Do conditional cash transfer programs shift votes

Electoral Studies 32 (2013) 864–874

Contents lists available at SciVerse ScienceDirect

Electoral Studies journal homepage: www.elsevier.com/locate/electstud

Do conditional cash transfer programs shift votes? Evidence from the Honduran PRAF Elizabeth Linos* Harvard University, Harvard Kennedy School, 79 JFK Str., Cambridge, MA 02138, USA

a r t i c l e i n f o

a b s t r a c t

Article history: Received 5 June 2012 Received in revised form 25 March 2013 Accepted 28 March 2013

How do national social programs influence local voting? This study utilizes the experimental set up of a conditional cash transfer program to show that small, targeted cash transfers can have large electoral effects. The Honduran PRAF program allocated an average of $18 per capita per year to poor households within municipalities that were randomly assigned to receive the program. Although the program was administered at the national level, the program increased an incumbent mayor’s re-election probabilities by 39%, without significantly influencing voting behavior in presidential elections. Moreover, the evidence suggests that transferring cash to poor households were more effective at increasing political support than interventions providing public goods for poor villages. Ó 2013 Elsevier Ltd. All rights reserved.

Keywords: Conditional cash transfers Local government Voting Latin American politics

1. Introduction In October 2006, Brazilian president Lula da Silva was re-elected as chief of state despite a series of corruption scandals and crises during his first term in office. News reporters attributed the election victory to Lula’s famous social welfare program called Bolsa Familia (see Kingstone, 2006; Reel, 2006). This program is one of an increasing number of Conditional Cash Transfer (CCT) programs in Latin America, which provide cash assistance to households under specified conditions, such as sending children to primary school and visiting health clinics regularly. Hailed as the “magic bullet in development” (Dugger, 2004), CCT programs have rapidly expanded with the financial support of the World Bank and the Inter-American Development Bank (IADB, 2006), because of their significant effects on health and education. The effects of CCT programs on voting, however, are not as evident. In many of the countries now introducing CCT programs, voting is traditionally linked to patron-client networks that would make each individual vote a product

* Tel.: þ1 617 895 8195. E-mail address: [email protected]. 0261-3794/$ – see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.electstud.2013.03.007

of “hereditary” party affiliation and clientelistic social transfers, suggesting no effect of a CCT program on elections. General pocketbook voting theories, on the other hand, present mechanisms through which a voter would reward political leaders for receiving a cash transfer, even if that cash is a product of an objectively-targeted social program. Moreover, it is unclear if a national program, implemented at a local level, would affect local or national electoral politics. To my knowledge, there is no available causal evidence on the effects of CCTs on local elections. Some initial unpublished studies on the effect of CCT programs on national elections come to mixed conclusions. The most rigorous by Manacorda et al. (2009) models the effect of the Uruguayan PANES program on reported support for the incumbent government. Using those voters who were on either side of the cut-off to receive the cash transfer, they conduct a quasi-experimental study that finds increased reported support for the incumbent government, for those that received the cash transfer. Whether this increase in government support reported by beneficiaries would translate into changes at the ballot box is yet to be seen. Two unpublished analyses of actual voting behavior in the 2000 Mexican elections, after the initiation of PROGRESA/Oportunidades display contradictory results. De la O (2006) reports an increase in voter turnout and an

E. Linos / Electoral Studies 32 (2013) 864–874

increase in votes going to the PRI (Partido Revolucionario Institutional), the party that had instituted the program. In contrast, Green et al. (2002) finds no effect on turnout and no change in vote shares in PROGRESA areas. These differences result from features of the Mexican CCT program that complicate the study of its effect on voting. First, the borders of localities receiving CCT funds did not coincide with electoral precincts. Second, the PRI, which instituted the program, already had overwhelming support in the poor areas that received the transfers. Other studies linking voting behavior to social programs display a potential positive correlation between the two, but cannot display a causal link. For example, an unpublished analysis of Brazil’s Bolsa Escola shows a positive correlation between incumbent mayor re-election and effective program delivery (De Janvry et al., 2008). This study contributes to the debate on whether social programs shift votes in national elections, while also opening the discussion on the causal effect of CCT programs on local elections. I study the case of the Honduran PRAF1 program, presenting a more precise econometric strategy than prior work on electoral results. Specifically, municipalities receiving funds in Honduras were randomly selected and were also electoral districts, offering a rare empirical opportunity to test whether an exogenous increase in some households’ income influences voting behavior. In addition, the availability of national and municipal electoral results allows for a more comprehensive understanding of the effect of such programs on local and national electoral politics. Given the experimental set up of the social program, the Honduran case permits us to study the causal effects of targeted cash transfers on voting for the Partido Liberal (PLH) which implemented the program initially, and then the main opposition, the Partido Nacional (PNH) which continued it. The Honduran PRAF program was implemented from 2000 to 2005 and transferred cash to households and communities in an effort to increase investment in human capital. Unlike previous programs in Honduras and much of Latin America, the government targeted municipalities based on poverty indices, rather than political affiliation. The 70 poorest municipalities in Honduras were targeted as potential beneficiaries. A public lottery then determined which municipalities would receive transfers and which municipalities would serve as the control group. Election data are available for the years before the implementation of the program, as well as for the years after its implementation in both municipal and presidential elections. Election data are supplemented by survey data that follow the same households in two time periods, which provide information regarding the socio-economic characteristics of the communities under study. The basic econometric specification of this paper uses a difference-in-difference model to consider how votes received by major parties and incumbents, aggregated at

1 The full name of the program is PRAF/BID II: Proyecto 1026. The acronym PRAF stands for Programa de Asignación Familiar, or Family Allowance Program. For the purposes of this study, the project will consistently be referred to as the PRAF program.

