Incentives to Vote, Political Preferences and Information

0 downloads 0 Views 2MB Size Report
Oct 21, 2011 - conducted a follow-up survey, where I gathered an objective measure of turnout. ..... in a district different than the one where she is registered, the voter will be ... Importantly, no major news outlet reported the changes in the fine, and .... a random sample of households.15 By clustering the randomization at ...
Incentives to Vote, Political Preferences and Information:



Evidence from a Randomized Experiment in Perú

Gianmarco León



Department of Agricultural and Resource Economics University of California, Berkeley

October 21, 2011

Job Market Paper Abstract Electoral institutions that encourage citizens to vote are widely used around the world. Yet little is known about the eects of such institutions on voter participation and the composition of the electorate. In this paper, I exploit an unusual institutional change in Peruvian voting laws to identify the eect of nes for abstention on voting. My ndings suggest that the elasticity of voting with respect to cost is -0.2, which implies that reducing the ne for not voting by 50% leads to a 10 percentage point reduction in turnout. Consistent with the theoretical model presented, the reduction in turnout is driven by voters who (i) are in the center of the political spectrum, (ii) are less interested in politics, and (iii) hold less political information. However, voters who respond to changes in the cost of abstention do not have dierent preferences for policies than those who vote regardless of the cost. Further, involvement in politics, as measured by the decision to acquire political information, seems to be independent of the level of the ne. Additional results indicate that an increase in the opportunity cost of voting (i.e. a reduction in the ne) reduces the incidence of vote buying and increases the price paid for a vote.



I am very greatful to Elisabeth Sadoulet, Alberto Chong, Alain de Janvry, Ernesto Dal Bó, Fred Finan, Mitch Homan,

Larry Karp, Valerie Koechlin, Jeremy Magruder and Ted Miguel.

Their insighful comments, suggestions, support, and

encouragment have been extremely important for the development of this project. Special thanks to Alex Solís, my ocemate, who has patiently heard the contents of this paper at least a thousand times, and always provided smart feedback. Participants in the UC Berkeley Development Lunch, ARE Development workshop, ARE Department Seminar, and seminars held at the University of San Francisco and GRADE (Lima-Perú) provided useful suggestions. Roberto Rodríguez was an outstanding eld assitant, and Alina Xu and David Arnold assisted with the data cleaning.

Financial support from the Institute of

Business and Economic Research (IBER) and Center for Evaluation and Global Action (CEGA) is greatly appreciated. The standard disclaimer applies. † Department of Agricultural and Resource Economics, University of California, Berkeley. Coordinates: 228B Giannini Hall, University of California, Berkeley, CA 94720-33. E-mail: [email protected].

1

Keywords: Voting Behavior; Incentives to Vote, Public Choice, Perú JEL Classication Codes: D71, D72, O53

2

1

Introduction

Thirty-three countries around the world encourage participation in elections through compulsory voting. Most of them implemented this institution believing that it would ensure that voters' preferences are adequately represented in elections. However, it is not clear the extent to which voting incentives aect turnout. More importantly, voting institutions may change the composition of the electorate and therefore the outcome of elections. For example, by mandating people to vote, we could be inducing less informed or uninterested voters into the polls, which might distort public choice. Further, participating in elections could generate incentives for voters to get more involved in the political process.

On the other hand,

under voluntary voting, abstainers could be over-represented in a particular group of the population, for example the poor. By not mandating them to vote, this group's preferences will not be reected in the policies enacted. However, both voting and enforcement institutions are costly, and if the objectives of higher participation and more involvement are not achieved, there could be signicant welfare losses. To understand how voting institutions aect the outcome of an election, it is important to rst explain what determines voters' decision to participate. More importantly, we need to know who is more likely to respond to incentives, the magnitude of voters' responsiveness, and the implications for public choice. In this paper I explore these questions through a rational choice model of voter behavior with imperfect information.

The model predicts that a reduction in the cost of abstention will decrease turnout, and

that this decrease will be more than proportional among (i) centrist voters, (ii) those who have a lower subjective value of voting, and (iii) voters who hold less political information. I test these prediction using random variation in the cost of abstention. Empirically, I exploit a change in Peruvian voting laws that reduced nes for abstention, and particularly the fact that knowledge about this reduction was not widespread. I study the 2010 municipal elections in Perú, where I generate experimental variation in the cost of abstention by informing voters in the treatment group about the new levels of the ne for not voting. Voters assigned to the control group were reminded about the ne, without any mention of the exact amount.

After the election, I

conducted a follow-up survey, where I gathered an objective measure of turnout. My ndings suggest that the elasticity of voting with respect to the cost is -0.2, implying that reducing the ne by 50% causes a 10 percentage point reduction in turnout. Extrapolating my results, this means that if voluntary voting were implemented (i.e. the ne was reduced to zero), turnout would decrease from 94.2% to about 74%, roughly

3

what we observe in countries where voting is voluntary. Consistent with the predictions of the model, the reduction in turnout is driven by voters in the center of the political spectrum, those less interested in politics, and by uninformed voters. Overall, I nd that the reduction in turnout is clearly related to specic voter characteristics. However, this change in the composition of the electorate does not necessarily imply that the outcome of the election will be aected. Poor people are not more likely to respond to changes in the ne. Interestingly, voters whose turnout decision is more sensitive to a change in the ne do not have distinct policy preferences. Further, voters who respond to the reduction in the ne by not voting do not show less involvement in politics, as measured by the amount of political information they acquire. Additional results indicate that a decrease in the cost of abstention reduces the incidence of vote buying by 20%, and increases the price politicians pay for the marginal vote by 76%. This last result is consistent with an exogenous shift in the supply of votes. Voting behavior has been studied by both economists and political scientists for a long time, yet there is no canonical model for understanding turnout decisions.

While theoretical research modeling

the determinants of voter turnout has peaked in the last decade, very few empirical studies have been conducted to study voter behavior, let alone to test the predictions of these models. This is especially the case in developing countries.

In this paper, I provide evidence supporting the prediction of one of

the models derived from the classic calculus of voting literature (Downs, 1957; Rikker and Ordeshook

1

1968).

The empirical results from this paper are closely related to several strands of the literature on voter behavior and electoral institutions. First, I contribute to the growing literature on the determinants of voter turnout (Gerber and Green, 2000, Gentzkow, 2006, Brady and McNulty, 2011).

1

2

My paper uses

Merlo (2006) and Martinelli (2007) provide excellent reviews of the theoretical models of turnout. The models available

in the literature can be classied as those that emphasize the probability of being pivotal as the main motivation to vote (Borgers, 2004; Ledyard, 1984; Palfrey and Rosenthal, 1985); those that argue that citizens are driven to the polls to fulll their civic duty and do the right thing (Harsanyi, 1980; Feddersen and Sandroni, 2006; Feddersen, Gailmard, and Sandroni, 2009, Coate and Conlin, 2004); and uncertainty voter models, which endogenize a component of the cost of voting (Deagan, 2006; Deagan and Merlo, 2009, Feddersen and Pesendorfer, 1996, 1999; Matsusaka, 1995).

2

Several of these papers use large scale eld experiments to identify the positive eects of dierent types of voter mobi-

lization campaigns on turnout in the United States (Gerber and Green, 2000, 2001 and Gerber et al., 2003). This literature has also shown that social pressure is an important extrinsic motivation for voting (Gerber et al. 2008) and that voting is habit forming: voting in one election signicantly increases the probability of going to the polls in the next election (Gerber et al., 2003). Another strand of the literature emphasizes that more informed voters are more likely to vote. Areas where the TV or radio expanded earlier were more likely to show higher turnout (Gentzkow, 2006, Lasen, 2005). This fact has been shown to hold with specic information campaigns at the individual level (Banerjee et al., 2010). More closely related to my paper, a few empirical studies use natural experiments to test whether changes in the cost of voting aect the likelihood of going to the polls in the election day. Brady and McNulty (2011) show that an increase in the cost of voting induced by an

4

experimental evidence from the eld to understand how an institutional change that exogenously reduces the cost of abstention aects turnout.

3

Unlike the previous literature, I am able to quantify the changes

in the cost of (not) voting at the individual level. Moreover, changes in the perceived ne are induced by a randomly assigned treatment, which allows me to causally interpret the eect on turnout. Further, I provide the rst estimates in the literature of the cost elasticity of voting, a parameter necessary for evaluating the dead weight loss associated with policy interventions that aect the cost of voting.

4

To a large extent, the lack of credible evidence on the eects of electoral rules on turnout decisions is due to the fact that there are not many changes in electoral rules around the world. When there are, it is nearly impossible to collect individual level information and especially objective measures of turnout. Further, these institutions apply to every voter, which limits our ability to causally interpret changes in behavior. In this paper, I take advantage of an institutional change about which few people were aware, and I generate experimental variation by randomizing its salience and updating voters' information set. In this sense, I contribute to the growing literature that uses eld experiments to understand voter behavior in developing countries.

5

Experimenting with the salience of an institutional change is a promising research

tool to get causal estimates from specic institutional features. New laws are passed frequently, and they are not always publicized for dierent reasons, or citizens are not aware of them because of selective and limited attention.

Even though it is nearly impossible to randomize an institution, we can experiment

with its salience. A third strand of literature closely related to my paper analyzes how policy making responds to changes in the electorate. For example, the standard median voter model predicts that any changes in the composition of the electorate aects who gets elected through a change in the characteristics of the median voter (Persson and Tabellini, 2000; Husted and Kenny, 1997). Miller (2008) and Fujiwara (2011) analyze specic events in which groups of the population with identiable policy preferences were enfranchised. As a consequence, they observe that policies respond to the new composition of the electorate. Unlike these studies, in the case I analyze, there is no reason to expect that the groups that stop going to the unexpected reduction in the number of polling stations in California's 2003 gubernatorial elections generated 3.03 percentage point reduction in polling place turnout, while absentee vote increases in 1.18 percentage points. Another commonly used source of exogenous variation is the presence of inclement weather conditions in the election day. These studies nd that, on average, an additional millimeter of rain tends to reduce turnout by 1 percentage point (Knack, 1994, Gomez et al., 2007, Hansford et al. 2010, Fraga and Hersh, 2010.). In terms of partisan eects, the results are mixed.

3 4

Laboratory experiments along these lines have been conducted by Gerardi et al. (2010). Examples of such policies are the increase of polling stations, transportation to the polling stations, electronic voting,

availability of ID cards, etc.

5

Pande (2011) provides a comprehensive survey of this literature.

5

polls have particular policy preferences. As such, though my ndings show that the reduction in the cost of abstention changes the composition of the electorate, citizens who stop voting do not have signicantly dierent policy preferences, which suggests that we shouldn't expect changes in the policies enacted. Finally, the results of the paper also speak to the growing literature analyzing vote buying in developing countries (Finan and Schechter, 2011; Vicente, 2008; Vicente and Wantchekon, 2009). My results provide evidence that a change in the cost of abstention leads to a shift in the supply of votes, thus reducing the incidence of vote buying, and increasing the price, making it more costly to politicians to aect the outcome of the elections. In the next section, I present a theoretical model to characterize voter behavior, and I motivate the empirical analysis.

Section 3 gives institutional background on the Peruvian electoral system and the

change in the law that reduced the ne for abstention. Section 4 explains the experimental design and the data that I use for the empirical analysis, which is presented and discussed in Section 5. Finally, Section 6 summarizes my ndings and discusses them.

2

The Model

In this section I present a slight variation of the basic model from Degan (2006), Merlo (2006), and Degan and Merlo (2011), in which I introduce an additional term of interest to motivate the empirical analysis. The theory builds on a rational choice model where the voting decision is based on a threshold strategy: if the cost of voting is greater than the benets, citizens go to the polls, otherwise, they abstain. More specically, I consider a two candidate election where voters share a common prior about the distribution of ideological positions in the population, but are uncertain about the position of the candidates. Each voter has an exogenous portion of the cost of voting and abstaining, represented by the utility derived from fullling one's civic duty and the ne for not voting, respectively. On the other hand, there is an endogenous portion of the cost of voting, which is the utility loss due to the possibility of making a voting

mistake, i.e. voting for a candidate whose ideological position is far from the voter's. The endogenous part of the cost of voting drives the comparative statics of the model, which predict that a reduction in the cost of abstention will reduce turnout. Lower turnout will be driven by voters who (i) are in the political center, (ii) have a lower subjective value of voting, and (iii) are uninformed. I assume that there are two candidates running in the election, which I denote by

6

J = {L, R}.

Each

candidate has a position

y

in a uni-dimensional policy (or ideological) space

Y = [−1, 1].

We can interpret

the ideological or policy space as left/right, where 0 represents the center. I assume that candidate always to the left of candidate

tive, the candidate's ideological positions are random variables

F (yL , yR |yL < yR ).

distributed on the support

[−1, 1].

(y˜L ,y˜R )

From the voter's perspec-

distributed according to a joint

Without loss of generality, I assume that

F (.)

is uniformly

The main source of heterogeneity between voters is the amount of

information they hold about the candidates, which I denote by

Ωi ∈ Ω = {F (yL , yR )}.

completely uninformed about the ideological position of the candidates, she only observes she has perfect information,

is

R, yL < yR .

Citizens are uncertain about the ideological position of the candidates.

probability distribution

L

Ωi = (yL , yR ).

If a voter is

F (.),

while if

Information is assumed to be an exogenous, individual level

parameter. Voters are also heterogeneous in the subjective benet they derive from voting, or from fullling their civic duty. This utility is represented by

di ,

which follows a uniform distribution on the support

Additionally, there is a cost of not going to the polls: a ne for not voting,

Mi .

[−1, 1].

Voters observe a noisy signal

about the level of the ne for not voting, and hence each voter has a dierent perceived ne (Mi

= M +εi ).

