TEXTO PARA DISCUSSÃO
No. 550
Campaign Advertising and Election
Outcomes: Quasi-Natural Experiment
Evidence from Gubernatorial
Elections in Brazil
Bernardo S. da Silveira
João Manoel Pinho de Mello
DEPARTAMENTO DE ECONOMIA
www.econ.puc-rio.br
Campaign Advertising and Election Outcomes:
Quasi-Natural Experiment Evidence from
Gubernatorial Elections in Brazil
Bernardo S Da Silveira† and João M P De Mello‡
Abstract
Despite the “minimal effects” conventional wisdom, the question of whether campaign advertising
influence elections outcome remains open. This is paradoxical because in the absence of a causal link
from advertising to candidate performance, it is difficult to rationalize the amounts spent on campaigns
in general, and on TV advertising in particular. Most studies using US data, however, suffer from
omitted variable bias and reverse causality problems caused by the decentralized market-based method
of allocating campaign spending and TV advertising. In contrast with received literature, we explore a
quasi-natural experiment produced by the Brazilian electoral legislation, and show that TV and radio
advertising has a much larger impact on election outcomes than previously found by the literature. In
Brazil, by law, campaign advertising is free of charge and allocated among candidates in a centralized
manner. Gubernatorial elections work in a runoff system. While in the first round, candidates’ TV and
radio time shares are determined by their coalitions’ share of seats in the national parliament, the two
most voted candidates split equally TV time if a second round is necessary. Thus, differences in TV and
radio advertising time between the first and second rounds are explored as a source of exogenous
variation to evaluate the impact of TV advertising on election outcomes. Estimates suggest that a one
percentage point increase in TV time causes a 0.241 percentage point increase in votes. Since TV
advertising is the most important item in campaign expenditures, this result sheds light on the more
general question of the effect of campaign spending on elections outcome.
KEYWORDS: Campaign Expenditures, Election Outcomes, Endogeneity, Quasi-Natural
Experiments
JEL CODES: D72, C33
†
‡
Department of Economics, New York University.
Departamento de Economia, PUC-Rio.
1
“More generally, for the vast subfield of voting behavior and elections, determining whether political
campaigns influence individual vote choice and election outcomes has become a Holy Grail.” Goldstein and
Ridout on the Annual Review of Political Science (2004)
I – INTRODUCTION
Political scientists, economists and policy makers have held a long interest in how political campaigns
affect voting behavior. The impact of campaign spending in general, and of electoral advertising in particular,
have attracted special attention (among many examples, see Welch (1981), Green and Krasno (1988),
Abramowitz (1991), Levitt (1994) on campaign spending, and Zaller (1996), Finkel (1993), and Goldstein and
Ridout (2004) on TV advertising). Policy implications are almost self-evident. If money buys elections, it calls
for regulation on campaign expenditure (or a more drastic centralized solution with public campaign funding),
assuming a policy goal of minimizing the influence of economic power on election results. If campaign spending
is irrelevant, then policy discussions on public financing, both in the United States and other large countries such
as Brazil, are grounded on a false premise. The effect of TV advertising is equally important, for campaign
managers and policy makers alike.
This paper estimates the impact of TV and radio advertising on the elections outcome and, in contrast
with most received literature, finds a large impact. Since the most important item on campaign spending is TV
advertising (Ansolabehere and Iyengar (1996)), the difference between campaign spending and TV advertising is
largely immaterial.1
Despite the strong policy interest, and a large body of literature, there is still no consensus on the issue.
Although there is a “minimal-effects” conventional wisdom that campaign spending and TV advertising have
little impact on elections outcome (Gerber (2004), Levitt (1994)), a paradox remains.2 If money and TV are
irrelevant, how can we rationalize the not-so-trivial sums spent on campaigns in general, and on advertising in
particular?3
1
It seems intuitive to use TV advertising as proxy for campaign spending. Political scientists, however, have done just the opposite, using
spending as a measure for advertising (see Goldstein and Ridout (2004)). This only reinforces our perception that these two variables are
one and the same.
2
Bartels (1993), for example, calls the literature on media effects “one of the most notable embarrassments of modern social science.”
3
In the 2000 election, roughly U$ 3 billion were raised and spent by candidates and party committees. How large this number is
debatable. Tullock’s puzzle posits that campaign contributions are low when compared with the federal governments’ gross investment
and consumption (some 590 billion), which is the (maximum) amount “up for grabs” by special interests. See Ansolabehere, Figueiredo
and Snyder (2003). Nevertheless, consider GM, the third largest advertiser in the US in 2006, for a sense of the magnitudes involved. GM
spent 2.2 billion dollars in measured media. US revenues were 201 billion. Thus, political campaign advertising intensity is not much
lower than GM’s. See Advertising Age, June, 25th, 2007.
Official Brazilian numbers are considerably lower. According to Transparência Brasil, an affiliated of Transparency International, total
official campaign receipts in the 2002 national elections were close to R$400 millions (roughly U$ 110 million at the 10/02 exchange
rate). However, there is evidence of significant covert campaign contribution and money rose by other illegal means. In 2005, for
example, a large scale political scandals unveiled a large scheme of funneling money from state enterprises to fund political campaigning.
See The Economist, June, 2nd, 2005 “Brazil bribery scandal: Jeffersonian democracy, tropical style.”
2
Even if campaign spending and TV advertising do impact election results, identifying its effect is far
from trivial, especially with US data. The reason lies in the decentralized, market-based method of campaign
financing, In this case, campaign funding, spending and TV advertising are a choice of political actors
(candidates and donors), and the observed quantums are equilibrium values. Starting from equilibrium, one
would expect a small effect of variations in any of these variables, otherwise it would not be an equilibrium.
Suppose one candidate, wealthy and unknown, faces a popular opponent whose support base is poor. Under
reasonable assumptions about the production function of votes as a function of advertising, the equilibrium has
the former with large disbursement strategy, and the later with a low disbursement, and votes are split. In this
case, the raw relationship between spending and votes is indeed zero, although spending influence voting.4 An
analogy could be made with an ordinary industry: using GM and FORD advertising outlays and their market
shares, one would probably find no relationships between these two variables. Why would they spend this
money?
From an econometric perspective, the decentralized allocation causes at least two serious problems for
measuring the relationship between campaign spending/TV advertising and election outcomes: reverse causality
and omitted factors.5 Reverse causality arises because stronger candidates should receive more contributions if
donors expect their money to have an impact on policy making.6 Electoral district and candidate unobserved
heterogeneity cause omitted variable bias, normally a bias away from zero. One example: more able candidates
may be better at raising money, or people may prefer to donate to more competent candidates. If ability is
unobservable to econometrician, then the campaign expenditures will capture (some of) the link from
competence to elections’ outcome, producing a bias away from zero in estimates.