865

the municipal level, changed differentially for treated municipalities after the initiation of the program. Qualitative research supplements these data. Based on two months of fieldwork and interviews in Honduras, this research also explores issues of program implementation. The analysis below shows that relatively small transfers can have large effects, as a program allocating approximately an average of 3.6% of annual household expenditure to qualifying individuals in randomly selected poor villages increased incumbent mayor’s re-election probabilities by 39%. A shift of this magnitude is particularly surprising in the context of prior research claiming that electoral volatility in Honduras is more stable than the Latin American average, and that partisanship is “frozen” and unrelated to short-term policy changes (Roberts and Wibbels, 1999). The program did not influence voting patterns for presidential elections. Moreover, this research suggests that transferring cash to poor households was more effective at increasing political support than interventions providing public goods for poor villages. In the following section, I present the Honduran electoral context and outline the PRAF program and implementation. I then present hypotheses that are tested econometrically, before presenting the results in the context of a wider voting literature. 2. Partisanship and pocketbooks: a framework for understanding voting While this study is based on the availability of a precise econometric model, whether a social program should have any effect on voting is grounded in the relevant literature on voting behavior in general. Below I present two relevant theoretical models whose basic hypotheses can be tested econometrically in this study.

2.1. Partisanship While there are five registered parties in Honduras,2 the overwhelming majority of votes is awarded to two parties, the Liberal Party (PLH), and the National Party (PNH). The cumulative vote share going to these two parties in presidential elections exceeds 95% of total votes. These two parties are also the oldest in Honduran democratic history, tracing their roots to intra-elite divisions of 19th century Honduras,3 and they have remained popular by drawing support across classes and social groups, through patronage networks (Roberts and Wibbels, 1999). McDonald and Ruhl (1989, 17), for example, claim that both the PNH and the PLH “are multi-class parties that benefit from widespread hereditary party attachment”. Indeed, the majority of research on Honduran elections cites clientelism as a major determinant of party preferences. Dix (1989,

2 These parties are: El Partido Liberal (PLH), El Partido Nacional (PNH), La Democracia Cristiana (DC), El Partido de Innovación y Unidad SocialDemócrata (PINU-SD), and El Partido Univicación Democrática (UD). 3 Taylor-Robinson (2006) notes that the parties were first established by competing international banana companies, and only became independent entities in 1929.

866


30) describes such parties as “non-ideological and pragmatic”, which rely on clientelism and “inherited” party loyalty for electoral support. The underlying implication is that Honduran politics do not follow a simple Meltzer– Richard framework, where party affiliation is associated with income or socio-economic status. The Honduran case, however, is not unique. Dix (1989, 30) cites political systems in Colombia, Uruguay, Honduras and Paraguay, in arguing that the “preponderant party form in contemporary Latin America comes considerably closer to reflecting the characteristics of the catch-all party than to replicating the more ideological, class-center mass parties of the West.”. In such cases, it becomes particularly costly for a client to switch their vote when a government performs badly. Specifically, since benefits are associated with supporting the party in power, a client would only benefit from a vote switch if the opposition party wins the elections. Moreover, a vote switcher would have to wait until all previous clients had already been served in patronage networks. Thus, it is more rational for a client of the party in power to express dissatisfaction through not participating in elections, rather than changing party loyalties (Taylor-Robinson, 1996, 333). The existing relevant literature on Honduras, therefore, would predict no change in voting behavior. As explained above, the literature on Honduras overwhelming links voting to patron-client networks, and hereditary voting, that is less elastic to changing wellbeing of citizens. If party affiliation is the leading determinant of the vote, there will be no empirical difference between treated and control communities, in the vote share allotted to each political party. Such literature, however, stems from an analysis of Honduran politics in the 1980s and the early 1990s, and it is unclear if such patron-client networks are as strong in the 21st century. Furthermore, it is possible that patron-client loyalties may have been more important in previous elections when presidential, municipal and legislative positions were all decided using the same ballot. Since a 1997 change in electoral rule, which allows for a split ballot between municipal and presidential voting, however, it is possible that not all voting decisions are made using the same criteria. It is possible, therefore, that some frameworks may be most relevant to presidential voting, while other factors determine municipal voting. 2.2. Pocketbook voting Equally relevant to the study of electoral behavior and social programs are rational choice models. According to such models, voters are rational actors, voting for the party that will provide the most benefits. In a simple two-party model, an individual considers her expected utility with party A, the incumbent party, and expected utility with party B, the opposition.

A B E Utþ1 E Utþ1 If this differential is positive, expected utility with party A is larger than with party B and the individual will vote for the incumbent party. (Downs, 1957, 38). This model, too, would

predict no change in voting patterns, because a rational, informed voter would understand that neither party had any say in who received funds. Since the program was randomized, if a household received a transfer, it was due to luck and not the competence of their mayor or president. As such, if the Honduran PRAF did shift votes, this rejects two crucial models: the mainstream explanation of Honduran voting as fixed and clientelistic, and the rational choice model that views voters as maximizing their utility. Yet, given a precedent of pork barrel politics, it is highly plausible that Hondurans would not trust that the social program was allocated according to a lottery. As a variant of the basic rational choice model, the pocketbook model argues that an evaluation of personal finances can be used as a shortcut in measuring expected utility with the incumbent. A pocketbook voter need not attribute the increase in disposable income to the incumbent government directly. That is, a voter need not be perfectly informed about the source of the change in income. As explained by Fiorina (1979, 5), “in order to ascertain whether the incumbents have performed poorly or well, citizens need only calculate the changes in their own welfare”. While Popkin (1994, 32) does not accept that voters only look to their personal experience when making an electoral decision, he accepts that “voters who see political reasons for changes in their pocketbooks will reward and punish political leaders”. This caveat applies to the Honduran PRAF given that a beneficiary household may be unaware of the randomized nature of the program, but still acknowledge its link to the government. Given a history of clientelistic politics, Honduran voters may seek to maximize their utility, in a context of patron-client networks. If voters are uninformed but utility maximizing, voters in beneficiary households should reward the incumbent for the initiation of the program. If individuals vote based on an evaluation of personal finances, a program such as the Honduran PRAF, which increases household income, should also affect the electoral choice. 2.3. Applying the framework to municipal versus national elections The two aforementioned models both apply to electoral politics in Honduras and bring forth contrasting predictions. On the one hand, if partisanship is “frozen”, we should see no difference between beneficiary communities and non-beneficiary communities at the ballot box. On the other hand, if Hondurans are pocketbook voters, those municipalities that received cash transfers should reward the incumbent government for their increase in disposable income. The available data allow for the study of electoral effects on two different levels of elections. While municipal and presidential elections happen on the same day, it is plausible that different frameworks apply to different levels of government. Fixed partisanship, for example, is more than an affirmation of patron-client networks. Like finances, partisanship can also serve as an information shortcut when other information on candidates is not readily available. If rural voters who live far from the capital have