The voter's problem can be conceptualized as a two stage maximization. First, she evaluates the costs and benets of abstention. If she decides to vote, she'll choose between the two candidates based on which has a higher probability of being closer to her own ideological position, given her information set:

M ax

ti [di − Ci (vi ; yi , Ωi )] − (1 − ti )Mi

(1)

ti ∈{0,1},vi ∈{L,R}

where, ti

∈ {0, 1} denotes the turnout decision, vi ∈ {L, R} is the candidate choice, and Ci (vi ; yi , Ωi ) is

the utility loss associated with making a voting mistake by choosing candidate set

vi ,

given the information

Ωi . There is a continuum of voters of measure 1, hence no voter can be pivotal. This means that all the

costs and benets of voting are realized at the time of the election. Each citizen has an ideological position

7

yi ∈ [−1, 1],

which she uses to evaluate candidate

y

6

based on a utility function of the form:

ui = −(yi − y)2

(2)

Uncertainty in the candidate's ideological position generates the possibility of making a mistake by voting for the wrong candidate, which carries a utility loss. Given the information held by citizen and her ideological position by:

(yi ),

the voter's expected utility loss of voting for candidate

L

i (Ωi )

will be given

7

Ci (L ; yi , Ωi ) = E [1 {ui (yL ) < ui (yR )} · (ui (yR ) − ui (yL )) | Ωi ]

(3)

Note that Equation (3) will be greater than zero only when a voting mistake occurs, i.e. vote for candidate

L

is cast while she should have voted for

R (ui (yL ) < ui (yR )).

informed, she will always vote for the right candidate, and thus

when a

If a voter is perfectly

Ci (L ; yi , Ωi ) = Ci (R ; yi , Ωi ) = 0.

Voters

who hold less information have a higher probability of making a voting mistake, and hence will face a utility loss. Working backwards through the voter's problem from Equation (1), I characterize the candidate choice:

vi∗ (yi , Ωi ) =

if

Ci (R ; yi , Ωi ) = Ci (L ; yi , Ωi ),

    L

if Ci (L ; yi , Ωi ) < Ci (R ; yi , Ωi )

   R

if Ci (R ; yi , Ωi ) < Ci (L ; yi , Ωi )

the citizen randomizes between the two options.

i

Simplifying the

L if f

(the expression is symmetric for the

Ci (L ; yi , Ωi ) − Ci (R ; yi , Ωi ) < 0

(5)

expression above, we have that citizen vote for candidate

(4)

will vote for candidate

R):

E [ui (yL ) − ui (yR ) | Ωi ] > 0 Plugging in the utility function from (2) in (4), it is straightforward to derive

6

τi ,

the ideological cut-o

Alvarez (1998) provides a justication of the use of a quadratic functional form in the context of a electoral environment

with uncertainty about the candidates' policy positions. All of the results in this section also hold for more general singleβ peaked pay-o functions of the form: ui = −|yi − y| , β ≥ 1

7

The expression for the utility loss of voting for candidate

R

8

is symmetric.

that completely determines the candidate choice:

 2  E yR − yL2 | Ωi τi = 2E [yR − yL | Ωi ] The optimal voting rule for voter i, position

vi∗ (yi , Ωi ) will thus be completely specied by the voter's ideological

(yi ), and her ideological cut-o (τi ).

and for candidate

R if f yi > τi .

(6)

This is, voter

i will choose to vote for candidate L if f yi < τi ,

If the information set held by citizen

i

is

Ωi = (yL , yR ),

be exactly the midpoint between the two ideological position of the candidates:

Ωi = F (.), the cut-o is zero.

Given the assumption on the distribution of

F (.), τi

τi =

the cut-o will

yL +yR , and when 2

will be symmetric, with

mean zero. Note that the previous formulation will always lead to sincere voting. Unlike other theoretical settings (Feddersen and Peserdorfer, 1996), there is no strategic voting in this model. Using this result, we can characterize the turnout decision, given that the utility loss of voting is

Ci (yi , Ωi ) ≡ Ci (vi∗ (yi , Ωi )):

ti (yi , Ωi ) =

    1

if Ci (yi , Ωi ) − di ≤ Mi

   0

if Ci (yi , Ωi ) − di > Mi

(7)

The model predicts that an exogenous change in the cost of abstention (Mi ) will cause lower turnout. Further, the composition of the electorate will change over the three dimensions of heterogeneity between voters. Upon a reduction in

Mi ,

1. Have an ideology closer to

we will observe that citizens who abstain will more likely be those who:

τi :

Note that the utility loss of voting

Ci (yi , Ωi ) peaks at the ideological cuto τi .

Intuitively, the closer

a citizen is to her ideological cut-o, the more likely she is to make a voting mistake for any pair

(yL , yR ).

Hence, the payo loss associated with voting is higher for voters closer to

Given that

τi

τi . 8

is symmetric with mean zero, voters with centrist ideology will face a higher expected

loss from voting, and thus (in expectation) will be in the margin.

2. Have a lower subjective benet of voting of

8

Ωi

(di ). di

follows a uniform distribution, which is independent

(and thus of the utility loss of voting). From Equation (7), it is clear that a lower

Take for example any two generic citizens,

(yL , yR ) for which both citizens long as ui (.) is strictly concave.

j

and

k

with ideological positions

make a voting mistakes by voting for

9

L,

yj < yk < τ .

di

implies a

For any candidate positions

the associated payo loss is higher for citizen

k

as

higher net cost of voting, and thus, for any ideology or information set the probability of voting is lower.

3. Have less information:

Ci (yi , Ωi )

is decreasing on

Ωi ,

implying that less informed people are more

likely to make a voting mistake, and hence have a higher expected utility loss of voting for any given

yi .

The predictions of the model will be tested in Section 5.

3

Institutional Background

Since 1933, voting in Perú, as in most Latin American countries, is mandatory for all citizens between 18 and 70 years old. Abstention is penalized with civil disenfranchisement. Citizens who are unable to show proof of voting (an ocial stamp in the ID card) are denied any public or private service for which ocial identication is required.

9

In order to get back full citizenship, ne has to be paid in the National Bank,

and once the payment is done, the bank ocial places a stamp on the ID card. De facto, enforcement is mixed: it is usually stronger at banks, the judiciary, public notary, passport or driver license oces, or the public registry, while it has been softer at lower levels of government or basic service delivery, such as

10

police stations, municipalities, birth or death registry, social programs, among others.

The high level of the ne for abstention has historically led to high turnout levels. For example, in the June 2006 presidential election, 87.7% of the population voted, while in the local elections held in

11

2002, turnout was 83.1%.

Until 2006 the ne was S/.144 (~US$50), which represented about 26% of the

minimum ocial monthly wage. That year, Congress started discussing whether or not to change voting to a voluntary regime, with strong proponents on both sides. A nal agreement was reached in August 2006, when they passed a law according to which voting was still mandatory, but the ne was reduced for

9

Civil disenfranchisement implies an eective ban to get ocial certicates from the national registrar, take part on any

judiciary or administrative process, sign a contract, take a government job, get a passport, be part of the social security system, get a driver's license, or in general identify themselves ocially (which includes doing any transaction in a bank, such as cashing a check). Not having voted in an election does not restrict the right to vote in any other election.

10

In Perú, the ocial ID card is used for voting, thus most of the adult population is registered to vote. Votes can only be

cast in person on the election day, and citizens can only vote in the district where they are registered. In case someone lives in a district dierent than the one where she is registered, the voter will be subject to the ne level of the later. Voting by mail or other mechanism for remote or delayed voting is non-existent.

11

The mild enforcement is reected in the percentage of the population that actually pays the nes. For example, in the

November, 2006 local elections, out of the 12.4% of abstainers, about 14.1% of them had paid their nes as of July, 2010. In urban districts, this proportion is higher. For example, in the region of Lima, the abstention rate was 11.87%, and out of the abstainers, 17.9% paid the ne as of July, 2010.

10

everyone, and this reduction was larger for citizens registered in the poorest districts. The poverty level of the district was determined based on a ranking generated by the national statistical institute (INEI). Overall, districts were classied in one of three poverty (and ne) levels:

abstainers

registered in non-poor districts (n=184) are subject to a ne of S/.72 (~US$25); those in poor districts (n=793), the ne was reduced to S/.36 (~US$12.5), while voters in extremely poor municipalities (n=852), it amounts to S/.18 (~US$6). Importantly, no major news outlet reported the changes in the ne, and no campaigns were conducted

12

to spread the information about the new ne structure.

In fact, most of the population is still uninformed

about the new ne structure, as will be shown in Section 4. The fact that electoral laws changed, and that very few people were informed about it, presented a unique opportunity to explore the eects of (dis-)incentives to vote on voter behavior, and to test the predictions of the model.

4

Experimental Design and the Data

The goal of the empirical analysis is to identify the eects of changes in the cost of abstention on turnout by comparing voters exposed to dierent levels of the ne. One way to address the question would be to compare voting behavior of citizens in districts with dierent level of the ne for abstention, however this strategy would face two major challenges. On one hand, the fact that voters are not informed about the new levels of the nes imply that the researcher would not observe any variation in the independent variable of interest (the perceived ne).

Even if this variation were observable, it would probably be

correlated with other relevant variables, such as information, or interest in politics, which leads to a downward bias in the estimated eects. Additionally, it would be impossible to disentangle the eect of district specic characteristics, such as the electoral context (candidates running for oce, availability of polling stations, etc.) or poverty level, from the one of the dierent ne levels. For example, given the well documented association between wealth and turnout (Matsusaka, 1995, Perea, 2002, Frey, 1971), if we compared turnout in the average poor district with the one in the average non-poor district, we would

12

El Comercio,

th the major newspaper in the country only published two very short articles about this on July 6 (when 20th , 2006 (the day after local elections were held). Additionally, the

the law was still under debate) and on November

government oces in charge of publicizing electoral rules and providing electoral information, the ONPE (National Oce of Electoral Processes) and the JNE (Electoral Jury), get a share of their annual revenues from the collection of these nes and use turnout as a performance indicator, hence they did not have incentives to publicize the new law. In 2004, the share of the budget of the ONPE coming from collection of nes was 24.5%, while for the JNE, this share was 30.5%. Informal conversations with government ocials at the time indicated that the heads of both oces were committed to keeping high turnout in elections, so no eorts were made to publicize the law.

11

not be able to know whether the dierences are due to wealth or dierences in the ne. One way to isolate the eect of district specic characteristics from dierent levels of the ne would be to compare districts that are just on the threshold between being classied as poor and non-poor, or between being extremely poor and poor. In expectation, districts that are just on both sides of each of the thresholds should be comparable in all relevant characteristics. Further, if we believe that the monetary cost of abstention matters at all for the decision to vote, had voters been informed about the reduction in the ne, we would observe a decrease in turnout in the elections that took place after the reduction in nes came in eect, i.e. the November, 2006 and October, 2010 local elections. On the other hand, this change in turnout would not be present in the elections that took place before the law came into eect, for example in the 2002 local elections. In Figure 1, I show the results of a regression discontinuity analysis for the last three local elections

13

(2002, 2006, 2010).

For each of these elections, I rank districts from richest to poorest, plotting their

turnout, and tting a cubic polynomial for municipalities in each of the three poverty levels. The vertical lines indicate the thresholds at which a district is categorized as Non-Poor, Poor, or Extremely Poor. It is clear from the gure that there is no statistically signicant dierence in turnout between districts located at each side of the thresholds, as one would expect if the population were informed about the new levels of the ne. Importantly, the dierence in turnout is not statistically signicant for the elections before the law came into eect (2002), or for the ones that took place after the nes were reduced (2006 and 2010). The results presented in Figure 1 can be interpreted as evidence that changes in the monetary cost of not voting do not inuence the decision to go to the polls. Alternatively, it could mean that the cost matters for turnout, but in this case, voters were not informed about the change in the ne for not voting. Voters decide whether or not to go to the polls based on their perceived level of the ne. If these beliefs are still aligned with the old level of the ne (which did not vary across poverty categories), we shouldn't expect to see a dierence at each threshold.

4.1

Experimental Design and Sample

I follow the later interpretation of the results from Figure 1. I designed the experiment to generate within district, individual level variation in the cost of abstention. I do this by randomly providing information

13

For the 2010 elections, I exclude from the sample the 10 districts where I run the experiment to allow a cleaner comparison.

The plots for 2002 and 2006 include these districts, but the basic results remain the same if I exclude them. The regression versions of the Figure are available upon request.

12

about the actual levels of the ne to voters in 10 districts in the Region of Lima just before the municipal

14

elections of October, 2010.

After the election, I re-interviewed all the subjects in the treatment and

control groups, and among other information, I collected objective measures of turnout. The advantage of this strategy is that I can compare the voting behavior of people who believe that the nes are still in their old level (control group) with those whose information set has been updated by the treatment. Within each district, I randomly sampled villages (in rural areas) or neighborhoods (in urban areas), and within each village we interviewed individuals eligible to vote (between 18 and 70 years old) from a random sample of households.

15

By clustering the randomization at the village level, I can make

comparisons within villages, thus isolating the eect of any district (and village) specic characteristic. The unit of observation is the individual, but the treatment status is determined at the household level, hence in the empirical analysis I allow for arbitrary correlation of the errors within the household by clustering at that level. One of the objectives of the experiment is to be able to estimate the cost elasticity of voting. One might be concerned that having only two points in the demand curve is not enough to estimate an elasticity. My sampling strategy allows me to make comparisons between people who face more than two dierent levels of the ne for abstention.

Specically, I choose districts in a single region (Lima) which were as

close as possible to each of the two thresholds that dene the poverty level of the district.