Papers that attempt to measure the impact of campaign spending or TV advertising on election outcomes
using cross-sectional data are unpersuasive for several reasons. First, with pure cross-sectional data, it is hard to
account to unobserved quality heterogeneity among candidates. Gerger (1998), for example, try to control for the
quality of contenders by including biographical information, such as dummies for whether the contender
previously occupied an elected office, which capture only one possible dimension of quality. Gerber (1998) also
use contender wealth to instrument for campaign expenditures, hoping to account for reverse causality. However,
successful business men and politicians share many (unobserved) characteristics. Thus, if quality is not
convincingly controlled for to begin with, wealth will correlate with unobservable determinants of elections
outcome. Welch (1981) use demographic characteristics of the district, such as inequality and educational
attainment, arguing that they determine donations but not voting behavior, which is quite debatable.
4
This “irrelevance result” arises in formal models. See Prat (2002) for an example.
Psychologists and political scientists mention another problem, not related to the decentralized allocation: measurement. Exposure is one
necessary condition. Having sufficient cognitive ability, which is almost never observable, is another. See Goldstein and Ridout (2004)
and section II.
6
If donations are purely ideological, or preference-based, then reverse causality does not arise.
5
3
Second, it is also challenging to account for all relevant electoral district heterogeneity. A republican
candidate running in a predominantly democratic district will face difficulty in raising money and getting votes.
Or democrats in predominantly democratic district will not bother to raise any money. Either way, if the district
political inclinations are not properly controlled for, results will again be biased, with the direction of the bias
undetermined.7
Welch (1981) uses an interesting approach to account for district heterogeneity, using party affiliation
and Gerald Ford’s share of votes in the district as controls. Although these variables most certainly capture some
district preferences, they are far from perfect. Much better if the same district is observed more than once, which
demands a panel approach.
The most convincing work in the early literature combines cross-sectional and time-series data. Levitt
(1994) examines situations in which the same pair of candidates faces each other more than once in elections for
the house of representatives. A panel strategy is adopted: an observation is pair candidate/pair – election. By first
differencing the data, Levitt (1994) can account for all time-invariant unobserved heterogeneity among
candidates and electoral districts. His estimates suggest a modest effect of campaign expenditures on the
outcome of elections.
There is however one major problem with Levitt’s strategy. If candidates’ strength vary over elections,
the procedure will not it solve the problem of reverse causality. More generally, the procedure only account for
time-invariant unobserved heterogeneity, and candidates and districts’ characteristics may change significantly
over elections. Relatively common occurrences, such as charges of corruption, sexual related scandals,
immigration, or increase minority mobilization, may considerably change voters’ perceptions and candidates’
ability to raise money. If elections are sufficiently apart in time, these concerns are less damaging. With the low
frequency of US legislative elections (every two years), one cannot dismiss the charges. 8
In this paper, we use the Brazilian electoral legislation to overcome the problems of omission and reverse
causality that plague most empirical studies in the literature. TV time allocation is centralized, producing a quasinatural experiment. Gubernatorial elections work in two-round system. The runoff round happens if no candidate
reaches 50% plus one of the votes in the first round, in which case the two most voted candidates runoff to
decide the winner. For a period of time before each round, TV and radio broadcasters are obliged by law to air
candidates’ advertising, free of charge. The law determines that, before the first round, time is allocated among
candidates according to the coalitions’ share of representation at the National parliament. Before the second
round, however, the two runner-off candidates split the time equally. Hence, the difference between first and
second round TV and radio time share is a source of exogenous variation in TV advertising time, over a short
7
While the former story would push estimates away from zero, the latter would bias estimates towards zero.
Levitt (1994) attempts to address these issues by including both year dummies, and dummies indicating whether the candidate was
involved in a scandal. Year dummies only account for time specific effects, and are useful only if voters’ perceptions towards evolved in
the same manner. Scandal dummies - derived from only one source (Congressional Quarterly) - suggest there were but seven candidates
involved in scandals over the sample, which contains 633 elections and 598 candidates, a number rather small to be plausible.
8
4
period of time, to estimate the effect of TV advertising, and consequently campaign expenditures, on election
outcomes.
As a preview, the estimated impact of TV advertising time is much stronger then what received literature
suggests. Using our preferred estimate, one percentage point change in the difference in TV advertising time
shares causes a 0.241 percentage point change in the difference in vote share. For a sense of practical relevance,
between first and second rounds in the gubernatorial elections in our sample, difference in TV time shares
change on average roughly 8 percentage points, which implies a change of 1.928 percentage point in the
difference in voting. This means that changes in average TV share time alone would be enough to reverse the
results 14% of the second rounds in the sample.9
In addition to evaluating how TV advertising time impacts elections outcome, we explore the crosssection variation in our sample to investigate how this impact varies with some city-level demographics. To
the best of the authors’ knowledge, this is novelty of this paper, made possible by the wide variation in city
demographics for same election event.10 The influence of TV advertising is stronger in larger, less educated,
more urban, poorer and more unequal cities. All four “slope shifters” have the expected sign, and are significant
both statistically and practically. These findings help reconcile our result that TV advertising matters with the
“minimal effects” conventional wisdom. Brazil is an urbanized, poorly education, highly unequal middle income
country. Hence, it scores 4.5 out of 5 in the characteristics that augment the influence of TV advertising on
election outcomes. Perhaps, special conditions must be met in order for campaign spending and TV advertising
to have an impact.
The paper is organized as follows. In section II, the Brazilian electoral system is described in detail, with
emphasis on the legislation on TV time allocation. Section III describes the data and section IV outlines the
estimation strategy. Results are in section V, and section VI concludes.
II – ELECTORAL ADVERTISING IN BRAZIL
In studies using U.S. data, the empirical evaluation of the impact of electoral advertising (the most
important by-product of campaign expenditures) on elections’ outcomes face two serious methodological
problems. First, campaign expenditures and electoral advertising are related to candidates’ characteristics such as
ability, competence, or how the candidate’s platform benefits special groups at the expense of voters. Not all, if
any, of these characteristics are observable. Thus, omission of relevant covariates is a problem. Second,
candidates that have a better chance of winning to begin with are more likely to receive campaign contributions.
9
In five out of thirty-six second rounds in our sample the difference was less than 1.928.
Although we consider gubernatorial elections, the electoral unit is city. See sections II and III.
10
5
If this ex-ante electoral strength is unobservable, then a reverse causality bias arises.11 In summary, campaign
spending (and consequently TV advertising) is determined in decetralized market system, and observed
quantities are equilibrium values produced by the strategic interaction of the choices of the agents.
