ndings and the causal interpretation behind them. First, we set the eligibility cutoff at arbitrary values below and above the actual threshold and re-ran the reduced form of the main empirical models employed to estimate the effects of eligibility on intent-to-vote. If by chance the analysis was picking a structural break in the relationship between the outcomes of interest and the SISBEN score, we might observe a discontinuous variation also at different values of the eligibility threshold. Table 6 shows that none of the coefficients for the reduced form results using the arbitrary eligibility cutoffs are statistically signi�cant across all the different model speci�cations. Considering that the cutoff for the �rst bracket of SISBEN determines eligibility to both FA and a small child feeding program for poor families with children between 6 months and 5 years old, another concern is that the actual effect on voting behavior may be a combination of the influence of both programs. To address this issue, we re-estimate the models on registration to vote excluding individuals that belong to households potentially eligible to the child feeding program based on the age structure of their children. Overall, the positive and statistically signi�cant effects of FA on registration also hold in this subsample (Table 7). The fact that program effects on the intent-to-vote remain for only FA participants suggests that the effects 19 of the child feeding scheme, if any, are unlikely to be large. 6 Conclusion This paper provides empirical evidence to support the notion that political participation and political views are responsive to targeted transfers. The issue is increasingly attracting the attention of policy makers and researchers particularly in a context where identifying policy instruments to balance the trade-offs of the direct and indirect effects of CCTs - including mobilizing people to participate in political processes as voters and electoral accountability- has become important. Using a unique Census dataset with individual level voter registration data, we show that ben- e�ciaries of FA are around 2.5 percentage points more likely than comparable non-bene�ciaries to vote. Increasing voter turnout is often considered a desirable outcome since it increases rep- resentation, a basic element that underpins a democracy. This argument is even stronger for countries like Colombia where voting is not compulsory and turnout rates are around 45 percent, lower than those observed in comparable democratic systems.19 The increase in intent-to-vote and actual turnout is explained by larger political participation among bene�ciary women. In many democracies women are less likely to participate in the political sphere, either as voters or as candidates running for public office. Suffrage was extended to women in many countries during the last century, yet women have traditionally voted less than men. Explanations to understand the gender gap in voter turnout range from cultural norms to limited decision power within the household, personal security concerns, mobility constraints, and access to fewer and poorer sources of information to understand the bene�ts of voting. Although CCTs are not conceived to directly overcome these obstacles, analyses have found that they improve the status of women in the form of exerting larger control over family 19 Authors’ calculation based on information from http://www.idea.int/vt/countryview.cfm?CountryCode= CO. 20 on resources, and increased exposure to public life and information (Attanasio et al. 2009, Acci´ Social, 2010). During the 2010 presidential election voters covered by FA not only voted more often, but also expressed a stronger preference (around 2 percentage points) for the official party that implemented and expanded the program. A possible explanation, consistent with principles of government accountability, is that bene�ciaries value the program given the bene�ts documented in the literature, and respond to these positive policy outcomes at the polls. Alternative models include reciprocity where voters support politicians who helped them in the past, and models of rational