Eectively,

I can compare both, people in the treatment and control group within a district, and voters in districts with dierent ne levels. Overall, I chose 10 districts such that I could have roughly the same number of observations at each side of the two thresholds. Table 1 provides descriptive statistics about the districts from which the sample was drawn, while Figure 3 shows the location of the districts in a map, indicating their poverty category. The baseline interview took place between one and four weeks before the municipal elections of October

3rd ,

2010. I included questions regarding household characteristics, composition and expenditures. I also

asked about basic demographics, political preferences, policy priorities for the district, knowledge about the current electoral process, and main sources of political information, past voting, and usage of public services. Importantly, I asked everyone whether they knew if there were consequences for not voting, if the respondent answered that there was a ne, I asked for the amount of the ne. At the end of the interview,

14

In municipal elections, voters elect the mayor for the district, the mayor for the province, and the regional president.

These are the three sub-national levels of government. In this paper, I use interchangeably district and municipality.

15

In the national census, the villages are called centro poblado.

13

the enumerator provided the treatment. If the household was chosen to be part of the treatment group, the enumerator read a script informing the respondent about the level of the ne in eect in the district where she was registered to vote.

16

In

order to reinforce the message, the enumerator showed a copy of the ocial newspaper where the law was published, also she gave the respondent a ier with the exact text on the script, and the message was sent by text messages during the week before the elections (for those who had a cell-phone). In order to avoid any salience eect, individuals in the control group received a reminder that voting is mandatory and that there is a ne for not voting (without mentioning anything about the amount of the ne).

17

Respondents

in the control group also received a ier repeating the script, and those with a cell phone received text message reminders. The follow-up survey was gathered between one and three weeks after the election. The main variable collected in the survey was whether or not each respondent voted in the elections.

I measured voting

through a self reported variable, but also gathered an objective measure of voting by asking each respondent

18

to show their ID card, where each enumerator conrmed if it had the ocial stamp or not.

Among the

2,276 respondents in the follow-up, only 5 of them refused to tell the enumerator whether they voted or not, while for 1,526 of them we were able to objectively measure the voting through their ID cards. There does not seem to be a tendency to lie about the vote. Out of those for whom I have the self reported and objective measures of voting, 6 respondents reported that they did not vote, and their ID cards had the ocial stamp, while the opposite happened in only 7 cases. Given the low lying rate, in order to maximize

16

Along the questionnaire, we asked every respondent in what district is she registered to vote. This information was cross

checked with the subject's ID. Every enumerator had a list of the 1,834 districts in the country, with their corresponding poverty level, so they were able to tell each respondent the exact level of the ne applicable the district where she was registered. The script for the treatment group was as follows:

Dear Sir/Madam,. On August 2006, Congress passed a law in which the nes for not voting were reduced (Ley No. 28859). According to this law, those who do not vote are no longer subject to a ne of S/.144, but the nes are now lower for everyone, and they vary according to the poverty level of the district where you vote. According to the information that you just provided me, if you do not vote in the upcoming elections you will be subject to a ne of S/.(AMOUNT IN THE DISTRICT WHERE SHE'S REGISTERED). 17

The exact script for the control group was as follows:

Dear Sir/Madam, In Perú, voting is mandatory by law, and not voting is subject to a sanction that implies a ne. 18

The option to pay the ne and get the ocial stamp in the ID card is only available once the full voting record is

centralized, which usually happens more than a month after the elections, and hence the only way in which the respondents could have the stamp was by having voted.

14

my sample size I dene my turnout variable based on the objective measure of voting for those who showed their ID, while I take the self reported values for those who did not. In the empirical analysis in the next section I show that the results are not aected by my choice. The survey also included questions about political preferences, information about the political process and sources of information, and a series of questions about vote buying.

4.2

Descriptive Statistics

Overall, I interviewed in the baseline and follow-up surveys 2,276 individuals from 1,668 households.

I

provide the descriptive statistics for the balanced sample of respondents in Table 1. 23% of the sample is registered to vote in an extremely poor district, 38.8% votes in a poor district, and the remaining 37% does so in a non-poor district. 42% of the sample is composed by male respondents, who are on average 40 years old, have about 9.6 years of education, and spend S/.255.1 (~US$94) per capita per month. The ideological position of the population is highly concentrated in the center, with 66.7% of the respondents placing themselves in the center of the political spectrum, while 8.3% are in the left and and the remaining 25.1% in the right.

19

Very few people (8.2%) declare themselves to be very interested in

politics, while 46.8% are somewhat interested, and 45.1% are not interested at all. The small involvement in politics is also reected in small proportion of people who declare to be very interested in the results or the campaign of the current election (39.9% and 10.5%, respectively). 44.3% of the sample reports to be somewhat interested in the results of the election, and 55.6% are somewhat interested in the campaign. Finally, 15.3% and 33.9% are not interested in the results or the campaign, respectively. It is important to note that, none of these questions were placed one after another, but rather as separate as possible, an most of them were asked in dierent modules of the questionnaire in order to avoid conrmatory bias in the responses. Political knowledge and information is measured in several ways. On the one hand, I included open questions asking respondents to name all the candidates and parties running in the election for the mu-

19

These measures are taken from a scale with ve values, where 1 represents the extreme left, and 5 the extreme right. I

dene centrists as those place themselves at 2, 3, or 4. The second measure of ideology that I use is an aggregation of several measures of policy preferences. In the survey, I asked respondents to name (in order) the rst ve policies that she would implement if she were elected mayor of the district.

This was an open question, and the enumerators had to place their

answers in one of twenty eight policy categories. For each of these categories, the policy preferences are ordered from not mentioned (zero) to most preferred (ve). I aggregate these questions by taking the rst principal component, and dividing the sample into quintiles. The center is dened by those in the quintiles 2, 3, and 4, while the rst and fth quintiles dene the ideological extremes.

15

nicipality where she is registered to vote. In order to get a uniform measure of knowledge, I express the knowledge indices as the ratio of the number of candidates (and/or parties) that the respondent is able to name, divided by the total number of candidates (and/or parties) running in the district's election. On average, respondents are able to name 38.8% of the candidates and 29% of the parties running. Additionally, I include questions about the political process in general. I asked 17 questions about knowledge of

20

the political structure of the country, and electoral rules.

On average, respondents were able to get right

9.3 questions (54.7%). Table 3 provides descriptive statistics for both the treatment and control group, showing that there are no statistically signicant dierences by treatment status in the relevant variables.

21

Even though there

wasn't a lot of time between the baseline and follow-up surveys, we weren't able to track down about 13% of the households from the baseline survey, which represents 19.8% of the respondents interviewed in the baseline. Table 17 in the Appendix shows the balance of variables between those who we were able to track and the ones that were only in the baseline survey. Among attrited individuals, there were slightly more men than women, and more people who showed interest in politics. Overall, the sample seems to be balanced. The main variable of interest is the perceived ne for abstention.

22

Given that the treatment was

randomly allocated, I should observe that the perceived nes are balanced between the treatment and control groups, within each poverty category. One concern is that some people could know the current level of the ne before the treatment was provided. If that were the case, still the pre-treatment levels of the perceived ne should be balanced between the treatment and the control groups, so this should not imply a bias in my estimates. Figure 2 shows the distribution in the baseline and follow-up surveys, for the control and treatment group, and by the poverty level of the district where each respondent is registered to vote. In each graph,

20

The questions include information about the length of the term for congressmen, president and mayor, whether it is

possible to be reelected, the ocial voting age, institutions in charge of the elections, ID cards, etc.

21

Table 18 shows the balance between the treatment and control groups when splitting the sample by poverty level of the

district where each respondent is registered to vote. Here we also see that the dierences between treatment and control are not signicant within each of the poverty levels.

The only variable that seem to be systematically unbalanced is the

proportion of voters who are in the left. The control group seems to have a higher proportion of leftists than the treatment group.

22

The question was structured in the following way: First, I asked whether the respondent knew what were the consequences

of not voting. If among the answers, the respondent mentioned a ne, I asked her if she knew how much was it. For people who did not mentioned the ne among the answers, I assume that she perceive that it is S/.0.

Also, if the respondent

mentioned a ne among the consequences of abstention, but did not remember the exact amount, I asked her to place the ne in a range, where each of the ranges provided include the new levels of the ne. For voters who chose one of the ranges, I use the median of each range as their perceived ne.

16

23

the vertical line represents that actual level of the ne.

Importantly, in the baseline survey, the average

respondent reports that the ne for not voting is S/.122.29, which is very close to its level before August, 2006. This conrms that the majority of the population was not informed about the change in the voting laws. There is signicant dispersion in the data, ranging from people who think that voting is voluntary (no ne), to those who think that the nes are around S/.300. The distributions of these perceptions do not dier by treatment status within each poverty level. Panel A of Table 4 shows the mean perceived

24

ne in each of the groups, as well as the t-tests for dierences in means.

The information treatment had an eect on the respondent's perceived ne, however there seems like not only those in the treatment group learned that the nes for not voting had decreased. For example, respondents registered in non-poor districts who received the treatment reported in the follow-up survey that the ne for not voting is S/.66.77, while those in the control group have an average perceived ne of S/.90, which is signicantly lower than the S/.126 reported in the baseline survey. The dierence between treatment and control groups among voters from non-poor districts is statistically signicant. For people voting in poor districts, I nd a similar pattern, where the distribution of perceived nes clearly moves to the left for both the treatment and the control group, but the former is centered at S/.42, which is close to the actual S/.36 stipulated for this group, while the control group reports on average that the ne is at S/.71, and this is signicantly dierent from treatment group.

Voters from extremely poor districts

learn more often about the new levels of the ne. While the treatment group reports a perceived ne of S/.19, the mean for control group is S/.36. This is also apparent from the Figure, where we see that the distribution of perceived nes shift to the left, for both the treatment and control groups. Overall, the treatment had the desired eect of informing the population about the new level of the nes, however the control group also learned about the new nes, and this is especially true for people voting in extremely poor districts. Among the sampled respondents, turnout is higher than the one observed in ocial statistics (see: Table 1). As Panel B in Table 4 shows, 94.2% of the respondents in my sample voted in the October 2010 elections. The eective reduction in the cost of not voting led to lower turnout. On average, respondents in the treatment group were 3.1% less likely to show up to vote the day of the elections. This result can

23

In the left panel, for the baseline survey, the vertical line represents the old level of the ne (S/.144), while in the graphs

in the right, the lines are set at the new levels of the ne: S/.72 for voters in non-poor districts, S/.36 for those in poor districts, and S/.18 for voters in extremely poor districts.

24

These results represent the direct eect of the treatment on the perceived nes, i.e. the rst stage of the regressions

without controls.

17

be interpreted as a reduced form eect, or the direct eect of the treatment on turnout. The magnitude of this eect also depends on how much did the nes decreased. For example, in non-poor districts the nes we ocially cut in half, from S/.144 to S/.72, however the dierence that we observe between respondents in the treatment and control group in the follow-up is of S/.24 (S/.66 and S/.90, respectively).

This

change leads to a dierence of 2.1 percentage points in turnout. Likewise, among voters in poor districts, the dierence in nes between treatment and control groups is of S/.29, which leads to 5.1% dierence in turnout. Finally, among voters in extremely poor districts, the perceived nes are S/.17 lower in the treatment than in the control group, which leads to a 1% dierence in turnout (not signicant). This low and non-signicant eect for the extremely poor is not surprising, since the treatment had no eect on the perceived nes for this group. Overall, the perceived nes for the extremely poor were on average lower for everyone, and as such, I do observe that in these districts, the average turnout is at least 2 percentage points lower than in the control group in poor and non-poor districts (93.5% versus 96.7% and 95.9%, respectively). Given that the experiment did not aect the perceived nes for the extreme poor, I drop them for all the subsequent analysis.

25

Summarizing, the descriptive data shown above supports the basic hypothesis that a reduction in the nes for not voting lead to lower turnout.

Next section outlines a more formal framework to test the

predictions of the model presented in Section 2.

5

Empirical Strategy and Results

5.1

Basic Facts

The empirical strategy implemented to test the predictions of the theoretical model from Section 2 follows from the experimental design.

The rst step will be to test whether a higher reduction in the nes

for abstention caused lower turnout.

For this, I compare turnout in the control group, for whom the

information treatment did not update their beliefs about the ne for not voting, with the those in the treatment group, whose information set was updated by the treatment, and thus their perceived nes should be lower. Formally, this test implies regressing turnout on the change in the perceived nes reported in the the

25

I have also run all the tables below including the extreme poor, and they are available upon request. All of the patterns

and main results remain unchanged.

18

baseline and the follow-up surveys. However, since some voters in the control group learned that the the nes for not voting were reduced and some treatment respondents did not comply with the treatment, this approach will likely yield biased estimates. My strategy exploits the exogenous variation provided by the treatment status in order to identify the eects of a change in the perceived nes on turnout. The local average treatment eect identied from the instrumental variables regressions will thus estimate the interest parameter for voters whose beliefs about the nes for not voting were updated. The rst part of the empirical analysis looks at the direct eect of the treatment on turnout.

The

reduced form equation is given by:

V oteij = α + β1 N onP oorij ∗ T reatij + β2 P oorij ∗ T reatij + β3 P oorij + γXij + δk + νij , where October

V oteij

3rd ,

is an indicator of whether voter i, registered to vote in district

2010. The treatment status is given by the indicator variable

(8)

j , voted in the election of

T reatij .

Given that there are

two distinct treatment groups depending on the poverty level of the district where voter

i

is registered to

vote (and thus two instruments), in all the regressions I separate the eect of the dierent treatment levels by interacting the treatment dummy with the poverty level of the district. of 1 when subject

i

N onP oorij

is registered to vote in a Non-Poor district and zero otherwise, while

takes the value

P oorij

indicate

whether the voter is from a Poor district. The inclusion of the dummies indicating the level of poverty of the district where voting (N oP oorij being the excluded category) allows to restrict the comparison to treatment and control units within the same level of the ne. Finally, I include some relevant controls that are likely to aect voting, such as age, log(Per Capita Expenditures), education and gender. These variables are included in the matrix

Xij .

Finally,

δk

denotes a xed eect at the level of the is the village

where interview took place (where the respondent lives), and

νij

is a random error term.

