Electoral Rules, Strategic Voting, Political Competition and Fiscal Performance: Evidence from Brazilian Municipalities§

Marcos Chamon†, João M P De Mello‡ and Sergio Firpo¥

Abstract

In this paper we investigate whether political competition at municipal level has a causal impact on fiscal outcomes. In order to establish causality we exploit a discontinuity in the Brazilian voting system. In cities with less than 200 thousand voters, mayors are elected with a plurality of the vote. In cities with more than 200 thousand voters, a run-off election takes place among the top two candidates if neither achieves a majority of the votes. At a first stage, we establish that tworound elections increase the number of candidates/political fragmentation. At a second stage, we use the discontinuity as a source of exogenous variation to infer causality from political competition to fiscal policy. While our first step results show that voters do not vote sincerely, causing party competition, our second stage results suggest that political competition induces more investment and less current expenses at the municipal level. KEY

WORDS:

Electoral

Systems;

Strategic

Voting;

Political

Competition; Regression Discontinuity; Fiscal Outcomes. JEL CODES: H72;D72;C14;P16

§

We would like to thank … International Monetary Fund [email protected] The views expressed in this paper are those of the authors and should not be attributed to the International Monetary Fund. ‡ Departamento de Economia, PUC-Rio: [email protected] ¥ Departamento de Economia, PUC-Rio: [email protected] †

1

I. Introduction Electoral rules have strong implications for party-formation and for how the political process plays out. For example, a plurality voting rule favors a two-party system(“Duverger’s Law”, Duverger 1954).1 Since electoral rules affect party formation/electoral competition, we should expect them to affect policy outcomes. This paper analyzes these implications in the context of municipal elections in Brazil. Voting is mandatory in Brazil.2 Mayoral elections take place every four years, with the election rules varying depending on the size of the electorate. In a city with more than 200,000 registered voters elections are in a two-round system. A run-off between the first-round winner and the runner-up takes places if the former receives less than 50% of valid votes.3 Otherwise there is only one round and the candidate that receives more votes is declared the winner, which, by “Duverger’s Law”, should lead to two candidate races. For example, suppose 60 percent of the electorate is left-leaning. If there is one leftleaning and one right-leaning party contesting the election, the former should easily win the election. But if there are two competing left-leaning parties, the right-leaning one may be able to achieve a plurality of the vote. In a two-round election, the presence of the third candidate should not affect the outcome. But with a plurality rule, the third candidate would be a “spoiler,” and in a well functioning system, the two left-leaning parties would collide and launch a single candidate.4 Our results point to a steep increase in party fragmentation once the 200 thousand voter threshold is crossed, with the vote share of the two main candidates declining by over 15 percent around that threshold. While documenting this party concentration effect is interesting on itself, our main set of results focuses on the implications for fiscal policy. One natural variable to focus on is the size of the government. However, the vast majority of expenditures in small Brazilian 1

For a formal proof of this result, please refer to Palfrey (1989). The electorate is composed of three groups. First, all citizens between 18 and 64 years are automatically registered, and voting is mandatory for registered voters. Second, between 16 and 18 registering in optional, but voting is mandatory once registered. Finally, voting is optional for registered voters older than 64 years. 3 Votes are valid if they are for a candidate or blank. A third category, null votes, and not considered valid. 4 Note that the presence of a run-off would not necessarily rule-out a right-leaning party victory in this example. Suppose there are four left-leaning parties each of which receives 15 percent of the vote, and two right-leaning parties that receive 20 percent of the vote. Then, a run-off would take place with the two right-leaning parties. 2

2

municipalities is financed by transfers from the federal government [QUANTIFY], making it exogenous to the municipal-level political process. As a result, we turn our attention to the composition of expenditures, namely the breakdown between investment and current expenditures. On the one hand, political fragmentation can lead to more current expenditures, as more players need to be co-opted by the political process. But on the other hand, the competitive pressures associated with lower political entry-costs can induce better fiscal policies by the incumbents. We show the latter effect dominates, and higher party fragmentation is associated with more investment and less current expenditures. A one percentage point increase in the share of votes for the 3rd or lower placed candidates increases the share of investment by 0.3 percent. Previous papers have documented the link between electoral rules and fiscal policy outcomes across countries. Persson and Tabellini (2004) show that presidential regimes and majoritarian rules lead to smaller governments than parliamentary regimes and proportional representation. Majoritarian rules also tilt the composition towards less transfer expenditures than proportional representation. This last result was also presented and formalized in Milesi-Ferretti et al (2002). While these studies are very careful when trying to identify the effects of the electoral rules, they suffer from the endogeneity problems inherent in this type of cross-country exercise. Our results contradict those findings, by showing that plurality rules lead to better fiscal performance, assuming investment is better than current expenditures. Our municipal setting provides far sharper identification than these cross-country studies, but faces an “external validity” critique (e.g. the finding may be applicable to small Brazilian municipalities, but not elsewhere). III. The 1st Stage: How does the 2nd Round Affect Political Competition? III.A Data and Preliminary Evidence III.A.1 Data We use two sources of data. Election data is published by the Tribunal Superior Eleitoral, the federal-level electoral authority. Election results, as well the number of registered voters, are available for a total of 16,498 races over three election cycles: 1996, 2000 and 2004. Data for the 1992 elections are not available. Fiscal data comes from the Secretaria do Tesouro Nacional, the National Treasury, which is subordinated to the 3

Ministry of Finance. From the Tesouro we have annual data on current spending of all Brazilian municipalities for the 1996-2005 period.

In the first semester of the election year the state-level electoral authority counts the number of registered voters per city to define where there will be the possibility of a second round.5 The first round takes place sometime in the beginning of October, and the second round in sometime between the end of October and mid November.6 Where there is only one round, it takes place the same day as the first round. Four mayoral elections have been held under these rules: 1992, 1996, 2000 and 2004. No other significant changes in the rules took place, except for the possibility of reelection (once) beginning in 2000. III.A.2 Descriptive Statistics and Preliminary Graphical Evidence

Table I contains some descriptive statistics on the size of the electorate.

5

The electoral authority rests with the Electoral Justice System, which is composed of a federal entity, Tribunal Superior Eleitoral (TSE), and 27 state entities, the Tribunais Eleitorais Regionais (TREs). The TSE creates regulation for presidential, gubernatorial and mayoral elections, and administer presidential elections. The TREs are responsible for the execution of gubernatorial and mayoral elections. Among the executive tasks are registering voters, resolving litigation among candidates, enforcing electoral legislation, and running the actual voting process. 6 Footnote with the precise dates for the 1996, 2000 and 2002 elections.

4

The sample is mainly composed of small cities: 95% of races had less than 54,000 voters. Out of 16,498 races, 256 took place in cities with more 150,000 voters, and 123 between 150,000 and 250,000 voters. Although there are few cities around the discontinuity threshold – 200thd voters, we do have enough observations to perform our analysis. Not surprisingly, however, precision is not very high is some cases.

5

Descriptive statistics summarize the background of the first stage. First, size of electorate and the number of candidates are positively related, which is expected as the size of the political market induces entry. Figure I shows that this relationship is concave, again as expected.