but poorly informed voters who extract signals of government preferences for social expenditures from their participation in the program (Manacorda et al 2011). Women in par- ticular have been found to vote more systematically in support of social policies (Gleason 2001; Aidt and Dallal 2008; Funk and Gathmann 2006). Another possible explanation is that FA was strategically targeted and motivated by clientelism and vote buying. The identi�cation strategy and data available do not allow us to distinguish between any of these competing interpretations, and these are interesting questions to be addressed in future research. Given the local nature of the estimator presented in this paper, it is not possible to fully account for all of the political support for the official party brought about by the FA program. Yet, contrary to anecdotal evidence presented by the Colombian media and considering the wide margin of victory exhibited by the elected President, it is highly unlikely that the overall change in political preferences attributed to FA explains the �nal outcome of the 2010 presi- dential election. The results, however, show that voters respond to targeted transfers and that these transfers can foster support for incumbents, thus making the case for designing political and legislative mechanisms that avoid successful anti-poverty schemes from being captured by political patronage. 21 References on Social, “El Camino Recorrido Diez A˜ Acci´ on,� Technical Report, Acci´ nos Familias en Acci´ on on 2010. Social, Departamento Nacional de Planeaci´ Aidt, Toke and Bianca Dallal, “Female Voting Power: The Contribution of Womens Suffrage to the Growth of Social Spending in Western Europe (18691960),� Public Choice, March 2008, 134 (3), 391–417. Attanasio, Orazio, Emla Fitzsimons, Ana Gomez, Maria I. Gutierrez, Costas Meghir, and Alice Mesnard, “Childrens schooling and work in the presence of a condi- tional cash transfer program in rural Colombia,� Economic Development and Cultural Change, 2010, 58 (2), 181–210. , Erich Battistin, and Alice Mesnard, “Food and Cash Transfers: Evidence from Colom- bia,� The Economic Journal, March 2009, 122 (559), 92–124. Banerjee, Abhijit V, Selvan Kumar, Rohini Pande, and Felix Su, “Do Informed Voters Make Better Choices? Experimental Evidence from Urban India,� Mimeo, Yale Univeristy 2011. Camacho, Adriana and Emily Conover, “Manipulation of Social Program Eligibility,� American Economic Journal: Economic Policy, 2011, 3 (2), 41–65. De la O, Ana L., “Do Conditional Cash Transfers Affect Electoral Behavior? Evidence from a Randomized Experiment in Mexico,� American Journal of Political Science, forthcoming. Drazen, Allan and Marcela Eslava, “Electoral Manipulation Via Voter-Friendly Spending: Theory and Evidence,� Journal of Development Economics, May 2010, 92 (1), 39–52. and , “Programmatic Targeting,� Working Papers, Universidad de Los Andes February 2012. Finan, Frederico and Laura Schechter, “Vote-Buying and Reciprocity,� Econometrica, March 2012, 80 (2), 863–881. Fiszbein, Ariel, Norbert R. Schady, Francisco Ferreira, Margaret Grosh, Nial Kelle- her, Pedro Olinto, and Emmanuel Skou�as, Conditional cash transfers: reducing present and future poverty, World Bank Publications, 2009. Funk, Patricia and Christina Gathmann, “What Women Want: Suffrage, Gender Gaps in Voter Preferences and Government Expenditures,� Working Paper, Universitat Pompeu Fabra July 2006. Gine, Xavier and Ghazala Mansuri, “Together We Will: Experimental Evidence on Female Voting Behavior in Pakistan,� Policy Research Working Paper Series 5692, World Bank 2011. 