It is not straight forward that we should expect a reduction in the ne for not voting to cause lower turnout. Gerber et al. (2003) shows that voting is habit forming, and voting in one election makes voters signicantly more likely to vote in the next election. In the Peruvian context, where mandatory voting has been in place for a long time, and turnout is consistently high, it could be the case that the habit eect is stronger that the monetary eect. Table 5 presents the reduced form estimates of the eects of the treatment on turnout. Overall, the monetary eect seems to dominate the habit eect. Treated voters in non-poor municipalities are 2.7 percentage points less likely to vote than the controls in this poverty

19

category (Column 1). Likewise, people who were informed about the reduction in nes from S/.144 to S/.36 (those in poor municipalities) showed up in the polling station 5.2 percentage points less often (Column 2). Pooling voters does not aect the magnitude of signicance of the results (Column 3). All the regressions shown include controls, village xed eects, and the standard errors are clustered at the

26

household level.

These results are very similar to the descriptive statistics shown in Panel A of Table 4.

Voters who were informed about a larger reduction in the ne through the treatment are less likely to vote. While in non-poor districts the ne was reduced in half, in poor districts it was cut to one fourth of its original level. This is reected in the estimated eects of the treatment on turnout. The coecient for the former is roughly double the one for voters from non-poor districts. The rst stage regression in the instrumental variable approach measures the eect of the treatment on the change in the perceived ne, and is given by:

4F ineij = α + β1 N onP oorij ∗ T reatij + β2 P oorij ∗ T reatij + β3 P oorij + γXij + δk + νij

, where

4F ineij = (F ine2 −F ine1 )ij

and baseline surveys for individual

i

(9)

represents the change in the perceived ne between the follow-up

who votes in district

j.

In this case the coecients

β1

and

β2

tell

us the dierence in the change in the perceived ne between the treatment and the control group. This comparison is made within the same poverty level of the district registered and between people who were interviewed in the same village. The results from the rst stage regression are displayed in Table 6. Column (1) present the results for voters registered in Non-Poor municipalities, and we see that voters in the treatment group registered in non-poor municipalities perceive that the ne for abstention was reduced in S/.18.8, compared to people in the control group. Similarly, the treatment eect for voters in poor districts is a reduction in the perceived ne of S/.30.5. Column (3) pools the results. Overall the results from Table 6 provide a strong rst stage for my IV strategy, with an F-statistic for the excluded instruments of 28.7 in the pooled specication. In the second stage, I look at the eect of the changes in the perceived ne, instrumented by the treatment status, on turnout. The regression equation is displayed in Equation (10).

26

The results are very similar when I do not include controls, or village xed eects.

20

V oteij = α + β1 4F ineij + β2 P oorij + γXij + δk + ij β1

(10)

is the estimated local average treatment eect of a change of S/.1 in the nes for not voting on the

likelihood of voting for those whose information was updated due to the treatment. The main identifying assumption in this regression is that the treatment only aects turnout through the change in the perceived nes, and hence the treatment is uncorrelated with

ij .

The fact that the treatment was randomized, and

that we see that the main variables in the analysis are not statistically dierent from each other between the treatment and control groups support this assumption. The instrumental variables results are presented in Table 7. I nd that an exogenous decrease in the perceived nes for not voting cause less people to attend to the polls. Column (1) shows that a reduction in the ne of S/.1 leads to a signicant decrease in the likelihood of voting of 0.14 percentage points among Non-Poor voters, while in Column (2) the estimates suggest that the reduction for voters in Poor districts is of 0.14 percentage points. Pooling the results, the average voter in my sample has a 0.16 percentage points lower likelihood of attending to the pools. The average voter, who perceived that the nes were reduced in S/.56.65 (45.7% from her initial perception of S/.124), has a 9.59% lower probability of voting in the elections, this is a reduction in turnout from 94.5% to 85.4%. With this information, I can infer that the price elasticity of voting is of about -0.21. Extrapolating these results to the whole population, we could say that driving the nes to zero could drive turnout to 74.7%, a level comparable to the one in presidential elections in the US or France, where voluntary voting is in place. To put these results in context with the previous evidence, Gerber et al. (2008) nds that reminders to vote emphasizing social pressure messages cause an increase in turnout between 4.8 and 8.1 percentage points. In my experiment, a reduction of S/.56.7 (~US$20) leads to a reduction in turnout of 9.1 percentage points.

27

Stretching the external validity of both experiments, we could say that people is willing to pay

about US$20 to keep their turnout information from being shared with friends or household members. In Table 8 I show the heterogeneity of the eects of the reduction in the ne on voting by several demographic characteristics. Overall, I nd the eect is constant between people of dierent ages, educational levels, expenditure levels, or for those who were aected by political violence. However, women

27

Gerber et al. (2008) found that sending mailings to people informing recipients that who votes is public information and

listing the recent voting record of each registered voter in the household had an eect of 4.8% on turnout. Listing not only the household's voting records but also the voting records of those living nearby led a 8.1% higher turnout.

21

seem to be signicantly more sensitive to changes in the perceived nes. Importantly, consistent with the constant elasticity found, poor voters are not more likely to respond to changes in the nes for not voting.

5.2

Robustness and Validity Checks

One potential concern with the interpretation of my result is that the elasticity of voting with respect to the cost might not be constant, and thus the extrapolation of my results will not be valid. When I compute the elasticity using the results from the separate estimations, I nd that for Non-poor voters it is -0.18, while for voters in Poor districts it is -0.21, and they are not statistically dierent from each other. This evidence supports the idea of a constant price elasticity. This result is in itself interesting for Peruvian representation system, since the largest reduction in the ne took place in the poorest districts, and hence it is lowering the representation from voters in poorest districts. Additionally, it is important to note that when I split the sample I am only using one instrument in each regression, rather than three. Still, the rst stage regressions have very strong predictive power, with F-statistics ranging between 14.7 and 41.03, which reinforces the idea that the previous results are not driven by the over identication of the rst stage regression. An important robustness check in place regards to the measurement of my dependent variable.

As

mentioned in the previous section, the dependent variable is constructed based on self reported and objective measures of voting. I run the main specication with both variables, with the available sample, and with matching sample sizes in Table 9. The results are very similar across the dierent specications and samples. In the sample for which I have both self reported and objective voting measures, turnout is higher, since more people who reported not voting were lees likely to show their ID cards. In this sample, the results using the self reported measure of voting is attenuated but still large and signicant. Table 10 presents a validity test for the eect of the treatment on the turnout. If the treatment did aect the perceptions about the magnitude of the nes, it should have aected turnout in 2010, but had no way of aecting past behavior.

Table 10 shows the results of running the same specications as in

Table 7, but using turnout in 2006 as a dependent variable. None of the specications yield statistically signicant eects of the change in the perceived nes on turnout in 2006. Also, it is reassuring to see that the coecients are very close to zero.

22

5.3

Ideological Position

The model predicts that voters with centrist ideological positions are more likely to abstain upon a reduction in the ne, since they could make a voting mistake .

Ideological positions are not randomly

distributed in the population, but the random variation in the cost of not voting provided by the treatment ensures that within each ideological position, I am able to observe voters who face dierent costs of abstention. Hence, even though I can not make causal inference from the level eect of the ideological position on turnout, I can causally interpret the variation in the cost of abstention induced by the treatment

within each ideological position category. This is, the interactive term between the change in the perceived nes and the ideological position, instrumented by the treatment dummies and their interactions, provide causal evidence of whether people with centrist ideologies are the more likely to react to a change in the cost of abstention, as the model predicts. More precisely, assuming that there are three ideological positions, left, center and right, which are denoted by

Pijl (l = 1, 2, 3),

the eect of the reduction in nes

on turnout, for each ideological position will be identied from equation (11).

V oteij = α+

3 X n=1

βn 4F ineij Pijn +

3 X n=1

βn1 Pijn P oorij +

3 X

βn2 Pijn N onP oorij +β10 P oorij +γXij +δk +ij

(11)

n=1

In order to compare people within the same ne level, the model in Equation (11) only allows comparisons within each poverty category by interacting all the relevant coecients with the poverty level dummies. The only eects that I am constraining to be constant across poverty categories are the control variables (Xij ). The coecients of interest in this case are going to be

βn ,

and if the predictions of the

model hold, we should observe that the coecients associated with the interaction between the indicators of political extremes with the change in the perceived nes will be zero (β1 and

β3 ).

On the other hand,

the coecient testing for the eects of changes in ne on voting among centrists voters (β2 ) should be positive, meaning that a larger decrease (increase) in the perceived ne causes lower (higher) turnout. Ideology is not unidimentional, and thus I measure it in two dierent ways to capture a broader range of ideological distributions. First, I use a self reported measure in which respondents placed themselves in a scale ranging from extreme left (1) to extreme right (5). I take the categories in the middle (2, 3 and 4) to represent the political center. The second measure is an aggregation of several measures of policy preferences. In the survey, I asked respondents to name (in order) the rst ve policies that she would implement if she were elected mayor of the district. This was an open question, and the enumerators had

23

to place their answers in one of twenty eight policy categories. For each of these categories, the policy preferences are ordered from not mentioned (zero) to most preferred (ve). I aggregate these questions by taking the rst principal component, and dividing the sample into quintiles. The center is dened by those

28

in the quintiles 2, 3, and 4, while the rst and fth quintiles dene the ideological extremes.

The Policy

Extreme 1 is related to preference for public goods, such as health and education infrastructure, roads, accessibility, etc. On the other hand, the Policy Extreme 2 is associated to public goods which are more easily appropriated by a private party (club goods), such as youth labor training, security, promotion of private investment, etc. It is important to mention that the questions that dene the ideological position of each voter were asked in the baseline survey, before the treatment was administered, so I am able to take them as predetermined. Table 11 shows the results of the specication laid out in Equation (11).

In column (1) I use the

standard measure of political ideology, and nd that the bulk of the eect of the changes in the ne on turnout observed in Table 7 comes from those voters who place themselves in the political center. On the other hand, voters on both political extremes seem to be insensitive to changes in the price of not voting. The results shown in column (2), using my second measure of ideological position are even more stark. Voters in the the second through fourth quintiles of the policy preference scale bear the whole eect of changes in the ne for not voting, while voters in the political extremes show eects close to zero and statistically insignicant. Overall, the results from Table 11 are consistent with the rst prediction of the model, and show that people in the political extremes are less likely to respond to incentives to not vote. This result has important implications in terms of how to structure the incentives to vote, and its potential eects on political competition and social conict.

If the political center reduced, we would

observe parties that are also bunched in the extremes, which could lead to higher degrees of social conict.

5.4

Interest in Politics / Subjective Value of Voting

Voters with a higher subjective value of voting (di ) need lower incentives to attend to the polls, compared to those who get lower utility gains from voting. The subjective benet of voting is an unobserved individual characteristic, so I use a battery of questions to approximate it. To avoid any type of conrmatory bias in the responses, none of these questions were placed one after another, but rather as separate as possible, an most of them were asked in dierent modules of the questionnaire. I asked people about their interest

28

The coecients for each policy item loading into the PC analysis are listed in Table 21.

24

in politics, interest in the results of the current political election, and in the campaign. As I show in Table 12, voters who are more interested in politics attend to the polls regardless of the change in the perceived ne. People who reports being somewhat interested in politics are less likely to vote when the ne for abstention is reduced, however, the eect is smaller in magnitude to the one we observe for voters who are not interested in politics. Similarly, voters who are very interested in the political campaign or in the results of the election are unlikely to respond to a reduction in the ne, while people who are somewhat interested have a signicant eect, but again, lower in magnitude than those with a low interest in the campaign or in the results of the election.

5.5

Political Information

The model also predicts that

Ci (yi , Ωi )

is decreasing on

Ωi ,

which implies that less informed voters are

more likely to make a voting mistake , and hence have a higher expected cost of voting for any given

yi .

Empirically, I can test this prediction by interacting dierent measures of political information with the changed in perceived nes, always relying on the treatment status to identify the LATE. More precisely, I run the test for this prediction through the following equation:

V oteij = α+β1 4F ineij +β2 4F ineij Inf oij +β3 Inf oij ∗P oorij +β4 Inf oij ∗N onP oorij +β5 P oorij +γXij +δj +ij

(12) As before, in Equation (12) I am only comparing people within poverty categories.

Following the

model, we should observe that the eect of reductions in the cost for not voting is steeper for people who have less precise information about politician's ideological stance. On the other hand, having perfect information about the politicians means that the voter can not make a voting mistake , and thus she should vote regardless of the cost of abstention.

Following this prediction, we should expect

positive, while for people with perfect information (Inf oij

= 1), β1 + β2

β2

to be

should be equal to zero.

In Table 13 I test this hypothesis using four dierent measures of political information.

I use four

normalized indices to proxy for political knowledge. The rst three of them measure the percentage of candidates or/and parties running for oce that the voter is able to name.

Additionally, I also use a

normalized political information score, ranging from zero to one, which uses information from seventeen questions about the electoral process, including knowledge about the electoral organisms, ocial voting

25

age, reelection rules in dierent instances of the government, etc. In all four columns of Table 13, the interaction between the information indices and the changes in the perceived ne is negative and signicant, meaning that people who have higher levels of information are less likely to change their turnout decision when they learn that the ne has been reduced. Moreover, the magnitude of this coecients line up remarkably well with the predictions of the model. People who is fully informed about the candidates and/or parties running in the local election, are unaected by the changes in the ne as the coecient of the interaction osets the direct eect. Previous evidence (Ferraz and Finan, 2008;Banerjee et al., 2011, Chong et al., 2011, Pande, 2011) shows that more informed voters are more likely to hold the elected ocials accountable and less likely to elect corrupt politicians, thus it is possible that by reducing the cost of not voting, and allowing less informed voters to select out of voting, we could increase the quality of elected ocials.

