Learn from thy Neighbor: Do Voters Associate

Corruption with Political Parties?∗

Arieda Muço†
Job Market Paper
Please find the latest version Here

January 6, 2017

Abstract
This paper exploits a randomized audit program to document how information about
a corrupt politician causes electoral spillovers on his party. I focus on two types of
spillovers: spillovers across types of elections (cross-electoral spillovers) and spillovers
across jurisdiction borders (cross-border spillovers). Using detailed data on radio an-
tenna location and coverage, I identify neighboring areas in the same media market.
Moreover, via machine learning and text-analysis tools, I take a data-driven approach
to create an index that ranks municipalities according to their corruption level. I un-
cover how information about corruption shapes voting decisions through the structure
and geography of the media market. I show that voters hold the party of the incumbent
politician accountable in four distinct ways. In municipalities where corruption occurs,
voters punish parties in (1) local and (2) national elections. Most importantly, I show
that news of a politician’s corruption affects his party in neighboring municipalities
that share the same media market, and these spillovers affect both (3) local and (4)
national elections. Ruling out other potential mechanisms, I show that these findings
are consistent with electoral accountability. Finally, a back-of-the-envelope calculation
suggests that an audit occurring at the local level causes the party’s vote support in
the audited neighborhood (audited municipality and its neighbors) to decrease by 10
percentage points.

∗
I am indebted to Manuel Arellano, Erik Lindqvist, Torsten Persson, Diego Puga, Giancarlo Spagnolo, and David Strömberg
for advice, support and encouragement at different stages of this project. I am extremely grateful to Felipe Carozzi for exten-
sive conversations and help. This paper has also benefited from conversations with Samuel Bentolila, Martina Björkman
Nyqvist, Guillermo Caruana, Kerem Coşar, Tore Ellingsen, Ruben Enikolopov, Stefano Gagliarducci, Mounir Karadja, An-
dreas Madestam, Mónica Martı́nez-Bravo, Anders Olofsgård, Elena Paltseva, Maria Perrotta Berlin, Maria Petrova, and Jakob
Svensson. I thank Reinaldo Luz, Fernanda Senra (discussant), Ricardo Lima (research assistant), and Robson Tigre for help
with understanding of the Brazilian institutions. The help of the staff at the CGU, especially Giuliana Biaggini, Adriano
Souza, Camilla Bezzera, Maria Rezende and Günther Lorenz at FMSCAN/FMLIST, is gratefully acknowledged. I also thank
participants in seminars: CEMFI, IIES, SSE, UAB, UCL, Stockholm University, Micro Workshop Stockholm, IFN Stockholm,
University of Gothenburg, Max Planck Institute and in conferences: SITE Corruption, Stockholm; Enter Jamboree, Carlos III;
IPSWD, Columbia University; ASWEDE, Gothenburg; SUDSWEc, Stockholm University; European Winter Meetings of the
Econometric Society, Edinburgh; Spanish Economic Association, Bilbao. All errors are my own.
†
Stockholm School of Economics. E-mail: [email protected]
1 Introduction
Corruption is not a victimless crime. It has been widely argued that corruption is to blame
for impediments in growth and economic development (Treisman (2000)). Despite a large lit-
erature on corruption, its causes and consequences are not yet fully understood. Institutions
that promote accountability, such as elections, are among the main prescriptions for fighting
malfeasance. Well-functioning political institutions, however, require an informed electorate.
Indeed, if provided with information, voters punish corrupt politicians.1 It is, however, an
open question whether the party of the corrupt politician is also sanctioned.
In this paper, I show that the actions of individual politicians have broad consequences for
their party. More precisely, I examine how news of individual malfeasance impacts the party
at the local level and how it spills over across jurisdictions. Highlighting important features
of information disclosure that have not been previously documented, I find that voters hold
the party accountable in four distinct ways. Within municipalities where corruption occurs,
voters punish parties in (1) local and (2) national elections. Most importantly, I show that
news of a politician’s corruption also affects neighboring municipalities that share the same
media market, and these spillovers affect both (3) local and (4) national elections. These
effects operate through voters’ recognition of party names.
To estimate the spillovers from individual corruption, I take advantage of a randomized
audit program that was initiated by the Brazilian federal government in 2003. This on-going
program aims at fighting corruption in local governments. A public lottery determines which
municipalities are audited. Once a municipality is selected, a team of auditors is sent within
a week to inspect. After an investigation has been concluded, the auditors provide a detailed
report which is made public.
In a country plagued by endemic corruption, audit investigations reveal corruption in
90 percent of the cases. I use the randomized timing of audits with respect to elections to
investigate the spillovers of political corruption. My empirical analysis consists of two main
parts. First, I investigate the effect of an audit on the electoral outcomes across jurisdictions.
Specifically, I examine the impact of these investigations on the audited municipality and its
neighbors, both in local and national elections.2
Second, I explore how voters respond to reported levels of corruption, given that the
1
See Ferraz and Finan (2008); Costas-Pérez, Solé-Ollé and Sorribas-Navarro (2012); Bobonis,
Cámara Fuertes and Schwabe (2016).
2
Local and national elections alternate each other ever two years.

1
intensity of the punishment may depend on the severity of the corruption disclosed. To do
so, I use the detailed assessment of the auditors and put together a dataset by scraping more
than 2000 PDF reports which span the period from 2003 to 2013. Moreover, via machine
learning tools, I take a data-driven approach to construct a corruption index that ranks
municipalities according to the amount of disclosed corruption.3
To build the corruption index, I create a severe irregularity count that records the number
of severe cases of irregularities found. From the audit reports, I also build several other
variables that correlate with the amount of corruption uncovered, namely the total amount
of irregularities found, the number of pages, the number of lines, and the number of images.
I then combine these variables in a principal component analysis (PCA). The main idea
behind the PCA is that the latent common factor measures the underlying corruption. This
methodology has the advantage that once the criteria for classification are set, the index does
not rely on subjective decision making.4
I start by documenting the effect of an audit in the audited municipality (benchmark
case) on vote shares in elections, both for the same level of government where information is
disclosed and for higher levels of government (cross-election spillovers). For this purpose, I
compare municipalities that are audited before elections with municipalities that are audited
after elections.5 To examine whether higher-level elections are affected by an audit investiga-
tion I focus on the parties that compete in local and national elections under the same party
label.6
Next, I turn to geographical spillovers – the effect of the audit on the electoral outcomes
in the neighboring municipalities – in elections for the same level of government where infor-
mation is disclosed (cross-border spillovers), and in elections for higher levels of government
(cross-border and cross-electoral spillovers). Following the recent political economy and me-
dia literature, which provides evidence of media being the relevant channel of information,
3
The irregularities fall under the responsibility of the politician in office.
4
Previous studies such as Ferraz and Finan (2008, 2011); Brollo et al. (2013); Litschig and Zamboni (2011);
Timmons and Garfias (2015) in Brazil, Larreguy, Marshall and Snyder Jr (2014) in Mexico, and Bobonis,
Cámara Fuertes and Schwabe (2016) in Puerto Rico have relied on hand-coded measures of the audit reports.
5
Given the randomization of the audit and its timing, these two groups should be homogenous. I check
that this is indeed the case by performing some randomization checks, as described in Section 4.1.
6
There are only two main parties that show up throughout my sample with the same party names, Partido
da Social Democracia Brasileira (PSDB) and Partido dos Trabalhadores (PT). These two parties have been
the main contestants in the race for president since the democratisation of Brazil. Similar to local elections,
I compare municipalities audited before presidential elections with municipalities audited after presidential
elections.

2
I look for an effect in the informed neighbors of the audited municipality.7 Using a unique
dataset detailing the location and coverage of the universe of FM antennas in the country, I
compute the fraction of the population covered by an antenna in each municipality. Informed
neighbors are defined as nearby neighbors of the audited municipality which are contiguous
(share a vertex or border) and have a positive population fraction covered by the antenna
signal emitted from the audited municipality.
The choice of radio rather than other potential media outlets is natural in the Brazilian
context. Media penetration in Brazil is dominated by television. However, in most TV
broadcasts, state and national channels prevail, making it unlikely for local news to be
transmitted by these outlets. Radio does not only have high penetration rates but is also
oftentimes locally provided. According to Latinobarómetro 40 percent of Brazilians use radio
as a source of political information. I do not take newspapers into account since Brazil has
one of the lowest rates of newspaper penetration in the world, with an average circulation of
only 42 newspaper copies per 1000 inhabitants, (Porto (1985); IBOPE (2012)).
When investigating geographical spillovers, I compare the group of informed neighbors
that receive information about the audit outcome (in neighboring municipalities) before elec-
tions with the group of informed neighbors that receive the information after elections. The
randomization of the audit makes these two groups comparable.8
Several findings emerge from this analysis. An audit reduces the vote support of the
incumbent party in local elections in the audited municipality, by 2-3 percentage points.
The reelection likelihood of the party remaining in power is also reduced. However, the
intensity of corruption does not affect the vote outcome. These findings are consistent with
the opposition having hard evidence to effectively attack the incumbent party. Indeed, in
the audited municipalities opposition parties are 20 percent more likely to increase their
campaign expenditure.
The electoral effects of an audit are not isolated to the audited municipality. They also
spillover to neighboring governments. In the informed neighboring municipalities, in local
elections, the audit itself does not affect the electoral outcomes but the electoral support
for the party is reduced when the level of disclosed corruption increases. These findings are
7
See for example, Strömberg (2004); DellaVigna and Kaplan (2007); Enikolopov, Petrova and Zhuravskaya
(2010); Snyder and Strömberg (2010); DellaVigna et al. (2014); Yanagizawa-Drott (2014); Bursztyn and
Cantoni (2015).
8
I show that these two groups are indeed comparable by performing balancing tests described in the
Appendix.

3
in line with the political science literature, which stresses that the candidate is the most
important figure in local elections (Pereira and Mueller (2003); Desposato (2004); Folke,
Persson and Rickne (2016)). Voters in neighboring municipalities do not seem to react to
an audit investigation because their local politician is not affected. The finding that the
level of disclosed corruption matters to voters in neighboring municipalities suggests that
the likelihood of voters learning about party characteristics increases with the amount of
disclosed corruption.
The electoral effects at higher levels of government are similar in the audited neighborhood
(audited municipality and its neighbors). In the neighborhood, the vote support for the party
is reduced by 4 percentage points in the audited municipality and by 3 percentage points in
the informed neighboring municipalities. These municipalities are also punished more harshly
if more corruption is revealed. Taken together, these findings indicate that voters sharing
the same information flow behave similarly. The strong electoral effect suggests that these
corruption scandals were triggering voters’ memories of the party’s wrongdoing at the local
level. Indeed, anecdotal evidence suggests that the presidential candidates have been using
these corruption cases to blame and shame each other in public debates.
Throughout this paper, I maintain the assumption that voters recognize party labels.
To test for this, I estimate the impact of an audit investigation on the main party in the
coalitions formed for presidential elections. The focus on the main party is necessary because
the coalition is recognized as the party of the presidential candidate. If the name of the party
is the relevant information for voters, there should be no effect of the audit on the vote share
of the coalition. Indeed, this hypothesis is confirmed by the data.
While the electoral effects at higher levels of government and the spillovers across border-
ing jurisdictions are novel contributions of this paper, the effect of the audit on local electoral
outcomes has previously been documented. For instance, using the Brazilian context, Ferraz
and Finan (2008) find that incumbent mayors face a reduction of the likelihood of reelection
by 7 percentage points in municipalities where at least two violations associated with cor-
ruption were reported. Larreguy, Marshall and Snyder Jr (2014) exploit the random timing
of the information release in Mexican municipalities to investigate the electoral outcomes
of the incumbent political party, given that in Mexico the politician cannot rerun for elec-
tions.9 The authors show that voter sanctioning increases with the increase in the severity of
9
In Mexico, the audit process is not random, but the timing of the release of the findings relative to
elections is random.