22 Gleason, Suzanne, “Female Political Participation and Health in India,� Annals of the Amer- ican Academy of Political and Social Science, 2001, 573 (1), 105–126. Imbens, Guido and Karthik Kalyanaraman, “Optimal Bandwidth Choice for the Regres- sion Discontinuity Estimator,� NBER Working Papers 14726, National Bureau of Economic Research February 2009. Manacorda, Marco, Edward Miguel, and Andrea Vigorito, “Government Transfers and Political Support,� American Economic Journal: Applied Economics, July 2011, 3 (3), 1–28. Nupia, Oskar, “Anti-Poverty Programs and Presidential Election Outcomes: Familias en on in Colombia,� Unpublished Manuscript, Universidad de Los Andes-CEDE June 2012. Acci´ Pop-Eleches, Cristian and Grigore Pop-Eleches, “Targeted Government Spending and Political Preferences,� Quarterly Journal of Political Science, June 2012, 7 (3), 285–320. Robinson, James A and Thierry Verdier, “The Political Economy of Clientelism,� CEPR Discussion Papers 3205, C.E.P.R. Discussion Papers February 2002. Rogoff, Kenneth, “Equilibrium Political Budget Cycles,� American Economic Review, March 1990, 80 (1), 21–36. World Bank-IEG, “Evidence and Lessons Learned from Impact Evaluations in Social Safety Nets.,� Technical Report 2011. Zucco, Cesar, “Cash-transfers and voting behavior: An empirical assessment of the political impacts of the Bolsa Familia program,� Social Development, 2009. 23 Figure 1: Cumulative Program Enrollment Figure 2: Histogram of SISBEN score All SISBEN Matched Sample: Men and Women Matched Sample: Women 24 Figure 3: Covariates at the Eligibility Threshold –Women Years of Schooling Cohabitating Household Size Number of Children Age Unemployed 25 Figure 4: Probability of Being an FA Bene�ciary Men and Women Women Figure 5: Probability and Timing of Registering to Vote Being registered to vote Men and Women Women Registered after the onset of FA Men and Women Women 26 Table 1: Summary Statistics Variable Obs. Mean Std. Dev. Min Max Panel A: Individual Level All Sample FA(%) 3,653,718 0.196 0.397 0 1 Eligible (%) 3,653,718 0.457 0.498 0 1 Register to vote (%) 3,653,718 0.932 0.251 0 1 Registered after the onset of FAa (%) 3,406,786 0.283 0.451 0 1 Female (%) 3,653,718 0.588 0.492 0 1 Age in years 3,653,718 38.070 5.599 25 46 Years of schooling 3,653,709 7.068 3.787 0 22 Panel B: Individual Level Female FA (%) 2,150,133 0.255 0.436 0 1 Eligible (%) 2,150,133 0.457 0.498 0 1 Register to vote (%) 2,150,133 0.955 0.208 0 1 Registered after the onset of FAa (%) 2,053,136 0.292 0.455 0 1 Age in years 2,150,133 37.839 5.678 25 46 Years of schooling 2,150,128 7.108 3.761 0 22 Panel C: Voting Booth Level Registered people 103,367 420.015 116.712 6 2,239 Number of votes 103,367 248.412 84.781 1 751 FA (%) 103,367 0.081 0.098 0 1 Eligible (%) 103,367 0.206 0.173 0 1 FA female 103,367 26.554 43.307 0 364 Non-FA female 103,367 81.845 96.730 0 613 Turnout (%) First round 51,965 0.607 0.192 0 1 Runoff 51,402 0.609 0.203 0 1 Votes for incumbent (%) First round 51,965 0.377 0.138 0 1 Runoff 51,402 0.507 0.144 0 1 Margin of victoryb (%) First round 51,965 0.152 0.212 -1 1 Runoff 51,402 0.279 0.232 -1 1 ıa Nacional, DNP-DDS. Note: a Conditional on being registered to vote. b Margin of Source: Registradur´ otes Incumbent−Runner U p Victory is de�ned as V V otes Incumbent+Runner U p . 27 Table 2: RD Regression of Registering to Vote: Men and Women (1) (2) (3) (4) Panel A: Registering to vote First Stage Eligibility 0.332*** 0.295*** 0.291*** 0.286*** (0.002) (0.003) (0.004) (0.004) Second Stage FA 0.016*** 0.019*** 0.025*** 0.019** (0.004) (0.006) (0.010) (0.009) R-Squared 0.003 0.003 0.004 Observations 3,653,709 1,630,197 299,494 101,086 Bandwidth 5 1 0.340 Functional form Quintic Cubic Quadratic I&K Panel B: Registering after the onset of FA First Stage Eligibility 0.332*** 0.309*** 0.290*** 0.285*** (0.002) (0.002) (0.003) (0.004) Second Stage FA 0.031*** 0.023*** 0.018 -0.024 (0.006) (0.007) (0.012) (0.016) R-Squared 0.026 0.029 0.022 Observations 3,406,777 1,524,958 279,344 90,553 Bandwidth 5 1 0.305 Functional form Quartic Quadratic Linear I&K Municipality �xed effects Yes Yes Yes Demographic covariates Yes Yes Yes Note: Standard errors in parenthesis. Sample in panel B is conditional on being registered to vote. Estimations in columns from (1) to (3): are clustered at the household level, include controls for age and schooling, and the functional form of the aligned SISBEN score is allowed to be different above and below the threshold. I & K uses an optimal bandwidth algorithm developed by Imbens and Kalyanaraman (2009). *** p < 0.01, ** p < 0.05, * p < 0.1. 