5.6

Extensions: Policy Preferences, Information Acquisition and Vote Buying

Overall, the results from Tables 11, 12, and 13 are consistent with the predictions of the theoretical model, and have important implications for the design of electoral incentives. Lower nes for not voting draw a lower share of the population the polls. This is particularly important for voters who are in the center of the ideological spectrum, those who have lower subjective value of voting (or who are less interested in politics), and uninformed voters.

The natural question that follows from these results regards the

implications for the aggregation of citizen preferences in electing a government. Electoral institutions in democratic societies are designed to maximize voter's representation in the government, and ensure that policies are catered towards the interests of the majority. Mandating citizens to participate in elections impose a cost on society, and hence this additional cost could be justied if by providing incentives to vote we are able to achieve a better representation of voter's preferences. Theoretical arguments on these grounds are mixed.

Depending on the assumptions over the type of

information available to voters, dierent authors have argued that compulsory voting can be welfare increasing or decreasing. For example, Krishna and Morgan (2011) present a theoretical model showing that under voluntary voting, information aggregation holds, and mandating people to vote imposes a net cost to society. Along the same lines, Borgers (2004) reach a similar conclusion based on a model with simple private value majoritarian elections. On the other hand, Ghosal and Lockwood (2009), argue that when voters have common preferences, compulsory voting Pareto dominates voluntary participation.

26

Even though I am not able to give direct evidence supporting any of these models, I can provide suggestive evidence that help us start thinking about the extent to which dierent incentive schemes to (not) participate in elections can aect policy outcomes. One way to address this issue is to analyze whether people who have a certain type of policy preferences is more likely to respond to the incentives to (not) vote. If that is the case, a reduction of the nes for abstention will lead to the under representation of people who have these preferences, and thus the policies preferred by this group will not be enacted (assuming perfect commitment by politicians). I do this by using the policy preference questions, aggregating them into 10 categories that represent broad policy issues, and analyze whether the supporters of each type of policy are more or less likely to respond to changes in the nes for not voting. I present the results from this analysis in Table 14. Overall, I see no statistical signicance in the interactive terms between policy preferences and changes in the ne for not voting. Moreover, most of the coecients are very close to zero, suggesting that voters with particular policy preferences are not over represented among those who stop voting. Even though the coecients found are not statistically signicant, some of them are big enough in magnitude that it is worth further consideration. For example, for voters who have preferences towards policies that promote agricultural activities (i.e. water projects, investment in improved seeds, etc.) I see a negative interaction coecient which is about of the same magnitude as the average eect for the population. Hence, eect of the changes in the ne are completely oset for this group, and they are not likely to stop voting when the nes are reduced.

Similarly, people who prefer policies related to institutional improvement,

such as ghting corruption, reorganization of bureaucratic procedures, etc. are likely to vote regardless of the changes in the ne.

On the other hand, people who support policies for the educational sector,

or those who prefer policies catered towards the youth and women are more sensitive than the average voter to changes in the ne. Potentially, the reduction in the nes for not voting would lead towards more agricultural and institutional policies, while educational and youth/women projects will not be carried through by elected ocials as often. Proponents of mandatory voting argue that mandating people to vote not only increases participation, but also involves people in the political process, by for example informing themselves about the electoral race.

The model that they have in mind is one similar to the one proposed here, but it endogenizes

information acquisition (see for example, Deagan, 2011, Oliveros, 2011). For costs of not voting suciently high, abstention will drop signicantly, and people demand more political information so they are less likely

27

to make a voting mistake. In the follow-up questionnaire, I included questions assessing the level of political information held by each respondent, so I am able to see whether people who perceives the costs of not voting to be lower are more likely to not acquire political information. In Table 15 I regress the change in the dierent measures of political information on the change in the perceived nes, instrumented by the treatment status. The results presented here suggest that people who face lower costs for not voting do not acquire more political information than people who perceive lower decreases in the nes for not voting.

29

Electoral processes in developing countries are often prone to vote buying (Vicente, 2008, Vicente and Wantchekon, 2009, Finan and Schechter, 2011). Vote buying represent a net loss for society, since it tends to distort voters preferences, and it can potentially change the results of an election. In an electoral system with mandatory voting, arguably those voters that would have stayed at home in the election day, but voted because of the high cost of abstention, are potentially easier targets for vote buying.

Using the

exogenous variation in the cost of not voting, I am able to test whether if a reduction in the opportunity cost of voting increases the amount of vote buying, and aects the priced paid for each vote. I do this by using information collected in the nal section of the follow-up survey, where I asked respondents if they were oered (and if they accepted) any gift or cash from someone associated with any candidate before the election took place, and if the money or gift was given directly to the person, or indirectly in a fair o a massive giveaway. Table 16 shows the eects of the change in perceived nes (instrumented by the treatment) on whether the voter accepted money for her vote, and the amount of money accepted. As a result of a reduction of the ne, the opportunity cost of voting is higher, and thus less voters will be willing to accept a payment for their vote, and those who do accept a payment will demand a higher amount. Eectively, my treatment presents an exogenous shift in the supply of votes. The results in Column (1) show that a decrease in the ne for abstention of S/.1 leads to a 0.1% lower likelihood of accepting money for the vote. The standard error is large, but the magnitude of the eect is non negligible.

This implies a 19% reduction in vote

buying due to the reduction in the ne for not voting. Column (2) shows the eects on the amount of money received directly from a candidate or her representatives before the election. ne of S/.1 leads to an increase in the price of the vote of S/.0.03.

29

A decrease in the

This implies that for the average

These results must be taken with a grain of salt. Even though around the elections is the time when voters are more

likely to get informed about the candidates, and the political process overall, we must have in mind that the average time between surveys was short, ranging between three and ve weeks.

28

voter, who perceived that the nes were reduced in S/.56, her vote became 76% more expensive than the average S/.2.2 for what she settled before.

As a robustness check for this result, in Column (3) I use

as a dependent variable the amount of money indirectly received by the voter. If there is a negotiation between the voter and the political operator about the price of the vote, I do not expect this negotiation to aect the amount received in a massive giveaway of money or souvenirs. Indeed, I nd a statistically and economically insignicant eect. Overall, I nd that the increase in the opportunity cost of voting leads to less vote buying, and when it happens, each vote becomes more expensive.

6

Summary and Discussion

Electoral institutions that encourage citizens to vote are widespread around the world.

Most of these

institutions were introduced in a spirit of democratization, hoping to achieve better representation, and to involve the citizenship in the political process. However, encouraging citizens to attend to the polls is costly, and if the desired eects are not achieved, these institutions could induce a welfare loss. In this paper, I explore this question through a rational choice model of voter behavior with incomplete information, and test its predictions using random variation in the cost of abstention. In order to be able to empirically identify the eect of changes cost of abstention on turnout, I exploit a change in Peruvian voting laws that reduced the nes for abstention. Information about the change in the cost for abstention was not widespread, so before the last municipal elections in Perú I provided voters in the treatment group with information about the new levels of the ne for not voting, and afterward observed their voting behavior, and observe their voting decisions after the election.

The research design allows me

to generate individual level variation in the cost of abstention, thus overcoming the problems faced by previous literature trying to identify the eects of changes in the cost of voting on turnout. The estimates in the paper imply that cutting the nes for not voting in half leads to a 10 percentage point reduction in turnout. Further, the experimental design allows me to back out the elasticity of voting with respect to the cost, which I nd to be about -0.2. This elasticity represent an important parameter to evaluate the potential welfare losses of policies that aim at changing the cost of voting, such as the increase in the number of polling stations, electronic voting, etc. Consistent with the theoretical model, voters in the extremes of the political spectrum, those more interested in politics, and more informed are insensitive to changes in the cost of voting.

29

Even though there is a change in the electorate due to the reduction in the ne for not voting, this does not necesarily imply that the outcome of the election will be aected. On average, voters who stop attending to the polls due to the reduction in the ne do not seem to have dierent policy preferences than the average voter who does not respond to the change in the cost of abstention. Further, the fact that some peope do not vote as a response to the treatment does not lead them to get less involved in politics, as measured by the amount of political information they acquire. Additional results indicate that an decrease in the ne for not voting decreases the incidence of vote buying, and at the same time increases the price paid by politicians to buy votes. This last nding is consistent with an exogenous shift in the supply of votes in the market. These results presented have strong implications for the design of electoral institutions. First, voters respond to monetary incentives to attend to the polls, and the extent in which they respond is nonnegligible. Second, the experimental evidence suggests that the objectives of mandatory voting, namely, insure representation and involvement in politics does not seem to be aected by the reduction in the incentives. If this result holds when the incentives are completely eliminated, mandatory voting would lead to a welfare loss to society. Further research is needed to explore this last nding in depth.

30

References [1] Aker, Jenny, Paul Collier, and Pedro Vicente (2011). Is Information Power?

Using Cell

Phones during an Election in Mozambique. Mimeo, Oxford University.

[2] Alvarez, R. Michael (1997). Information & Elections. Ann Arbor: The University of Michi-

gan Press.

[3] Banerjee, Abhijit, Selvan Kumar, Rohini Pande and Felix Su (2010). Do Informed Voters Make Better Choices? Experimental Evidence from Urban India. Mimeo, MIT.

[4] Blais, Andre (2000).  To Vote or Not to Vote? . Pittsburgh, PA: University of Pittsburgh Press.

[5] Blais, Andre and Robert Young (1999).  Why do people vote? An experiment in rationality .

Public Choice, Vol. 99: 3955.

[6] Borgers, Tilman (2004).  Costly Voting , American Economic Review, Vol. 94(1): 57-66.

[7] Brady, Henry and John McNulty (2011).  Turning Out the Vote: The Costs of Finding and Getting to the Polling Place , American Political Science Review, Vol. 5(1): 1-20.

[8] Chong, Alberto, Ana L. De La O, Dean Karlan and Leonard Wantchekron (2011). Information Dissemination and Local Governments' Electoral Returns, Evidence from a Field Experiment in Mexico. Mimeo.

[9] Coate, Stephen, and Michael Conlin (2004).  A Group Rule-Utilitarian Approach to Voter Turnout: Theory and Evidence, American Economic Review, Vol.94, 1476-1504.

[10] Degan, Arianna (2006). "Policy Positions, Information Acquisition, and Turnout", Scandina-

vian Journal of Economics, Vol. 108(4): 669-682.

[11] Degan, Arianna and Antonio Merlo (2009). "Do Voters Vote Ideologically?", Journal of Eco-

nomic Theory, Vol. 144(5): 1868-1894.

[12] Degan, Arianna and Antonio Merlo (2011). "A Structural Model of Turnout and Voting in Multiple Elections", The Journal of The European Economic Association.

31

[13] Downs, Anthony (1957).  An Economic Theory of Democracy . New York: Harper and Row.

[14] Feddersen, Timothy, and Wolfgang Pesendorfer (1996).  The Swing Voter's Curse. American

Economic Review, Vol. 86(3): 40824.

[15] Feddersen, Timothy, and Wolfgang Pesendorfer (1999).  Abstention in Elections with Asymmetric Information and Diverse Preferences. The American Political science Review, Vol. 93(2): 38198.

[16] Feddersen, Timothy, Sean Gailmard, and Alvaro Sandroni (2009).  Moral Bias in Large Elections: Theory and Experimental Evidence. American Political Science Review, Vol. 103(2): 175-192.

[17] Feddersen, Timothy J, and Alvaro Sandroni (2006). A Theory of Participation in Elections ,

American Economic Review, Vol. 96(4): 1271-1283

[18] Ferraz, Claudio and Frederico Finan (2008). Exposing corrupt politicians: The eects of Brazil's publicly released audits on electoral outcomes. Quarterly Journal of Economics, Vol.123 (2), 703745.

[19] Finan, Frederico, and Laura Schechter (2011). Vote buying and Reciprocity. Econometrica,

forthcoming.

[20] Fraga, Bernard, and Eitan Hersh (2010). Voting Costs and Voter Turnout in Competitive Elections, Quarterly Journal of Political Science, Vol. 5: 339356

[21] Frey, Bruno (1971). Why do high income people participate more in politics?, Public Choice, Vol. 11(1): 101-105.

[22] Fujiwara, Thomas (2011). Voting Technology, Political Responsiveness, and Infant Health: Evidence from Brazil, Mimeo, University of British Columbia.

[23] Gentzkow, Matthew (2006). Television and Voter turnout. Quarterly Journal of Economics, Vol. 121(3): 931-972.

[24] Gerardi, Dino, Margaret A. McConnell, Julian Romero, Leeat Yariv (2011). Get Out the (Costly) Vote: Institutional Design for Greater Participation. Mimeo, Harvard University.

32

[25] Gerber, Alan and Donald Green (2000)  The Eects of Canvassing, Telephone Calls, and Direct Mail on Voter Turnout:

A Field Experiment, American Political Science Review,

Vol.94(3), 653- 663.

[26] Gerber, Alan and Donald Green (2001),  Do Phone Calls Increase Turnout? A Field Experiment, Public Opinion Quarterly, Vol. 65(1), 75-85.

[27] Gerber, Alan and Donald Green (2002). Reclaiming the Experimental Tradition in Political Science. In Helen V. Milner and Ira Katznelson, eds., Political Science: The State of the

Discipline, 3rd Edition. New York: W.W. Norton & Co., pp. 805-832.

[28] Gerber, Alan, Donald Green, and Ron Shachar (2003). Voting May Be Habit-Forming: Evidence from a Randomized Field Experiment. American Journal of Political Science, Vol. 47(3): 540-550.

[29] Gerber, Alan, Donald Green, and Christopher Larimer (2008).  Social Pressure and Voter Turnout: Evidence from a Large Scale Field Experiment , American Political Science Review, Vol. 102(1): 33-48.

[30] Gerber, A., Green, D. P. and Nickerson, D. (2003),  Getting out the Vote in Local Elections: Results from Six Door-to-Door Canvassing Experiments, Journal of Politics, Vol. 65, 10831096.