4
disclosed malfeasance. Bobonis, Cámara Fuertes and Schwabe (2016) investigate, in Puerto
Rico, how news about the incumbent politician affects his reelection chances. In Puerto Rico
the audit is not random, but the timing of municipal audits is predetermined with respect
to an election. The authors find that incumbent mayors in municipalities with a negative
audit outcome are less likely to run for reelection. It also leads to a significant short-term
reduction in municipal corruption but they do not exhibit decreased levels of corruption in
subsequent audits.
This paper contributes to several strands of the literature. First, it relates to the broad
literature on effective policies designed to reduce corruption (Reinikka and Svensson (2005);
Björkman and Svensson (2009); Gonzalez (2016)). Second, it relates to the literature that
considers audits to be an effective anti-corruption tool in developing countries (Di Tella and
Schargrodsky (2003) and Olken (2007)). Differently from these studies, I show that the effects
of audits are magnified due to spillovers. Third, similarly to the growing empirical literature
on the electoral effects of access to media (Strömberg (2004); Gentzkow (2006); DellaVigna
and Kaplan (2007); Enikolopov, Petrova and Zhuravskaya (2010); Snyder and Strömberg
(2010); Adena et al. (2015)), my study confirms the finding that the media structure plays
an important role in electoral accountability. In addition and similar to the work by Larreguy,
Marshall and Snyder Jr (2014), my findings stress the relevance of local media (rather than
national media outlets). Fourth, the paper also contributes to the literature on collective
blame, providing an empirical test to the theoretical contribution pioneered by Tirole (1996).
Overall the findings in this paper are also relevant towards understanding why national
parties might have incentives to better monitor and improve selection among local candidates.
At least in the short run, politicians’ behaviour has an impact on the party across government
borders and levels, even in a context like Brazil where party labels, in local politics, are
considered to be weak.
The rest of this paper is structured as follows. Section 2 discusses the institutional set-
tings. Section 3 describes the data sources and provides summary statistics. Section 4
outlines the empirical analysis and provides results on the effect of an audit in the audited
municipality. Section 5 considers geographical spillovers. Section 6 provides additional find-
ings. Section 7 addresses robustness and different placebo tests. Section 8 discusses the
findings of the paper and Section 9 concludes the paper.

5
2 Institutional Setting
2.1 The Audit Program
The randomization of municipal audits in Brazil was initiated in 2003 with the objective
of fighting corruption in local governments. The program – which continues to date – is
implemented by the Controladoria Geral da União (CGU) and consists of random selection,
for audit purposes, of Brazilian local governments. Under examination are the transfers
allocated by the federal government to each municipality.
Before the lottery is held, the CGU publishes the list of eligible municipalities jointly
with the lottery numbers attributed to each municipality. Once the municipalities to be
audited have been chosen, a team of auditors is sent to inspect the use of the transfers,
following guidelines from CGU headquarters. In the first rounds of the program, all federal
transfers were audited. As the program proceeded, it was decided that only a subset of funds
would be inspected. Funds under inspection are also randomly selected and announced
jointly with eligible municipalities and lottery numbers attributed. Education and Health
are the areas that receive more audits, they are also the areas which have the most cases
of mismanagement.10 State capitals and municipalities with populations above (or below)
certain population thresholds are also excluded. If a municipality is drawn, that municipality
is re-eligible for selection after a predetermined number of rounds.11 The rules for exclusion
and the funds to be audited are made publicly available the week before the lottery is held.
After several days of investigation in the municipality, the auditors compile a report of their
findings.12
Anecdotal evidence suggests that media covers issues related to the audit outcomes. This
implies that citizens learn about the audit outcomes either directly – by observing the auditors
in the municipality – or indirectly – from various media sources. For example, in one of the
biggest national newspapers, Folha de S. Paulo, some of the mayors were mainly mentioned
due to the exposure via the audit program.13 Published on October 1 2004, two days before
the first round of municipal elections, the headlines of several articles cite: “Suspects of fraud
10
25 percent of the overall irregularities in the sample have occurred in the funds related to Education.
11
In the sample, a municipality is non-eligible for some rounds. The maximum number of rounds for
exclusion during my sample period has been 12 rounds. More information on the population cutoffs in
Appendix A1.
12
In the first page of the audit reports the auditors state the days they have spent in the municipality and
the average duration is around one week.
13
Available here: http://acervo.folha.com.br/fsp/2004/10/01/125/.

6
[will cause an] overthrow of 199 mayors” The title of another article “40 mayors accused of
corruption are running for reelection” and a third article talking about mayors contesting
the CGU investigations where one of the incumbent mayors running for reelections refers to
the negative outcome of the audit in his municipality:

“People knew the opposition would take advantage over this in the elections...”

Another mayor complains about the release of the audit before the municipal elections stating:

“It is an irresponsibility from the CGU to point out at not proven irregular-
ities in electoral period, so they can be used by the opposition like these [the
irregularities] were crimes or diversion of funds.14 It is a disaster because my
opponent is from PT.15

Between 2003 and 2013, namely the audit years I use in this analysis, a total of 2053 audits
have occurred and more than 400 municipalities have been audited twice and 30 have been
audited three times. The timing of the audits is illustrated in Figure 1. Excluding the
pilot audits in 2003, where one municipality per state was audited, the number of audited
municipalities increased from 50 to 60 municipalities per round after the year 2005.

2.2 Brazilian Municipalities

Brazil is divided into 5,570 municipalities and 27 federal units comprising 26 states and the
Federal District, where the national capital, Brası́lia, is located.16 Municipal governments
correspond to the lowest tier of public administration in the country. They are composed of an
executive branch chaired by the mayor (prefeito), a vice-mayor, and a multi-member legisla-
tive chamber (câmara municipal ) constituted of council members also known as vereadores.
Municipalities vary greatly in population, size, wealth and literacy rates.
The Brazilian Constitution guarantees the independence of municipal governments. Lo-
cal governments are responsible for providing certain goods and services including primary
education and healthcare in coordination with the state and the federal authorities. While
larger municipalities can generate up to 40 percent of their revenues via taxes, smaller size
14
The mayor of this municipality, as all mayors, was given the chance to reply to the draft of the report
and provide justifications or proofs to overturn the findings. However, for most of the findings, he was not
able to provide any credible justification. (Municipality of Laguna in Santa Catarina, audited in lottery 8.)
15
PT is the presidential party throughout my sample period.”
16
Table A2 provides a more detailed description of the subdivision of the states in the country.

7
municipalities depend heavily on intergovernmental transfers. Municipalities with a popula-
tion size below 50,000 inhabitants generate roughly 80 percent of their revenues via federal
transfers (e.g Farvacque-Vitkovic and Kopanyi (2014)).

2.3 Elections and Political Parties

National and municipal elections in Brazil are, respectively, held every four years. Thus, there
are elections every two years, alternating between national and municipal elections. Elections
occur simultaneously throughout the country on the first Sunday of October. Voting is
compulsory in Brazil in all types of elections and all Brazilian citizens cast their votes under
the same electoral rules.17 If voters cannot participate in an election, a justification form can
be filled out at election centres and post offices. Otherwise, a fine must be paid. Election
timing in the sample joint with the lottery rounds are shown in Figure 1.

[Figure 1 Here]

In municipal elections, voters are called to elect the mayor and members of the municipal
chamber, vereadores. The chamber of local legislators acts as a supervisory authority and the
size is directly related to the size of the population in the municipality. The voting system
used to elect the municipal government is majority ruling. Exceptions are municipalities
with a population size above 200,000 voters, for which the mayoral elections require a second
round. Mayors face term limits, they can only get elected for two consecutive terms. If a
mayor has already served two consecutive terms, he is eligible for reelection after a term out
of office. National legislators and the president are elected in general elections.18
Brazil has a multi-party system both at the federal and the local level. More than
30 parties participate in local elections each term; however, only few are relevant at the
national level. Since 1994, only two parties have dominated the race for president: Partido
da Social Democracia Brasileira (PSDB) and Partido dos Trabalhadores (PT). Since 1995,
the president has been from either of these parties. They are also the only parties that
have a presidential candidate throughout the sample period.19 In presidential elections, the
17
Participation in the elections is compulsory for literate citizens between 18 and 70 years old. It is
non-compulsory for Brazilians aged 16-17 or over 70 or illiterate citizens of any age.
18
The president is elected through a two-round system. I consider the votes in the first round for municipal
and presidential elections.
19
Fernando Henrique Cardoso (PSDB) was in power from 1995 to 2003. PT has held the presidency since
2003 to 2016. First with Luis Inácio Lula Da Silva (2003-2011) and latter with Lula’s handpicked successor
Dilma Rousseff (2011-2016).

8
identities of these two parties are strong and well-recognized. At the local level, the political
power is highly fragmented between numerous parties.20
PT is a relatively minor player at the local level, holding slightly more than 3 percent
of local governments in the 2001 electoral term while PSDB ruled over a sixth of Brazilian
municipalities in the same electoral term. Figure 2 shows the distribution of PT and PSDB
in the Brazilian territory during the electoral term of 2001.

[Figure 2 Here]

In Brazil, party identities are considered to be weak (Desposato (2004); Pereira and
Mueller (2003); Folke, Persson and Rickne (2016)). It is not uncommon for mayors and other
candidates in the municipal council to switch parties from one election to the next.21 For
instance, in the municipal elections of the year 2000, 62 percent of the reruning candidates
switched party.
Although the political party system revolves around the candidate, there are three main
reasons to believe that parties may be recognized as corrupt if their candidates misbehave.22
First of all, in order to be eligible, all candidates must be supported by a party, making the
link between politician and party relevant.23 Second, when citizens cast their vote, the party
of the candidates is quite visible on the ballot. Indeed, the ballot first gives information about
the party (and the party number) and then the list of candidates for each party, members of
the municipal council and mayor.24 An illustrative example is shown in Figure 3.