28 Table 3: RD Regression of Registering to Vote: Women (1) (2) (3) (4) Panel A: Registering to vote First Stage Eligibility 0.412*** 0.367*** 0.364*** 0.355*** (0.003) (0.003) (0.005) (0.006) Second Stage FA 0.015*** 0.019*** 0.030*** 0.026*** (0.003) (0.005) (0.008) (0.008) R-Squared 0.006 0.005 0.006 Observations 2,150,128 962,683 179,212 54,105 Bandwidth 5 1 0.305 Functional form Quintic Cubic Quadratic I&K Panel B: Registering after the onset of FA First Stage Eligibility 0.412*** 0.381*** 0.364*** 0.356*** (0.003) (0.002) (0.004) (0.005) Second Stage FA 0.035*** 0.032*** 0.026** 0.008 (0.006) (0.007) (0.011) (0.014) R-Squared 0.030 0.031 0.024 Observations 2,053,131 922,466 171,116 75,919 Bandwidth 5 1 0.428 Functional form Quartic Quadratic Linear I&K Municipality �xed effects Yes Yes Yes Demographic covariates Yes Yes Yes Note: Standard errors in parenthesis. Sample in panel B is conditional on being registered to vote. Estimations in columns from (1) to (3): are clustered at the household level, include controls for age and schooling, and the functional form of the aligned SISBEN score is allowed to be different above and below the threshold. I & K uses an optimal bandwidth algorithm developed by Imbens and Kalyanaraman (2009). *** p < 0.01, ** p < 0.05, * p < 0.1. 29 Table 4: Fixed Effect Regression: Vote Turnout Dependent variable: Vote Turnout in Presidential Elections First Round Runoff OLS 2SLS OLS 2SLS (1) (2) (3) (4) Panel A: All Voters Proportion FA 0.070*** 0.085*** 0.089*** 0.090*** (0.010) (0.010) (0.010) (0.011) R2 W ithin 0.41 0.41 0.42 0.42 Observations 52,089 52,089 51,458 51,458 Panel B: Female Voters Prop. FA female 0.061*** 0.070*** 0.084*** 0.094*** (0.007) (0.006) (0.008) (0.007) 2 R W ithin 0.40 0.40 0.42 0.42 Observations 52,089 52,089 51,458 51,458 Note: OLS robust standard errors in parenthesis. 2SLS standard errors in parenthesis. All regressions include polling station �xed effects and booth level controls. Booth level controls include average char- acteristics at the booth for: female proportion (panel A), age, household size, children, cohabitation, education level, employment status, homemaker, student status, renter, pensioner, and length of time between the SISBEN interview and the elections (length), and all of the previous variables interacted with the distance variable. Standardized coefficients reported. *** p < 0.01, ** p < 0.05, * p < 0.1. Table 5: Fixed Effect Regression: Political Support Dependent variable: Votes for Incumbent Margin of Victory Party Candidate (%) (incumb-ru)/total votes OLS 2SLS OLS 2SLS (1) (2) (3) (4) Panel A: All Voters Proportion FA 0.068*** 0.098*** 0.072*** 0.072*** (0.007) (0.007) (0.007) (0.007) 2 R W ithin 0.61 0.61 0.56 0.56 Observations 103,367 103,367 103,367 103,367 Panel B: Female Voters Prop. FA female 0.075*** 0.100*** 0.081*** 0.110*** (0.006) (0.005) (0.006) (0.004) R2 W ithin 0.61 0.61 0.55 0.55 Observations 103,367 103,367 103,367 103,367 Note: OLS robust standard errors in parenthesis. 2SLS standard errors in parenthesis. Proportion FA corresponds to the proportion of people at the booth who are FA bene�ciaries. Proportion FA female corresponds to the proportion of women at the booth who are FA bene�ciaries. Standardized coefficients reported. *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions include polling station, election round �xed effects and booth level controls. Booth level controls are de�ned as in Table 4. 30 Table 6: RD Regression of Registering to Vote at Points Above and Below the Eligibility Threshold (1) (2) (3) (4) (5) (6) (7) (8) Panel A: -/+ 1 point 1 point below the threshold 1 point above the threshold Reduced Form Eligibility -0.001 -0.002 -0.003 -0.002 0.002 -0.001 0.001 0.004 (0.001) (0.002) (0.003) (0.002) (0.002) (0.002) (0.003) (0.003) R-Squared 0.004 0.004 0.006 0.004 0.003 0.004 Observations 2,150,128 887,855 224,795 79,032 2,150,128 1,050,610 138,245 57,059 Bandwidth 5 1 0.348 5 1 0.427 Panel B: -/+ 5 points 5 point below the threshold 5 point above the threshold 31 Reduced Form Eligibility 0.000 0.001 0.001 0.003 0.000 0.001 0.001 0.001 (0.001) (0.002) (0.003) (0.003) (0.001) (0.001) (0.002) (0.002) R-Squared 0.004 0.006 0.005 0.004 0.002 0.001 Observations 2,150,128 753,488 139,546 60,738 