Number of Candidates and HHI vs Electorate Size

0

2500

3000

3500

4000 HHI

Number of Candidates 5 10

4500

5000

15

Fractional Polynomial Fit - 95% Confidence Interval

0

2000000

4000000 Electorate

6000000

8000000

The number of candidates increases considerably around the discontinuity threshold: from an average of 4.63 in cities whose electorate is between 150thd and 200thd, to 5.33 in cities with electorate between 150thd and 200thd. Same pattern arise for the median. This is particularly interesting given the overall concave relationship between electoral size and number candidates depicted in figure I. Voting concentration across candidates as measured by the Herfindahl-Hirschman Index (HHI) follows a similar pattern. Concentration falls monotonically with electoral size. A large drop in HHI arises around the discontinuity point 200thd, again interestingly given the fact that HHI falls with electorate at rapidly decreasing rates. However, this is not at

6

all surprising since HHI is a decreasing convex function of the number of candidates. In fact, HHI mirrors the number of candidates as a function of electorate. More interestingly, the same result arises when concentration is measured by the percentage of votes received by the 3rd placed candidate or lower. Considering only races with more than 2 candidates, this percentage rises from 18.11% in cities with electorate size between 150thd and 200thd to 23.11% in cities with electorate size between 200thd and 250thd. Again, both are significantly higher than in the whole sample (14.73% on average) We provide some preliminary graphical evidence of the behavior of concentration around the discontinuity threshold. Imbens and Lemieux [2007] propose a histogram-type procedure. Let c = 200thd. We construct 8 bins by dividing the [100thd,200thd] interval into five mutually exclusive equal-sized sub-intervals of width 20thd, and by dividing the [200thd,300thd] interval into three intervals: (200thd,225thd], (225thd,255thd] and (255thd,300thd]. The asymmetry is due to the rapidly decreasing number of observations for larger electorate sizes. The larger bin width in the (100thd,200thd] interval guarantees at least 20 observations per bin. For each bin, we compute the average HHI and % of votes received by the 3rd or lower placed candidates, and attribute this number for all values in the bin width. Figure II shows the result.

7

Figure II: Behavior Around 200thd Political Outcomes %Votes for Candidates placed 3rd or lower

100000

17.2

3500

18.4

3600

19.6

3700

20.8

HHI

3800

22

23.2

3900

24.4

4000

HHI

150000

200000 250000 Electorate Size

300000

100000

150000

200000 250000 Electorate Size

300000

In the [100thd, 200thd] interval, the HHI index remains roughly constant around 3930. Concentration falls by almost 10% in the (200thd,225thd] interval, and fluctuates around this level over the (200thd,300thd] interval. Results are even stronger for the % of votes received by the 3rd or lower placed candidates. Of course, these are unconditional differences. As figure I shows, the number of candidates increase with electoral size, which may drive results completely.

In sub-section III.C we provide preliminary

evidence on the conditional impact of the presence of the second round. In section V, we provide conditional evidence on the impact of second around the 200thd threshold. The last piece of preliminary evidence we show refers to the distribution of electorate, the “forcing variable”. An implicit identifying in the Sharp Regression Discontinuity Design is that the 200thd rule is in fact exogenous. This means cities do not manipulate their electorate number to be above or below the threshold.

8

Although it is hard to prove this assertion directly, one can check whether the density of electorate presents a discontinuity at 200thd. A discontinuity at 200thd would raise the suspicion that cities were manipulating the electorate size.8 [100thd,300thd] interval.

Histrogram and Estimated Density of Electorate

0

5

2.000e-06

Percentage 10

6.000e-06 Estimated Density

15

20

.00001

[100thd,300thd] interval

100000 125000 150000 175000 200000 225000 250000 275000 300000 Size of Electorate

Figure III: Kernel Density Estimate with Epanechnikov Kernel

As expected, the histogram shows that the frequency drops almost monotonically with electorate size. The bins [175000,187500] and [187500,200000]

are

slightly

higher

than would one expect, but the difference is within normal fluctuation in the figure. The figure also contains the kernel density estimate of the size of electorate. No noticeable jump occurs at or around 200thd.

8

The head count is done by the “Tribunal Regional Eleitoral”, the state-level electoral authority and not by a city-level body. Thus, it is rather inconceivable that cities would be able to manipulate the counting.

9

III.A.3 OLS Evidence 1 on the First Stage: Second Round and Political Fragmentation

While descriptive statistics suggest that the concentration of voting falls around 200thd, this is far from conclusive as evidence linking the second-round system to strategic voting and party fragmentation. One reason is the mechanical relationship between the number of candidates and the HHI. The number of candidates increases with electoral size, and concentration falls with the number of candidates.9 Here, we check whether the pattern suggested in table I remains if we control for the electoral size. We also verify if the association between size of electorate around 200thd and concentration arises for other measures, such as % of votes of the third-placed candidate in the first round, for instance. The model is:

CONCENTRATION r = β 0 + β1 DUM 200thd r +

β 3CANDIDATEsr + φ (ELECTORATE r ) + ΣYEARr + ε r

(1)

where r is electoral race, i.e., a city in an election year. The coefficient of interest is β1, associated with DUM200thdr, the treatment dummy. DUM200thdr assumes the value 1 for cities with more than 200thd voters, and zero otherwise. CANDIDATESr is the number of candidates in the first-round race. ELECTORATEr is the number of registered voters in race r. φ (•) is a flexible polynomial of ELECTORATEr, which determines the presence of the second round. Finally, YEARr is a set of election year dummies (one for 1996 and one for 2000). We use two different measures of CONCENTRATION: HHI and the percentage of votes for all candidates except the first and second placed candidates in the first round. Only races in which at least three candidates ran since there is no reason why the presence of the second round should make any difference if there are only two

9

Candidacy decisions may be endogenous to the electoral system. In a two-round system, candidacy is “cheap” n the first-round. In this case, the second-round system induces party fragmentation not only because it induces sincere voting, but because it induces more candidacy. See next section for further discussion.

10

first-round candidates.10 The function φ (•) includes the logarithm and six powers of ELECTORATE . Table II shows the results.

Results in table are further indicative that something happens around 200thd. Start with panel A, which shows regressions whose dependent variable is the HHI. When the whole sample of races with more than two candidates is considered, no break in concentration in 10

Races with only one candidate were also excluded.

11

intercept arises. On the other hand, when the sample is restricted to cities around the threshold point, a large shift in intercept arises. Consider the sub-sample of cities with electorate sizes between 125thd and 275thd voters. The estimated coefficient on HHI is DUM200thd is -0.16, meaning that, after controlling for a flexible polynomial of ELECTORATE and the number of candidates, voting in races in two-round cities is on 16% less concentrated than in first-round cities. The difference is even larger when the 175thd-225thd interval is considered suggesting the difference is in fact due to the threshold around 200thd. Very similar results arise with the percentage of votes of all candidates placed third or lower, a dependent variable is similar but not equal to HHI.11 Finally, the difference is not an artificial misspecification result. In the last column, the treatment dummy is replaced by a faux treatment dummy as if elections in cities with electorates between 150thd-200thd were in the second round system. Results disappear.

III.B A Regression Discontinuity Approach

Results in last section are suggestive of an impact of the presence of the second round on political and fiscal outcomes. In this section, we estimate non-parametrically the causal effect of the presence of the second round. We implement two different procedures: a flexible polynomial of electorate, and local regression around the discontinuity.

III.B.1 Flexible Polynomial Procedure

The procedure consists on estimating the following two models:

11

Since we only consider races with more than two candidates, there is no direct link between the number of candidates and the percentage of votes for third and lower placed candidates. 14 The model was also estimated with quadratic and linear φs. Results are very similar and we omit them.

12

CONC r− = β 0− + β1− Candidates r + φ − (ELECTORATE r ) + Σ −YEARr + ε r for ELECTORATE ∈ [200thd − d ,200thd )

(2)

and CONC r+ = β 0+ + β1+ Candidates r + φ + (ELECTORATE r ) + Σ +YEARr + ε r for ELECTORATE ∈ [200thd ,200thd + d )

(3)

where r is a race (pair city-election year) i is a city. CONC dependent variables capture concentration of voting across candidates, such as the HHI index across candidates, and the percentage of votes received by third or lower placed candidates. Candidates and YEAR are as defined in model (1). φ is a fractional polynomial of electorate, the forcing variable.14 For comparison, we estimate (2) and (3) for three different subsamples, d = 25thd, and 75thd. The impact of the presence of the second round is measured by assessing the size of the discontinuity of f (ELECTORATE ) = βˆ0 + φˆ(ELECTORATE ) + Σˆ X r evaluated at 200thd. In other words, the impact is:

βˆ1, PO = CONC + (200thd ) − CONC - (200thd )

(4)

Finally, we check whether some discontinuity is present at 150thd, where there is no reason for concentration to jump. This placebo evaluates if the estimated discontinuity is an artifact of our specification.