little information about national candidates, they may vote on party affiliation (Popkin, 1994). In contrast, they may have more direct information, and be more amenable to switch parties, when evaluating mayoral candidates. Since separate municipal elections were only introduced in 1997, it is possible that the “rules” associated with voting in presidential elections need not apply in the newly instituted municipal elections. That is, a voter could maintain fixed party affiliations in presidential elections, but still vote as a pocketbook voter in municipal elections. The pocketbook voter model, then begs the question: who is the incumbent that should be rewarded? Officially, and practically, the Honduran PRAF was administered through the national government. Indeed, the randomization process ensured that local governments played no role in whether a municipality received the program or not. As such, if an electoral reward to the incumbent is in order, through a pocketbook voter model, it should be directed to the national government. Still, there are numerous reasons why we may see an electoral effect at the municipal level. First, voters may have difficulty distinguishing between levels of government and reward all incumbents in the election. Second, it is possible that voters mistakenly attributed the program success to municipal governments. Given that the lottery to assign municipalities was public, it might be difficult for municipal leaders to explicitly claim credit for the program. However, previous large-scale social programs, such as the FHIS (Honduran Social Investment Fund) of the 1990s were administered through municipal governments, where municipal leaders “exercised significant power in deciding how to invest the funds” (Moore, 2008). This phenomenon is not uncommon. Much of the literature on Brazilian decentralization shows that state governors became very influential “political brokers” in the 1980s through social policy programs and distribution of funds (see Samuels and Abrucio, 2000; Ames, 2001). As such, it is not inconceivable that voters would reward municipal level governments if they mistakenly thought municipal leaders played an important political role in bringing the social program to their area. Similarly, beneficiaries may have mistakenly rewarded municipal leaders for the implementation of the program. Arceneaux and Stein (2006) show some evidence of electoral effects on officials, even for events outside their control, such as flooding, if voters attribute some functional responsibility to officials. Finally, it is possible that the wide-ranging authority of national governments, as compared to the limited authority of mayors, made the PRAF program very salient to mayoral but not to national elections. If this is the case, the empirical results should show an increase in vote share going to the incumbent mayor in treated municipalities, but show no such effect in presidential elections. The randomized nature of the field experiment allows for the isolation of the electoral effect of the PRAF program on both levels of voting. That is, any difference in vote share in the treated communities can be causally linked to the PRAF program. The availability of data for both local and national elections allow for a more nuanced understanding of how different electoral behavior may apply to different elections.

867

3. The PRAF program The PRAF program included two types of transfers to rural communities. The first intervention targeted individual households and provided direct cash transfers, on the condition that they send their children to school and take their infants to local health centers at regular intervals. Initially, it was proposed that the educational voucher for primary school-aged children would cover the total costs of going to school, L. 828 (US$ 58.07) and the nutritional voucher for infants and pregnant women would be set at L.660 (US$ 46.30) per year per child (IFPRI, 2000). According to some estimations, the average amount received is $18 per capita per year (Bouillon and Tejerina, 2006), and reflects less than 3.6% of total expenditures of the target population. Mexico’s Oportunidades, in contrast, covers approximately 20% of the target population’s total expenditures (IFPRI, 2003, 7). Thus, if this small sum could prompt behavioral change, one might expect larger electoral effects in programs with larger payments. The second type of transfer targeted the community as a whole. This transfer provided cash to parent associations and health care teams, who were responsible for improving local schools and health clinics. In order to evaluate which type of intervention is most effective, the targeted municipalities were separated into four groups. Group 1 (G1) would only provide householdlevel cash transfers or vouchers. Group 3 (G3) would only receive public good transfers. In order to examine the interaction of public and private interventions, Group 2 (G2) would include both vouchers to mothers and transfers to service providers. As with most impact evaluations, a control group (G4) was also included in the design, so as to examine what would have happened in the absence of the program (Fig. 1). The random allocation of the targeted municipalities to one of the four intervention groups occurred at a public event on October 13, 1999. The 70 municipalities were first divided into groups of 14, according to their rank in the height-for-age ordering. Each group then had a lottery, where their intervention was assigned. The result was a stratified, random assignment of the 70 municipalities (IFPRI, 2000, 33). This random assignment allows for a precise econometric model that isolates the causal effect of the Honduran PRAF on voting. Yet, the effects of the PRAF program on voting behavior should also be affected by the practical implementation of the program.4 Despite frequent complaints regarding the initial selection of beneficiaries, the receipt of household cash transfers was described as relatively prompt and consistent by beneficiaries and administrative staff. Making the same conclusion, a study conducted by the International Food Policy Research Institute (IFPRI, 2002) found that 76–79% of cash transfers had been provided

4 The econometric analysis of the PRAF is complemented by two months of qualitative interviews and fieldwork with participants, administrators, community members, and community leaders, regarding the implementation of the program.

868


Household Cash Transfers?

Public Good Transfers?

Yes

No

Yes

G2

G3

No

G1

G4

Before Program After Program Difference Across Time

Treatment

Control

T0 T1

C0 C1

Difference Across Groups T0 – C0 T1 – C1

T0 - T 1

C0 - C1

(T1 – C1) – (T0 – C0)

Fig. 2. Creating a difference-in-difference estimation of program effect. Fig. 1. Group assignment according to type of transfer.

adequately by 2001. The same does not apply to the receipt of community-level transfers. Beneficiaries and administrators complained that they did not receive the full amount they had been promised, or did not receive any transfers at all. By the end of 2002, the PRAF program had only provided 17% of planned health center transfers, and approximately 7% of school supply transfers (IFPRI, 2002). 4. Quantitative analysis 4.1. The empirical model This study allows for estimations using a differencein-difference (DID) model. This estimate uses two differences to isolate the effect of the treatment on municipalities. As a first step, I measure the difference in voting behavior between the treatment (T) and control (C) groups. Using the second difference, that compares voting behavior before and after (T0 and T1) the initiation of the program, I can effectively isolate the effect of the treatment itself. The following figure shows the DID estimation, (T1 C1) (T0 C0) (Fig. 2). The program could influence votes going to the incumbent party, votes going to the opposition party, or participation rates.5 The main dependent variable is: vote share going to a created variable “incumbent”, which is the PLH for 2001 and the PNH for 2005 and which may be of either party in municipal elections, depending on the municipality and the year. In order to capture the effect of being in a treatment group after the program has been implemented, I first create three separate binary variables to indicate each of the treatment groups: H refers to those municipalities who only received household cash transfers, H-P to those who received both household-level transfers and public goods transfers, and P refers to those municipalities who only received municipality-wide public goods transfers. I then create a time-specific binary variable, AFTER. AFTER takes the value one in the elections years of 2001 and 2005, referring to the years after the implementation of the PRAF program, and zero otherwise. The coefficient on the interaction between each treatment group and AFTER measures the additional effect of being assigned to a treatment group once the program has been initiated, and reflects the