[31] Gerber, Alan, Dean Karlan, and Daniel Bergman (2009). "Does the Media Matter? A Field Experiment Measuring the Eect of Newspapers on Voting Behavior and Political Opinions".

American Economic Journal: Applied Economics, Vol.1 (2): 3552.

[32] Ghosal, Sayantan and Lockwood, Ben (2009).  Costly voting when both information and preferences dier: is turnout too high or too low?. Social Choice and Welfare, Vol 33: 25-50.

[33] Giné, Xavier, and Ghazala Mansuri (2001). Together we Will: Experimental Evidence on Female Voting Behavior in Pakistan. World Bank Policy Research Working Paper N.5692.

[34] Gomez, Brad, Thomas Hansford, and George Krause (2007).  The Republicans Should Pray for Rain: Weather, Turnout, and Voting in U.S. Presidential Elections , Journal of Politics,

Vol. 69: 649-63.

33

[35] Hansford, Thomas, and Bradt Gomez (2010). Estimating the Electoral Eects of Voter Turnout , American Political Science Review, Vol. 104(2): 268-288.

[36] Harsanyi, John (1980).  Rule Utilitarianism, Rights, Obligations and the Theory of Rational Behavior, Theory and Decision, Vol. 12: 115-133.

[37] Husted, Thomas, and Lawrence Kenny (1997). The Eect of the Expansion of the Voting Franchise on the Size of Government. The Journal of Political Economy, Vol. 105(1): 54-82

[38] Knack, Steve (1994). Does Rain Help Republicans? Theory and Evidence on Turnout and Vote. Public Choice, Vol. 79: 187-209

[39] Krasa, Stephan and Mattias Polborn (2008).  Is Mandatory Voting Better than Voluntary Voting? . Games and Economic Behavior, Vol. 66: 275-291.

[40] Krishna, Vijay and John Morgan (2011).  Voluntary Voting: Costs and Benets . Mimeo,

UC, Berkeley.

[41] Lassen, David Dreyer (2005). The Eect of Information on Voter Turnout: Evidence from a Natural Experiment. American Journal of Political Science, Vol. 49(1): 103-118.

[42] Ledyard, John (1984).  The Pure Theory of Large Two-Candidate Elections, Public Choice, Vol.44, 7-41.

[43] Martinelli, Cesar (2006). Análisis Económico de la Conducta de los Votantes. El Trimestre

Económico, Vol. 73: 211-237

[44] Matsusaka, John (1995).  Explaining Voter Turnout Patterns: An Information Theory, Pub-

lic Choice, Vol. 84, 91-117.

[45] Merlo, Antonio (2006).  Whither Political Economy? Theories, Facts and Issues, in Richard Blundell, Whitney Newey and Torsten Persson (eds.), Advances in Economics and Econo-

metrics, Theory and Applications: Ninth World Congress of the Econometric Society, Vol. I: 381-421, Cambridge: Cambridge University Press.

[46] Miller, Grant (2008). Women's Sufrage, Political Responsiveness, and Child Survival in American History. Quarterly Journal of Economics, Vol. 123(3), 1287-1327.

34

[47] Oliveros, Santiago (2011). "Abstention, ideology and information acquisition". Mimeo, UC

Berkeley.

[48] Palfrey, Thomas, and Howard Rosenthal (1985).  Voters Participation and Strategic Uncertainty, American Political Science Review, Vol. 79, 62-78.

[49] Pande, Rohini (2011). Can Informed Voters Enforce Better Governance?

Experiments in

Low Income Democracies, Annual Review of Economics, Vol. 3.

[50] Perea, Eva Anduiza (2002). Individual characteristics, institutional incentives and electoral abstention in Western Europe. European Journal of Political Research, Vol. 41: 643673.

[51] Persson, Torsten and Guido Tabellini (2000). Political Economics:

Explaining Economic

Outcomes, Cambridge: MIT Press.

[52] Riker, William and Peter Ordeshook (1968). A Theory of The Calculus of Voting., American

Political Science Review, Vol. 62: 25-42.

[53] Stromberg, David (2004). Radio's Impact on Public Spending. Quarterly Journal of Eco-

nomics, Vol.119(1): 189-221.

[54] Vicente, Pedro (2008). Is Vote-buying Eective? Evidence from a Field Experiment in West Africa, Mimeo, Oxford University.

[55] Vicente, Pedro, Leonard Wantchekon (2009). Clientelism and Vote-buying: Lessons from Field Experiments in African Elections, Oxford Review of Economic Policy, Vol. 25(2), pp. 292-305.

35

Figure 1: Discontinuity Analysis: Turnout Across Poverty Levels

Effect non−voting fine law on turnout

.7

.84

Turnout − 2006 Local Elections .86 .88 .9

Turnout − 2002 Local Elections .75 .8 .85

.9

.92

Effect non−voting fine law on turnout

0

500 1000 1500 Poverty ranking (richest to poorest)

2000

0

500 1000 1500 Poverty ranking (richest to poorest)

2000

.82

Turnout − 2010 Local Elections .84 .86 .88

.9

Effect non−voting fine law on turnout

0

Notes:

500 1000 1500 Poverty ranking (richest to poorest)

2000

This gures plot the ocial turnout rates at the district level in the 2002, 2006, and 2010 municipal elections.

Districts are ranked from richest to poorest, and the vertical lines indicate the thresholds at which a district is categorized as Non-Poor, Poor, or Extremely Poor.

36

Figure 2: Perceived nes, by treatment and poverty status

Baseline

Follow-up

Voters from Non−Poor Districts

Density 0

0

.002

.005

Density .004

.01

.006

.015

Density: Perceived fines−Follow−up

Voters from Non−Poor Districts

.008

Density: Perceived fines−Baseline

50

100 150 Perceived Fines

250

0

50

Treatment

100 150 Perceived Fines Control

200

250

200

250

Treatment

Perceived fines−Baseline

Perceived fines−Follow−up

Voters from Poor Districts

Voters from Poor Districts

Density 0

0

.002

.005

Density .004

.01

.006

.008

Control

200

.015

0

50

100 150 Perceived Fines Control

200

250

0

50

Treatment

100 150 Perceived Fines Control

Treatment

Density: Perceived fines−Follow−up

Voters from Extremely Poor Districts

Voters from Extremely Poor Districts

Density 0

0

.002

.005

Density .004

.01

.006

.008

Density: Perceived fines−Baseline .015

0

0

50

100 150 Perceived Fines Control

200

250

Treatment

0

50

100 150 Perceived Fines Control

37

200 Treatment

250

Figure 3: Geographic location of the districts in the survey

38

39

88.9% 88.4% 87.5% 86.7% 3.1% 3.5% 3.8% 2.6% 10 5

2006-3 2010 2006-3 2010 2006 2010

4,291 4,291 4,297 4,615

2006-1 2006-2 2006-3 2010

2006-1 2006-2 2006-3 2010

1767 251 632

173 Non-poor

3.7% 5.7% 11 13

9.1% 7.4%

92.6% 92.2% 91.1% 90.0%

24,668 24,668 25,006 26,873

8170 248 930

176 Non-poor

Lima Cañete Imperial

9.8% 5.4% 13 20

7.0% 9.3%

92.1% 91.4% 90.1% 88.2%

17,577 17,577 18,183 20,630

6429 240 756

191 Poor

Lima Huaura Santa Maria

13.2% 12.3% 7 5

8.4% 6.8%

95.0% 94.8% 94.5% 93.7%

2,200 2,200 2,255 2,525

897 241 776

201 Poor

Lima Cañete San Antonio

9.2% 8.0% 8 12

11.5% 14.3%

93.9% 93.7% 92.5% 89.9%

5,564 5,564 6,478 13,640

240 668

841 Poor

Lima Huarochiri San Antonio

Table 1: Descriptive Statistics, Districts sampled

Lima Lima Punta Hermosa

2.5% 0.7% 9 2

5.5% 11.6%

91.2% 90.5% 91.6% 91.4%

872 872 1,017 1,367

717 146 418

876 Poor

Lima Huaura Leoncio Prado

4.9% 4.5% 8 7

8.9% 10.2%

86.0% 88.0% 83.1% 82.0%

543 543 568 683

311 93 264

938 Poor

Lima Yauyos Huantan

34.3% 4.3% 6 6

8.9% 9.0%

77.1% 79.0% 82.5% 83.6%

724 724 778 921

461 166 460

978 Ext. Poor

Lima Cajatambo Huancapon

5.4% 4.4% 7 6

13.9% 15.5%

73.6% 78.3% 75.4% 79.0%

1,857 1,857 1,891 1,820

791 166 462

1042 Ext. Poor

Lima Cajatambo Cajatambo

Notes: 2006-1 and 2006-2 refers to the rst and second round of the presidential elections held in April and June 2006, respectively. 2006-3 refers to the municipal elections held in November, 2006, the rst elections under the new levels of the nes.

Num. of candidates running for the local government

Invalid Votes (%)

Blank Votes (%)

Turnout

Registered voters

Electoral variables

Number of HHs Sampled HHs Sampled Individuals

Poverty Ranking Poverty Category

Region Province District Poverty

2.0% 39.3% 9 6

12.3% 4.6%

86.5% 85.5% 88.7% 85.4%

490 490 550 664

593 121 308

1047 Ext. Poor

Lima Huaura Checras

Table 2: Summary Statistics for Full Sample

Variable

Mean Std. Dev. Min. Max.

Perceived Fine (Baseline)

N

122.292

57.325

0

300

2276

0.424

0.494

0

1

2276

39.885

13.358

18

70

2276

9.586

4.063

0

21

2276

5.19

0.871

-2.303

8.849

2276

Aected by political violence

0.156

0.363

0

1

2276

Center

0.667

0.472

0

1

2202

Left

0.083

0.275

0

1

2202

Right

0.251

0.434

0

1

2202

Policy Extreme 1 (Pub. goods)

0.207

0.406

0

1

2276

Policy Center

0.598

0.491

0

1

2276

Policy Extreme 2 (Club goods)

0.195

0.396

0

1

2276

Very Interested in politics

0.082

0.274

0

1

2244

Interested in politics

0.468

0.499

0

1

2244

Not Interested in politics

0.451

0.498

0

1

2244

Very Interested in the results of this election

0.399

0.49

0

1

2276

Interested in the results of this election

0.443

0.497

0

1

2258

Not Interested in the results of this election

0.153

0.36

0

1

2276

Very Interested in the campaign of this election

0.105

0.307

0

1

2254

Interested in the campaign of this election

0.556

0.497

0

1

2254

Not Interested in the campaign of this election

0.339

0.473

0

1

2254

Name recall- Candidates running

0.388

0.35

0

1

2276

0.29

0.317

0

1

2276

Gender Age Yrs. of education Log(PC Expenditures)

Name recall- Parties running Name recall- Candidates+Parties running

0.339

0.315

0

1

2276

Political information score

0.547

0.179

0

1

2276

40

Table 3:

Variable

Balance Between the Treatment and Control Group

Perceived Fine (Baseline) Gender Age Yrs. of education Log(PC Expenditures) Aected by political violence Center Left Right Policy Extreme 1 (Pub. goods) Policy Center Policy Extreme 2 (Club goods) Very Interested in politics Interested in politics Not Interested in politics Very Interested in the results of this election Interested in the results of this election Not Interested in the results of this election Very Interested in the campaign of this election Interested in the campaign of this election Not Interested in the campaign of this election Name recall- Candidates running Name recall- Parties running Name recall- Candidates+Parties running Political information score

Obs. Treatment Control T - C P-value 2275 2275 2275 2275 2275 2275 2201 2201 2201 2275 2275 2275 2243 2243 2243 2275 2257 2275 2253 2253 2253 2275 2275 2275 2275

124.233 0.419 39.916 9.593 5.176 0.142 0.671 0.076 0.253 0.213 0.614 0.173 0.080 0.459 0.461 0.390 0.433 0.170 0.106 0.549 0.345 0.390 0.297 0.343 0.545

120.362 0.428 39.874 9.576 5.203 0.170 0.662 0.090 0.248 0.202 0.580 0.216 0.083 0.477 0.440 0.407 0.454 0.136 0.105 0.563 0.332 0.387 0.283 0.335 0.550

-3.871 0.009 -0.041 -0.017 0.027 0.027 -0.009 0.014 -0.005 -0.010 -0.034 0.043 0.003 0.018 -0.021 0.017 0.020 -0.034 -0.001 0.014 -0.013 -0.002 -0.014 -0.008 0.005

(0.107) (0.658) (0.941) (0.920) (0.452) (0.072) (0.641) (0.228) (0.796) (0.542) (0.101) (0.009) (0.822) (0.385) (0.320) (0.406) (0.335) (0.024) (0.951) (0.510) (0.514) (0.874) (0.297) (0.540) (0.503)

Notes: The table includes all subjects interviewed in the baseline and follow-up surveys. Table 18 in the Appendix shows the balance by poverty status.

41

Table 4: Turnout and Perceived nes, by Treatment and Poverty Status

Total Treatment Control T - C P-value PANEL A: Turnout Non-Poor

0.948

0.938

0.959

-0.021

(0.175)

Poor

0.940

0.913

0.967

-0.054

(0.001)***

Extreme Poor

0.935

0.930

0.940

-0.010

(0.641)

Total

0.942

0.927

0.958

-0.031

(0.002)***

Non-Poor

126.5

123.8

129.4

-5.605

(0.144)

Poor

122.1

122.3

122.0

0.230

(0.951)

Baseline

PANEL B: Perceived Fines

Extreme Poor

115.9

111.9

120.0

-8.066

(0.132)

Total

122.3

120.4

124.2

-3.871

(0.107)

Non-Poor

78.5

66.8

91.0

-24.197

(0.000)***

Poor

57.3

42.1

71.2

-29.047

(0.000)***

Extreme Poor

27.9

19.4

36.6

-17.199

(0.000)***

Total

58.2

46.1

70.2

-24.111

(0.000)***

Non-Poor

-48.0

-57.0

-38.5

-18.593

(0.000)***

Poor

-64.8

-80.1

-50.9

-29.277

(0.000)***

Extreme Poor

-88.0

-92.5

-83.4

-9.133

Total

-64.1

-74.2

-54.0

-20.239

Follow-up

Change

(0.121) (0.000)***

Notes: The actual changes that occurred were: for people voting in Non-poor districts, S/.72 (from S/.144 to S/.72); for those voting in Poor districts, S/.108 (from S/.144 to S/.36); and for people registered to vote in Extremely Poor districts, S/.126 (from S/.144 to S/.18).