In Brazil, in contrast, TV and radio advertising is determined in a centralized manner. The Brazilian
electoral legislation produces an exogenous variation in TV advertising time, and allows us to circumvent the
problems of omitted variable and reverse causality that plague most studies in the literature. Before continuing, a
(very) short digression to describe the Brazilian political system.
Brazil is a presidential federal republic comprising 26 states and one federal district, where the capital
Brasília is. The executive branch of the federal government is headed by the President of Brazil, while the
legislative branch consists of two houses – the Senate and the Chamber of Deputies. In each state government, as
well as in the government of the federal district, the head of the executive branch is a governor, and the
legislative branch consists of the State Assembly. The president, the governors, the members of the Senate, the
Chamber of Deputies, and the State Assemblies are elected by direct ballot every four years for four-year terms,
except for the senators, who serve eight-year terms.12 Brazil has a multi-party system. The current number of
active political parties in the country is close to thirty. Of course, only some of such parties are relevant at the
national level – as data will show in the next section.
Gubernatorial elections work in a runoff system.13 All properly registered candidates participate on the
first round. If no candidate reaches 50% plus one votes, there is a second round, in which the two most voted
first-round candidates participate. Second-round winner is the one who receives most valid votes.14 There is a
three-week interval between the first and the second round. Table I shows the dates of first and second rounds of
the three gubernatorial elections in our sample: 1998, 2002, and 2006.
11
The task of controlling for strength is in fact more complicated than just described. Strength of candidacy unveils over the electoral
process, so controlling for ex-ante strength might not be enough. One would need a measure of strength previous to every time a donation
was made.
12
Each state has three senators, two of them coincide, and overlap with the other: at every four year cycle either one or two, but never all
three, seats are disputed.
13
Mayoral elections in cities larger than 200thd inhabitants also work on a runoff system. We concentrate on gubernatorial elections for
two reasons. First, different cities are several observations on the same pair election-state. For mayoral elections, we would have only one
observation for each pair election-city. Therefore data from mayoral elections are intrinsically noisier. Second, and more importantly, TV
advertising should have a larger impact in elections in large, geographically disperse elections. For cities, particularly smaller ones, other
means of campaign communications may be more important than TV and radio advertising. Additionally, presidential elections are also
on the runoff system, but then there is data available for only three election years.
14
Valid votes are those for any candidate, and blank votes. The other category is null votes.
6
Table I – Calendar of Gubernatorial elections in Brazil
Year
First round
Second round
1998
October, 4
th
October, 25th
2002
October, 6th
October, 27th
2006
October, 1st
October, 29th
Source: Tribunal Superior Eleitoral (TSE) and Lei N° 9.504, September 1997.
Federal law mandates that, over a period from 45 to 60 days preceding elections, part of the TV and radio
daily grids are allocated free of charge to political advertising.15 In case of runoff, political advertising is again
mandatory, but for a period of roughly two weeks before elections.16 In the first round, time is not equally
allocated among candidates. From 1998 onwards, air time was allocated among parties or coalitions according to
the following criteria: a) one third of the time equally divided among candidates; and b) two-thirds of time
proportional to the number of representatives in the Chamber of Deputies In case of a collation, the total number
of representatives in the coalition is considered.17
In case of second round, the law mandates that the time be equally split among the two runner-off
candidates. The difference in TV advertising time between rounds is explored as a source of exogenous variation
in advertising time.
III –DATA AND DESCRIPTIVE STATISTICS
Two pieces of electoral data are used, both publicly available from the national electoral authority, the
Tribunal Superior Eleitoral (TSE). First, the voting record of all candidates in three different gubernatorial
elections (1998, 2002 and 2006), at the city-level. Second, the composition of the coalitions, which allows us to
compute TV and radio advertising time for all three gubernatorial elections, according to the law described
15
From 1998 onwards, political advertising was aired over a period of 45-day that ended in the Friday before election, that is, two days
before voters go to the ballots. Gubernatorial advertising were aired on Mondays, Wednesdays and Fridays, in two blocks of 25 minutes
each, one at lunchtime and another at prime night time.
16
From 1998 onwards, political advertising is transmitted starting 48 hours after the first-round results are announced officially by the
local electoral authority (the TREs), and again end on the Friday before elections. Advertising was aired daily, in two blocks of 20
minutes each, one at lunchtime and another at prime night time.
17
There is considerable party switching in Brazil. Thus TV time allocation depends on which point of the legislature one considers. Until
the 2004 elections, the relevant representation was at the beginning of the current legislature. For the 2006 elections, representation during
the 2002 election was considered. Thus, actual representation at the time of elections is potentially different from the representation that
determined the TV time allocation. As we will see below, this poses no challenge to our identification strategy.
7
above.
18
Air time is used as a proxy for advertising exposure.19 City-level demographics are from the 2000
census.
Although we consider gubernatorial elections, the unit of observation is a city. There are two major
advantages in using cities as a unit of observation. First, for the same election, we have several observations of
voting behavior, which dramatically increases precision of estimation. Second, it provides much more variation
at the voting unit, which allows us to investigate how the impact of TV advertising varies with demographics
such as educational attainment, income, inequality, etc.
The sample consists of all gubernatorial races in which a second round was necessary, during three
elections: 1998, 2002 and 2006. In total, there are 35 gubernatorial races in the sample, which took place in 18
different states and the Distrito Federal.20 Table II contains information on the distribution of the sample, across
states and municipalities.
18
For the elections prior to 2006, we only observe the composition at the end of the elected legislature. Since representatives can switch
parties in the months between they are elected and when they are sworn in, which is the one relevant to compute TV time allocation.
Hence, we observe TV time allocation with some noise. This should not be a reason for concern since party switching in this period,
although it happens, is not such a relevant phenomenon.
19
Clearly, being aired is a necessary, but not sufficient, condition to expose candidates to TV advertising. A finer measure of exposure is
Gross Point Ratings (GRP), the number of times a TV set was turned at the time the advertising is aired. Unfortunately this information is
private to broadcasters and perhaps candidates.
20
Distrito Federal, which includes Brasília (the federal capital) and several other cities (called satellite-cities), has the same legal status as
a state. Results are unchanged if the Distrito Federal is excluded.