5 In regressions not shown, the effect of the program on participation rates was analyzed. While participation was highly correlated to the election year, there was no significant difference between treated and control communities, with regards to participation rates. That is, the PRAF program did not affect participation rates in any significant manner. Thus, this issue is not discussed further in the article.

difference-in-difference estimation presented above. In order to account for possible variations in the voting by year, I include dummy variables for three of the four elections years.6 This year variation only applies to regressions of presidential elections, where data are available for 2001 elections and 2005 elections. Moreover, since randomization was stratified by the malnutrition level of the municipality, the different municipal malnutrition categories are not randomly allocated to the four groups. Thus, I must control for the effect of being in a given malnutrition category on voting behavior. Acknowledging that the 280 or 140 observations only apply to 70 distinct municipalities, I cluster my results at the municipal level. The resulting econometric specification is presented below: V it ¼ a0 þ a1 H þ a2 H P þ a3 P þ b1 H AFTER þ b2 H P AFTER þ b3 P AFTER þ b4 S þ b5 Y þ εit (1) where Vit ¼ vote share received by the incumbent party in municipality i, in period t; H ¼ municipalities receiving individual-level transfers (Group 1); H-P ¼ municipalities receiving both individual and public goods transfers (Group 2); P ¼ municipalities receiving public goods transfers (Group 3); A ¼ dummy variable indicating whether the observations corresponds to one of the election years after the program has been implemented, that is 2001 and 2005; S ¼ stratum dummy variables for each of the five malnutrition categories; Y ¼ dummy variables for the election years. The variables that isolate the effect on voting are the coefficients b1, b2, and b3. For reasons of simplification, I assume that individuals in a given municipality are only influenced by the experience of their own municipality. This is a plausible assumption, due to the physical separation between villages. Even if villagers were influenced by the experience of others, their main contact would be with people of the same municipality. This was also the justification for using the municipality as the unit for randomization. 4.2. Data description The econometric analysis uses two data sets. The main data set covers election statistics as collected by the Honduran National Electoral Tribunal: Tribunal Supremo

6 In regressions not presented below, I also interacted the interaction term TREAT * AFTER with the year 2005, in order to allow the electoral effect of the program to vary by election year. This had a dismissible effect on results.


Electoral (TSE). Included are election results from the 1993, 1997, 2001, and 2005 presidential elections, as well as from municipal elections in 1997 and 2001.7 Thus, for each of the 70 municipalities, there are four observations regarding presidential elections, and two observations regarding municipal elections. The total number of observations is 280. All election results are aggregated at the municipal level, and include seven electoral variables: the vote share going to each of the five political parties in Honduras, the abstention rate, and the number of registered voters. Vote share is defined as the number of votes going to a given party, as a percentage of the total number of valid votes. The second data set was designed by the International Food Policy Research Institute (IFPRI)8 specifically for the evaluation of the PRAF program. The survey was given to individuals in 80 randomly selected “dwellings”9 in each of the 70 targeted municipalities. The total number of households surveyed is 587410 (Glewwe et al., 2003). The survey identifies each municipality’s group assignment (Group 1–Group 4), and it includes socio-economic indicators at the household and individual level. Also noteworthy is that this is a panel data set following the same households in 2000 and 2002. Since the PRAF program was implemented in 2000, the data on the 1993 and 1997 elections, as well as the data in the 2000 survey refer to information before the implementation of the program. The 2001 and 2005 election data and the 2002 follow-up survey all relate to post-implementation information. The merged data set includes election statistics from the 70 municipalities and is complemented by a variety of socio-economic indicators that are aggregated at the municipal level, such as average daily wage, literacy, and remittance availability. Therefore, any econometric analysis has the municipality as the unit of analysis. Although not ideal, this is an appropriate unit, considering that program randomization and targeting also happened at the municipal level. If the PRAF program had targeted individual households, we would be interested in analyzing the effects of this program on their individual voting behavior. However, since treatment was assigned to the municipality (and not to the household), we are interested in how the municipality’s voting behavior has changed. Moreover, the use of the municipality as the unit of analysis captures spillover effects to non-beneficiary households that live in treated communities. This is especially relevant in municipalities where supply-side transfers were given to schools and hospitals (Groups 2 and 3). The lack of individual-level data, however, does mean that I cannot describe how individual voters react

7 Honduras initiated a split ballot for municipal and presidential elections in 1997, so the notion of municipal elections is not applicable in 1993. The data regarding municipal elections in 2005 are not currently available. 8 While IFPRI designed the survey, ESA Consultores were responsible for the compilation of the database used. 9 A “dwelling” signifies a physical house, rather than a household. More than one household could theoretically dwell under the same roof. 10 This number is larger than 5600, the number of dwellings surveyed, because in some cases there were more than one household living in each dwelling.