[Figure 3 Here]

Third, candidates may be non-eligible for election either due to term limits or due to
other impediments. Parties, on the other hand, have a longer life span than candidates.
20
The local level does usually have the highest fragmentation. The multitude of parties frequently requires
the parties to form coalitions, at any level of the government, as no party has any chance of gaining power
alone. The biggest players in local elections are: PTB, PT, PMDB, PSDB, PP and, PFL, respectively
standing for Partido Trabalhista Brasileiro, Partido dos Trabalhadores, Partido do Movimiento Democrâtico
Brasileiro, Partido da Social Democracia Brasileira, Partido Progressista, Partido da Frente Liberal, now
called Democratas.
21
According to Desposato (2006), the strongest motives for a party change are: access to distributive
resources, electoral opportunities, and compatible policy positions.
22
This has been the case even during recent events in Brazilian politics, where citizens were linking most
PT politicians to corruption due to recent political scandals in the country.
23
They must have been members of the party for at least two years before elections.
24
The campaign for the candidates is mostly done with party numbers.

9
2.4 Local Media
As in many other countries in the world, television and radio are the most used media
outlets. Television has the largest penetration on a national scale. However, in the year
2000, only 10 percent of the municipalities broadcasted local channels, whereas 34 percent
of the municipalities had a local radio station according to Perfil dos Munı́cipios Brasileiros
(2001). According to Latinobarómetro, a public-opinion survey, 40 percent of the Brazilian
population uses radio as a source of political information. I use information on FM radio
coverage to define neighborhoods for audited municipalities.
The use of local radio is motivated by the fact that it is more likely for these outlets
to report local news. It is quite rare for local news to be mentioned extensively on state
or national TV. The focus on FM radios mainly arises because of reception quality. While
Brazil still uses many AM transmitters, which travel farther distances, audio quality is often
poor whereas FM broadcasting is capable of a better sound quality.
I do not take into account newspapers as Brazil has one of the lowest levels of newspaper
penetration in the world, with only 42 newspaper copies per 1000 inhabitants (Porto (1985);
IBOPE (2012)). Community radio stations are also excluded from the analysis. While they
certainly constitute another source of local information in the Brazilian municipalities, most
stations have low power transmitters with a maximum broadcast range of one kilometre
(Boas and Hidalgo (2011)).25

3 Data Sources
I collect and combine data from different sources. Three most important sources used are:
(i) audit data from which I build a measure of severity of irregularities found by the auditors,
(ii) electoral data, and (iii) a measure of antenna coverage. These data are described in detail
in this section.

3.1 Audit Reports

The audit reports consist of more than 2000 PDF documents downloaded from the Contro-
ladoria Geral da Uniaõ (CGU) - for the period 2003-2013. Each audit starts with a description
of the purpose of the auditors in the municipality, the number of days the auditors spent in
25
Hence, it would not be possible to construct informational neighborhoods, as their signals are unlikely
to cross municipal boundaries.

10
that town and the funds investigated as well as all irregularities found. Although the struc-
ture of the reports has been altered from time to time, some features remain similar across
lotteries. Within the first five pages, a table, or list, is included, the total amount of funds
investigated and the number of programs in each area. From lottery 8 onwards, a summary
of irregularities was also introduced in all reports (see Figure 5).26 For each inspection, there
is a detailed description of the findings in the text. Several pictures are also included in the
report to further illustrate the wrongdoing.27

[Figure 5 Here]

For example, in Figure 6 a simulation of a bidding procedure is documented by the

auditors in the municipality of Alvaraes in the state of Amazonas (lottery 2). The call for
bids was for cleaning products, while all three bids submitted intended to deliver food.28
Except from the bid layout being (exactly) the same, the bids also coincide also on the type
of the orthographic mistakes. The only slight change the auditors point out is the different
use of font.

[Figure 6 Here]

Another fraud in a procurement process is the example in the municipality of Afonso

Cunha in Maranhão (lottery 27) where three firms “participated” in the call for bids. The
auditors visited the locations of the address registered as the firm domicile and none of them
were located as indicated in the bids.

[Figure 7 Here]

In the text, for each irregularity found is also included a statement by the mayor (if
provided). There is also a further comment by the auditors in which they explicitly write if
they accept the justifications or if they maintain the original statement. Most of the original
statements remain unaltered.29
26
The summary in lotteries 34 to 38 is somewhat different because the auditors include a description
paragraph instead of the irregularity list.
27
There is a high variation in the type of irregularity documented such as roads not constructed, abandoned
construction sites to poor conditions in the schools and kinder-gardens, non safe storage of food and medicine.
28
For the validity of the procurement process, the law (Law number 8.666/93) requires the submission of
at least three valid bids from different suppliers.
29
Only a small fraction of the justifications are (partially) accepted.

11
3.1.1 Text Classification

From the audit reports, it is possible to have municipalities ranked according to the severity
of the committed irregularities. Processing more than 2000 reports (a total of 172,768 pages)
is challenging for two main reasons: the first, easier to overcome, is that the information in
these documents is not stored in a homogeneous format, not only across lotteries but also
within the same lottery in the same state. The second is that the amount of training data
available – data from which the algorithm should learn – is very limited. When using a
machine learning classifier (in supervised learning), it is extremely important to have access
to a sufficiently large amount of labeled data for the algorithm to learn so that accurate
out-of-sample predictions are made.30
Taking advantage of the summary that the auditors provide from the 8-th round of the
lottery, I construct a classifier that is able to recognize for each element in the summary if
an irregularity found is more severe than others on the same list. To automatise the process
(of classifying a total of 104,337 phrases), I created a list of rules to distinguish more severe
from less severe cases. Hand-written rules are recognized and extensively used in the machine
learning literature in cases where the amount of classified documents available is limited (see
Manning, Raghavan and Schütze (2010)). The accuracy of these models can be high.31
The classifier works in the following way: given a set of inputs (phrases from a document),
call it S, the purpose is to classify each input of the two classes, C = {c1 , c2 }, indicating
whether the irregularity is severe or not. The predicted output would be c ∈ C. To get enough
knowledge about the domain of the data with the help of a Brazilian research assistant, I
read and classified around 200 reports. The classification was used to test the performance
of the classifier and compile the list of rules. Lotteries from 20 - 33 were also used – in the
cases where the auditors pointed out the majors irregularities – to build the list of rules for
classification.
The list of rules consists of stemmed words, or composition of words, bigrams and other
n-grams. All accents are removed from the original documents, and no accents are included in
the list of rules. This ensures that if a word is miswritten, i.e missing accents, the algorithm
is still able to recognize that word. Upper case letters have also been replaced by lower case
30
Labeled data is jargon in Machine Learning literature to indicate classified documents by a human. Good
practice of the procedure requires that the subset of documents for human classification is selected from a
random sample.
31
Examples are provided in Jacobs and Rau (1990) who report 92 percent precision and a 88.5 percent
recall rate while Hayes and Weinstein (1990) report 94 percent recall and 84 percent precision.

12
letters.
Words like “Fraud”, “Collusion”, “Fake”, which indicate a severe irregularity themselves,
are included as simple words. In other cases, I search for combinations of words.32 Given
that in some instances the meaning of the sentence is the same but the wording is different,
such as “Firm not located”, “Non existing firm”, “Ghost Firm” (for an example of a non-
located/non-existing firm, see Section 3.1) I ensure these are included as synonyms in the
list of rules.33
Below, I provide an illustration of the approach taken. Suppose, for example, that the
summary written by the auditors for municipality X is the following:

1. Indication of fraud in the procurement process.

2. Resources (from a specific program) used differently than what was fore-
seen by the program.

3. Payments for non executed services.

4. Non actualisation of pupil’s cadastral information.

5. Delay of book delivery in a rural school in the municipality.

Out of five irregularities found by the auditors, three are more severe. For each summary (in
each document), I record the total of amount of the most severe irregularities (three in this
example).34 This measure is a count of severe irregularities found in the municipality.35 The
list of rules seems to be quite accurate as it also overlaps with the overview that the auditors
have of the most corrupt activities in Brazilian municipalities. An example of this can be
found in the final report for the year 2004, where the auditors summarize their annual work.

Indicators of fraud in the procurement process, availing of ghost firms; non ex-
isting or fake bills; overpricing; lack of proof of expenditures; manipulation of
32
For instance, when a good was supposed to be provided but could not be located by the auditors, I search
for “good not loc”. The algorithm, in this case, will capture if the auditors have written the phrase in plural
form such as “goods not located”.
33
The list of rules has a total of 140 n-grams.
34
In this simple case, the vocabulary would be: fraud; resource differ, payment non execut.
35
Classifying the reports as either corrupt or not corrupt is very restrictive. In the random sample that
the RA classified, only 7 percent of the municipalities had no instances of corruption.

13
 the procurement process, paralysed construction, non existing object and non
reached objectives.36

3.1.2 Principal Component

The irregularity count described above may be subject to measurement error because the
classifier may fail to recognize a severe irregularity. For example, if the auditors indicate
that some goods are not located in the municipality, they write “Bens nao localizados”, the
regular expression, composed of stemmed words, to classify this phrase would be “ben nao
local”.37 This regular expression will also capture a phrase like “beneficiarios nao localizados”,
indicating that the beneficiaries have not been located in the municipality, which is not a
severe irregularity in itself.38
To ensure that the irregularity count indeed measures what it should, I create another
index, which is the first principal component of a series of variables extracted from the
audit reports. The variables are: the number of pages, the number of lines, the number of
images included in the PDF, the number of total irregularities – the count of all irregularities
summarized for each report.39 The main idea behind this is that the latent common factor
will express a dimension of the underlying corruption.
Some preprocessing is needed to use the PCA, i.e the standardisation of the variables.
This is done so that variables have comparable scale values and allows us to deal with
cases in which variables have different scaling. The first principal component and the severe
irregularity count are highly correlated, the correlation coefficient of 0.86. The correlation
between both measures is shown in Panel A of Figure 8.

3.1.3 Corruption index

The corruption index, which I use for the main analysis, is the first principal component of
all the variables described above (in Section 3.1.2) and the measure of the severe irregularity
36
Available here: http://www.cgu.gov.br/sobre/auditorias/arquivos/2004-a-2001/relatorio_
gestao_cgu_2004.pdf.
37
It is a sequence of characters that define a search pattern, mainly for use in pattern matching with
strings, or string matching.
38
Usually, the beneficiaries are not found because they have moved to another municipality or because they
may have passed away etc. There are some cases when beneficiaries of certain funds were not located and
members of the municipal council were making use of these funds. In the latter case, it would make sense to
classify this irregularity as more severe.
39
Also in this case, I discuss in the Appendix how I get the total number of irregularities in the first rounds
of the lottery (lottery 3-7).

14
count (see Section 3.1.1). The distribution of the measure is shown in Panel B of Figure 8.