42

Table 5: Reduced Form - Eect of Treatment on Voting Dep. Var: Voted in the 2010 Election Non-Poor Treatment: Fine S/.72

Poor

-.027

-.026

(0.015)∗

(0.015)∗

Treatment: Fine S/.36 Gender Age Yrs. of education Log(PC Expenditures)

-.052

(0.016)∗∗∗ -.0009

0.018

(0.016)

(0.016)

0.001

0.001

(0.0007)

(0.0006)∗∗

0.002

0.004

(0.002)

(0.003)

0.004

0.011

(0.008)

(0.013)

Votes in Poor district

Village FE Mean dep. var. Obs.

R2

All

-.053

(0.016)∗∗∗ 0.013

(0.011) 0.001

(0.0005)∗∗∗ 0.004

(0.002)∗∗ 0.007

(0.008) 0.0006

(0.022)

Y

Y

Y

0.9482

0.9410

0.9446

850

882

1732

0.102

0.093

0.049

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α + β1 N onP oorij ∗ T reatij + β2 P oorij ∗ T reatij + β3 P oorij + γXij + δk + νij

43

Table 6: First Stage - Eect of Treatment on Changes in Perceived Fine Dep. Var: Non-Poor Treatment: Fine S/.72

Age

Perceived Fines Poor

-18.807

All -19.317

(4.905)∗∗∗

(4.854)∗∗∗

Treatment: Fine S/.36 Gender

4

-30.465

-2.962

(4.946)

(4.692)∗∗∗

-2.135

-2.839

(4.741)

0.333

0.409

(0.201)∗

-30.340

(4.756)∗∗∗

(0.182)∗∗ -.753

(3.393) 0.363

(0.133)∗∗∗

Yrs. of education

0.266

(0.74)

(0.703)

(0.499)

Log(PC Expenditures)

-4.101

-1.684

-2.369

(3.524)

(3.532)

Votes in Poor district

Village FE Mean dep. var. Obs.

R2

-.243

(2.520) 0.09

(6.270) Y

Y

Y

-48.00

-64.99

-56.65

851

882

1733

0.116

0.106

0.11

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

4F ineij = α + β1 N onP oorij ∗ T reatij + β2 P oorij ∗ T reatij + β3 P oorij + γXij + δk + νij

44

Table 7: IV - Eect of Change in Perceived Fines on Turnout Dep. Var: Voted in the 2010 Election Non-Poor

4

Perceived Fine

Poor

0.0014

0.0017

(0.0009)∗

(0.0006)∗∗∗ 0.022

All 0.0016

(0.0005)∗∗∗

Gender

0.0034

(0.0175)

(0.017)

(0.0124)

Age

0.0005

0.0008

0.0007

Yrs. of education

0.0013

(0.0024)

(0.0031)∗

(0.002)∗∗

Log(PC Expenditures)

0.0101

0.0142

0.0109

(0.0008)

(0.0007)

0.0056

(0.0108)

(0.0145)

Votes in Poor district

Village FE Mean dep. var. Obs. F-statistic

0.018

(0.0005)

0.0042

(0.0087) -.0041

(0.0218) Y

Y

Y

0.9482

0.9410

0.9446

850

882

1732

14.68

41.03

28.66

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α + β1 4F ineij + β2 P oorij + γXij + δk + ij

45

Table 8: Eect of Changes in Perceived Fine on Turnout, by Demographic Characteristics Dep. Var: Voted in the 2010 Election (1)

4

Perceived Fine

4

Fine*Age

4

Fine*Male

4

Fine*Yrs. Educ.

4

Fine*Log(PC Expenditures)

4

Fine*Aected by violence

Controls Village FE Obs.

0.0008

(0.0013)

(2) 0.0025

(0.0008)∗∗∗

(3) 0.0051

(0.0025)∗∗

(4) 0.004

(0.0023)∗

(5) 0.0017

(0.0005)∗∗∗

0.00002

(0.00004)

-.0021

(0.001)∗∗ -.0003

(0.0002) -.0005

(0.0004) -.0007

(0.0013) Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

1732

1732

1732

1732

1732

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α + β1 4F ineij + β2 4F ineij Xij + β3 Xij ∗ P oorij + β4 Xij ∗ N onP oorij + β5 P oorij + γXij + δk + ij

46

Table 9: Robustness: Eect of Changes in Perceived Fine on Turnout - Dierent Measures of Turnout Dep. Var: Voted in the 2010 Election Available Sample Benckmark

4

Perceived Fine

Gender

0.0016

Votes in Poor district

Village FE Obs. F-statistic

0.001

Sticker 0.0015

(0.0005)∗∗∗

(0.0005)∗∗

(0.0005)∗∗∗

0.018

0.0142

0.0104

0.0018

0.0109

0.001

0.0002

0.0005

0.0002

(0.0005)

Log(PC Expenditures)

0.0015

Comparable Sample Self Reported

(0.0005)∗∗∗

0.0007

Yrs. of education

0.0013

Sticker

(0.0005)∗∗∗ (0.0124)

Age

Self Reported

0.0042

(0.0122)

(0.0005)∗

(0.0126)

(0.0005)

(0.0116)

(0.0004)

(0.0128) (0.0005)

(0.002)∗∗

(0.002)∗∗

0.0049

0.0014

0.0025

0.0014

0.0109

0.0081

0.0115

0.0069

0.0118

-.0029

0.0101

0.0118

0.0104

(0.0087) -.0041

(0.0218)

(0.0082) (0.0213)

(0.0021)

(0.0075) (0.0231)

(0.002)

(0.0064) (0.0221)

(0.0021)

(0.0075) (0.0232)

Y

Y

Y

Y

Y

1732

1729

1130

1127

1127

28.6595

28.2653

17.2611

16.8161

16.8161

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Columns (4) and (5) use only the sample of observations for which both outcomes are available. Regression equation:

V oteij = α + β1 4F ineij + β2 P oorij + γXij + δk + ij

47

Table 10: Robustness: Eect of Changes in Perceived Fine on Past Turnout Dep. Var: Voted in the 2006 Election

4

Perceived Fine

Gender Age Yrs. of education Log(PC Expenditures)

Non-Poor

Poor

All

-.0016

0.0007

0.00006

-.0090

0.0212

0.0117

(0.001)

(0.0006)

(0.0175)

(0.0157)

0.0049

0.0023

(0.0011)∗∗∗

(0.0009)∗∗

0.0112

0.0078

(0.0029)∗∗∗

(0.0023)∗∗∗

-.0104

0.0175

(0.0117)

(0.0153)

Votes in Poor district

Village FE Mean dep. var. Obs.

(0.0005)

(0.0109)

0.0035

(0.0007)∗∗∗ 0.0085

(0.0017)∗∗∗ 0.004

(0.0083) 0.0042

(0.0216)

Y

Y

Y

0.9459

0.9444

0.9451

758

791

1549

F-statistic

11.92

32.33

23.44

R2

-.0735

-.0141

0.0441

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

2 + γXij + δk + ij . V otet−1 = α + β1 4F ineij + β2 P ovij ij

reported, and it refers to turnout in the November, 2006 municipal election.

48

The dependent variable is self

Table 11: Eect of Changes in Perceived Fine on Turnout, by Political Preferences Dep. Var: Voted in the 2010 Election (1)

4

Fine*Left

4

Fine*Center

4

Fine*Right

4

Fine*Policy Extreme 1 (Pub. Goods)

4

Fine*Policy Center

4

Fine*Policy Extreme 2 (Club Goods)

(2)

-.0009

(0.0026) 0.0015

(0.0006)∗∗∗ 0.0009

(0.0008)

0.001

(0.0013) 0.002

(0.0007)∗∗∗ 0.0006

(0.0009)

Controls

Y

Village FE Obs.

Y

Y

Y

1665

1732

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α +

P3

n=1

βn 4F ineij Pijn +

P3

n=1

βn1 Pijn P oorij +

dummy variable representing political preferences

P3

n = 1, 2, 3

n=1

βn1 Pijn N onP oorij + β10 P oorij + γXij + δk + ij , Pijn

for individual

i

interviewed in village

k.

is a

In Column (1), Left,

Center, and Right are self reported variables indicating positions in the ideological scale, which ranges from 1 to 5. People choosing 1 and 5 are categorized as Left or Right, respectively, while 2, 3 and 4 are considered in the Center. The second measure of ideological positions (used in Column(2)) is an aggregation of several measures of policy preferences. I use responses from a question where I asked respondents to name (in order) the rst ve policies that she would implement if she were elected mayor of the district. For each of these categories, the policy preferences are ordered from not mentioned (zero) to most preferred (ve). I aggregate these questions by taking the rst principal component, and dividing the sample into quintiles. The center is dened by those in the quintiles 2, 3, and 4, while the rst and fth quintiles dene the ideological extremes: Policy Extreme 1 (Pub. Goods), Policy Extreme 2 (Club Goods), respectively. The results from the principal component analysis is shown in Table 21 in the Appendix.

49

Table 12: Eect of Changes in Perceived Fine on Turnout, by Interest in Politics Dep. Var: Voted in the 2010 Election (1)

4

Fine*Very interested in politics

4

Fine*Interested in politics

4

Fine*Not interested in politics

4

Fine*Very interested in results

4

Fine*Interested in results

4

Fine*Not interested in results

4

Fine*Very interested in pol. campaign

4

Fine*Interested in pol. campaign

4

Fine*Not interested in pol. campaign

(2)

(3)

0.0001

(0.0018)

0.0012

(0.0007)∗ 0.0018

(0.0007)∗∗∗ 0.0007

(0.0006)

0.0018

(0.0007)∗∗∗ 0.0039

(0.002)∗∗ 0.0023

(0.002)

0.0009

(0.0005)∗ 0.0023

(0.001)∗∗

Controls

Y

Y

Y

Villafe FE

Y

Y

Y

1713

1717

1714

Obs.

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α +

P3

n=1

n + βn 4F ineij Iij

P3

n=1

n P oorij + βn1 Iij

dummy variable representing interest in politics

n = 1, 2, 3

P3

n=1

k n N onP oorij + β10 P oorij + γXij + δk + ij , Iij βn1 Iij

for individual

50

i

interviewed in village

k.

is a

Table 13: Eect of Changes in Perceived Fine on Turnout, by Political Information Dep. Var: Voted in the 2010 Election (1)

4

Perceived Fine

4

Fine*Candidate recall

4

Fine*Party recall

4

Fine*Candidate and Party recall

4

Fine*Pol. Info. Score

(2)

0.0024

(0.0008)∗∗∗

0.0022

(0.0007)∗∗∗

(3) 0.0024

(0.0008)∗∗∗

(4) 0.0079

(0.0031)∗∗

-.0023

(0.0012)∗∗ -.0022

(0.0011)∗ -.0027

(0.0012)∗∗ -.0113

(0.0053)∗∗

Controls

Y

Y

Y

Y

Village FE

Y

Y

Y

Y

1732

1732

1732

1732

Obs.

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α + β1 4F ineij + β2 4F ineij Inf oij + β3 Inf oij ∗ P oorij + β4 Inf oij ∗ N onP oorij + β5 P oorij + γXij + δj + ij . The information variables are indices ranging from zero to one. The candidate and/or party recall represent the proportion of candidates/parties running in the election in the unicipality where the voter is registered. Additionally, I included a battery of 17 questions related to the features of the political system, mandatory ages for voting, term limits at dierent levels of the government, etc. The political information score represents the proportion of questions that the respondent was able to answer correctly.

51

Table 14: Eects by policy preferences Dep. Var.: Voted in the 2010 Election Coe. on

4

Perceived Fine

Coe. on

4

Perceived Fine*Policy

Policy Health

0.0019 ∗∗

(0.0008) Education

0.0009 (0.0005)

Infrastructure Order and Security

0.0012 ∗

0.001

0.0007 (0.0012)

0.0022 ∗∗∗

0.0016 ∗∗∗

(0.0005) Agriculture

0.0022 ∗∗∗

(0.0007) Youth/Women

0.0013 ∗∗

(0.0006) Cleaning/Environment

0.0013 ∗∗

(0.0005) Institutions

0.0018 ∗∗∗

(0.0006) Social/work programs

(0.001)

(0.0011) (0.0007) Promote micro-enterprises/training

-.0005 (0.0009)

0.0017 ∗∗∗

(0.0006)

-.0012 (0.001) 0.0002 (0.0012) -.0020 (0.0008)

∗∗

0.0013 (0.0011) 0.0007 (0.001) -.0010 (0.001) -.0004 (0.001)

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation:

V oteij = α + β1 4F ineij + β2 4F ineij P olicyij + β3 P olicyij P oorij + β4 P olicyij N onP oorij + β5 P oorij + γXij + δk + ij . The coecients shown in each row come from separate regressions. Policy preferences include: (1) Health: Infrastructure, health professionals, and training for health workers; (2) Education: Infrastructure, teachers, and training for teachers; (3) Infrastructure: Roads and access to them, sewage, water, electricity and telecommunications infrastructure, build markets, churches, community building, main square; (4) Order and Security: Trac, more policemen in the streets, ght drugs and gangs; (5) Promote micro-enterprises/training: promote micro/small rms, train local entrepreneurs, promote private investment, promote tourism; (6) Agriculture: Build dams and irrigation infrastructure, technical assistance to agriculture, seed banks, support livestock farmers; (7) Youth/Women: Women empowerment and equality, youth policies, sporting events; (8) Cleaning/Environment: street cleaning, increase green areas, promote recycling; (9) Institutions: Transparency in managing the municipality, ght corruption, modernize the bureaucracy, participatory decision-making, land titling; (10) Social/work programs: Job training programs, help those in poverty, food aid, child care, generate jobs. For each of these categories, the dependent variable is a dummy indicating whether the respondent named at least one of the policies in this category as one of her ve priorities for the district.