869

to the PRAF program, without addressing the ecological inference problem. Therefore, this quantitative analysis will only present municipal level effects of the PRAF program. 4.3. Summary statistics The validity of the regressions rests on the assumption that the treatment assignment of the PRAF program was truly randomized. Randomization allows for the effect of treatment on voting behavior to be isolated, since all other factors that may influence voting should be randomly distributed between all four groups. Thus, any significant difference in voting patterns in the four groups can be attributed to treatment assignment. In order to ensure that the impact of treatment can be isolated, baseline voting behavior should not be differentiable by group assignment. Table 1a depicts the differences between each assignment in the 1993 and 1997 elections, so as to ensure that there is no difference between treatment groups. There are large changes in abstention rates between 1993 and 1997, suggesting that some external factor influenced political mobilization between 1993 and 1997. Yet, as expected, these changes are distributed relatively equally between groups. That is, there are no statistically different results between groups in the 1993 and 1997 elections. This, then, provides a sound base for the use of this randomization to isolate the effect of the PRAF program in the 2001 and 2005 elections. Table 1b presents some socio-economic characteristics of the households under study, disaggregated by group, before the initiation of the program. While this does not have a direct influence on voting behavior, the table describes the population in each treatment group. It also shows that the municipalities are relatively similar overall. For two variables, average daily pay in the community and the average age children start work, we see statistically significant differences between assignment groups. However, as long as voting behavior before the initiation of the program is similar, which it is, these differences should not distort the results of the study. Hence, the difference-indifference econometric specification can be effectively used to isolate the effect of the PRAF program on voting behavior. 4.4. Econometric results 4.4.1. Effect on presidential elections Table 2 disaggregates the effect of being in each of the different treatment groups after the initiation of the program, on presidential elections. The interaction variables between each of the groups and the variable AFTER (H * AFTER, H-P * AFTER, P * AFTER) measure the electoral effect of being in a treatment group, after the initiation of the program. There is no significant effect on voting behavior in any of the three treatment groups after the initiation of the program. As one might expect, the vote share going to the two main parties is highly related to the election year. Ultimately, however, the quantitative analysis shows no discernable effect of participation in the

870


Table 1a Baseline election results disaggregated by group assignment.

Presidential Elections 1993 % PL vote % PN vote Abstention Rate 1997 % PL vote % PN vote Abstention Rate Municipal Elections 1997 % PL vote % PN vote Abstention Rate

HOUSEHOLD

HOUSEHOLD þ PUBLIC

PUBLIC

CONTROL

P-value of difference

.418 (.082) .543 (.083) 25.9 (9.35)

.429 (.083) .530 (.076) 24.5 (7.53)

.423 (.131) .540 (.129) 21.3 (9.23)

.453 (.069) .509 (.066) 23.6 (6.81)

.639 .606 .524

.456 (.078) .500 (.082) .186 (.058)

.456 (.079) .486 (.089) .184 (.063)

.429 (.098) .519 (.106) .160 (.064)

.469 (.090) .483 (.085) .158 (.088)

.684 .711 .493

.494 (.070) .455 (.064) .177 (.075)

.455 (.093) .472 (.103) .183 (.065)

.445 (.087) .497 (.094) .169 (.064)

.465 (.082) .460 (.077) .175 (.062)

.371 .602 .951

Notes: 1. In the first four columns, the mean value is given followed by the standard deviation in parentheses. 2. The last column presents the P-value of the difference between the means, as calculated by an F-test.

PRAF program on voting decisions for presidential elections. 4.4.2. Effect on municipal elections The effects of the program on municipal elections are shown in Tables 3–5. The regressions consider the differential change in voting behavior between the treatment groups and the control group, between 1997 and 2001. The coefficients that measure the effect of being in a treatment group on voting behavior in 2001 are the three interaction terms: H * AFTER, H-P * AFTER, and P * AFTER, presenting the electoral effect on each of the three types of treatment. The dependent variable measured is the vote share going to the incumbent mayor. In the years under study, the incumbent mayor could belong to either party, irrespective of which party was implementing the program at the presidential level. The results show a significant difference in voting behavior for municipalities only receiving direct cash transfers (Group 1 municipalities). Specifically, the incumbent’s party vote share is 6.6% points higher in municipalities with direct cash transfers, after the initiation of the PRAF program. Thus, the incumbent mayor’s party is rewarded for receipt of cast transfers interventions. In municipalities with both household-level and municipality-wide public goods transfers, as well as in municipalities with solely public goods transfers, there is no statistically significant effect of the PRAF program. An Ftest was used to determine whether these two coefficients were statistically different from each other. Even though municipalities with both types of transfers (Group 2) seem to be responding more negatively to the PRAF program, the coefficient on solely public goods transfers (Group 3) is not statistically different from the coefficient on municipalities that received both interventions. That is, there are no evident conclusions regarding the electoral effect of the PRAF program in those municipalities. In order to understand the practical magnitude of the effect in municipalities that received household cash transfers, Table 4 presents the probability of the incumbent party re-winning the elections. I run a probit model on the

binary dependent variable of winning an election for the incumbent mayor.11 That is, Table 4 shows the probability of the incumbent mayor winning the election in 2001, after the receipt of transfers. In line with Table 3, being in a municipality that receives household cash transfers after the initiation of the PRAF program increases the likelihood that the incumbent mayor’s party will be re-elected by 38.9% points. 4.4.3. Incumbency effect or party effect? Until now, the effects on incumbent mayors have been analyzed as the main dependent variable. Since many of the incumbent mayors however were of the PNH party, it is difficult to discern whether voters are specifically voting for the PNH party or for the incumbent mayor. Nonetheless, an “incumbency effect” would be the driving force if the incumbent vote share went up in treated municipalities, even when controlling for party affiliation. Conversely, a “party effect” would be the driving force if the PNH vote share went up in treated municipalities, even when controlling for municipalities that elected their incumbent. I, therefore, place the two effects into one regression by making a new dependent variable, VOTE 2001, which refers to the vote share received by both the PLH-party candidate, and the PNH-party candidate. The independent variables then include a binary variable that distinguishes which party the observation is referring to, as well as if that party was the incumbent in the given year. The new econometric specification is presented below:

(2) VOTE 2001it ¼ a0 þ a1VOTE 1997 þ b1PN þ b2INCUMB þ b3H þ b4H-P þ b5P þ b6H * INCUMB þ b7H-P * INCUMB þ b8P * INCUMB þ b9H * PN þ b10H-P * PN þ b11P * PN þ b12 STRAT þ εit

11 Since in every case where the PL did not win, the PN won the municipality, there is no reason to present the results on the PN winning an election. The probability of the PNH winning an election is equal and opposite to the probability of the PLH winning an election.


871

Table 1b Household and individual-level characteristics of targeted municipalities.