III.B.2 Local Linear Regression Procedure

Given the nature of the problem we have, that is, estimation of a local effect (around the discontinuity of 200thd voters), we also used a method that puts more weight into observations that are closer to the discontinuity point. Note that the previous estimated polynomials provide a global approximation of the functions of interest but not a local

13

approximation. Thus, we proceed as in Hahn, Todd and van der Klaauw (2001) who propose estimation using local linear regressions. We first create two variables: uˆ p , which is simply the residuals from the regression of

POLITICAL on Candidates and dummies of election year. We can then regress uˆ p on

electorate but using local linear regressions for both sides of the discontinuity separately. Define Xi = ELECTORATEi – 200K. Let a + and a − be the intercepts of the following regressions::

2 N ⎡a + ⎤ ⎞ ⎛ ˆ¨ 1{X i > 0}× ⎜ u p,i − a − b × ( X i )⎟ ×K h ( X i ) ∑ ⎢ + ⎥ = arg min a,b ⎝ ⎠ i =1 ⎣b ⎦

(4)

2 N ⎡a − ⎤ ⎞ ⎛ ˆ¨ { } ( ) X u a b X arg min 1 0 = < × − − × ⎜ p,i ∑ ⎢ −⎥ i i ⎟ ×K h ( X i ) a,b ⎠ ⎝ i =1 ⎣b ⎦

(5)

where Kh is a kernel function (the Gaussian kernel, for example) and 1{•} is the indicator function. Now define: CONCENTRATION + = lim+ a + and CONCENTRAT ION − = lim− a − X →0

X →0

(6)

POLITICAL- by doing the same but for the 1{ELECTORATE − 200 K < 0} dummy. The non-paramentric local regression discontinuity estimate of the impact of the presence of the second round on concentration is:

βˆ1, NP = CONCENTRATION + − CONCENTRATION -

(7)

Standard errors are obtained by bootstrapping.

III.B.3. Results

14

III.B.3.A Flexible Polynomial Procedure

The flexible polynomial procedure is implemented for two political dependent variable, HHI and the 100 – (%1st + %2nd), and two political variables, total spending in 2002 and 2003. Procedure in equations (3)-(4) is implemented for d = 75thd, 50thd and 25thd. For all dependent variables, we run a placebo experiment with a faux threshold at 150thd.

Percentage of Votes for the 3rd placed candidate or lower

Figures VII-XI show the results when the dependent variable in (2) and (3) is the percentage of votes received by all but the first two placed candidates. The shaded area around the curves is a 90% confidence interval.

Figure IV: Fractional Polynomio Phi

-8

Residuals of %3rd or lower -6 -4 -2 0 2 4 6 8

10

d = 75thd

125000

150000

175000

200000 225000 250000 275000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

15

Figure V: Fractional Polynomial Phi Residuals of %3rd or lower -20 -16 -12 -8 -4 0 4 8 12 16 20

d = 25thd

175000

187500

200000 212500 225000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

Figure VI: Fractional Polynomial Phi

Residuals of %3rd or lower -10 -8 -6 -4 -2 0 2 4 6

8

10

Placebo at 150thd

100000

125000

150000 175000 200000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

16

The flexible polynomial estimates confirm the preliminary evidence in tables I and II and figure II. The HHI as a function of electorate size has a big jump at 200thd. The jump is higher the more the sample is restricted to be around. For d = 75thd, the jump is roughly 9 percentage points, and for d = 25thd the jump is 20 percentage points. Taking d = 75thd as a benchmark, the causal impact of the presence of the second round is to reduce the percentage of votes for the two first-placed candidates by 9 points. With the sample [125thd, 275thd] this represents roughly one standard deviation (see table I). The 90% confidence intervals do not intersect or intersect very little (when d= 75thd). Finally, no significant discontinuity arises at 150thd.

HHI

Figures VII-IX show the results when the log of HHI is the dependent variable in (2) and (3). The shaded area around the curves is a 90% confidence interval.

Figure VII: Fractional Polynomio Phi

-.2

Residuals of Log HHI -.12 -.04 .04

.12

d = 75thd

125000

150000

175000

200000 225000 Electorate Residuals of Log of HHI on year dummies and # of candidates.

250000

275000

17

Figure VIII: Fractional Polynomio Phi

-.4

-.2

Residuals of Log HHI 0 .2 .4

.6

d = 25thd

175000

187500

200000 212500 Electorate Residuals of Log of HHI on year dummies and # of candidates.

225000

Figure IX: Fractional Polynomio Phi

Residuals of Log HHI -.2 -.15 -.1 -.05 0 .05 .1

.15

.2

Placebo at 150thd

100000

125000

150000 175000 Electorate Residuals of Log of HHI on year dummies and # of candidates.

200000

18

Results are unchanged when the dependent variable in (2) and (3) is HHI.

III.B.3.B Local Linear Regression

Percentage of Votes for the 3rd placed candidate or lower

Figures X-XII show the results when percentage of votes of the 3rd placed candidate or lower is the dependent variable in the local regression procedure.

Figure X: Regression Discontinuity Plots

-4

Residuals of %Less than Second -2 0 2 4

6

d = 75thd

125000

150000

175000

200000 Electorate

225000

250000

275000

19

Figure XI: Regression Discontinuity Plots

-10

Residuals of %Less than Second -5 0 5

10

d = 25thd

175000

187500

200000 Electorate

212500

225000

Figure XII: Regression Discontinuity Plots

-4

-3

Residuals of %Less than Second -2 -1 0 1 2 3 4 5

6

Placebo at 150thd

100000

125000

150000 Electorate

175000

200000

20

Local regression results, if anything, are stronger than the fractional polynomio ones. Considering the sample of electorate sizes between 125thd and 275thd (d = 75thd, figure X), βˆ1, NP = CONCENTRAT ION + − CONCENTRAT ION − = 8.82 percentage points . The bootstrapped standard error is 4.10, meaning we can rejected the null hypothesis that the difference is zero at the 5% now. Considering the sample of electorate sizes between 175thd and 225thd (d = 75thd, figure XI), results are again stronger (10.02 percentage points, with a bootstrapped standard error of 5.00). Finally, no relevant discontinuity arises at 150thd, our placebo experiment.

V.B.2 LOGHHI

Figures XIII-XV show the results when HHI is the dependent variable in the local linear regression procedure.

Figure XIII: Regression Discontinuity Plots

-.1

Residuals of Log HHI -.05 0 .05

.1

d = 75thd

125000

150000

175000

200000 Electorate

225000

250000

275000

21

Figure XIV: Regression Discontinuity Plots

-.2

Residuals of Log HHI -.1 0 .1

.2

d = 25thd

175000

187500

200000 Electorate

212500

225000

Figure XV: Regression Discontinuity Plots

-.1

Residuals of Log of HHI -.05 0 .05

.1

Placebo at 150thd

100000

125000

150000 Electorate

175000

200000

22

Results are again very similar to those from the flexible polynomial procedure. Figures XIII and XIV imply a 15% to 18% reduction in concentration associated with the second round. Bootstrapped standard errors are 8% and 9%, respectively. Again no significant discontinuity arises at 150thd (figure XV).