8
Table II – Sample Distribution
State
1998
Amapá
X
Ceará
Distrito Federal
X
Goiás
X
Maranhão
Mato Grosso do Sul
X
Minas Gerais
X
Pará
X
Paraíba
Paraná
Pernambuco
Rio de Janeiro
X
Rio Grande do Norte
Rio Grande do Sul
X
Rondônia
X
Roraima
X
Santa Catarina
São Paulo
X
Sergipe
X
†In 2006, the state of Rio de Janeiro had 92 municipalities
††In 2006, the state of Rio Grande do Sul had 466 municipalities
†††In 2002, the state of Rondônia had 52 municipalities
Source: Tribunal Superior Eleitoral (TSE)
2002
2006
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
# of municipalities
16
184
1
242
217
77
853
143
223
399
185
91†
167
467††
53†††
15
293
645
75
The sample includes cities from all 5 Brazilian regions, and includes very heterogeneous states, from
Southern European-like states Rio Grande do Sul and Paraná to the very under-developed Northeastern states
Maranhão and Ceará. Incidentally, it includes the three most populous states (Rio de Janeiro, São Paulo and
Minas Gerais). In general, heterogeneity poses challenges for estimation as it makes omitted variable bias all
more likely. In our case, the procedure accounts for all time-invariant unobserved heterogeneity (which is
incidentally all the relevant unobserved heterogeneity because the time lapse between cross-section observations
is only three weeks). In this case, heterogeneity is desirable since it improves external validity. Finally,
heterogeneity allow for estimating a different impact of advertising according to characteristics of the cities (see
section V).
Table IIII presents the distribution of Federal Chamber of Representatives by party, and by elected
legislatures (1994, 1998 and 2002).
9
Table III: Elected members of the Federal Chamber of Deputies, by party
Year
Party
PPR
PDT
PT
PTB
PMDB
PSC
PL
PPS
PFL
PMN
PRN
PP
PSB
PSD
PV
PRP
PSDB
PC do B
PPB
PRONA
PSL
PST
PSDC
Total
Herfindhal-Hirschman Index
C4
C2
Source: Tribunal Superior Eleitoral (TSE)
1994
1998
2002
51
34
50
32
107
3
13
2
89
4
1
34
15
3
1
1
63
10
0
0
0
0
0
513
1227
60%
38%
0
25
59
31
83
2
12
3
105
2
0
0
18
3
1
0
99
7
60
1
1
1
0
513
1403
68%
40%
0
21
91
26
76
1
26
15
84
1
0
0
22
4
5
0
70
12
48
6
1
3
1
513
1179
63%
34%
Some important characteristics of the Brazilian political system arise from table III. There are four main
parties al the national level: Center-right Partido da Frente Liberal (PFL)21, centrist Partido do Movimento
Democrático Brasileiro (PMDB), center-left Partido da Social Democracia Brasileira (PSDB), and leftist Partido
dos Trabalhadores (PT). Additionally, there are at least four other relevant middle sized parties (PTB, PPB,
PL and PSB), and several marginal parties. Thus the Brazilian political structure, which has not changed
significantly over the sample period, is not concentrated, which is confirmed by the low HerfindahlHirschman Index. C4 and C2 ratios, however, are somewhat high, which suggest more concentration than
low HHI and the large number of parties suggest. Therefore, first-round advertising time is neither too
concentrated nor too dispersed. This not-so-concentrated structure is ideal as an empirical setting. Were
21
PFL has been renamed Democratas (DEM) in the first semester of 2007.
10
the Brazilian system similar to the American one, i.e., bipartisan with very marginal small parties, there
would be very little variation in TV and radio time between rounds.22 Were the system very dispersed, time
differences between the first and second rounds would all be very large, and it would hard to identify the effect of
advertising on elections’ outcomes.
While TV advertising time is determined at the national level, local politics is what ultimately matters for
gubernatorial elections. Table IV presents correlations between pairs of national and state level distributions of
parliamentary seats.23
Table IV - Pairwise Correlation: Brazilian Chamber of Deputies and in the State
Assemblies†
Year
1998
1994
Amapá
Distrito Federal
Goiás
Mato Grosso do Sul
Minas Gerais
Pará
Rio de Janeiro
Rio Grande do Sul
Rondônia
Roraima
São Paulo
Sergipe
0.61
0.34
0.83
0.82
0.80
0.87
0.61
0.60
0.46
0.28
0.85
0.89
Amapá
Ceará
Distrito Federal
Mato Grosso do Sul
Pará
Paraíba
Paraná
Rio Grande do Norte
Rio Grande do Sul
Rondônia
Roraima
Santa Catarina
São Paulo
Sergipe
2002
0.48
0.68
0.63
0.75
0.87
0.70
0.87
0.73
0.62
0.80
0.55
0.86
0.91
0.83
Goiás
Maranhão
Pará
Paraíba
Paraná
Pernambuco
Rio de Janeiro
Rio Grande do Norte
Rio Grande do Sul
Santa Catarina
0.77
0.57
0.72
0.82
0.93
0.88
0.70
0.65
0.74
0.89
†: Correlation between the seat distribution of the State Assembly and the National Chamber of Deputies
Source: Tribunal Superior Eleitoral (TSE)
Data shows that TV advertising time – determined by national political strength – also reflect state-level
political strength. Although there is some variation (correlation coefficients vary from 0.28 (Roraima in 1994) to
0.93 (Paraná in 2002)), twenty-three out of thirty-six of the estimated correlations are above 0.70. Therefore, it is
difficult to identify the impact of TV using only first round variation in TV time.
Table V contains descriptive statistics on the number of candidates per gubernatorial election, and table
VI presents the first-round TV and radio times of the first and second placed candidates in elections that required
a second round.
22
In fact, there would be very little use to two rounds.
Tables A.I, A.II and A.III in appendix A contain the distribution of seats among parties in the State Assemblies of state-election year
pairs in our sample
23
11
Table V - Descriptive statistics: number of 1st
round candidates
Year
1998
2002
2006
Number of Candidates
Mean
Mean ± 1 Standard Deviation
6.23
7.86
7.90
(3.64, 8.82)
(4.90,10.81)
(5.82,9.98)
Source: Tribunal Superior Eleitoral (TSE)
The average number of first round candidates, from 6.23 to 7.90, reflects the dispersion of the Brazilian
party system. Note that the mean number of candidates minus one standard deviation is never less than 3.64,
which is important for identification because otherwise there would not be sufficient variation in TV and radio
time between rounds.