Female Age Can Read? Daily Pay in Community Average age start work Households with a child between 6 and 16 yrs old Number of people living in one house Households with bathroom within the house Households with electricity Households with access to Remittance

HOUSEHOLD

HOUSEHOLD þ PUBLIC

PUBLIC

CONTROL

P-value of difference

.498 22.7 .516 541 11.9 .312 6.39 .053 .160 .335

.512 23.3 .561 553 11.5 .328 6.23 .059 .188 .323

.492 22.7 .577 55.1 11.3 .291 6.00 .104 .120 .264

.493 23.8 .559 125 11.2 .295 6.26 .053 .195 .346

.633 .409 .320 .000 .070 .330 .322 .420 .866 .341

(.011) (.480) (.021) (130) (.196) (.009) (.123) (.013) (.033) (.028)

(.012) (.427) (.035) (113) (.171) (.021) (.123) (.019) (.040) (.021)

(.017) (.841) (.038) (12.4) (.292) (.014) (.175) (.029) (.064) (.034)

(.012) (.561) (.016) (35.5) (.218) (.011) (.161) (.013) (.031) (.037)

Notes: 1. In the first four columns, the mean value is given followed by the standard deviation in parentheses. 2. The last column presents the P-value of the difference between the means, as calculated by an F-test.

where VOTE 2001 ¼ vote share going to both parties in 2001; VOTE 1997 ¼ vote share going to both parties in 1997; PN ¼ a dummy variable that equals 1 if the vote share refers to the PNH party, and 0 if the vote share refers to the PLH party; INCUMB ¼ a dummy variable that equals 1 if the vote share refers to the incumbent, and 0 otherwise; H ¼ municipalities receiving household-level cash transfers (Group 1); H-P ¼ municipalities receiving household and public goods transfers (Group 2); P ¼ municipalities receiving public goods transfers (Group 3). Since the significant increase in vote share occurs in municipalities receiving household transfers, the coefficients that distinguish an incumbency effect from a party effect are b6 and b9 respectively. The coefficient b6 measures the differential effect in municipalities receiving household transfers on the vote share going to the incumbent mayor, irrespective of party affiliation. If this is significant, the results can be interpreted as an increase in support for the incumbent mayor’s party. The coefficient b9 measures the differential effect in municipalities receiving Table 2 Effect of being in a given treatment group on vote share in presidential elections. Y ¼ % vote in presidential elections

Y ¼ % Incumbent vote

HOUSEHOLD HOUSEHOLD þ Public PUBLIC HOUSEHOLD * AFTER HOUSEHOLD L PUBLIC * AFTER PUBLIC * AFTER 1997 2001 2005 Strat1 Strat2 Strat3 Strat4 Constant N R2

.011 .004 .004 L.013 L.006 L.010 .073*** .081*** .046*** .004 .007 .002 .004 .527*** 278 .171

(.011) (.010) (.011) (.013) (.012) (.014) (.019) (.017) (.011) (.007) (.008) (.007) (.006) (.011)

Notes: 1. In each cell, the coefficient is given, followed by the robust standard error in parenthesis. 2. *** denotes significance at the 1% level, ** at the 5% level, and * notes at the 10% level. 3. Results are clustered by municipality

household transfers, on the vote share going to the PNHparty candidates, irrespective of incumbency status. If this is significant, the results can then be interpreted as an increase in support for the PNH-party candidate. Table 6 presents the results of this regression. As is clear in the table, both the coefficient measuring incumbency effect and the coefficient measuring party effect are highly significant when placed in the same regression. Interestingly, however, the effect is larger than in previous regressions. The vote share going to the incumbent in municipalities with household transfers increases by 6.9% points. Similarly, PNH-party candidates receive a 9.8% point increase in support. The larger magnitudes can be attributed to the high correlation between being an incumbent mayor and being a PNH-party candidate. Thus, in the cases where an incumbent mayor was reelected and was the PNH-party candidate, I do not have other information that might discern which effect was driving the results. Moreover, the small sample size of PNHparty mayors who were not the incumbent means that we cannot make any further conclusions. Internal validity constraints prohibit an ex-post analysis of such a small subsample of individuals, so this additional result is presented in favor of transparency and to incite further research on this issue.

5. Discussion: pocketbooks matter, but when? This discussion briefly explores the differential effect of CCT on municipal and national elections.12 In municipalities that received household-level transfers (Group 1), the incumbent mayor enjoyed increased electoral support, as compared to control municipalities. Vote share going to the

12 I discount the possibility that different individuals are voting in municipal and presidential elections. Considering that the elections happen on the same day, and the two ballots are given to the voter at the same time, it is highly unlikely that a significant number of voters in a municipality would only vote in one of the elections. Of course, it is possible that some voters may only be interested in voting in one election. However, given the significant electoral effect on municipal elections, it is important to question why the same voter might vote differently in municipal versus presidential elections. That is, we must explain what determines the municipal electoral choice, as opposed to the presidential electoral choice.

872


Table 3 Effect of being in a given treatment group on voting in municipal elections. Y ¼ % vote in municipal elections

Y ¼ % Incumbent vote

HOUSEHOLD HOUSEHOLD þ PUBLIC PUBLIC HOUSEHOLD * AFTER HOUSEHOLD L PUBLIC * AFTER PUBLIC * AFTER AFTER PROGRAM Strat1 Strat2 Strat3 Strat4 Constant N R2

.021 .025 .045 .066*** L.015 .000 .016 .004 .024 .035* .010 .491*** 135 .159

(.020) (.026) (.029) (.024) (.024) (.029) (.018) (.025) (.022) (.019) (.018) (.019)

Notes: 1. In each cell, the coefficient is given, followed by the robust standard error in parenthesis. 2. *** denotes significance at the 1% level, ** at the 5% level, and * notes at the 10% level. 3. Results are clustered by municipality.

incumbent mayor increased by 6.6% points, which translated into an increased probability of 38.9% of being re-elected. No similar electoral reward can be discerned in presidential elections. This pattern rejects the basic model of fixed partisan voting that permeates the literature on Honduran electoral behavior. Instead, the large electoral reward to incumbent mayors implies that Honduran voters are willing to vote based on changes in disposable income, affirming the pocketbook model as a potential explanation of voting behavior. Yet, the non-effect on presidential elections suggests that an increase in disposable income does not necessarily translate into a general reward for the incumbent government, as the basic pocketbook model might suggest. The question, then, arises: when do pocketbooks matter most? One explanation follows Popkin’s (1994) caveat that changes in income must be attributed to politicians in order

Table 4 Effect of Being in a treatment group on the probability of a mayor winning the elections. Y ¼ Probability winning elections (0,1)

Y ¼ Incumbent wins

HOUSEHOLD HOUSEHOLD þ PUBLIC PUBLIC HOUSEHOLD * AFTER HOUSEHOLD L PUBLIC * AFTER PUBLIC * AFTER AFTER PROGRAM Strat1 Strat2 Strat3 Strat4 N Pseudo R2

.237 .074 .146 .389*** L.058 .167 .144 .082 .072 .221 .126 136 .104

(.160) (.153) (.185) (.071) (.207) (.226) (.106) (.157) (.156) (.146) (.139)

Notes: 1. The marginal effect is presented, followed by the robust standard error in parentheses. 2. *** notes significance at the 1% level, ** notes significance at the 5% level, * notes significance at the 10% level.