IV. The 2nd Stage: Using Second Round as Exogenous Variation to Estimate the Causal Impact of Political Competition on Fiscal Outcomes

Section III showed that the presence of the second round induces more first-round electoral competition. In this section we use this variation to estimate the causal impact of political competition on fiscal outcomes. Let i be a city and t an year in an administration cycle, and τ is the last period of the administration cycle, i.e., the election year. Hence

τ ≥t. We are interested in estimating the following structural equation:

FISCALit = β 0 + β1 E [POLITICAL COMPETITION iτ | t ] + ΒX it + ε it

(8)

where β1 is the causal, structural effect of the expected POLITICAL COMPETIION at the election year has on FISCAL outcome variables at the year t is the administration cycle preceding elections at year τ. X it is a vector of controls, which in our case is polynomial of the size of electorate. Before further elaborating on the empirical strategy to recover the structural parameter β1, we make two digressions. First, we outline the theoretical reasons why political competition impacts fiscal variables. Second. we argue that we need an identification strategy beyond estimating (??) by OLS, i.e., why E [ε it POLITICAL COMPETION it ] ≠ 0 , and therefore a simple OLS estimation strategy would fail to recover the structural, causal impact of POLITICAL COMPETIION on FISCAL.

23

IV.A Theoretical Reasons Why Political Competition Causes Fiscal Outcomes

The political economy literature contains numerous reasons why political competition causes fiscal outcomes. The following example inspired in Rogoff [1990] is illustrative. Assume, as it is common in the literature, that policy makers maximize an objective function composed of both social welfare and the probability of being in office.15 Perfectly foresighted voters want to elect the candidate (incumbent or challengers) with most “managerial ability”, say, as the capacity to deliver public goods at the lowest cost. However, they do not observe ability, only the result of fiscal policy. In this setting, less than completely benevolent incumbents have an incentive to spend away from the socially optimal level to signal ability and consequently influence voters’ decisions in their favor. Consider now admittedly extreme situation in which the incumbent anticipates no competition, i.e., no challenger will run. In this case there is no reason to distort fiscal policy. Suppose, at the other extreme, that competition is infinite, in the sense that the probability of reelection is zero. Again, there is no reason to distort fiscal policy. In general, suppose that N equally competitive candidates are running for office against an incumbent.16 In general, the propensity to distort fiscal policy will depend on N.

IV.B Why is E[ε it POLITICAL COMPETION it ] ≠ 0 ?

There are at least two reasons why the structural parameter β1 in (8) cannot be estimated by

Ordinary

Least

Squares

(OLS):

the

joint

determination

of

E [POLITICAL COMPETITION iτ | t ] and FISCALit , and measurement error.

15 16

Assumptions about voters’ behavior are From a theoretical perspective, it is only necessary that one of the candidates is from the situation.

24

Consider the political economy of fiscal policy. Again assume that policy makers maximize an objective function composed of both social welfare and the probability of being in office. With rational, forward-looking but imperfectly informed voters, the equilibrium fiscal policy is distorted away from the social optimum. From the incumbent’s perspective, the reason to distort fiscal policy away from the social optimum is precisely because manipulating fiscal policy changes the probability of reelection. Affecting the probability of reelection is only relevant if there is competition for office. By manipulating fiscal policy, incumbents can: 1) affect entry decisions by potential candidates; and 2) weaken the case of actual entrants. This causes OLS bias for the following reason. Consider the following example above. Thus, an increase in N, for example, will change the expected level of electoral competitiveness, affecting the equilibrium fiscal policy. This is our β1. Now the new equilibrium fiscal policy will in turn change political competitiveness through the channels described above, which in turn will change incumbent’s incentives, and so on. In other words, expected political competition and fiscal outcomes are jointly determined by equation (8) and the following equation:

E [POLITICAL COMPETITION iτ | t ] = γ 0 + γ 1FISCALit + ΓZ it + uit

(9)

There is an additional, econometric, reason for E [ε it POLITICAL COMPETION it ] to be different from zero. The variable POLITICAL is measured with error by construction. The relevant level of POLITICAL COMPETITION variable is the incumbent’s expectation of how competitive the political environment will be in next elections. Save

rare circumstances, this variable is unobservable.17 The alternative taken by the literature on political cycles (which we emulate) is using actual, realized political competition. In other words:

17

It is conceivable that one could use opinion polls during the administration cycle. However, these polls are not conducted at a sufficient number of cities to implement any quantitative empirical procedure.

25

POLITICAL COMPETITION iτ = E [POLITICAL COMPETITION iτ | t ] + υit

where υit is uncorrelated with E [POLITICAL COMPETITION iτ | t ] . In this case, measurement error causes attenuation bias.

IV.C The Empirical Strategy and Results

Last section shows that estimating the structural parameter β1 consistently involves finding exogenous variation in political competition. Section III showed that , DUM 200thd a dummy for cities with electorate above 200thd, belongs to Z it in equation (9). In other words, the first-stage regression establishes convincingly that the presence of the second round does increase political competition. Under the identifying assumption that DUM 200thd only impacts fiscal policy through its effect on political competition, DUM 200thd is a source of exogenous variation to estimate β1. Therefore we estimate (8) using DUM 200thd as an instrument. We implement a weighted two stage least squares. Observations from cities with electorate size between 125thd and 275thd receive weights inversely proportional to the difference between electorate size and 200thd. All other observations are discarded ˆ are defined as the (receive zero weight). More precisely, our estimates βˆ0 , βˆ1 and Β solution to the following system of equations:

26

∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit )DUM 200thd it = 0 ∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit ) = 0 ∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit )Xit = 0

where Wit =

1 . ELECTORATEit − 200thd

Two fiscal dependent variables are considered: the log of investment as a share of total spending and the log of current spending as a share of total spending. Since yearly data is quite noisy, the dependent variables are the total share over the administration cycle. Two cycles are considered: 1997-2000 and 2001-2004. Data from the 1993-1994 is rather unreliable and are discarded.18 Finally, political competition is measured by the two firststage concentration measures in section III: the Herfindahl-Hirschman Index and the share of votes for the third placed candidate or lower. Results are in table III.

18

Data from the pre-stabilization period (pre July1994) is quite erratic, TO BE COMPLETED…

27

Table III shows two facts. Political competition increases the share of investment in total spending, and reduces the share of current expenditures. This is true regardless of the measure of political competition. Consider column (1). A one-percent increase in the share of votes for the 3rd placed candidate or lower causes a 0.322 increase in the share of investment in public spending. On average, the share of all but the winner and runner-up is 18%, with a standard deviation of 12 percentage points. Half a standard (6pp) represents 33% of the mean (18pp). Hence, a half-standard-deviation increase staring from the mean causes a 13% increase in the share of investment. When log HHI is used as a regressor, the impact is some 25% increase in the share of investment. In column (2), only the 1997-2000 cycle is considered. Some 20% of the observations for 2003 and 2004 are missing, which raises suspicion on selection problems. If results were markedly different if only 1997-2000 is considered, we would be concerned. They are not. In column (3) we present the OLS results for comparison. Under OLS, political competition seemingly does not have an effect on investment. Besides being compatible with the hypothesis that political competition falls with investment, this results is

28

compatible with the attenuation bias caused by the measurement error of political competition.

V. Discussion

TO BE COMPLETED

VI. Conclusion

TO BE COMPLETED

References

Hahn, J., P. Todd and W. Van der Klaauw, “Identification and Estimation of Treatment Effects with Regression Discontinuity Design,” Econometrica, Vol. 69, pp. 201-209, 2001. Imbens, G. and T. Lemieux, “Regression Discontinuity Designs: a Guide to Practice,” NBER Working Paper 13039, 2007. Rogoff, K., “Equilibrium Political Budget Cycles,” American Economic Review, Vol. 80, pp. 21-36, 1990.