Table VI - Descriptive statistics: TV time and vote shares
Year
1998
2002
2006
Candidate‡
1st placed in the 1st
round
nd
2 placed in the 1st
round
st
1 placed in the 1st
round
nd
2 placed in the 1st
round
st
1 placed in the 1st
round
2nd placed in the 1st
round
Mean time share in
the first round of the
elections†
Mean vote share in
the first round of the
elections†
Mean vote share in
the second round of
the elections†
31%
43%
52%
27%
37%
48%
29%
41%
53%
23%
33%
47%
34%
44%
51%
21%
37%
49%
†Averages were computed attributing equal weights across state
‡ Statewide votes
Source: Tribunal Superior Eleitoral (TSE)
Summary statistics on TV time and vote share already contain the story of the paper. In the first-round,
first placed candidates have, on average, more TV time share than second placed candidates. Inspection of
columns (2) and (3) show that the difference in vote share between 1st and 2nd placed in the first round is larger
12
than that of the second round, in relative and absolute terms. Since second round TV time is equally split,
averages suggest a positive correlation between TV time and voting performance. Evidently, means can disguise
relevant heterogeneity and non-linearities that may spuriously produce the results. Estimation strategy outlined in
section IV solves these (potential) problems.
IV – EMPIRICAL STRATEGY
Identification of advertising (and campaign expenditures) effects suffer from omitted variable and
reverse causality biases. In the context of campaign expenditure, Levitt (1994) attempts to solve these problems
by first-differencing the same pair of candidates over elections, hoping to control for time-invariant unobserved
heterogeneity among candidate, such as political strength and candidate “quality”. Two questions arise. First,
campaign expenditures are still not randomly assigned to candidates. As a consequence, the question becomes
whether it is reasonable to assume that quality and political strength can be safely assumed to be constant over an
election period.
Similarly to Levitt (1994), we use the strategy of first-differencing the same pair of candidates over two
election cycles, which controls for all time-invariant unobserved heterogeneity among candidates. There are,
however, two major differences. First, in our case the two election cycles are in fact two rounds in the same
election. This is a particularly interesting feature of the data. Since candidates rematch over a three-week period,
odds are that relevant characteristics are constant. Second, but equally important, second round TV time
allocation is determined by legislation, not by choice of the participants in the political game. Therefore, the
difference in TV time between rounds is exogenously given, after controlling for candidate fixed-effects.
Therefore, the two major concerns about Levitt´s procedure – long interval between election cycles and nonrandom resource allocation – are solved.
The details of the empirical implementation are as follows. The data is organized in panel structure, in
which the cross-sectional unit if a pair election e (defined by the pair year (t) – city (i)), and the time-series unit
is a round r. All cross-sectional observations (elections) are observed twice, once for each round. For example,
the first round in the city of Santos of the 1998 gubernatorial election in the state of São Paulo is one observation.
The second round is another. Rounds r ∈ {1,2}, and election e is the set of year-city that belong to the set defined
by the cells of table II.24
In the following formulation, A and B refer to the statewide first round winner and runner-up,
respectively. Define votes _ Aer as the share of votes of the first round winner in the round r of election e. For
24
Not all cities appear three times in the sample for two reasons. First, cities are created (or extinguished) over the 1998-2006 period (see
table II). Second, and more importantly, in not all states, and consequently in not all cities, a second round was required in all three
election years.
13
example, votes _ Ae 2 is the share of votes of the first round winner in the second round of election e.
Analogously, define votes _ Ber as the share of votes of the first round runner-up in the round r of election e.
Define also TVtime _ Aer and TVtime _ Ber as the share of advertising time in round r of election e allocated to
the first round winner and runner-up, respectively.
Finally, define dif _ voteser e dif _ TVtimeer as
dif _ voteser = votes _ Aer − votes _ Ber
and
dif _ TVtimeer = TVtime _ Aer − TVtime _ Ber
Note that, while by definition,
∑ dif _ votes
e∈State
( dif _ votese1 < 0 ) .
∑ dif _ votes
e∈State
e2
e1
≥ 0 , it may be that the runner-up wins in some cities
≥ 0 can be negative, as long as the runner-up comes back in the second
round. Since (also by definition), TVtime _ Ae 2 = TVtime _ Be 2 = 0.5 , dif _ TVtempoe 2 = 0 for all e.
The goal is to investigate how the distribution of time affects the elections´ outcome. We impose a linear
relationship among the variables. The specification is
dif _ voteser = α + γ ⋅ dif _ TVtimeer + ω ⋅ round r + λ ⋅ X e + ε er
(1)
In equation (1), round r is a dummy that assumes the value 1 for second rounds, and X e is a vector of
characteristics of election e. They control for a wide range of the pair city-election year characteristics. Example
of such characteristics are (total) campaign expenditures, average candidate quality, city characteristics, election
year specific effects. Finally, ε er contain all other determinants of the voting outcomes.
γ is the parameter of interest: it captures the effect of the difference in TV time shares on the difference
in voting performance. We test the hypothesis that γ > 0. The coefficient ω captures changes between rounds in
average voting behavior. A positive ω means a “widening gap” effect: the first round winner increases her
advantage in the second round.
There are two problems in estimating equation (1). First, omission of relevant factors: not all relevant
determinants of voting behavior are included in X e . The quintessential example is the intrinsic difference in
quality between candidates. Being unobservable, this common determinant of election outcome ( dif _ votos er )
14
and TV time ( dif _ TVtempoer ), and it belongs to ε er . By first differencing the data over rounds, X e disappears
from (1), and no omission bias arises, as long as the identification hypothesis that X i is constant between rounds
is valid, a reasonable assumption given the short interval of time between rounds. In fact, after first-differencing
the data:
∆(dif _ votese ) = γ ⋅ ∆(dif _ TVtimee ) + ω ⋅ ∆(turno r ) + ∆(ε er )
(2)
where ∆ is the difference between the second and first round.
Another potential problem is reverse causality. TV time is allocated according to the party’s (or
coalition’s) representation at the Federal Congress. As long as state-level and national political power are related
(and they are) It is conceivable that (1) is jointly determined with the following equation:
dif _ TVtimeer = β + κ ⋅ dif _ voteser + λX e + ν er
(3)
It is political strength at the national level four years before elections (actually the current elected legislature) that
causes TV time. If political strength changes significantly over the course of four years, or if national level
politics were orthogonal to state-level, then first-round TV time allocation is as good as random, and one could
estimate (1) directly, without first-differencing the data. Tables IV and VI, however, suggest otherwise.
Identification is again by using the dynamics of first and second round within the same election. Reverse
causality arises because intrinsically stronger (or more able) candidates receive more first-round TV time.
Nevertheless, insofar as one controls for candidate fixed-effect (which as included in Xe) the difference in TV
time between the first and second round is not determined by the difference in voting performance. In other
words, after Xe is accounted for, dif _ voteser does not belong to (3), i.e., κ = 0. First-difference does just
that, and reverse causality becomes a lesser problem.
One problem remains, however. Judging by the first round TV shares, data suggests that first round
winners are slightly stronger than runner-ups at the national level (see table V). Therefore, TV time increases
more to weaker candidates. Although intrinsic strength is controlled for when the data is first-differenced, nonlinearities in return of TV time may affect results. This possibility is taken care of by conveniently restricting the
sample to those elections in which the runners-up had more TV time than the first-round winner.