[Figure 8 Here]

The first principal component of these series is the only eigenvalue with a value larger
than 1 and it retains around 80 percent of the explained variation. A visualisation of the
variance explained by each eigenvalue and the cumulative variance is shown in Figure 9.

[Figure 9 Here]

The combination all variables that indicate irregularities is done with the main objective
to minimize classification error. This measure is dependent on the set of rules I use to
construct the irregularity count; thus, it may considered to be partially subjective. However,
it has two main advantages: i) it is replicable and, ii) once the rules are set, no subjective
judgment is involved during the classification process, thus making it time efficient and less
prone to human mistakes.
I test the performance of the corruption index in two different ways.40 First, I use the
sub-sample of municipalities in which the auditors indicate the main problems found in each
town. The classifier seems to be scoring correctly among couples of municipalities to which
the auditors attribute the same number of irregularities. The reason why the classifier is able
to rank among municipalities to which the auditors have scored similarly, for each report,
the auditors rank the major irregularities within that same municipality, thus making it
non comparable to another municipality.41 To further clarify this point, consider the case
of audited municipalities Andarai in the state of Bahia (lottery 21) and Guatapara in the
state of São Paulo (lottery 22). The auditors find 35 irregularities in the first case and
27 irregularities in the second. They point at 3 major irregularities and 3 administrative
irregularities in both cases. However, according to human coding, Andarai is more corrupt
than Guatapara.42 The corruption index is able to capture this difference.43
40
These are two tests with external coding. I have also tested my measure with the measure the RA and
I created. Indeed, the list of rules was updated until I reached a high correlation with our own hand-coded
measure.
41
For this reason and due to the fact that the coding of the auditors is for a limited sub-sample of the data,
I use my measure rather than the coding of the auditors. My sample spans from lottery 3 to 38, making it
preferable to have a uniform coding throughout the sample period.
42
The language used by the auditors is also harsher.
43
Other examples are municipalities audited in the same lottery (round 21) but in different states such as
Gameleira in the state of Pernambuco and Bom Jesus do Araguaia in Mato Grosso where for each municipality,
there are 11 administrative irregularities where the first is again more corrupt than the second municipality

15
 I also test my corruption measure versus hand-coded measures in the literature. More on
this test can be found in Section 10.4 in the Appendix.

3.2 Electoral Data

Electoral data are obtained from Tribunal Superior Eleitoral (TSE), for municipal elections
in 1996, 2000, 2004, 2008 and 2012 and municipal-level outcomes of presidential elections in
1998, 2002, 2006. The vote share is defined by TSE as the number of valid votes scored by
the party over the total of valid votes in the municipality. A vote is considered valid if it
can be attributed to a candidate (or party). For all elections, the change in the parties’ vote
share over time can be measured, unless the party participates in one election but not in the
other.
I match the political party in the municipality from one time period to the next, in this
way I can follow the evolution of the votes of the party over time. A political party might
participate in one electoral race without necessarily having participated in the previous. In
municipal elections, 43 percent of the parties running in 1996 rerun in the 2000 elections and
46 percent of the parties in the 2000 elections rerun in the 2004 elections. Instead, 59 percent
of the incumbent parties run for reelections from 1996 to 2000 and 57 percent show up for
reelection from 2000 to 2004.

3.3 Antenna Coverages

Geocoded information on FM antenna locations and signal coverage maps are obtained from
FMSCAN.44 Coverage data are in the form of a map raster or a grid file indicating cells
where the signal strength is such that it can be received by a radio. A map of the antenna
coverage is shown in Figure 4.

[Figure 4 Here]

I convert the raster file into multiple polygons. Each polygon contains information about
a single antenna location and coverage. Next, I intersect each of the files containing antenna
in the pair. Other examples are couples such as: Braga in Rio Grande do Sul (lottery 20) and Meleiro in
Santa Catarina audited in lottery 23, Feira Nova do Maranhão in Maranhão and Sao Valerio da Natividade
in the state of Tocantins lottery 21.
44
Information on any country antenna location on their webpage: fmscan.org. I thank Maria Petrova
and Ruben Enikolopov for this reference. I cross-check antenna locations with the locations on antennas
provided by the National Agency of Telecommunications in Brazil (Agência Nacional de Telecomunicações,
(ANATEL). Additional data on media presence at the municipal level from 1999 to 2009, are obtained from
Perfil dos Munı́cipios Brasileiros.

16
information (roughly 6000 antennas) with a file containing information of the population
density, for each municipality, at each 20 square kilometres. In this way, I am able to
compute the fraction of municipal population and the area that each antenna covers.
To compute the population density at each 20 square kilometres, I retrieve population
density estimates of the year 2000 from the Center for International Earth Science Information
Network CIESIN-Columbia University (2005). They come in the form of 2.5 arc-minute grid
cells. To the gridded population, I overlay a map of Brazil’s municipal borders which I
download from the IBGE website. In this way, I can identify which points on the grid that
are reached by a radio signal and I can compute exactly the population density covered by
each FM antenna.45

3.4 Other Controls

To obtain information about socio-economic and demographic characteristics, I combine data
from two main sources, namely Instituto Brasileiro de Geografia e Estatı́stica (IBGE) and In-
stituto de Pesquisa Econômica Aplicada (IPEA). The population census carried out in Brazil
in 2000 provides several socio-economic and demographic characteristics used as controls in
the regressions. Some of these variables are per capita income, income inequality, population
density, share of the population that lives in urban areas, and literacy rate. Yearly popula-
tion and municipal GDP estimates are retrieved from IBGE. For the rest of the variables, I
use census variables and intercensal estimates. Spatial variables such as shapefiles, latitude,
longitude of the population centroid are also retrieved from IBGE. Using the above, I am
able to compute the distance to the nearest state capital, the distance to the federal capital
and the vector of contiguous neighbors for each municipality.

3.4.1 Summary Statistics

Table 2 presents some summary statistics of municipal characteristics measured pre-treatment

(in year 2000). The municipalities in this sample have a smaller population when compared to
the Brazilian average (30,833 inhabitants) also have slightly larger areas (the Brazilian mean
municipal area is 1529 km2 ). The average vote share for the party winning the municipal
elections also winning the first round is 55.22 percent. For municipalities where a second
45
All geographic computations in this paper are carried out using ArcGIS. I exclude non-local radios and
radios for which the coverage goes beyond their contiguous neighbors. The rationale for the latter is that
usually antennas covering more municipalities than immediate neighbors are not considered local radios.

17
round is included, I use the first round data.

[Table 2 Here]

The average number of antennas in the sample is similar to the country average, while
the number of contiguous neighbors is smaller, with a country average of roughly six. Other
characteristics such as male population, literacy rates, illumination, electricity, Human Devel-
opment Index (HDI) and income per capita are similar to the country average. The political
variables are indicators taking value one if the local government, in the electoral term of
2001, was headed by a mayor from that party and zero otherwise.

4 Benchmark Case
I exploit variation in the timing of audits to document the effect of the audit in the audited
municipality across levels of government. For each electoral term, I compare audited munic-
ipalities before elections with audited municipalities after elections.46

Municipal Elections – To study the effect of the audits on the audited municipalities, I
define municipalities audited before municipal elections as the treatment group. The control
group consists of municipalities audited within the first eight months of the new administra-
tion. Therefore, the control group is composed of municipalities that are: (a) audited during
the same electoral term as the municipalities audited before elections, or (b) the handling of
funds mainly corresponds to the previous administration. The outcomes of interest in this
case are the difference in the shares of votes and a dummy variable taking value one if the
incumbent audited party is reelected in municipal elections. I exclude municipalities where
the party wins with a vote share above 90 percent. This implies that less than 4 percent of
the total sample is removed.47

Presidential Elections – When turning to the presidential elections, the question of in-
terest becomes: How does an audit at the local level affect the electoral outcomes of the
party in the presidential elections? To answer this question, I focus on local parties which
46
The same empirical strategy has been used also by Ferraz and Finan (2008); Larreguy, Marshall and
Snyder Jr (2014); Bobonis, Cámara Fuertes and Schwabe (2016)
47
If I include these municipalities, the point estimates remain similar but specifications where the share of
votes is the dependent variable are noisier.

18
have a national counterpart. Therefore, the attention is restricted to the two main players
in national politics, namely PT and PSDB. The reason is that the rest of the parties sup-
port the presidential candidates of PT and PSDB or enter in coalition with other parties.48
The treatment group is composed of audited municipalities ruled either by PT or PSDB in
the period before presidential elections and the control group is composed of municipalities
audited after the presidential elections.49 The dependent variable is the difference in vote
share. In the presidential elections, I only use one cross section – that of the municipalities
audited before and after the elections of 2006. The first differences for this cross section are
constructed by using the electoral outcomes of the year 2002.

4.1 Test of Balance

Before proceeding with the main analysis, I perform several randomization checks despite
the randomization carried out by the CGU. I check if treatment and control groups differ
significantly on observable characteristics by mean comparison for both samples. In Table
1, the groups show no significant difference on most characteristics. However, there are two
main differences that require particular attention: the corruption index and the PT dummy.
The first emerges due to the difference in constructing the corruption index which concerns
rounds 3-7 (see Section 10.3 in the Appendix). However, there are no significant differences
when I exclude those lotteries. In order to maximise the sample size, I include these lotteries
in the main sample for the analysis. In the Appendix, I also show that the main results in
the paper remain unaltered, excluding these lottery rounds.
Another non-balanced characteristic is the indicator variable PT. The test for equality
of means for these coefficients is rejected at conventional significance levels (see the corre-
sponding column in Table 1), indicating that mayors from PT are more likely to be audited
before municipal elections and less likely to be audited before presidential elections. This
would be a source of concern if the audit was aimed to harm PT’s municipal elections and
favour them during presidential elections. However, this scenario seems unlikely as it would
make more sense for the party that has more control over the CGU throughout the sample
period, to consistently refrain from harming itself before elections and attempting to harm
its opponents in all types of elections.
48
In the 2002 and 2006 elections, respectively, 6 and 7 coalitions took place in municipal elections.
49
The treatment group is the set of municipalities audited in the period April 2003-July 2006 and the
control group consists of all audited municipalities in the period from May 2007 to the end of the year 2011.
These results are robust to varying the time span of the control group.

19
 To address these potential concerns, I downloaded all publications regarding the random
audits in the Official Gazette. This information is also available on the CGU website. In this
way, we have the list of eligible municipalities and the numbers allocated to each municipality.
As shown in Figure 10, the allocation of numbers, from 0 to 9, across lotteries from 3-31
mimics a uniform distribution. Lottery 7 is missing because the set of numbers attributed is
missing in the corresponding publication.50 This is a useful test in case one might be worried
that numbers are assigned in such a way that it would increase the chances of selecting some
municipalities and reducing the chances of selecting others.