29

Marcos Chamon†, João M P De Mello‡ and Sergio Firpo¥

Abstract

In this paper we investigate whether political competition at municipal level has a causal impact on fiscal outcomes. In order to establish causality we exploit a discontinuity in the Brazilian voting system. In cities with less than 200 thousand voters, mayors are elected with a plurality of the vote. In cities with more than 200 thousand voters, a run-off election takes place among the top two candidates if neither achieves a majority of the votes. At a first stage, we establish that tworound elections increase the number of candidates/political fragmentation. At a second stage, we use the discontinuity as a source of exogenous variation to infer causality from political competition to fiscal policy. While our first step results show that voters do not vote sincerely, causing party competition, our second stage results suggest that political competition induces more investment and less current expenses at the municipal level. KEY

WORDS:

Electoral

Systems;

Strategic

Voting;

Political

Competition; Regression Discontinuity; Fiscal Outcomes. JEL CODES: H72;D72;C14;P16

§

We would like to thank … International Monetary Fund [email protected] The views expressed in this paper are those of the authors and should not be attributed to the International Monetary Fund. ‡ Departamento de Economia, PUC-Rio: [email protected] ¥ Departamento de Economia, PUC-Rio: [email protected] †

1

I. Introduction Electoral rules have strong implications for party-formation and for how the political process plays out. For example, a plurality voting rule favors a two-party system(“Duverger’s Law”, Duverger 1954).1 Since electoral rules affect party formation/electoral competition, we should expect them to affect policy outcomes. This paper analyzes these implications in the context of municipal elections in Brazil. Voting is mandatory in Brazil.2 Mayoral elections take place every four years, with the election rules varying depending on the size of the electorate. In a city with more than 200,000 registered voters elections are in a two-round system. A run-off between the first-round winner and the runner-up takes places if the former receives less than 50% of valid votes.3 Otherwise there is only one round and the candidate that receives more votes is declared the winner, which, by “Duverger’s Law”, should lead to two candidate races. For example, suppose 60 percent of the electorate is left-leaning. If there is one leftleaning and one right-leaning party contesting the election, the former should easily win the election. But if there are two competing left-leaning parties, the right-leaning one may be able to achieve a plurality of the vote. In a two-round election, the presence of the third candidate should not affect the outcome. But with a plurality rule, the third candidate would be a “spoiler,” and in a well functioning system, the two left-leaning parties would collide and launch a single candidate.4 Our results point to a steep increase in party fragmentation once the 200 thousand voter threshold is crossed, with the vote share of the two main candidates declining by over 15 percent around that threshold. While documenting this party concentration effect is interesting on itself, our main set of results focuses on the implications for fiscal policy. One natural variable to focus on is the size of the government. However, the vast majority of expenditures in small Brazilian 1

For a formal proof of this result, please refer to Palfrey (1989). The electorate is composed of three groups. First, all citizens between 18 and 64 years are automatically registered, and voting is mandatory for registered voters. Second, between 16 and 18 registering in optional, but voting is mandatory once registered. Finally, voting is optional for registered voters older than 64 years. 3 Votes are valid if they are for a candidate or blank. A third category, null votes, and not considered valid. 4 Note that the presence of a run-off would not necessarily rule-out a right-leaning party victory in this example. Suppose there are four left-leaning parties each of which receives 15 percent of the vote, and two right-leaning parties that receive 20 percent of the vote. Then, a run-off would take place with the two right-leaning parties. 2

2

municipalities is financed by transfers from the federal government [QUANTIFY], making it exogenous to the municipal-level political process. As a result, we turn our attention to the composition of expenditures, namely the breakdown between investment and current expenditures. On the one hand, political fragmentation can lead to more current expenditures, as more players need to be co-opted by the political process. But on the other hand, the competitive pressures associated with lower political entry-costs can induce better fiscal policies by the incumbents. We show the latter effect dominates, and higher party fragmentation is associated with more investment and less current expenditures. A one percentage point increase in the share of votes for the 3rd or lower placed candidates increases the share of investment by 0.3 percent. Previous papers have documented the link between electoral rules and fiscal policy outcomes across countries. Persson and Tabellini (2004) show that presidential regimes and majoritarian rules lead to smaller governments than parliamentary regimes and proportional representation. Majoritarian rules also tilt the composition towards less transfer expenditures than proportional representation. This last result was also presented and formalized in Milesi-Ferretti et al (2002). While these studies are very careful when trying to identify the effects of the electoral rules, they suffer from the endogeneity problems inherent in this type of cross-country exercise. Our results contradict those findings, by showing that plurality rules lead to better fiscal performance, assuming investment is better than current expenditures. Our municipal setting provides far sharper identification than these cross-country studies, but faces an “external validity” critique (e.g. the finding may be applicable to small Brazilian municipalities, but not elsewhere). III. The 1st Stage: How does the 2nd Round Affect Political Competition? III.A Data and Preliminary Evidence III.A.1 Data We use two sources of data. Election data is published by the Tribunal Superior Eleitoral, the federal-level electoral authority. Election results, as well the number of registered voters, are available for a total of 16,498 races over three election cycles: 1996, 2000 and 2004. Data for the 1992 elections are not available. Fiscal data comes from the Secretaria do Tesouro Nacional, the National Treasury, which is subordinated to the 3

Ministry of Finance. From the Tesouro we have annual data on current spending of all Brazilian municipalities for the 1996-2005 period.

In the first semester of the election year the state-level electoral authority counts the number of registered voters per city to define where there will be the possibility of a second round.5 The first round takes place sometime in the beginning of October, and the second round in sometime between the end of October and mid November.6 Where there is only one round, it takes place the same day as the first round. Four mayoral elections have been held under these rules: 1992, 1996, 2000 and 2004. No other significant changes in the rules took place, except for the possibility of reelection (once) beginning in 2000. III.A.2 Descriptive Statistics and Preliminary Graphical Evidence

Table I contains some descriptive statistics on the size of the electorate.

5

The electoral authority rests with the Electoral Justice System, which is composed of a federal entity, Tribunal Superior Eleitoral (TSE), and 27 state entities, the Tribunais Eleitorais Regionais (TREs). The TSE creates regulation for presidential, gubernatorial and mayoral elections, and administer presidential elections. The TREs are responsible for the execution of gubernatorial and mayoral elections. Among the executive tasks are registering voters, resolving litigation among candidates, enforcing electoral legislation, and running the actual voting process. 6 Footnote with the precise dates for the 1996, 2000 and 2002 elections.

4

The sample is mainly composed of small cities: 95% of races had less than 54,000 voters. Out of 16,498 races, 256 took place in cities with more 150,000 voters, and 123 between 150,000 and 250,000 voters. Although there are few cities around the discontinuity threshold – 200thd voters, we do have enough observations to perform our analysis. Not surprisingly, however, precision is not very high is some cases.

5

Descriptive statistics summarize the background of the first stage. First, size of electorate and the number of candidates are positively related, which is expected as the size of the political market induces entry. Figure I shows that this relationship is concave, again as expected.