Another potential problem arises from changes in coalitions between the first and the second rounds.
Ideally, everything else would be constant between the first and the second round, except for TV time. Clearly, at
15
least one important thing changes: the number of contestants. What happens to other contestants? Normally, they
support one of the second round contenders, either formally or informally.26 Therefore, political support is not a
variable that is stays constant over rounds (does not belong to Xe). Can political support from defeated candidates
produce bias in the results? Since it is likely to affect voting performance, the answer is yeas if political support
also systematically related to determinants of TV time. The most conceivable scenario is the following: suppose
first round runners-up are more likely to receive support from defeated, and their TV time increase
comparatively to first round winners (table V provide some weak evidence in favor of this later conjecture).
Then, increase in TV time will capture some of the political support effect.
While there is evidence to support the conjecture that defeated candidates tend to support runners-up
(intuition would suggest otherwise), we cannot dismiss it with the available data. To assess whether it could
affect the results, a robustness test is implemented. We select a sub-sample of election in which political was less
relevant, where relevance is defined as the presence of pivotal third place in the first round. Let votes _ C e1 be
the number of votes of the third placed candidate in the first round. Define
⎧1, if votos _ C i − dif _ votosi ,1 > 0
C _ pivotal e1 = ⎨
⎩0, otherwise
In elections in which C _ pivotal e1 = 1, first-round third placed candidates’ support has the
potential of changing the second round outcome. In this case, second round alliances tend to be more
important phenomenon. By restricting the sample to elections e in which C _ pivotal e1 = 0, the possibility
that changes in TV advertising time capture political support is significantly reduced.
V – RESULTS
V.I Main Results
Table VII contains the main results of the paper. The model is estimated with a set of pair election-
city dummies.27
26
By formal support we mean a situation in which parties announce an agreement that normally involves positions in an eventual future
administration. Informal support is when defeated candidates simply declare vote, usually without any form of reciprocity.
27
Since there are only two time-series observations (first and second round) results are equal to first-differencing and doing the so-called
“fixed-effect procedure” (deviation from the average over time). See Woodridge (????).
16
Table VII - Dependent Variable: Vote share_A - Vote share_B†
OLS Results (First round
only)
(1)
Election-City Dummies
Number of Observations
7925
(2)
0.019
(7.97)***
0.394
(29.87)***
0.072
(37.79)***
Yes
15850
R
0.04
0.12
Second round
Time share A - Time share B
Constant
2
0.302
(17.69)***
0.08
(25.96)***
Fixed-effects results (First and second
rounds)
(3)††
0.026
(8.12)***
0.415
(25.00)***
0.117
(47.92)***
Yes
9074
(4)†††
0.029
(9.71)***
0.241
(14.55)***
-0.013
(5.89)***
Yes
5510
(5)††††
0.018
(4.67)***
0.453
(23.15)***
0.114
(36.97)***
Yes
10340
0.16
0.07
0.16
† White-Huber t-statistics in parentheses
†† Sample restricted to elections in which C_pivote < 0
††† Sample restricted to elections in which in the first round winner had the smaller (between A and B ) share in TV advertising
time
†††† Sample restricted to elections in which in the first round winner had the higher (between A and B ) share in TV advertising
time
*** = significant at the 1% level
** = significant at the 5% level
* = significant at the 10% level
Source: Tribunal Superior Eleitoral (TSE)
Column (1) presents the estimates of model (1): only first-round data to estimate a model that relates TV
advertising time to votes directly, by Ordinary Least Squares (OLS). According to this estimates, which controls
neither for candidate fixed-effect nor does it account for reverse causality, if the difference in TV advertising
share increases by 1 percentage point, then the difference in vote share raises by 0.302 percentage point. This
parameter is very precisely estimated (p – value <<< 1%).
Column (2) contains the first version of model (2). Data is differenced between rounds, and the model is
again estimated by OLS. The effect of TV advertising is, if anything stronger (0.394 percentage points), than in
column (1). In column (3), exclude all elections in which the third candidate is pivotal. Results are similar:
0.415 percentage point, which indicates that results are nor driven by unobserved political support. In both cases,
coefficients are precisely estimated.
Notice that these results are not mechanically produced by that runners-up tend to “catch up”. In fact,
there is a “widening gap” effect: the estimated coefficient on the second round dummy is always positive and
17
highly significant statistically. In other words, the absolute difference (in percentage points) between first-round
runners-up and winners increases in the second round.
Estimates in column (2) and (3), although qualitative in line with the estimate in column (1), are
somewhat intriguing. Most theories of omission and reverse causality suggest that sign of the bias is away from
zero. The reason why first-difference estimates are larger than those of model (1) lies on non-linearities
generated by the construction of the experiment. First-round runners-up, by definition, start behind in the second
round (7 percentage points on average, see table VI). Furthermore, inspection of table VI shows that (as
expected) runner-ups have less TV share in the first round. This implies that, on average, candidates that start
behind in the race will gain TV share relative to their second-round opponents. If TV advertising has decreasing
returns to scale, non-linearities will work towards inflating the estimates.
We assess the relevance of non-linearities by estimating (2) for two different sub-samples of elections:
one of elections in which the first-round winners had less first round TV share then runners-up, and vice-versa.
Results are in columns (4) and (5). Indeed, when runners-up had more TV time (0.241, column (4)), the effect of
TV advertising is lower than in column (1), which is compatible with the theoretical bias induced by omission
and reverse causality. When runners-up had less TV time (column (5)), results are even stronger than in column
(3). For this reason, our preferred estimate of TV advertising time is in column (4): a one percentage point
increase in the share of TV time causes a 0.241 percentage point increase in vote share difference.
To assess whether this effect is practically relevant, consider the TV time distribution in 1998, when the
difference in share between runners-up and winner was smallest: 4 percentage points (table VI). Thus, in 1998
TV time share difference fell, on average, from 4 to 0 percentage points. This implies a reduction in vote share
difference of 0.241 × 4 = 0.964 , that is, roughly one percentage point. In 1998, the average vote share
difference in the first and second rounds was 6 and 4 percentage points, respectively. Hence, half of the
closing gap in voting between rounds can be attributed to closing gap in TV shares.
V.II Demographic Determinants of TV advertising effect
In this subsection we investigate some possible determinants of the impact of media exposure. Five
culprits are examined: income per capita, income distribution, a measure of educational attainment, level of
urbanization, and size measured by population. Using the 2000 census, we match election-year data with
demographics of the city, and interact these five variables with dif _ voteser . Table VIII contains some
descriptive statistics on the cities in our sample. Results of the regressions procedures are in table VIII.