[Figure 10 Here]

Moreover, I link these data to the audit data and for each lottery, I am able to compute
the fraction of selected numbers per lottery over the total of times that a specific number is
allocated to a municipality. These fractions are depicted in Figure 11. Even in this case, the
distributions seem to be uniform across lotteries. This piece of evidence may be useful to
attenuate possible concerns about a potential randomization failure.
Third, I run a linear probability model to estimate whether a municipality headed by
PT is more or less likely to be audited. I also restrict the sample to municipalities where
mayors are either from PT or PSDB. Results seem to support the fairness of the lottery. This
procedure is discussed in detail in the Appendix.

[Figure 11 Here]

Fourth, the literature has established (and treats) these audits as random.51 . In light of the
above evidence, I have no reason to believe that the randomization was manipulated and used
for political purposes. Anyhow, to deal with imbalances on the party dummies, I include
them in all regressions throughout the paper. For the rest of the characteristics, I present
specifications with and without controls. Results are similar across specifications.

4.2 Empirical Strategy: Benchmark Case

Given that more than 2000 audits have occurred throughout the sample period, I am able
to use both cross section and time variation in my analysis. The estimating equation is
expressed in first differences and can be written as:
50
As in all publications, the number of municipalities per state and those that are non-eligible because of
a previous type of audit (random or not) are clearly stated.
51
Examples regarding this consensus in the literature can be found in Ferraz and Finan (2008, 2011); Brollo
et al. (2013); Litschig and Zamboni (2011); Silva (2010)

20
 ∆Epit = α + β0 Auditpit−s + νs + XiT δ + ηp + τt + εpit (1)

where ∆Epi is the change in the electoral outcome of interest for party p in municipality
i in elections, local or national. Auditpit−s is also an indicator variable. It takes value one
if the municipality i with mayor from party p was audited some time s before elections and
value zero if audited after elections. Note that the model is estimated in first differences, and
the audit dummy can also be written as ∆Auditpit−s given that at time t − 1 the municipality
had no audit. Estimating the model in first differences allows me to take into account time-
invariant characteristics in the municipality. νs is a set of state dummies needed due to the
fact that the lottery randomization is stratified at the state level. Xi is a set of controls as
predetermined to the audit, measured in the year 2000. ηp are the party fixed effects, τt are
time fixed effects and εpit is the idiosyncratic error term. The identifying assumption requires
that once it is conditioned for state fixed and time effects, Auditpit−s is uncorrelated to the
error term. Conditioning for time effects is necessary due to multiple time periods pooled
together.
β0 is the causal effect of the audit on the electoral outcomes of municipality i. We expect
the audit to have no effect, in case voters have correct priors about the party. If individuals
interpret the fact that the auditors show up in the municipality as a signal of a wrongdoing
occurring, a negative effect of the audit on electoral outcomes of the party is expected.
A point worth stressing is that the unit of observation in my analysis is the party in the
municipality. I follow the evolution of the electoral outcome of the incumbent party over
time. A further note regards the feature of the data: In municipal elections, given that few
parties show up for more than two consecutive terms, I cannot fully exploit the longitudinal
feature of the data. The same applies to municipal elections in case of geographical spillovers.
The design is simpler for presidential elections (audited municipalities and neighboring mu-
nicipalities). In presidential elections, both parties show up in both periods, thus allowing
me to take first differences.
In municipal elections, the focus is on incumbent political parties which rerun in the
upcoming elections. To ensure that parties audited before elections do not differentially exit,
I estimate a linear probability model of the exit dummy on the treatment dummy.52 Results
shown in Table A7 in the Appendix, suggest that there are no differences in exits of incumbent
52
Party exits, although they might be a consequence of the policy, would cause selection in the sample.

21
parties in municipalities.53 This exercise is not necessary for presidential elections, because
both parties are always present throughout the duration of the sample period.
To understand the role of corruption of the politician on the party’s electoral outcome,
I interact the main regressor with the corruption index, explained in Section 3.1.3. The
equation of interest becomes:

∆Epit = α + β0 Auditpit−j + β1 Icit−1 + β2 Auditpit−j × Icit−1 + νs + XiT δ + τt + ηp + εpit (2)

where Icit−1 is the corruption index in municipality i committed by the incumbent mayor,
from party p and his administration. The parameter of interest is β2 which measures the
effect of the audit conditional on the levels of irregularities found in the municipality. The
standard errors are clustered at the same level where the treatment occurs, namely at the
municipality level.

4.3 Results: Benchmark Case

In panel A of Table 3, results from Equation 1 are shown. Columns (1) and (2) show a linear
probability model where I regress the dummy for reelection of the party on the audit dummy
in municipality i. In columns (3) and (4) and columns (5) and (6), the outcome variable is
the change in vote shares of the party in municipal and presidential elections.

[Table 3 Here]

It appears that if the incumbent administration, headed by the mayor, is audited before
elections, the mayor’s party is less likely to be reelected as the ruling party in that munici-
pality. The support for the party, expressed as the share of valid votes in the municipality,
is also reduced. The coefficient remains similar in size and magnitude with he inclusion of
different controls (shown in even numbered columns). Parties associated with an incumbent
politician audited before elections are 9 percent less likely to be reelected as the ruling party
and they obtain 2 percentage points lower vote shares than parties associated with politicians
audited after elections. The elections under consideration are close elections.54
53
I estimate a linear probability model on the likelihood of exit of the political party from the municipality.
The coefficient in this case is not significant suggesting the lack of selection of parties into elections. Parties
audited before elections are not more likely to exit from the electoral race. The coefficient goes in the opposite
direction and is not statistically significant.
54
Indeed, 70 percent of the sample are composed of parties that, in the first round, received a vote share
between 40 and 60 percent of the valid votes in the municipality.

22
 In presidential elections, point estimates suggest different magnitudes when compared to
municipal elections. In municipalities where the audited mayor – before elections – is either
from PSDB or PT, the vote share obtained by this party in the presidential elections is lower
than in similar municipalities audited after elections. The size of the coefficients lies around
4 percentage points. The effect is both statistically and economically significant.55
These result imply that audits in local governments do not only have consequences for
the electoral outcomes of the incumbent party in local elections but also on higher levels of
government, such as the presidential elections.
In Panel B of the same table, results from Equation 2 are shown. Columns (1)-(4),
correspond to outcomes in municipal elections. In columns (1) and (3), I interact the policy
dummy with the corruption index. In columns (2) and (4), I present a more flexible functional
form where the policy dummy is interacted with terciles of the corruption index, allowing
the slope to vary for each tercile. The interaction terms are not significant in any of the
specifications, indicating that the effect of the pre-electoral audit in municipal elections does
not depend on corruption levels. The F-stat on the joint significance of the main effects and
the interaction term, except from column (4), indicates that interaction terms are jointly
significant.
Coefficient estimates for presidential elections are shown in columns (5) and (6). In column
(5), the corruption index enters the equation linearly. The coefficient on the interaction term
in column (5) is statistically significant. The F-stat on the interaction terms fails to reject the
null. In column (6), the treatment dummy is interacted with terciles of the corruption index.
The coefficient on the third tercile indicates that moving from the lower tercile, omitted
category, to the upper tercile, the votes of the incumbent party are reduced by 9 percentage
points. The reduction in votes seems to be statistically and economically significant.
Taken together, these results indicate that: i) electoral accountability does also operate
through party identities, and ii) that this effect spills over to different levels of government.
The latter has important implications for the role of parties in selecting and monitoring their
candidates. Apparently, individual malfeasance has strong repercussions on the party.
The findings in municipal elections may seem puzzling when compared to those by Ferraz
and Finan (2008). In their case, the exercise leads to negative but insignificant coefficients
55
I have attempted to estimate the presidential outcomes on the sub-sample of municipalities (with mayors
from PT and PSDB). I do not have much power for this specification. However, the coefficient remains similar
in magnitude. Moreover, I add controls for the timing of municipal elections. Results remain unaltered.

23
(in most specifications) for the audit dummy on electoral outcomes. They find electoral
punishment of the candidate, depending on the disclosed corruption levels. It should be kept
in mind that there are some key differences between their work and mine: First, they focus on
the incumbent candidate, the most important figure in Brazilian politics. I instead focus on
the party, less important at the political stage (60 percent of the candidates in the elections
of 1996 showed up in those of the year 2000 with a different party label). Second, they
focus on all eligible candidates, not only those rerunning for elections, imputing a zero to
the reelection dummy (the main regressor) if the candidate does not show up for reelection.
I follow the parties that rerun for election, either with the same politician or with another
candidate in case the politician does not rerun or is non-eligible. Third, they only use one
cross section of municipalities whereas I extend the sample for three different cross-sections.
As a robustness check, I also show results using only one electoral term, meaning the same
time period where they base their analysis. In doing so, only 1/2 of their sample coincides
with mine.

5 Geographical Spillovers
To understand whether and how the electoral effects of audits spill over on neighboring
municipalities, a definition of a relevant neighborhood is required.56 The media literature
(DellaVigna et al. (2014); Enikolopov, Petrova and Zhuravskaya (2010); Yanagizawa-Drott
(2014); Bursztyn and Cantoni (2015)) suggests that radio and TV play an important role in
information transmission to citizens; hence, I use information on FM transmitters to define
the treated areas, namely informed neighbors.
I define as informed neighbor a municipality that shares a border or a vertex with the
audited municipality and receives antenna coverage from the latter. In this way, I am able
to compare informed neighbors of an audited municipality before elections with informed
neighbors of an audited municipality after elections. This empirical strategy is similar in
spirit to the one used in the benchmark case.57 Given the randomization in audits, both
groups should be homogeneous.
In the case of local elections, I study the electoral effect of an audit on the audited party
56
Several reasonable candidates come to mind, for example: municipalities within a certain distance of
the audited town, municipalities within the same region, trading partners, municipalities sharing common
administrations such as courts, etc.
57
I exclude neighbors which were audited themselves because I want to understand the impact of the audit
on the neighbors if they were not audited.

24
in neighboring municipalities. The sample consists of informed neighbors that at the time
of the audit have had a positive vote share of the party. Note that in this case, the party
is not necessarily the party of the incumbent politician. When turning to the effects on
the presidential elections, the sample is restricted to all informed neighbors of an audited
municipality – with a head of government from either PT or PSDB. The reason is that in
presidential elections, both major parties have a positive vote share throughout the country.
In order to avoid that the same municipality shows up in a treatment and a control group
simultaneously – in municipal and presidential elections – I proceed in the following way: if a
municipality is the neighbor of two different audited municipalities, I allocate the municipality
to either treatment or control as a function of the antenna coverage the municipality receives
from the audited neighbor. The municipality is assigned to the treatment group if it receives
more coverage from a town audited before elections and in the control group if it receives
more coverage from a town audited after elections. If a municipality receives equal coverage
by two different towns, one before elections and the other after, the municipality is allocated
to the treatment group as it is supposed to learn about the audit before elections.
To ensure that the groups are indeed comparable, I perform randomization checks dis-
cussed in Section 10.5 in the Appendix.