Number of Candidates and HHI vs Electorate Size

0

2500

3000

3500

4000 HHI

Number of Candidates 5 10

4500

5000

15

Fractional Polynomial Fit - 95% Confidence Interval

0

2000000

4000000 Electorate

6000000

8000000

The number of candidates increases considerably around the discontinuity threshold: from an average of 4.63 in cities whose electorate is between 150thd and 200thd, to 5.33 in cities with electorate between 150thd and 200thd. Same pattern arise for the median. This is particularly interesting given the overall concave relationship between electoral size and number candidates depicted in figure I. Voting concentration across candidates as measured by the Herfindahl-Hirschman Index (HHI) follows a similar pattern. Concentration falls monotonically with electoral size. A large drop in HHI arises around the discontinuity point 200thd, again interestingly given the fact that HHI falls with electorate at rapidly decreasing rates. However, this is not at

6

all surprising since HHI is a decreasing convex function of the number of candidates. In fact, HHI mirrors the number of candidates as a function of electorate. More interestingly, the same result arises when concentration is measured by the percentage of votes received by the 3rd placed candidate or lower. Considering only races with more than 2 candidates, this percentage rises from 18.11% in cities with electorate size between 150thd and 200thd to 23.11% in cities with electorate size between 200thd and 250thd. Again, both are significantly higher than in the whole sample (14.73% on average) We provide some preliminary graphical evidence of the behavior of concentration around the discontinuity threshold. Imbens and Lemieux [2007] propose a histogram-type procedure. Let c = 200thd. We construct 8 bins by dividing the [100thd,200thd] interval into five mutually exclusive equal-sized sub-intervals of width 20thd, and by dividing the [200thd,300thd] interval into three intervals: (200thd,225thd], (225thd,255thd] and (255thd,300thd]. The asymmetry is due to the rapidly decreasing number of observations for larger electorate sizes. The larger bin width in the (100thd,200thd] interval guarantees at least 20 observations per bin. For each bin, we compute the average HHI and % of votes received by the 3rd or lower placed candidates, and attribute this number for all values in the bin width. Figure II shows the result.

7

Figure II: Behavior Around 200thd Political Outcomes %Votes for Candidates placed 3rd or lower

100000

17.2

3500

18.4

3600

19.6

3700

20.8

HHI

3800

22

23.2

3900

24.4

4000

HHI

150000

200000 250000 Electorate Size

300000

100000

150000

200000 250000 Electorate Size

300000

In the [100thd, 200thd] interval, the HHI index remains roughly constant around 3930. Concentration falls by almost 10% in the (200thd,225thd] interval, and fluctuates around this level over the (200thd,300thd] interval. Results are even stronger for the % of votes received by the 3rd or lower placed candidates. Of course, these are unconditional differences. As figure I shows, the number of candidates increase with electoral size, which may drive results completely.

In sub-section III.C we provide preliminary

evidence on the conditional impact of the presence of the second round. In section V, we provide conditional evidence on the impact of second around the 200thd threshold. The last piece of preliminary evidence we show refers to the distribution of electorate, the “forcing variable”. An implicit identifying in the Sharp Regression Discontinuity Design is that the 200thd rule is in fact exogenous. This means cities do not manipulate their electorate number to be above or below the threshold.

8

Although it is hard to prove this assertion directly, one can check whether the density of electorate presents a discontinuity at 200thd. A discontinuity at 200thd would raise the suspicion that cities were manipulating the electorate size.8 [100thd,300thd] interval.

Histrogram and Estimated Density of Electorate

0

5

2.000e-06

Percentage 10

6.000e-06 Estimated Density

15

20

.00001

[100thd,300thd] interval

100000 125000 150000 175000 200000 225000 250000 275000 300000 Size of Electorate

Figure III: Kernel Density Estimate with Epanechnikov Kernel

As expected, the histogram shows that the frequency drops almost monotonically with electorate size. The bins [175000,187500] and [187500,200000]

are

slightly

higher

than would one expect, but the difference is within normal fluctuation in the figure. The figure also contains the kernel density estimate of the size of electorate. No noticeable jump occurs at or around 200thd.

8

The head count is done by the “Tribunal Regional Eleitoral”, the state-level electoral authority and not by a city-level body. Thus, it is rather inconceivable that cities would be able to manipulate the counting.

9

III.A.3 OLS Evidence 1 on the First Stage: Second Round and Political Fragmentation

While descriptive statistics suggest that the concentration of voting falls around 200thd, this is far from conclusive as evidence linking the second-round system to strategic voting and party fragmentation. One reason is the mechanical relationship between the number of candidates and the HHI. The number of candidates increases with electoral size, and concentration falls with the number of candidates.9 Here, we check whether the pattern suggested in table I remains if we control for the electoral size. We also verify if the association between size of electorate around 200thd and concentration arises for other measures, such as % of votes of the third-placed candidate in the first round, for instance. The model is:

CONCENTRATION r = β 0 + β1 DUM 200thd r +

β 3CANDIDATEsr + φ (ELECTORATE r ) + ΣYEARr + ε r

(1)

where r is electoral race, i.e., a city in an election year. The coefficient of interest is β1, associated with DUM200thdr, the treatment dummy. DUM200thdr assumes the value 1 for cities with more than 200thd voters, and zero otherwise. CANDIDATESr is the number of candidates in the first-round race. ELECTORATEr is the number of registered voters in race r. φ (•) is a flexible polynomial of ELECTORATEr, which determines the presence of the second round. Finally, YEARr is a set of election year dummies (one for 1996 and one for 2000). We use two different measures of CONCENTRATION: HHI and the percentage of votes for all candidates except the first and second placed candidates in the first round. Only races in which at least three candidates ran since there is no reason why the presence of the second round should make any difference if there are only two

9

Candidacy decisions may be endogenous to the electoral system. In a two-round system, candidacy is “cheap” n the first-round. In this case, the second-round system induces party fragmentation not only because it induces sincere voting, but because it induces more candidacy. See next section for further discussion.

10

first-round candidates.10 The function φ (•) includes the logarithm and six powers of ELECTORATE . Table II shows the results.

Results in table are further indicative that something happens around 200thd. Start with panel A, which shows regressions whose dependent variable is the HHI. When the whole sample of races with more than two candidates is considered, no break in concentration in 10

Races with only one candidate were also excluded.

11

intercept arises. On the other hand, when the sample is restricted to cities around the threshold point, a large shift in intercept arises. Consider the sub-sample of cities with electorate sizes between 125thd and 275thd voters. The estimated coefficient on HHI is DUM200thd is -0.16, meaning that, after controlling for a flexible polynomial of ELECTORATE and the number of candidates, voting in races in two-round cities is on 16% less concentrated than in first-round cities. The difference is even larger when the 175thd-225thd interval is considered suggesting the difference is in fact due to the threshold around 200thd. Very similar results arise with the percentage of votes of all candidates placed third or lower, a dependent variable is similar but not equal to HHI.11 Finally, the difference is not an artificial misspecification result. In the last column, the treatment dummy is replaced by a faux treatment dummy as if elections in cities with electorates between 150thd-200thd were in the second round system. Results disappear.

III.B A Regression Discontinuity Approach

Results in last section are suggestive of an impact of the presence of the second round on political and fiscal outcomes. In this section, we estimate non-parametrically the causal effect of the presence of the second round. We implement two different procedures: a flexible polynomial of electorate, and local regression around the discontinuity.

III.B.1 Flexible Polynomial Procedure

The procedure consists on estimating the following two models:

11

Since we only consider races with more than two candidates, there is no direct link between the number of candidates and the percentage of votes for third and lower placed candidates. 14 The model was also estimated with quadratic and linear φs. Results are very similar and we omit them.

12

CONC r− = β 0− + β1− Candidates r + φ − (ELECTORATE r ) + Σ −YEARr + ε r for ELECTORATE ∈ [200thd − d ,200thd )

(2)

and CONC r+ = β 0+ + β1+ Candidates r + φ + (ELECTORATE r ) + Σ +YEARr + ε r for ELECTORATE ∈ [200thd ,200thd + d )

(3)

where r is a race (pair city-election year) i is a city. CONC dependent variables capture concentration of voting across candidates, such as the HHI index across candidates, and the percentage of votes received by third or lower placed candidates. Candidates and YEAR are as defined in model (1). φ is a fractional polynomial of electorate, the forcing variable.14 For comparison, we estimate (2) and (3) for three different subsamples, d = 25thd, and 75thd. The impact of the presence of the second round is measured by assessing the size of the discontinuity of f (ELECTORATE ) = βˆ0 + φˆ(ELECTORATE ) + Σˆ X r evaluated at 200thd. In other words, the impact is:

βˆ1, PO = CONC + (200thd ) − CONC - (200thd )

(4)

Finally, we check whether some discontinuity is present at 150thd, where there is no reason for concentration to jump. This placebo evaluates if the estimated discontinuity is an artifact of our specification.