18
Table VIII - Descriptive Statistics, cities in the
sample
Average
Standart Deviation
Population
33065
212273
%Rural†
0.398
0.238
Gini††
0.548
0.057
Years of Schooling†††
4.411
1.214
Per capita income††††
2963
5.875
† %Rural is given by (Rural Populationi / Populationi)
†† Gini belongs to the interval [0 , 1]
††† Years of Schooling is the average number of years of schooling
†††† Annual per capita income in thousands of 2000 dollars
Source: Instituto Brasileiro de Geografia e Estatística (IBGE)
Table IX Dependent Variable: Vote share_A - Vote share_B
Second round
Time share A - Time
share B
Log(Population)* (Time
share A - Time share B)
(1)
0.019
(7.69)***
0.400
(30.55)***
0.031
(3.57)***
%Rural†† * (Time share A
– Time share B)
(2)
0.018
(7.36)***
0.397
(30.40)***
Number of Observations
R2
(4)
0.013
(5.66)***
0.397
(31.91)***
(5)
0.013
(5.58)***
0.391
(31.53)***
-0.191
(6)
0.012
(5.21)***
0.421
(33.09)***
0.052
(4.12)***
-0.452
(7.99)***
0.593
(3.20)***
-0.228
(19.68)***
(15.11)***
0.036
(0.860)
Gini†††† * (Time share A Time share B)
Years of Schooling†††† *
(Time share A - Time
share B)
Log(Income)††††† *
(Time share A - Time
share B)
Constant
(3)
0.016
(6.43)***
0.405
(30.50)***
2.414
(13.65)***
0.072
(38.37)***
15674
0.127
0.073
(38.05)***
15472
0.128
0.076
(40.38)***
15674
0.146
0.078
(43.62)***
15674
0.179
-0.301
-0.050
(19.30)***
(2.51)**
0.078
(43.09)***
15674
0.177
0.079
(45.36)***
15472
0.206
† White-Huber t - statistics in parentheses. All variables normalized to have a 0 sample average.
†† %Rurali originally computes as Rural Populationi / Populationi.
††† Gini originally in the [0 , 1] interval.
†††† Years of Schooling is the average number of schooling years observed in municipality i.
††††† Log(Incomei) is the logarithm of the per capita income in municipality i.
*** = significant at the 1% level
** = significant at the 5% level
Source: Tribunal Superior Eleitoral (TSE) and Instituto Brasileiro de Geografia e Estatística (IBGE)
19
All results are as expected. TV exposure has more influence in less educated, more populous, more rural,
poorer and more unequal cities. Electricity penetration is not universal in rural areas in Brazil, which reduces TV
exposure. Notice that in columns (2), the coefficient associated with rural is positive (but not
indistinguishable from zero), while in column (6) it is positive and significant. Since urban places tend
to be poorer and less educated, the negative coefficient would only arise after controlling for income or
education, or both. Larger cities favor impersonal, mass media means of communication. Similarly to the effect
of rural, larger cities tend to be richer and better educated. Thus the coefficient on Log(Population) should go up
from columns (1) to (6). Better educated people have access to sources of information other than television
(newspapers, magazines, etc). Finally TV has a stronger impact in poor and unequal cities. Similarly to
education, richer people are likely to have better access to information. After controlling for income per capita,
inequality (higher Gini) should increase the impact of TV: after a certain level, income is not likely to improve
acquisition of information. Notice that, despite including dif _ voteser six times, parameters are still very
precisely estimated. This is due to the larger number of observations and the large variation in the
regressors (see table VIII).
VI –CONCLUSION
From received literature, there is a conventional wisdom that campaign spending and media have
“minimal effects” on elections outcome. In this paper, using a quasi-experiment data, we present evidence
contrary to the conventional wisdom: TV and radio exposure in gubernatorial elections in Brazil have a
significant effect on elections outcome, both practically and statistically. The magnitude - half the average
closing gap from first to second round in gubernatorial elections - suggests that TV advertising cannot always
predict elections outcome, but is one of the major factors in elections.
How can one reconcile our results with a large literature that says differently? The empirical literature on
campaign spending and media effects suffer from either one of two problem It is either not persuasive that the
omission and reverse causality problems are properly solved or, in the way of solving, throw so much variation
away that it is difficult to estimate anything. In contrast with the literature we have access to a source of
significant exogenous variation.
Another avenue for reconciliation is in table IX. Brazil is poorly education, middle income, urbanized,
with the presence of very cities and, least but not least, with a very unequal income distribution. According to
table IX, Brazil scores 4.5 out of 5 in the factors that magnify the impact of TV advertising. This is certainly a
issue worth further research
20
REFERENCES
ABRAMOWITZ, A. I. “Incumbency, Campaign Spending and Decline of Competition in U.S. House Elections”,
Journal of Politics 53 (1991), 34 – 56.
ANSOLABEHERE, S. AND IYENGAR, S. “Going Negative: How Political Advertisements Shrink and
Polarize the Electorate,” The Free Press, New York, 1996.
ANSOLABEHERE, S., FIGUEIREDO, J. AND SNYDER, J. “Why is There So Little Money in U.S. Politics?”
The Journal of Economic Perspectives 17 (2003), 105-130.
BARTELS, L. “Messages Received: the Political Impact of Media Exposure,” American Political Science
Review 87 (1993), 267-285.
GERBER A. “Estimating the Effect of Campaign Spending on Senate Election Outcomes using Instrumental
Variables,” American Political Science Review 92 (1998), 401-411.
GOLDSTEIN, K. AND RIDOUT T. “Measuring the Effects of Televised Political Advertising in the United
States,” Annual Review of Political Science 7 (2004), 205-226.
GREEN, D. AND KRASNO, J. “Salvation for the Spendthrift Incumbent: Reestimating the Effects of Campaign
Spending in House Elections,” American Journal of Political Sciences 32 (1988), 884 – 907.
LEVITT, S. “Using Repeat Challengers to Estimate the Effect of Campaign Spending on Election Outcomes in
the U.S. House,” Journal of Political Economy 102 (1994), 777 – 798.
PRAT, A. “Campaign Spending with Office-Seeking Politicians, Rational Voters and Multiple Lobbies,” Journal
of Economic Theory 103 (2002), 162 – 189.
PRAT, A. “Rational Voters and Political Advertising,” working paper, London School of Economics, London,
2004.
WELCH, W. P. “Money and Votes: a Simultaneous Equation Model,” Public Choice 36 (1981), 209 – 234.