5.1 Empirical Strategy: Geographical Spillovers

Similar to the empirical strategy used in the Benchmark Case (see Section 4.2), the estimating
equation is the following:

∆Epjt = α1 + γ0 Auditpit−s + νs + XjT δ + τt + ηp + εpjt (3)

In this case, i is the audited municipality and j corresponds to the informed neighbor.
Epjt is the difference in the vote share in the municipality j, Auditpit−j is an indicator taking
value one in case there was an audit in a nearby municipality i at time s previous to elections
and value zero in case the audit in town i occurred after elections.
All remaining variable definitions are as in Equation 1. γ0 captures the average impact
of the audit in a nearby town i, with incumbent politician from party p, on the votes of that
party in the closest upcoming elections in the informed town j.
When testing for the differential effect of the audit – conditioning on levels of corruption
in the audited municipality – the above equation becomes:

25
 ∆Epjt = α1 + γ0 Auditpit−s + γ1 Icit−1 + γ2 Auditpit−s × Icit−1 + νs + XjT δ + τt + ηp + εpj (4)

where Icit−1 is the level of irregularities committed by the municipal administration in the
nearby municipality. The interest lies in the interaction term, γ2 which captures the effect
of the audit in the nearby municipality conditioning on the corruption level found in that
municipality.
When estimating Equations (3) and (4), standard errors are clustered at the Mesoregion
level. Mesoregion is a subdivision of the Brazilian states, by IBGE, which clusters together
municipalities by geographic proximity and common characteristics. Brazilian states are
subdivided into a finer level, also constructed for statistical purposes, namely Microregions.58
Although Micro and Mesoregions were created for statistical purposes and do not constitute
an administrative area, they are frequently mentioned in the audit reports.59
To check on the robustness of the inference, I also cluster standard errors at the Microre-
gion level and also correct for spatial dependence. A further check is carried allowing for
spatial dependence among neighboring municipalities 100 kilometres apart, following Conley
(1999) with a linear decaying in the spatial weighting (a la Bartlett).60

5.2 Results: Geographical Spillovers

Results for the geographical spillovers are presented in Table 4. Panel A presents results from
Equation (3). Columns (1) and (2) contain the electoral outcomes for municipal elections
whereas (3) and (4) contain those for the presidential elections.

[Table 4 Here]

In local elections, there is no effect of the audit in the neighboring town on the votes
of that party on the informed neighbors. The coefficient γ0 is not statistically significant,
with and without inclusion of controls. In presidential elections, the coefficient is statistically
significant and negative. If there is an audit in a nearby town, the party loses votes in
58
See Table A2 for more on the subdivision of the Brazilian states.
59
Some municipalities within the same Mesoregion have joint federal programs. Auditors also mention
municipalities in the micro or Mesoregion to compare prices among municipalities or to mention that some
beneficiaries of different funds, such as Bolsa Escola, have moved to municipalities within the same e.g.
Micro/Mesoregion.
60
I use the version of the STATA command by Hsiang (2010). Results available upon request.

26
presidential elections. The behaviour of the voters in neighboring municipalities is similar
to the behaviour of the voters in the audited municipality itself. However, there is a small
difference with the coefficient estimates in the audited municipality. The coefficient on β1 is
larger than γ1 , the corresponding estimate in the case of geographical spillovers. A t-statistics
on the equality rejects the null hypothesis that these coefficients have the same magnitude.
In Panel B, variations from Equation (4) are shown. The coefficients of interest are the
interaction terms, either on the linear specification columns (1) and (3) or with a more flexible
functional form, columns (2) and (4). For municipal elections, in both types of specification,
the corruption index enters the equation with a negative sign. Moving from the first to the
second tercile of the distribution of the corruption index decreases the votes of the party
by roughly 8.3 percentage points while moving from the first to the third by 9 percentage
points. In the presidential sample, moving from the first to the third tercile, the vote share is
reduced by 9.8 percentage points. The coefficient of interest, in these separate regressions, is
negative and statistically significant, indicating that for higher levels of the corruption index
the political party is punished harsher in nearby municipalities.
In municipal elections, the findings suggest that voters do not change behaviour once
an investigation in the nearby town occurs. Indeed, in nearby municipalities the incumbent
mayor is not audited. The votes under considerations are those of any party which has
a neighbor whose politician was audited before elections. However, the coefficient on the
interaction term indicates that voters do learn about party characteristics as they became
more salient.
Turning to cross-border and cross-electoral spillovers, a similar argument holds. Here
again the results confirm the similar behaviour of citizens in neighboring municipalities with
those in the audited municipality itself, highlighting that when voting for higher levels of
government, areas sharing the same information network behave similarly.
From the sample size, it seems that on average, for each audited town, two neighboring
municipalities are affected. A back-of-the-envelope calculation suggests that in the audited
neighborhood (audited municipality and the informed neighbors), the vote share of the au-
dited party goes down by 10 percentage points. This is a sizeable effect both politically and
economically.

27
6 Are party labels recognized?
To understand whether party names are indeed recognized, I take advantage of coalition
formation in national elections. I test if an audited party in municipal elections harms the
electoral prospect of the main party in the coalition. If voters recognize party labels, no effect
on the coalition is expected. If the party is linked to candidates with a high enough amount
of disclosed corruption, it might as well be the case that voters vote less for the coalition.
To answer this question, I use as the definition coalition parties that have been in the
same coalition with either PT or PSDB in the previous presidential term.61 The estimating
equation is similar to the above, the only difference is the main regressor.

∆Epit = α0 + β0 Auditrit + νs + XiT δ + ηp + εpit (5)

where Auditrit is an indicator taking value one if party r in coalition with party p is audited
before elections and zero if audited after elections. Results are shown in Table 5. Columns
(1)-(3) present results for the baseline regression with progressive addition of controls. In
columns (4)-(6) the policy dummy is again interacted with the irregularity index. In column
(7) and (6) I interact the policy dummy with the quadratic term of the irregularity index.

[Table 5 Here]

Notice that columns (1) and (2) indicate that the audit of one of the parties on the
coalition seem to have no effect on the votes of the main party in the coalition. This also
seems to be the case when I interact the investigation dummy with the corruption index
or with different terciles of the corruption index. These findings suggest that indeed party
names seem to convey the relevant information to the voters.62

6.1 Heterogeneous treatment

In the case voters truly react to news about corruption, it is important to understand if voters
change behaviour relative to the timing of information provision. To answer this question, I
61
In 2002, parties that were in coalition with PT are PC do B, PCB, PL and PMN. Instead, the only party
in coalition with PSDB was PMDB. In this exercise, I do not consider areas where the mayor was from PCB
as it was part of another coalition in the 2006 elections. These parties are more or less the main coalition
parties throughout the sample period under consideration.
62
Results, not shown, are very similar in the neighboring municipalities.

28
allow for the treatment to be heterogeneous with regards to the timing of the audit before
presidential elections. The estimating equation is stated below:

∆Epit = α + β0 Auditpiy−1 + β1 Auditpiy−2 + β2 Auditpiy−3 + νs + XiT δ + ηp + εpit (6)

Auditpiy−1 is an indicator variable taking value one in case the investigation occurred one
year before presidential elections and zero otherwise. Auditpiy−2 takes value one in case the
audit happened two years before elections and Auditpiy−3 three years before elections. The
remaining variables are defined as in Equation 1. Even in this case, in order to understand
if the electoral punishment varies with the level of irregularities found in the municipality, I
interact all pre-electoral audit dummies with the corruption index.
Results with different variations of Equation 6 are in shown in Table 6.

[Table 6 Here]

The F-test of joint significance rejects the null of all coefficients being zero in each speci-
fication. These findings indicate that regardless of the timing of the audit, voters use audit
information when deciding whom to vote for in national elections. The same pattern seems
to be confirmed in case the policy dummies are interacted with the corruption index.

7 Placebos and Robusteness

7.1 Robusteness
To make sure that my estimates are robust to different sample selections, I run different
robustness tests among different samples that I use for the main analysis.

7.1.1 Choice of the control group

Municipal Elections – As mentioned above, in municipal elections in the audited municipality,

the treatment group is composed of municipalities audited before elections and the outcome of
the audit is also announced before elections. The control group is composed of municipalities
audited after elections, either during the same mayoral term, or those audited within the first
months of the subsequent administration. The latter is justified due to the fact that most of
the irregularities belong to the previous administration. However, it may also be the case that
some of the municipalities in the control group are misclassified. To deal with this concern,

29
I restrict the sample to municipalities audited just before the municipal elections of the year
2004 and just after the same election during the same electoral term. The reason why I only
use the electoral term of 2004 is because it is the only electoral term in which municipalities
are audited before and after elections during the same municipal administration. In the rest
of the electoral term, an audit after election coincides with a new administration.
Results are provided in the top panel of Table 7. Coefficients and standard errors are a bit
larger for the difference in votes. The coefficients on the reelection dummy are also slightly
larger. These differences when compared to coefficients shown in the main specification are
due to the differences in the samples used.

[Table 7 Here]

Presidential Elections – To test the robustness of the estimates in the presidential sample,
I consider an alternative strategy which is selecting treatment and control groups from the
subset of municipalities audited before elections. I compare municipalities where the mayor
is either from PT or PSDB with municipalities in which the mayor is from any other party.63
Results are shown in the bottom panel of Table 7. Apparently, there are not any major
differences between the point estimates in these samples and those that I use in the main
specifications. The coefficient on β1 is similar across those presented in panel B of Table 7.

7.1.2 Excluding the first lottery rounds

Given that the auditors did not provide a list of irregularities found in lotteries from 3-7, I
exclude these lotteries in order to test the robustness of the estimates. This further robustness
test demonstrates that there are no significant differences among the coefficient estimates in
the full sample and those in the samples where lotteries 3-7 are excluded. Tables are shown
in Tables 8 and Table 9, respectively, for the effect of the audit in the audited municipality
and its neighbors.

[Table 8 Here]

[Table 9 Here]
63
In order to be able to assign parties to outcome variables, I use a random random number generator
which attributes a label to each municipality audited before elections, either PT or PSDB.

30
7.2 Placebos
To ensure that the empirical strategy is valid, I employ two different placebo strategies. The
first consists of using a placebo outcome variable. In the second I create a placebo treatment.
The first consists of estimating Equations (1) and (3), I use the differences in the previous
vote shares as the outcome variable. Namely, the differences between vote shares in the
years of 2000 and 1996, with respect to municipal election, and those between 2002 and 1998
in presidential elections. Note that these elections took place before the audit program was
initiated so it is expected that the treatment has no effect on previous vote outcomes. Results
for the benchmark case are shown in Table 10 and those for the geographical spillovers are
shown in Table 11. The coefficient of interest is not statistically significant; thus it is not
possible to reject the null hypothesis of no effect of the audit on previous votes.