52

Table 15: Eects of Fines on Information Acquisition Dep. Var.:

4 4

Perceived Fine

Gender Age Yrs. of education Log(PC Expenditures) Votes in Poor district

Village FE Obs. F-Statistic

R2

Candidate Recall -.0002

(0.0005) -.0236

4

Party Recall -.0005

(0.0005) -.0371

4

Cand.+Party Recall -.0004

(0.0005) -.0304

(0.0125)∗

(0.0137)∗∗∗

(0.0121)∗∗

-.0003

0.0006

0.0001

(0.0005) -.0036

(0.0017)∗∗ -.0176

(0.0006) -.0065

(0.0018)∗∗∗ -.0106

(0.0005) -.0051

(0.0016)∗∗∗ -.0141

Pol. Info Score 3.00e-06

(0.0003) -.0265

(0.0092)∗∗∗ -.0009

(0.0004)∗∗ -.0061

(0.0013)∗∗∗ -.0046

(0.0094)∗

(0.0094)

-.0212

0.0104

(0.0141)

(0.0124)

(0.0137)

Y

Y

Y

Y

(0.0131)

(0.0084)∗

4

-.0054

(0.0066) -.0054

1733

1733

1733

1733

28.675

28.675

28.675

28.675

0.006

0.0064

0.0061

0.0248

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation: 4Inf oij = α + β1 4F ineij + β2 P oorij + γXij + δk + ij , where 4Inf oij represents

the

change in the political information between the baseline and follow-up surveys. The dependent variable is the dierence in the information measures between the follow-up and baseline surveys.

53

Table 16: Eects of Fines on Vote buying Dep. Var:

4

Perceived Fine

Gender Age Yrs. of education Log(PC Expenditures) Votes in Poor district

1=Accepted Money

Amount Accepted

Amount Accepted

Amount Accepted

or a Gift

Directly

Indirectly

Total

-.0303

0.0043

-.9115

0.0005

-.0010

(0.0009) -.0365

(0.0227) -.0014

(0.0009) -.0010

(0.0161)∗ (0.4431)∗∗ -.0379

(0.0221)∗ -.0277

(0.0071)

(0.1748) -.0208

(0.0084)∗∗ -.0189

-.0276

(0.0178) -1.1150

(0.4829)∗∗ -.0541

(0.0236)∗∗ -.0280

(0.0032)

(0.0637)

(0.0196)

(0.0684)

-.0082

0.1046

0.0956

0.2389

-.0335

0.1663

0.1884

(0.0143) 0.0758

(0.027)∗∗∗

(0.3056)

(0.3879)

(0.1242)

(0.1856)

(0.3389) (0.443)

Controls

Y

Y

Y

Y

Village FE

Y

Y

Y

Y

0.287

2.20

0.818

3.25

Mean dep. var. Obs. F-statistic

1733

1733

1733

1733

28.675

28.675

28.675

28.675

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in 2 parentheses. Regression equation: Yij = α + β1 4F ineij + β2 P ovij + γXij + δk + ij . In Column (1), Yij is an indicator for whether voter i accepted money from a politician or his/her representative for her vote. In Column (2) through (4), it measures the amount of money accepted (directly or indirectly) to buy a vote.

54

APPENDIX Table 17: Balance Between Attrited and non-Attrited

Variable Gender Age Yrs. of education Log(PC Expenditures) Aected by political violence Center Left Right Policy Extreme 1 Policy Center Policy Extreme 2 Very Interested in politics Interested in politics Not Interested in politics Very Interested in the results of this election Interested in the results of this election Not Interested in the results of this election Very Interested in the campaign of this election Interested in the campaign of this election Not Interested in the campaign of this election Name recall- Candidates running Name recall- Parties running Name recall- Candidates+Parties running Political information score (baseline)

Obs. Non-Attriters Attriters NA - A P-value 2838 2838 2838 2838 2838 2838 2754 2754 2754 2838 2838 2838 2795 2795 2795 2838 2814 2838 2809 2809 2809 2837 2837 2837 2838

55

125.764 0.482 39.180 9.619 5.225 0.167 0.670 0.071 0.259 0.171 0.609 0.221 0.065 0.443 0.492 0.375 0.455 0.164 0.112 0.512 0.377 0.401 0.308 0.355 0.561

122.292 0.424 39.885 9.586 5.190 0.156 0.667 0.083 0.251 0.207 0.598 0.195 0.082 0.468 0.451 0.399 0.443 0.153 0.105 0.556 0.339 0.388 0.290 0.339 0.547

-3.472 -0.059 0.706 -0.034 -0.035 -0.011 -0.004 0.012 -0.008 0.037 -0.011 -0.026 0.016 0.025 -0.041 0.024 -0.012 -0.010 -0.007 0.045 -0.038 -0.013 -0.019 -0.016 -0.014

(0.199) (0.012) (0.265) (0.860) (0.409) (0.529) (0.872) (0.354) (0.685) (0.052) (0.634) (0.168) (0.205) (0.290) (0.081) (0.307) (0.618) (0.544) (0.653) (0.058) (0.091) (0.436) (0.212) (0.289) (0.096)

Table 19: Robustness: Main Regressions, Without Controls (1)

(2)

(3)

Panel A: Reduced Form Dep. Var: Voted in the 2010 Election Treatment: Fine S/.72 Treatment: Fine S/.36

R2

-.0250

-.0217

(0.0149)∗

(0.0152)

-.0532

-.0533

(0.0162)∗∗∗ 0.0391

-.0258

(0.015)∗ -.0527

(0.016)∗∗∗

(0.0161)∗∗∗

0.0181

0.0487

Panel B: First Stage Dep. Var: Treatment: Fine S/.72 Treatment: Fine S/.36

R2

-19.5131

4

Perceived Fine

-18.5018

(4.8591)∗∗∗

(5.1395)∗∗∗

-30.5384

-29.1100

-19.3167

(4.8544)∗∗∗ -30.3400

(4.7246)∗∗∗

(4.7584)∗∗∗

(4.6921)∗∗∗

0.104

0.0506

0.1098

Panel C: IV Dep. Var: Voted in the 2010 Election

4

Perceived Fine

0.0016

0.0016

(0.0005)∗∗∗

(0.0005)∗∗∗

0.0016

(0.0005)∗∗∗

Controls

N

Y

Y

Village FE

N

N

Y

0.9445

0.9445

0.9445

-56.65

-56.65

-56.65

28.7586

25.2301

28.6595

1732

1732

1732

Mean Vote 2010 Mean

4

Perceived Fine

F-statistic Obs.

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equations: Reduced Form:

V oteij = α + β1 4F ineij + β2 P oorij + γXij + δk + ij

56

57 0.446 0.475 0.319 0.468 0.208 0.094 0.518 0.387 0.266 0.214 0.240 0.547

Very Interested in the results of this election

Interested in the results of this election

Not Interested in the results of this election

Very Interested in the campaign of this election

Interested in the campaign of this election

Not Interested in the campaign of this election

Name recall- Candidates running

Name recall- Parties running

Name recall- Candidates+Parties running

Political information score

0.077

Policy Extreme 1 (Pub. goods)

Not Interested in politics

0.232

Right

Interested in politics

0.067

Left

0.078

0.701

Center

Very Interested in politics

0.116

Aected by political violence

0.374

5.482

Log(PC Expenditures)

0.548

10.275

Yrs. of education

Policy Extreme 2 (Club goods)

38.150

Age

Policy Center

0.372

129.415

Treat

Gender

Perceived Fine (Baseline)

Variable

0.561

0.258

0.224

0.292

0.370

0.525

0.104

0.153

0.474

0.371

0.468

0.456

0.076

0.410

0.522

0.066

0.241

0.100

0.659

0.153

5.532

10.348

38.121

0.403

123.810

(0.217)

(0.343)

(0.582)

(0.216)

(0.610)

(0.832)

(0.637)

(0.039)

(0.878)

(0.112)

(0.819)

(0.773)

(0.912)

(0.294)

(0.438)

(0.537)

(0.768)

(0.087)

(0.197)

(0.111)

(0.338)

(0.776)

(0.974)

(0.358)

(0.144)

Non-Poor Control Di. p-val

0.555

0.410

0.369

0.452

0.363

0.550

0.087

0.156

0.427

0.410

0.478

0.458

0.064

0.083

0.671

0.246

0.194

0.055

0.751

0.149

5.187

9.739

39.443

0.423

122.040

Treat

0.555

0.384

0.339

0.428

0.297

0.622

0.081

0.122

0.461

0.413

0.431

0.500

0.069

0.141

0.610

0.249

0.205

0.087

0.708

0.185

5.194

9.746

38.901

0.418

122.270

(0.947)

(0.265)

(0.225)

(0.370)

(0.038)

(0.031)

(0.746)

(0.150)

(0.308)

(0.927)

(0.167)

(0.214)

(0.787)

(0.007)

(0.060)

(0.912)

(0.679)

(0.078)

(0.165)

(0.148)

(0.898)

(0.979)

(0.547)

(0.871)

(0.951)

Poor Control Di. p-val

Table 18: Balance Between Treatment and Control Group, by poverty level

0.522

0.388

0.302

0.474

0.249

0.596

0.155

0.138

0.391

0.466

0.410

0.479

0.111

0.015

0.619

0.366

0.383

0.121

0.496

0.172

4.684

8.291

43.448

0.485

119.959

Treat

0.523

0.382

0.289

0.475

0.326

0.531

0.143

0.131

0.410

0.456

0.410

0.474

0.116

0.026

0.628

0.347

0.325

0.077

0.598

0.172

4.694

8.080

44.182

0.485

111.894

(0.97

(0.80

(0.60

(0.96

(0.04

(0.12

(0.70

(0.82

(0.64

(0.81

(0.99

(0.90

(0.86

(0.38

(0.84

(0.64

(0.16

(0.09

(0.01

(0.99

(0.89

(0.55

(0.52

(0.99

(0.13

Extreme Poor Control Di. p

Table 20: Robustness: Main Regressions, Including the Voters from Extreme Poor Districts Reduced Form Voted in 2010

4

First Stage

4

in Perceived Fine

Perceived Fine

Treatment: Fine S/.72 Treatment: Fine S/.36 Treatment: Fine S/.72 Treatment: Fine S/.36 Treatment: Fine S/.18 Gender Age Yrs. of education

IV Voted in 2010 0.0015

(0.0005)∗∗∗ -.0208

-19.3585

(0.0157)

(4.8621)∗∗∗

-.0508

-30.1273

(0.0161)∗∗∗

(4.6858)∗∗∗

-.0208

-19.3585

(0.0157)

(4.8621)∗∗∗

-.0508

-30.1273

(0.0161)∗∗∗

(4.6858)∗∗∗

-.0091

-8.5888

(0.0201)

(5.9851)

0.01

-4.8665

(0.0099)

(2.9300)∗

0.0015

0.3473

(0.0004)∗∗∗

(0.1188)∗∗∗

0.004

-.2727

(0.0015)∗∗∗

0.0174

(0.0111)

0.0009

(0.0005)∗∗ 0.0043

(0.4196)

(0.0016)∗∗∗

Log(PC Expenditures)

0.0014

(0.006)

(2.1271)

-.9997

0.0029

Votes in Poor district

0.0341

-2.1440

0.0303

-.1294

-18.7409

(8.6968)∗∗

(0.0547)∗

0.9424

-64.115

0.9424

2273

2273

2273

0.0544

0.132

Votes in Extreme Poor district

Mean dep. var. Obs.

(0.023)

(6.1572)

(0.0528)∗∗

F-statistic

R2

(0.0068) (0.0226) -.1038

19.57

* signicant at 10%; ** signicant at 5%; *** signicant at 1%. Robust standard errors clustered at the household level in parentheses. Regression equation for these regressions follow the structure detailed in the main text in equations (7),(9), and (10), but including an indicator for voting in an extremely poor district, and the corresponding interactions.

58

Table 21: Coecients for Policy Preference First Principal Component Policy issues

Coecients

Health: infrastructure

-0.116

Health: personnel and services

-0.145

Education: infrastructure

-0.114

Education: teachers and services

-0.085

Transport: Ordering transit

0.024

Transport: Infrastructure (roads, access, etc.)

-0.362

Basic services: Water, electricity, sewage, communications

-0.478

Promote tourism

-0.062

Economics: Support micro and small enterprises

-0.027

Economics: Training to local enetrepeneurs

-0.025

Economics: Agriculture - technical assistance, and training to local producers

-0.271

Economics: Agriculture - infrastructure projects for agriculture

-0.113

Economics: promote private investment

-0.020

Youth: Sport activities and infrastructure

-0.026

Youth: Labor training programs

0.024

Women: empowerment and programs

-0.003

Social: More participation/participatory budgets

-0.013

Security: More policemen

0.153

Security: Fight gangs and drugs in the streets

0.212

Environment: Cleaning the district / Garbage trucks

0.027

Environment: More green areas

-0.073

Environment: Recycling of solid residues

-0.010

Institutional: Transparency in procedures

-0.020

Institutional: Modernize procedures

-0.029

Infrastructure: Markets, public buildings

-0.052

Social: Children and elderly programs, school lunches, etc.

-0.027

Social: work for the poor

-0.022

Housing: titling,

-0.036

59