2,150,128 977,578 302,325 101,665 Bandwidth 5 1 0.440 5 1 0.323 Functional form Quintic Cubic Quadratic I&K Quintic Cubic Quadratic I&K Municipality �xed effects Yes Yes Yes Yes Yes Yes Demographic covariates Yes Yes Yes Yes Yes Yes Note: Standard errors in parenthesis. Estimations in columns (1) to (3) and (5) to (7) are clustered at the household level, include controls for age and schooling, and functional form of the aligned SISBEN score is allowed to be different above and below the threshold. I & K uses an optimal bandwidth algorithm developed by Imbens and Kalyanaraman (2009). *** p < 0.01, ** p < 0.05, * p < 0.1. Table 7: RD Regression of Registering to Vote: Women Excluding Potential Bene�ciaries to the Child Feeding Program (1) (2) (3) (4) Panel A: Registering to vote First Stage Eligibility 0.388*** 0.353*** 0.355*** 0.349*** (0.003) (0.004) (0.006) (0.006) Second Stage FA 0.014*** 0.020*** 0.038*** 0.037*** (0.004) (0.006) (0.010) (0.009) R-Squared 0.006 0.005 0.006 Observations 1,698,882 755,043 139,245 42,098 Bandwidth 5 1 0.307 Functional form Quintic Cubic Quadratic I&K Panel B: Registering after the onset of FA First Stage Eligibility 0.388*** 0.362*** 0.348*** 0.347*** (0.003) (0.003) (0.004) (0.005) Second Stage FA 0.037*** 0.037*** 0.032** 0.023 (0.008) (0.009) (0.013) (0.014) R-squared 0.022 0.024 0.017 Observations 1,617,859 721,687 132,634 75,924 Bandwidth 5 1 0.551 Functional form Quartic Quadratic Linear I&K Municipality �xed effects Yes Yes Yes Demographic covariates Yes Yes Yes Note: Standard errors in parenthesis. Sample in panel B is conditional on being registered to vote. Estimations in columns from (1) to (3): are clustered at the household level, include controls for age and schooling, and the functional form of the aligned SISBEN score is allowed to be different above and below the threshold. I & K uses an optimal bandwidth algorithm developed by Imbens and Kalyanaraman (2009). *** p < 0.01, ** p < 0.05, * p < 0.1. 32 Appendix Table A1: RD Regression of Registering to Vote: Men (1) (2) (3) (4) Panel A: Registering to vote First Stage Eligibility 0.200*** 0.187*** 0.186*** 0.187*** (0.003) (0.003) (0.005) (0.004) Second Stage FA 0.026** 0.013 0.013 0.005 (0.013) (0.016) (0.028) (0.020) R-Squared 0.002 0.002 0.001 Observations 1,503,581 667,514 120,282 61,599 Bandwidth 5 1 0.514 Functional form Quintic Cubic Quadratic I&K Panel B: Registering after the onset of FA First Stage Eligibility 0.200*** 0.187*** 0.186*** 0.187*** (0.003) (0.003) (0.005) (0.004) Second Stage FA 0.026** 0.013 0.013 0.005 (0.013) (0.016) (0.028) (0.020) R-Squared 0.002 0.002 0.001 Observations 1,503,581 667,514 120,282 61,599 Bandwidth 5 1 0.390 Functional form Quartic Quadratic Linear I&K Municipality �xed effects Yes Yes Yes Demographic covariates Yes Yes Yes Note: Standard errors in parenthesis. Sample in panel B is conditional on being registered to vote. Estimations in columns from (1) to (3): are clustered at the household level, include controls for age and schooling, and the functional form of the aligned SISBEN score is allowed to be different above and below the threshold. I & K uses an optimal bandwidth algorithm developed by Imbens and Kalyanaraman (2009). *** p < 0.01, ** p < 0.05, * p < 0.1. 33 Table A2: Fixed Effect Regression of Political Support by Electoral Round Dependent variable: Votes for Incumbent Margin of Victory Party Candidate (%) (incumb-ru)/total votes OLS 2SLS OLS 2SLS (1) (2) (3) (4) Panel B: First Electoral Round Proportion FA 0.065*** 0.090*** 0.069*** 0.048*** (0.007) (0.008) (0.007) (0.008) 2 R W ithin 0.39 0.39 0.45 0.45 Observations 51,965 51,965 51,965 51,965 Panel C: Second Electoral Round Proportion FA 0.069*** 0.106*** 0.073*** 0.096*** (0.009) (0.011) (0.008) (0.009) 2 R W ithin 0.48 0.48 0.55 0.55 Observations 51,402 51,402 51,402 51,402 Note: OLS robust standard errors in parenthesis. 2SLS standard errors in parenthesis. Booth level controls are de�ned as in Table 4. Standardize coefficients reported. *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions include polling station �xed effects and booth level controls. Booth level controls are de�ned as in Table 4. 34