[Table 10 Here]

[Table 11 Here]

A difference between these samples used in the placebo and those used in the main
analysis is due to the fact that some municipalities did not yet exist in 1996 or 1998.64
Another difference in the samples, which only refers to the municipal sample, arises from
the fact that the incumbent party in 1996 has not participated in the municipal elections
of 2000, not making it possible to generate the dependent variable, namely the difference in
vote shares.
The second placebo consists in randomly allocating treatment across municipalities. This
is done via simulation of a uniform distribution. I (randomly) assign municipalities to the
treatment group for which the value of the random variable is above the mean and to control
the control group municipalities for which the variable assumes a value below the mean.
These findings indicate that the null hypothesis is not rejected more often than in 5 percent
of the cases. Figure 13 depicts the distribution of the coefficient from the placebo treatment
and the vertical line indicates the true coefficient estimate.

[Figure 13 Here]

The random allocation of the placebo treatment has no effect on the vote share in the
municipality. Both placebo tests confirm the validity of the empirical strategy. I continue to
interpret my findings as valid.
64
New municipalities were born during the 1990’s from the split of several municipalities.

31
8 Mechanism
8.1 The Audit Investigation
The evidence provided in this paper supports the idea that voters react to new information
in neighboring governments, and at higher levels of government. These are novel findings.
Previous analyses had restricted the attention to the local elections within the audited mu-
nicipality. Moreover, these results show how party labels become important even in a setting
where the link between the candidate and the party is considered to be weak.
It is important to note that the coefficient estimates point (except in the case of cross-
border spillovers) at a negative impact of the audit on the electoral outcomes. The negative
electoral effect of audits may seem puzzling at first; the information disclosed regarding mu-
nicipal practices is not necessarily negative. However, it is worth bearing in mind that the
content of the audit reports shows that corruption and other irregularities are widespread in
Brazilian municipalities. In the sample, roughly 90 percent of municipalities have committed
a major irregularity. This is in line with Ferraz and Finan (2008), Brollo et al. (2013) and
Litschig and Zamboni (2011).

The negative effect of audits may arise for several reasons. Some of them are discussed
shortly here below:
Media Bias: Radios may be biased towards only transmitting the negative news dis-
closed in the audits and forgo positive news. Evidence in this direction for Brazil can be
found in Aldé, Mendes and Figueiredo (2008) which shows that most of the information on
candidates reported by the press in 2006 presidential elections was negative (for both PT
and PSDB).

Cognitive Bias: Individuals might only recall negative information and forget positive
news, such as negativity bias (surveyed in Skowronski and Carlston (1989)). This behaviour
has been shown to extend into belief formation about candidates in Lau (1982).

Negative Campaigning: Opponents might use negative audits in a more effective way
as compared to the use of positive audit outcomes by the audited party itself.65 In any
65
This relates to the widespread use of negative campaigning in Brazilian national elections, which has been
documented in Sampaio (2013) for the 2010 presidential dispute and Steibel (2005) for previous presidential

32
case, for voters to learn about the audit contents, the information needs to be disclosed
either by the media or in popular discussion. Anecdotal evidence suggests that voters in-
deed learn about the corruption scandals unveiled from the randomized audit program. As
mentioned in Section 2, these audits have been used actively in the campaign prior to local
elections. There is also anecdotal evidence indicating that local audit outcomes were used
for campaign purposes in presidential elections. For example, in a BBC report of October
2006, some of the voters’ views indicate they do indeed learn about the corruption scandals.66

If one or several of the above mechanisms are in place, they could lead to a negative
average effect of audits on electoral outcomes, as negative information would be used dispro-
portionately in the ballot box. Without data on media content or cognitive bias, I cannot
test for either of the first two arguments. However, I provide some evidence in line with
the argument that the opposition parties might have been using these audits for campaign
purposes.
For this purpose, I use data on campaign expenditures of parties and candidates for the
electoral term of 2004, 2008 and 2012. I regress the logarithm of the total expenditure of
each party on the audit dummy. In Panel A of Table 12, columns (1)-(4) show the spending
pattern of the opposition parties in the audited municipalities. The sign and magnitude of
this coefficient indicate that indeed, in local elections in audited areas, the opposition parties
increase their campaign expenditures. The increase is roughly 20 percent more than in areas
where information about the audit outcome is revealed after elections.
Panel B, columns (5)-(8) show the spending pattern of the opposition parties that partic-
ipated in the electoral race. Apparently, these parties are decreasing spending. These results
seem to point at the fact that parties are now diverting resources to safer areas, where the
winning likelihood is higher.
Regarding the presidential elections, I cannot proceed in the same way for the main
reason that party spending is not detailed at the municipality level. However, in presidential
elections, the main findings of this paper go in line with the possibility that the voters are
reminded about the party’s behaviour at the local level. It is likely that the big scandals
at the national level, where the main parties were blaming and shaming one another about
races.
66
http://kochhars.com/2/hi/talking_point/6091768.stm and television debate http://eleicoes.
uol.com.br/2006/campanha/ultnot/2006/10/08/ult3750u1308.jhtm

33
corruptness, served as a reinforcing mechanism to voters’ memory about the corruption of
the parties at the local level. After all, they were used by the presidential candidates to
attack each other during the presidential (televised) debate.67

8.2 Alternative Mechanism

Vote-Buying: The findings in this paper are also consistent with an alternative interpreta-
tion. The effect on vote reduction may not occur because voters punish corruption but for the
reason that now the incumbent party is no longer able to buy votes. After all, vote-buying is
present at any level of government in Brazil.68 This scenario is difficult to test without data
on actual vote-buying.
However, I am providing two pieces of evidence that may rule out this alternative story.
The first is related to municipal elections. In this case, I refer to the findings from the
campaign expenditures. As it can be seen from Panel A in Table 12, there is no significant
decrease in the amount spent before elections by the incumbent parties. This evidence is
in line with the hypothesis that the incumbent parties, in the audited municipalities, are
not more likely to be financially constrained after an audit. Potentially, their vote-buying
ability remains unaltered. However, in the event that the incumbent party is not able to
do “effective” vote-buying, because now voters are more willing to sell their vote to a less
corrupt party, this may still be considered as a form of electoral punishment. A more careful
investigation of the vote-buying mechanisms at the local level, with better data, is left to
future research.

[Figure 12 Here]

At the presidential level, the additional evidence is brought using data from the Latino-
barómetro surveys. In the wave of the year 2006, Latinobarómetro included a question related
to vote-buying. The question is: Have you known of someone in the last elections who was
pressured or received something to change his vote in a certain way?
67
see http://news.bbc.co.uk/2/hi/6095820.stm and http://www.washingtonpost.com/wp-dyn/
content/custom/2006/09/28/CU2006092800771.html
68
The most clamorous vote-buying scandal during the period of my study and the history of the country
is the vote-buying scandal that broke out on June 6, 2005 when Brazilian Congressional Deputy Roberto
Jefferson told the Brazilian newspaper Folha de S.Paulo that the ruling Partido dos Trabalhadores (PT) had
paid a number of Congressional deputies monthly salaries (around USD 12,000 at the time) to vote for the
legislations that were favoured by the ruling party.

34
 In Table 13, I provide evidence that individuals in areas audited before presidential elec-
tions, and its neighbors, provide a similar answer to individuals in the control group. This
piece of evidence does not rule out the possibility completely; however, the results go in the
expected direction.

[Figure 13 Here]

Parallel Auditing: A second concern, which only regards neighboring municipalities, is the
following: It may be the case that voters in neighboring municipalities are not responding
to information about the party but they are responding to the disclosed corruption level
in their municipality. Indeed, a potential worry might be that in case the auditors find
enough corruption in a municipality, the CGU may start parallel investigations in these
municipalities. This does not seem to be a potential explanation for the following reasons:
First, the vote shares under consideration in the neighboring municipalities in local elections
are not those of the incumbent mayor. Second, I use data from the publication of the lottery
rules. In these documents, the CGU also states the municipalities that are not to be audited
because they have been subject to (non random) audits. The amount of these municipalities
is small, and they are thus unable to influence the share of votes (in the subsample where the
incumbent party is of the same party as in the audited municipality). In light of the above
evidence, I continue to interpret my findings as evidence of electoral punishment.

9 Conclusions
This paper shows the extent to which information about corruption of local politicians af-
fects the electoral outcomes of the party, not only in the same government to which the
information corresponds, but it also spills over across jurisdictions and at higher levels of
elections. The findings are consistent with the hypothesis that information regarding a party
member at one level of government may be used by voters to decide for whom to vote across
levels of elections and borders of government. Within the municipality where the wrongdoing
occurs, the party of the corrupt incumbent experiences a vote reduction and a decline in the
likelihood of reelection. Voters punish the party also in national elections in the audited
neighborhood (audited municipality and its neighbors), suggesting that areas sharing the
same media market behave similarly. In neighboring municipalities, in local elections, voters
punish the party only if the disclosed corruption is high enough. These results have two

35
main implications. First, they show that tools meant to control corruption can have positive
accountability spillovers. Second, they suggest that political parties might have incentives to
better select and monitor their local members.

References
Adena, Maja, Ruben Enikolopov, Maria Petrova, Veronica Santarosa, and Eka-
terina Zhuravskaya. 2015. “Radio and the Rise of The Nazis in Prewar Germany.” The
Quarterly Journal of Economics, 130(4): 1885–1939.

Aldé, Alessandra, Gabriel Mendes, and Marcus Figueiredo. 2008. “Tomando partido:
imprensa e eleições presidenciais em 2006.” Polı́tica & Sociedade, 6(10): 153–172.

Björkman, Martina, and Jakob Svensson. 2009. “Power to the People: Evidence from a
Randomized Field Experiment on Community-Based Monitoring in Uganda.” The Quar-
terly Journal of Economics, 124(2): 735–769.

Boas, Taylor C, and F Daniel Hidalgo. 2011. “Controlling the airwaves: Incumbency ad-
vantage and community radio in Brazil.” American Journal of Political Science, 55(4): 869–
885.

Bobonis, Gustavo J, Luis R Cámara Fuertes, and Rainer Schwabe. 2016. “Moni-
toring Corruptible Politicians.” American Economic Review, 106(8): 2371–2405.

Brollo, Fernanda, Tommaso Nannicini, Roberto Perotti, and Guido Tabellini.

2013. “The Political Resource Curse.” American Economic Review, 103(5): 1759–1796.

Bursztyn, Leonardo, and Davide Cantoni. 2015. “A tear in the iron curtain: The impact
of western television on consumption behavior.” Review of Economics and Statistics.

Center for International Earth Science Information Network CIESIN-

Columbia University, Center. 2005. “Gridded Population of the World, Version 3
(GPWv3).”

Conley, Timothy G. 1999. “GMM estimation with cross sectional dependence.” Journal
of econometrics, 92(1): 1–45.