21
APPENDIX A
Table A.I - Number of elected members of the State Assemblies in the 1994 elections, by
party and by state
State
Party
PPR
PDT
PT
PTB
PMDB
PSC
PL
PPS
PFL
PMN
PRN
PP
PSB
PSD
PV
PRP
PSDB
PC do B
PPB
PRONA
PSL
PST
PSDC
PGT
PT do B
Total
Amapá
Distrito
Federal
Goiás
Mato
Grosso
do Sul
Minas
Gerais
Pará
Rio de
Janeiro
Rio
Grande
do Sul
Rondônia
Roraima
São
Paulo
Sergipe
0
1
0
2
2
0
2
0
4
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
13
0
1
5
1
1
0
1
1
0
0
1
6
0
0
0
0
2
0
0
0
0
0
0
0
0
19
3
0
2
1
8
0
3
0
3
0
0
2
0
3
0
0
3
1
0
0
0
0
0
0
0
29
0
2
2
3
4
0
1
0
3
0
0
3
0
0
0
0
1
0
0
0
0
0
0
0
0
19
0
5
5
6
9
0
4
1
6
1
0
9
1
1
1
0
6
0
0
0
0
0
0
0
0
55
6
1
3
2
9
0
2
0
3
0
0
2
0
0
0
0
2
0
0
0
0
0
0
0
0
30
4
9
4
2
5
2
3
0
1
2
0
3
3
1
0
0
10
1
0
1
0
0
0
0
0
51
11
7
4
8
9
0
0
0
0
0
0
0
3
0
0
0
1
1
0
0
0
0
0
0
0
44
0
3
1
1
3
2
1
0
1
3
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
17
3
0
0
5
0
1
0
0
1
0
0
1
0
1
0
0
1
0
0
0
0
0
0
0
0
13
7
2
13
6
17
0
4
0
5
0
0
0
1
2
1
1
12
1
0
1
0
0
0
0
0
73
3
1
2
1
3
0
0
0
5
1
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
18
Source: Tribunal Superior Eleitoral (TSE)
22
Table A.II - Number of elected members of the State Assemblies in the 1998 elections, by party and
by state
Party
PPR
PDT
PT
PTB
PMDB
PSC
PL
PPS
PFL
PMN
PRN
PP
PSB
PSD
PV
PRP
PSDB
PC do B
PPB
PRONA
PSL
PST
PSDC
PGT
PT do B
Total
Amapá
Ceará
Distrito
Federal
Mato
Grosso
do Sul
Pará
Paraíba
Paraná
0
3
1
1
3
0
2
0
2
0
0
0
3
2
0
0
1
0
0
0
1
0
0
0
0
19
0
1
3
2
5
1
1
4
2
0
0
0
1
0
0
0
18
1
2
0
0
0
0
0
0
41
0
1
4
2
4
1
1
1
1
0
0
0
1
1
0
0
1
0
1
0
0
0
0
0
0
19
0
2
1
3
4
0
1
2
2
0
0
0
0
0
0
0
6
0
0
0
0
0
0
0
0
21
0
3
3
3
7
0
3
1
3
0
0
0
1
2
0
0
7
0
4
0
0
0
0
0
0
37
0
2
3
0
16
0
0
0
4
0
0
0
0
0
1
0
5
0
1
0
1
0
0
0
0
33
0
3
4
9
7
1
0
0
12
0
0
0
2
0
0
0
5
0
7
0
0
0
0
0
0
50
State
Rio
Grande
do
Norte
0
1
1
1
8
0
2
0
4
0
0
0
1
0
0
0
1
0
2
0
0
0
0
0
0
21
Rio
Grande
do Sul
Rondônia
Roraima
Santa
Catarina
São
Paulo
Sergipe
0
6
11
9
9
0
0
0
2
0
0
0
1
0
0
0
2
0
11
0
0
0
0
0
0
51
0
3
2
2
4
2
1
0
2
0
0
0
0
0
0
0
3
0
2
0
0
0
0
0
0
21
0
4
0
2
2
0
0
0
3
0
0
0
0
0
0
0
1
0
4
0
3
0
0
0
0
19
0
1
5
1
9
0
0
0
8
0
0
0
0
0
0
0
3
0
9
0
0
0
0
0
0
36
0
6
13
4
8
0
5
3
11
0
0
0
2
0
1
1
20
2
9
2
0
0
0
0
0
87
0
1
0
1
5
0
0
1
3
1
0
0
2
0
0
0
3
0
2
0
0
0
0
0
0
19
Source: Tribunal Superior Eleitoral (TSE)
23
Table A.III - Number of elected members of the State Assemblies in the 2002
elections, by party and by state
State
Party
PPR
PDT
PT
PTB
PMDB
PSC
PL
PPS
PFL
PMN
PRN
PP
PSB
PSD
PV
PRP
PSDB
PC do B
PPB
PRONA
PSL
PST
PSDC
PGT
PT do B
Total
Goiás
Maranhão
Pará
Paraíba
Paraná
Pernambuco
Rio de
Janeiro
Rio
Grande
do Norte
Rio
Grande
do Sul
Santa
Catarina
0
0
3
0
8
0
1
1
3
0
0
0
1
1
0
1
10
1
4
0
0
1
1
0
0
36
0
4
2
2
2
1
1
1
13
0
0
0
1
6
0
0
3
0
1
0
0
0
0
1
0
38
0
2
5
4
7
0
4
1
0
0
0
0
1
2
0
0
6
1
2
0
0
2
0
0
0
37
0
1
4
2
8
0
1
1
3
0
0
0
2
0
0
0
10
0
2
0
0
0
0
0
0
34
0
5
8
3
7
1
2
2
7
0
0
0
2
0
0
1
4
0
4
0
2
0
0
0
0
48
0
3
4
1
6
2
2
1
7
0
0
0
4
2
1
0
4
1
3
0
1
0
1
0
0
43
0
3
8
1
11
2
3
1
4
0
0
0
11
0
3
0
4
1
5
2
2
0
0
0
2
63
0
1
2
1
3
0
1
0
4
0
0
0
2
0
0
0
0
0
7
0
0
0
0
0
0
21
0
6
12
6
8
0
0
3
1
0
0
0
2
0
0
0
3
1
8
0
0
0
0
0
0
50
0
0
8
2
7
0
1
0
7
0
0
0
0
0
0
0
2
0
8
0
0
0
0
0
0
35
Source: Tribunal Superior Eleitoral (TSE)
24
Departamento de Economia PUC-Rio
Pontifícia Universidade Católica do Rio de Janeiro
Rua Marques de Sâo Vicente 225 - Rio de Janeiro 22453-900, RJ
Tel.(21) 35271078 Fax (21) 35271084
www.econ.puc-rio.br
[email protected]