III.B.2 Local Linear Regression Procedure

Given the nature of the problem we have, that is, estimation of a local effect (around the discontinuity of 200thd voters), we also used a method that puts more weight into observations that are closer to the discontinuity point. Note that the previous estimated polynomials provide a global approximation of the functions of interest but not a local

13

approximation. Thus, we proceed as in Hahn, Todd and van der Klaauw (2001) who propose estimation using local linear regressions. We first create two variables: uˆ p , which is simply the residuals from the regression of

POLITICAL on Candidates and dummies of election year. We can then regress uˆ p on

electorate but using local linear regressions for both sides of the discontinuity separately. Define Xi = ELECTORATEi – 200K. Let a + and a − be the intercepts of the following regressions::

2 N ⎡a + ⎤ ⎞ ⎛ ˆ¨ 1{X i > 0}× ⎜ u p,i − a − b × ( X i )⎟ ×K h ( X i ) ∑ ⎢ + ⎥ = arg min a,b ⎝ ⎠ i =1 ⎣b ⎦

(4)

2 N ⎡a − ⎤ ⎞ ⎛ ˆ¨ { } ( ) X u a b X arg min 1 0 = < × − − × ⎜ p,i ∑ ⎢ −⎥ i i ⎟ ×K h ( X i ) a,b ⎠ ⎝ i =1 ⎣b ⎦

(5)

where Kh is a kernel function (the Gaussian kernel, for example) and 1{•} is the indicator function. Now define: CONCENTRATION + = lim+ a + and CONCENTRAT ION − = lim− a − X →0

X →0

(6)

POLITICAL- by doing the same but for the 1{ELECTORATE − 200 K < 0} dummy. The non-paramentric local regression discontinuity estimate of the impact of the presence of the second round on concentration is:

βˆ1, NP = CONCENTRATION + − CONCENTRATION -

(7)

Standard errors are obtained by bootstrapping.

III.B.3. Results

14

III.B.3.A Flexible Polynomial Procedure

The flexible polynomial procedure is implemented for two political dependent variable, HHI and the 100 – (%1st + %2nd), and two political variables, total spending in 2002 and 2003. Procedure in equations (3)-(4) is implemented for d = 75thd, 50thd and 25thd. For all dependent variables, we run a placebo experiment with a faux threshold at 150thd.

Percentage of Votes for the 3rd placed candidate or lower

Figures VII-XI show the results when the dependent variable in (2) and (3) is the percentage of votes received by all but the first two placed candidates. The shaded area around the curves is a 90% confidence interval.

Figure IV: Fractional Polynomio Phi

-8

Residuals of %3rd or lower -6 -4 -2 0 2 4 6 8

10

d = 75thd

125000

150000

175000

200000 225000 250000 275000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

15

Figure V: Fractional Polynomial Phi Residuals of %3rd or lower -20 -16 -12 -8 -4 0 4 8 12 16 20

d = 25thd

175000

187500

200000 212500 225000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

Figure VI: Fractional Polynomial Phi

Residuals of %3rd or lower -10 -8 -6 -4 -2 0 2 4 6

8

10

Placebo at 150thd

100000

125000

150000 175000 200000 Electorate Residuals of a regression of %3rd or lower on year dummies and # of candidates.

16

The flexible polynomial estimates confirm the preliminary evidence in tables I and II and figure II. The HHI as a function of electorate size has a big jump at 200thd. The jump is higher the more the sample is restricted to be around. For d = 75thd, the jump is roughly 9 percentage points, and for d = 25thd the jump is 20 percentage points. Taking d = 75thd as a benchmark, the causal impact of the presence of the second round is to reduce the percentage of votes for the two first-placed candidates by 9 points. With the sample [125thd, 275thd] this represents roughly one standard deviation (see table I). The 90% confidence intervals do not intersect or intersect very little (when d= 75thd). Finally, no significant discontinuity arises at 150thd.

HHI

Figures VII-IX show the results when the log of HHI is the dependent variable in (2) and (3). The shaded area around the curves is a 90% confidence interval.

Figure VII: Fractional Polynomio Phi

-.2

Residuals of Log HHI -.12 -.04 .04

.12

d = 75thd

125000

150000

175000

200000 225000 Electorate Residuals of Log of HHI on year dummies and # of candidates.

250000

275000

17

Figure VIII: Fractional Polynomio Phi

-.4

-.2

Residuals of Log HHI 0 .2 .4

.6

d = 25thd

175000

187500

200000 212500 Electorate Residuals of Log of HHI on year dummies and # of candidates.

225000

Figure IX: Fractional Polynomio Phi

Residuals of Log HHI -.2 -.15 -.1 -.05 0 .05 .1

.15

.2

Placebo at 150thd

100000

125000

150000 175000 Electorate Residuals of Log of HHI on year dummies and # of candidates.

200000

18

Results are unchanged when the dependent variable in (2) and (3) is HHI.

III.B.3.B Local Linear Regression

Percentage of Votes for the 3rd placed candidate or lower

Figures X-XII show the results when percentage of votes of the 3rd placed candidate or lower is the dependent variable in the local regression procedure.

Figure X: Regression Discontinuity Plots

-4

Residuals of %Less than Second -2 0 2 4

6

d = 75thd

125000

150000

175000

200000 Electorate

225000

250000

275000

19

Figure XI: Regression Discontinuity Plots

-10

Residuals of %Less than Second -5 0 5

10

d = 25thd

175000

187500

200000 Electorate

212500

225000

Figure XII: Regression Discontinuity Plots

-4

-3

Residuals of %Less than Second -2 -1 0 1 2 3 4 5

6

Placebo at 150thd

100000

125000

150000 Electorate

175000

200000

20

Local regression results, if anything, are stronger than the fractional polynomio ones. Considering the sample of electorate sizes between 125thd and 275thd (d = 75thd, figure X), βˆ1, NP = CONCENTRAT ION + − CONCENTRAT ION − = 8.82 percentage points . The bootstrapped standard error is 4.10, meaning we can rejected the null hypothesis that the difference is zero at the 5% now. Considering the sample of electorate sizes between 175thd and 225thd (d = 75thd, figure XI), results are again stronger (10.02 percentage points, with a bootstrapped standard error of 5.00). Finally, no relevant discontinuity arises at 150thd, our placebo experiment.

V.B.2 LOGHHI

Figures XIII-XV show the results when HHI is the dependent variable in the local linear regression procedure.

Figure XIII: Regression Discontinuity Plots

-.1

Residuals of Log HHI -.05 0 .05

.1

d = 75thd

125000

150000

175000

200000 Electorate

225000

250000

275000

21

Figure XIV: Regression Discontinuity Plots

-.2

Residuals of Log HHI -.1 0 .1

.2

d = 25thd

175000

187500

200000 Electorate

212500

225000

Figure XV: Regression Discontinuity Plots

-.1

Residuals of Log of HHI -.05 0 .05

.1

Placebo at 150thd

100000

125000

150000 Electorate

175000

200000

22

Results are again very similar to those from the flexible polynomial procedure. Figures XIII and XIV imply a 15% to 18% reduction in concentration associated with the second round. Bootstrapped standard errors are 8% and 9%, respectively. Again no significant discontinuity arises at 150thd (figure XV).