Table 5 Effect of being in a given treatment group on voting in municipal elections (robustness check). Y ¼ % vote in 2001 municipal elections

Y ¼ % incumbent vote

% PL vote in 1997 Elections % PN vote in 1997 Elections % Incumbent vote in 1997 Elections HOUSEHOLD HOUSEHOLD D PUBLIC PUBLIC Average Age Average Femalea % Can Read Education Level Average Age People Start Work % Access to Remittance % Enrolled in School this year Average School Satisfaction Average Health Center Satisfaction Strat1 Strat2 Strat3 Strat4 Constant N R2

– – .360*** .044** .003 .023 .006 .454* .073 .016 .002 .116* .053 .034 .020 .000 .006 .010 .028 .112 66 .430

(.087) (.021) (.019) (.018) (.005) (.232) (.127) (.039) (.005) (.067) (.120) (.034) (.030) (.022) (.027) (.030) (.018) (.240)

Notes: *** notes significance at the 1% level, ** notes significance at the 5% level, * notes significance at the 10% level. a The marginal effect is presented, followed by the robust standard error in parentheses.

for there to be an electoral effect. Given a strong tradition of patronage politics in Honduras, voters may have believed that mayors played a role in this selection process, and rewarded them accordingly. This mistake on the part of voters is plausible since by 2001, transfers were still provided by PRAF employees and service providers (rather than banks), so the governmental source of the transfer may have been less obvious than in the later years of the program. A more rigorous qualitative survey of beneficiaries regarding their understanding of the program design would be necessary to rule out the possibility of misinformation. Such research has been conducted in Mexico and Jamaica. Specifically, in her comprehensive qualitative study on Mexico’s PROGRESA, Adato et al. (2000, 41) shows that beneficiaries have an “adequate” understanding of program design and requirements. It is not clear whether this level of information would erase all suspicions that mayors influenced the selection process. A second explanation of when pocketbooks matter most reflects the importance of informational shortcuts. Previous claims about “frozen” party cleavages in Honduras are based on voting in presidential elections (Dix, 1989; TaylorRobinson, 2001, 2006; McDonald and Ruhl, 1989). The absence of an effect of the transfer program on national elections would be consistent with this literature if fixed partisanship is used as an informational shortcut. If information is less available on the performance of presidential incumbents, voters may use a candidate’s party affiliation to roughly estimate their expected utility from that candidate. In real terms, this means unchanging partisanship in national elections. The same voter, however, may be more amenable to switching his or her vote at the local level,

E. Linos / Electoral Studies 32 (2013) 864–874 Table 6 Comparing the incumbency effect with the party effect on municipal elections. Y ¼ Vote share in 2001 municipal elections Vote Share 1997 PNH (0,1) INCUMBENT (0,1) HOUSEHOLD HOUSEHOLD þ PUBLIC PUBLIC HOUSEHOLD * INCUMBENT HOUSEHOLD PUBLIC * INCUMBENT PUBLIC * INCUMBENT HOUSEHOLD * PNH HOUSEHOLD PUBLIC * PNH PUBLIC * PNH Party Strat1 Strat2 Strat3 Strat4 Constant N R2

.399*** .008 .041** .071*** .014 .036 .069** .004 .057* .098*** .037 .027 .010 .016 .015 .001 .315*** 138 .399

(.010) (.014) (.018) (.019) (.015) (.023) (.028) (.034) (.030) (.029) (.034) (.030) (.007) (.011) (.011) (.004) (.042)

Notes: 1. The marginal effect is presented, followed by the robust standard error in parentheses. 2. *** notes significance at the 1% level, ** notes significance at the 5% level, * notes significance at the 10% level.

rewarding a mayor for improvements in disposable income. Finally, it is possible that pocketbooks matter in both national and local elections but that other issues overshadowed the national elections. Domestic security, gang violence and kidnapping were very salient issues in the 2001 presidential elections (CID/Gallup Latin America, 1999), while both poverty and crime figured prominently in the 2005 presidential race (BBC News, 2005). Since the mayor would have no direct authority over such issues, it is possible that a household transfer was more salient at the municipal level, and rewarded accordingly. 6. Summary and implications for further research The explanations above can only provide plausible scenarios that should be affirmed or rejected through further research. The following explanation, however, is the most convincing. Firstly, Hondurans in treated municipalities may be uninformed, pocketbook voters that believed that mayors must have played a role in their receipt of transfers. This explanation matches both the non-effect in presidential elections, as well as the large incumbency effect in municipal elections. Secondly, Hondurans may be party loyalists in presidential elections, but pocketbook voters in municipal elections. This approach is confirmed by other studies on Honduras and may provide insight into the use of information shortcuts. Ultimately, the above empirical analysis shows that the PRAF program did not provide any electoral benefits to the party that initiated it. If governments cannot expect any short-term electoral rewards for implementing this type of social program, governments have little incentive to target social programs according to objective criteria such as malnutrition status or financial need. If incumbent mayors