IV. The 2nd Stage: Using Second Round as Exogenous Variation to Estimate the Causal Impact of Political Competition on Fiscal Outcomes

Section III showed that the presence of the second round induces more first-round electoral competition. In this section we use this variation to estimate the causal impact of political competition on fiscal outcomes. Let i be a city and t an year in an administration cycle, and τ is the last period of the administration cycle, i.e., the election year. Hence

τ ≥t. We are interested in estimating the following structural equation:

FISCALit = β 0 + β1 E [POLITICAL COMPETITION iτ | t ] + ΒX it + ε it

(8)

where β1 is the causal, structural effect of the expected POLITICAL COMPETIION at the election year has on FISCAL outcome variables at the year t is the administration cycle preceding elections at year τ. X it is a vector of controls, which in our case is polynomial of the size of electorate. Before further elaborating on the empirical strategy to recover the structural parameter β1, we make two digressions. First, we outline the theoretical reasons why political competition impacts fiscal variables. Second. we argue that we need an identification strategy beyond estimating (??) by OLS, i.e., why E [ε it POLITICAL COMPETION it ] ≠ 0 , and therefore a simple OLS estimation strategy would fail to recover the structural, causal impact of POLITICAL COMPETIION on FISCAL.

23

IV.A Theoretical Reasons Why Political Competition Causes Fiscal Outcomes

The political economy literature contains numerous reasons why political competition causes fiscal outcomes. The following example inspired in Rogoff [1990] is illustrative. Assume, as it is common in the literature, that policy makers maximize an objective function composed of both social welfare and the probability of being in office.15 Perfectly foresighted voters want to elect the candidate (incumbent or challengers) with most “managerial ability”, say, as the capacity to deliver public goods at the lowest cost. However, they do not observe ability, only the result of fiscal policy. In this setting, less than completely benevolent incumbents have an incentive to spend away from the socially optimal level to signal ability and consequently influence voters’ decisions in their favor. Consider now admittedly extreme situation in which the incumbent anticipates no competition, i.e., no challenger will run. In this case there is no reason to distort fiscal policy. Suppose, at the other extreme, that competition is infinite, in the sense that the probability of reelection is zero. Again, there is no reason to distort fiscal policy. In general, suppose that N equally competitive candidates are running for office against an incumbent.16 In general, the propensity to distort fiscal policy will depend on N.

IV.B Why is E[ε it POLITICAL COMPETION it ] ≠ 0 ?

There are at least two reasons why the structural parameter β1 in (8) cannot be estimated by

Ordinary

Least

Squares

(OLS):

the

joint

determination

of

E [POLITICAL COMPETITION iτ | t ] and FISCALit , and measurement error.

15 16

Assumptions about voters’ behavior are From a theoretical perspective, it is only necessary that one of the candidates is from the situation.

24

Consider the political economy of fiscal policy. Again assume that policy makers maximize an objective function composed of both social welfare and the probability of being in office. With rational, forward-looking but imperfectly informed voters, the equilibrium fiscal policy is distorted away from the social optimum. From the incumbent’s perspective, the reason to distort fiscal policy away from the social optimum is precisely because manipulating fiscal policy changes the probability of reelection. Affecting the probability of reelection is only relevant if there is competition for office. By manipulating fiscal policy, incumbents can: 1) affect entry decisions by potential candidates; and 2) weaken the case of actual entrants. This causes OLS bias for the following reason. Consider the following example above. Thus, an increase in N, for example, will change the expected level of electoral competitiveness, affecting the equilibrium fiscal policy. This is our β1. Now the new equilibrium fiscal policy will in turn change political competitiveness through the channels described above, which in turn will change incumbent’s incentives, and so on. In other words, expected political competition and fiscal outcomes are jointly determined by equation (8) and the following equation:

E [POLITICAL COMPETITION iτ | t ] = γ 0 + γ 1FISCALit + ΓZ it + uit

(9)

There is an additional, econometric, reason for E [ε it POLITICAL COMPETION it ] to be different from zero. The variable POLITICAL is measured with error by construction. The relevant level of POLITICAL COMPETITION variable is the incumbent’s expectation of how competitive the political environment will be in next elections. Save

rare circumstances, this variable is unobservable.17 The alternative taken by the literature on political cycles (which we emulate) is using actual, realized political competition. In other words:

17

It is conceivable that one could use opinion polls during the administration cycle. However, these polls are not conducted at a sufficient number of cities to implement any quantitative empirical procedure.

25

POLITICAL COMPETITION iτ = E [POLITICAL COMPETITION iτ | t ] + υit

where υit is uncorrelated with E [POLITICAL COMPETITION iτ | t ] . In this case, measurement error causes attenuation bias.

IV.C The Empirical Strategy and Results

Last section shows that estimating the structural parameter β1 consistently involves finding exogenous variation in political competition. Section III showed that , DUM 200thd a dummy for cities with electorate above 200thd, belongs to Z it in equation (9). In other words, the first-stage regression establishes convincingly that the presence of the second round does increase political competition. Under the identifying assumption that DUM 200thd only impacts fiscal policy through its effect on political competition, DUM 200thd is a source of exogenous variation to estimate β1. Therefore we estimate (8) using DUM 200thd as an instrument. We implement a weighted two stage least squares. Observations from cities with electorate size between 125thd and 275thd receive weights inversely proportional to the difference between electorate size and 200thd. All other observations are discarded ˆ are defined as the (receive zero weight). More precisely, our estimates βˆ0 , βˆ1 and Β solution to the following system of equations:

26

∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit )DUM 200thd it = 0 ∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit ) = 0 ∑Wit (FISCALit − β 0 + β1POLITICAL COMPETITION iτ + ΒXit )Xit = 0

where Wit =

1 . ELECTORATEit − 200thd

Two fiscal dependent variables are considered: the log of investment as a share of total spending and the log of current spending as a share of total spending. Since yearly data is quite noisy, the dependent variables are the total share over the administration cycle. Two cycles are considered: 1997-2000 and 2001-2004. Data from the 1993-1994 is rather unreliable and are discarded.18 Finally, political competition is measured by the two firststage concentration measures in section III: the Herfindahl-Hirschman Index and the share of votes for the third placed candidate or lower. Results are in table III.

18

Data from the pre-stabilization period (pre July1994) is quite erratic, TO BE COMPLETED…

27

Table III shows two facts. Political competition increases the share of investment in total spending, and reduces the share of current expenditures. This is true regardless of the measure of political competition. Consider column (1). A one-percent increase in the share of votes for the 3rd placed candidate or lower causes a 0.322 increase in the share of investment in public spending. On average, the share of all but the winner and runner-up is 18%, with a standard deviation of 12 percentage points. Half a standard (6pp) represents 33% of the mean (18pp). Hence, a half-standard-deviation increase staring from the mean causes a 13% increase in the share of investment. When log HHI is used as a regressor, the impact is some 25% increase in the share of investment. In column (2), only the 1997-2000 cycle is considered. Some 20% of the observations for 2003 and 2004 are missing, which raises suspicion on selection problems. If results were markedly different if only 1997-2000 is considered, we would be concerned. They are not. In column (3) we present the OLS results for comparison. Under OLS, political competition seemingly does not have an effect on investment. Besides being compatible with the hypothesis that political competition falls with investment, this results is

28

compatible with the attenuation bias caused by the measurement error of political competition.

V. Discussion

TO BE COMPLETED

VI. Conclusion

TO BE COMPLETED

References

Hahn, J., P. Todd and W. Van der Klaauw, “Identification and Estimation of Treatment Effects with Regression Discontinuity Design,” Econometrica, Vol. 69, pp. 201-209, 2001. Imbens, G. and T. Lemieux, “Regression Discontinuity Designs: a Guide to Practice,” NBER Working Paper 13039, 2007. Rogoff, K., “Equilibrium Political Budget Cycles,” American Economic Review, Vol. 80, pp. 21-36, 1990.

29