873

will be rewarded at the ballot box, the central government has an incentive to target the program such that municipalities will receive transfers only if their mayor is affiliated with the central government. Indeed, while the Honduran PRAF program as described in this study ended in November of 2005, the next president of Honduras initiated his own version of a Conditional Cash Transfer program. This new version of the program does not follow the older version’s targeting scheme. That is, the national government has selected the communities to which funds will be transferred, using alternative internal criteria. Ultimately, the Honduran PRAF suggests that correctlytargeted social programs can still provide electoral support for incumbents. That is, real social policy can “unfreeze” voting behavior even in countries where politics are considered deeply hereditary. However, the attribution of responsibility and the practical role of subnational institutions mediates how much political capital a national government can gain from such wide-reaching national programs. This, then, opens doors for further research regarding national programs and subnational voting behavior. Recent literature in the US context shows similar results. Mettler (2010) discusses how national government policies are submerged so that individual voters cannot appropriately attribute responsibility. Continued research on the topic is crucial given that the potential electoral rewards of such programs will surely influence their future implementation. Ultimately, this study suggests that Honduran voters are neither perfectly informed, nor perfectly loyal when making an electoral decision about social policy. As such, the quantitative analysis of the PRAF program contributes to a literature that is only beginning to emerge, regarding the electoral effects of large-scale social programs, experienced at the local level. Acknowledgments This paper would not be possible without the invaluable advice and support of Michael Kremer, Dan Levy, and Katerina Linos on previous versions of this research. The author also thanks Jinu Koola, Ayah Mahgoub and two anonymous reviewers for their insightful commentary on this paper. References Adato, M., Coady, D., Ruel, M., 2000. An Operations Evaluation of PROGRESA from the Perspective of Beneficiaries, Promotoras, School Directors, and Health Staff. International Food Policy Research Institute, Washington, DC. Ames, B., 2001. The Deadlock of Democracy in Brazil. University of Michigan Press, Ann Arbor. Arceneaux, K., Stein, R., 2006. Who is responsible when disaster strikes? The attribution of responsibility for a natural disaster in an urban election. Journal of Urban Affairs 28, 43–54. BBC News, November 24, 2005. Q&A: Honduras votes. BBC News Online. http://news.bbc.co.uk/1/hi/world/americas/4465792.stm (accessed 03.02.07.). Bouillon, C.P., Tejerina, L., 2006. Do we know what works? A systematic review of impact evaluations of social programs in Latin America and the Caribbean. In: Working Paper Presented at the Inter-American Development Bank Annual Meeting, Minas Gerais, Brazil. CID/Gallup Latin America. Opinión Pública No. 41: Honduras, 1999. CID Latin America. http://www.cidgallup.com/docs/opiniones/honop41. pdf (accessed 03.02.07.).

874


De Janvry, A., Finan, F., Sadoulet, Elisabeth, 2008. Local Electoral Accountability and Decentralized Program Performance. Working paper. Available at. http://are.berkeley.edu/wsadoulet/papers/Bolsa708.pdf. De la O, A.L., 2006. Do Poverty Relief Funds Affect Electoral Behavior? Evidence from a Randomized Experiment in Mexico. Working paper. Massachusetts Institute of Technology. Dix, R.H., 1989. Cleavage structure and party systems in Latin America. Comparative Politics 22 (1), 23–37. Downs, A., 1957. An Economic Theory of Democracy. Harper & Row, New York. Dugger, C.W., January 3, 2004. To help poor be pupils, not wage earners, Brazil pays parents. New York Times. Fiorina, M., 1979. Retrospective Voting in American National Elections. Yale University Press, New Haven. Glewwe, P., Olinto, P., de Souza, P.Z., 2003. Evaluating the Impact of Conditional Cash Transfers on Schooling in Honduras: An Experimental Approach. Working Paper. University of Minnesota. Green, D., Palmquist, B., Schickler, E., 2002. Partisan Hearts and Minds: Political Parties and the Social Identities of Voters. Yale University Press, New Haven, CT. IADB (Inter-American Development Bank), 2006. Honduras: country indicators. http://www.iadb.org/countries/cof.cfm?language¼English&id_ country¼HO&parid¼2 (accessed 18.10.06.). IFPRI (International Food Policy Research Institute), 2000. Second Report: Implementation Proposal for the PRAF/IDB Project – Phase II. International Food Policy Research Institute, Washington, DC. IFPRI (International Food Policy Research Institute), 2002. Proyecto PRAF/ BID Fase II: Impacto Intermedio: Sexto Informe. Report Submitted to PRAF. International Food Policy Research Institute, Washington, DC. IFPRI (International Food Policy Research Institute), 2003. Proyecto PRAF/ BID Fase II: Impacto Condicional: Séptimo Informe. Report Submitted to PRAF. International Food Policy Research Institute, Washington, DC.

Kingstone, S., October 30, 2006. Lula bounces back from the ropes. BBC News. Online edition. http://news.bbc.co.uk/2/hi/americas/6100928. stm (accessed 08.01.07.). Moore, C., 2008. Assessing Honduras’ CCT Programme PRAF, Expected and Unexpected Realities. International Poverty Centre Country Study. Available at: http://www.ipc-undp.org/pub/IPCCountryStudy15.pdf. Mettler, S., 2010. The Submerged State: How Invisible Government Policies Undermine Government. University of Chicago Press, Illinois, Chicago. McDonald, R.H., Ruhl, J.M., 1989. Party Politics and Elections in Latin America. Westview, Boulder, CO. Manacorda, M., Miguel, E., Vigorito, A., 2009. Government Transfers and Political Support. Working Paper. Available at. http://www.econ. berkeley.edu/emiguel/pdfs/miguel_uruguay.pdf. Popkin, S.L., 1994. The Reasoning Voter. University of Chicago Press, Chicago. Reel, M., October 29, 2006. Cash aid program Bolsters Lula’s reelection prospects. Washington Post. Roberts, K.M., Wibbels, E., 1999. Party systems and electoral volatility in Latin America: a test of economic, institutional, and structural explanations. American Political Science Review 93 (3), 575–590. Samuels, D., Abrucio, F.L., 2000. Federalism and democratic transition: the ‘New’ politics of the governors in Brazil. Publius 30, 43–62. Taylor-Robinson, M.M., 1996. When electoral and party institutions interact to produce caudillo politics: the case of Honduras. Electoral Studies 15, 327–337. Taylor-Robinson, M.M., 2001. Old parties and new democracies: do they bring out the best in one another? Party Politics 7 (5), 581–604. Taylor-Robinson, M.M., 2006. In: Helmke, G., Levistky, S. (Eds.), The Difficult Road from Caudillismo to Democracy: The Impact of Clientelism in Honduras in Informal Institutions and Democracy: Lessons from Latin America. John Hopkins University Press, Baltimore, MD.