36
Costas-Pérez, Elena, Albert Solé-Ollé, and Pilar Sorribas-Navarro. 2012. “Cor-
ruption scandals, voter information, and accountability.” European Journal of Political
Economy, 28(4): 469–484.

DellaVigna, Stefano, and Ethan Daniel Kaplan. 2007. “The Fox News Effect: Media
Bias and Voting.” The Quarterly Journal of Economics, 122(3): 1187–1234.

DellaVigna, Stefano, Ruben Enikolopov, Vera Mironova, Maria Petrova, and

Ekaterina Zhuravskaya. 2014. “Cross-Border Media and Nationalism: Evidence from
Serbian Radio in Croatia.” American Economic Journal: Applied Economics, 6(3): 103–
132.

Desposato, Scott W. 2004. “The impact of federalism on national party cohesion in Brazil.”
Legislative Studies Quarterly, 29(2): 259–285.

Desposato, Scott W. 2006. “Parties for rent? Ambition, ideology, and party switching in
Brazil’s chamber of deputies.” American Journal of Political Science, 50(1): 62–80.

Di Tella, Rafael, and Ernesto Schargrodsky. 2003. “The role of wages and auditing dur-
ing a crackdown on corruption in the city of buenos aires*.” Journal of law and economics,
46(1): 269–292.

Enikolopov, Ruben, Maria Petrova, and Ekaterina Zhuravskaya. 2010. “Media and
political persuasion: Evidence from Russia.” American Economic Review, 101(7): 3253–85.

Farvacque-Vitkovic, Catherine D, and Mihaly Kopanyi. 2014. Municipal finances: A

handbook for local governments. World Bank Publications.

Ferraz, Claudio, and Frederico Finan. 2008. “Exposing Corrupt Politicians: The Effects
of Brazil’s Publicly Released Audits on Electoral Outcomes.” The Quarterly Journal of
Economics, 123(2): 703–745.

Ferraz, Claudio, and Frederico Finan. 2011. “Electoral Accountability and Corrup-
tion: Evidence from the Audits of Local Governments.” American Economic Review,
101(4): 1274–1311.

Folke, Olle, Torsten Persson, and Johanna Rickne. 2016. “The Primary Effect: Pref-
erence Votes and Political Promotions.” American Political Science Review.

37
Gentzkow, Matthew. 2006. “Television and voter turnout.” The Quarterly Journal of
Economics, 931–972.

Gonzalez, Robert M. 2016. “Social Monitoring and Corruption in Developing Countries.”

Hayes, Philip J, and Steven P Weinstein. 1990. “CONSTRUE/TIS: A System for

Content-Based Indexing of a Database of News Stories.” Vol. 90, 49–64.

Hsiang, Solomon M. 2010. “Temperatures and cyclones strongly associated with economic
production in the Caribbean and Central America.” Proceedings of the National Academy
of sciences, 107(35): 15367–15372.

IBOPE. 2012. “Media Book 2012: Hábitos de Mı́dia e Investimento Publicitário em 2011.”
IBOPE.

Jacobs, Paul S., and Lisa F Rau. 1990. “SCISOR: Extracting information from on-line
news.” Communications of the ACM, 33(11): 88–97.

Larreguy, Horacio A, John Marshall, and James M Snyder Jr. 2014. “Revealing
Malfeasance: How Local Media Facilitates Electoral Sanctioning of Mayors in Mexico.”
National Bureau of Economic Research.

Lau, Richard R. 1982. “Negativity in political perception.” Political Behavior, 4(4): 353–
377.

Litschig, Stephan, and Yves Zamboni. 2011. “Audit risk and rent extraction: evidence
from a randomized evaluation in Brazil.” WP.

Manning, Christopher D., Prabhakar Raghavan, and Hinrich Schütze. 2010. Intro-
duction to Information Retrieval, Cambridge University Press. 2008. ISBN-13 978-0-521-
86571-5. Cambridge Univ Press.

Olken, Benjamin A. 2007. “Monitoring Corruption: Evidence from a Field Experiment in

Indonesia.” Journal of Political Economy, 115(2).

Pereira, Carlos, and Bernardo Mueller. 2003. Partidos fracos na arena eleitoral e par-
tidos fortes na arena legislativa: a conexão eleitoral no Brasil. SciELO Brasil.

Porto, Mauro. 1985. Mass media and politics in democratic Brazil.

38
Reinikka, Ritva, and Jakob Svensson. 2005. “Fighting corruption to improve schooling:
Evidence from a newspaper campaign in Uganda.” Journal of the European Economic
Association, 3(2-3): 259–267.

Sampaio, Thiago. 2013. “Ataques e contra-ataques: campanhas negativas na disputa pres-

idencial de 2010.” mimeo.

Silva, Pedro. 2010. “Learning to fear the Inspector-General: measuring spillovers from
anti-corrupt policies.”

Skowronski, John J, and Donal E Carlston. 1989. “Negativity and extremity biases in
impression formation: A review of explanations.” Psychological bulletin, 105(1): 131.

Snyder, James M, and David Strömberg. 2010. “Press Coverage and Political Account-
ability.” Journal of Political Economy, 118(2): 355–408.

Steibel, Fabro B. 2005. “Campanha negativa: construindo o objeto de estudo.” Contem-

poránea.

Strömberg, David. 2004. “Radio’s impact on public spending.” The Quarterly Journal of
Economics, 189–221.

Timmons, Jeffrey F, and Francisco Garfias. 2015. “Revealed Corruption, Taxation,

and Fiscal Accountability: Evidence from Brazil.” World Development, 70: 13–27.

Tirole, Jean. 1996. “A theory of collective reputations (with applications to the persistence
of corruption and to firm quality).” The Review of Economic Studies, 63(1): 1–22.

Treisman, Daniel. 2000. “The causes of corruption: a cross-national study.” Journal of

public economics, 76(3): 399–457.

Yanagizawa-Drott, David. 2014. “Propaganda and conflict: Evidence from the Rwandan
genocide.” The Quarterly Journal of Economics, 129(4): 1947–1994.

39
 Figure 1: Audit and Election Timeline

Notes: In the horizontal axes lottery rounds and the corresponding elections are depicted. The vertical axes show the number
of municipalities drawn per round.

40
 Figure 2: Distribution of PT and PSDB

Notes: Municipalities that had a mayor either from PT (in Red) or PSDB (in Blue) during the 2001 electoral term.

41
 Figure 3: Parties and Candidates

ELEIÇÕES MUNICIPAIS TREINAMENTO

JUSTIÇA
ELEITORAL
TREINAMENTO
SEU VOTO PARA
VEREADOR

?
1 2 3

4 5 6

Número: 7 8 9
Nome: ____
0
Partido: ____
Aperte a tecla: BRANCO CORRIGE CONFIRMA
VERDE para CONFIRMAR este voto
LARANJA para REINICIAR este voto

PARTIDOS LISTA DE CANDIDATOS

PEsp VEREADORES PREFEITO VICE-PREFEITO

91 PARTIDO
DOS ESPORTES

42

Vôlei Basquete Atletismo Futebol Tênis Natação Judô

91001 91002 91003 91004 91005 91

92 PMus
VEREADORES PREFEITO VICE-PREFEITO

PARTIDO DOS
RITMOS MUSICAIS

Rock Sertanejo Frevo Música Popular Reggae Forró Samba

Brasileira
92001 92002 92003 92005 92
92004

PProf
93
VEREADORES PREFEITO VICE-PREFEITO

PARTIDO
DAS PROFISSÕES

Operário Pintor Cozinheira Médica Enfermeira Bombeira Professora

93001 93002 93003 93004 93005 93

94 PFest
VEREADORES PREFEITO VICE-PREFEITO

PARTIDO DAS
FESTAS POPULARES

Páscoa Oktoberfest Folia de Reis Boi-Bumbá Cavalhadas Festa Junina Carnaval

94001 94002 94003 94004 94005 94

PFolc
95
VEREADORES PREFEITO VICE-PREFEITO

PARTIDO
DO FOLCLORE

Curupira Saci-Pererê Iara Lobisomem Mapinguari Boto Boitatá

Cor-de-Rosa
95001 95002 95003 95004 95005 95

Notes: Example made by the TSE. It depicts the voting machine and the list of candidates and parties. Voters are required to
enter the numbers associated with the pictures. The first vote is for the members of the municipal chamber and second vote is
for the mayor.

42
 Figure 4: Radio Coverage

Notes: FM antenna coverages in the Brazilian territory.

43
 Table 1: Balance Test (Audited Municipalities)

Municipal Elections Presidential Elections

Treated Control t-stat Treated Control t-stat
Corrupt. Index -0.02 0.17 -1.62 0.09 -0.10 0.98
Corrupt. Index (Reduced Sample) 0.04 0.17 -1.04 0.28 -0.10 1.59
Pop. (2000) 24452.06 27508.28 -1.01 24026.82 32007.67 -1.49
Literacy rate 78.22 78.14 0.09 76.82 78.62 -1.45
Frac. Urban 0.60 0.60 -0.07 0.59 0.60 -0.51
Frac. Men 0.51 0.51 1.19 0.51 0.51 0.86
Frac. Illumination 0.22 0.23 -0.39 0.21 0.22 -1.21
Electricity TV 75.21 75.50 -0.22 71.00 73.31 -1.04
HDI 0.70 0.70 0.20 0.69 0.70 -1.29
Income per capita 171.22 168.84 0.38 160.26 170.15 -1.08
Contiguous Nbd 5.82 6.11 -2.23 6.07 5.89 0.86
Nr Antennas 2.71 2.80 -0.52 2.87 2.63 0.92
Latitude -15.65 -15.93 0.48 -14.22 -15.31 1.27
Longitude -45.92 -45.86 -0.14 -47.02 -46.44 -0.86
Max Distance 83.76 82.32 0.31 101.36 93.57 0.79
Min Distance 16.17 17.07 -1.08 19.05 17.76 0.67
Dist. State capital 232.43 249.25 -1.56 262.96 242.16 1.22
Dist. Federal capital 1160.46 1111.18 1.64 1115.39 1074.70 0.83
Surface 1279.32 1793.75 -1.08 2269.20 2001.69 0.37
Electorate 15784.11 17611.44 -0.98 15853.48 21267.95 -1.56
Residual Vote 0.07 0.07 -0.90 0.14 0.14 -0.95
Munic. Council 9.42 9.38 0.52 11.02 11.52 -1.64
PT 0.11 0.07 2.32 0.19 0.38 -4.37
PSDB 0.16 0.17 -0.23
PTB 0.05 0.05 0.33
PMDB 0.23 0.26 -0.98
PP 0.09 0.11 -1.03
PFL 0.14 0.16 -0.72
Obs 594 379 181 206
Notes: Treated is the group of municipalities audited before elections. Control is the group of municipalities audited after
elections. Population and Literacy rate population and literacy rate in the municipality. Frac. Urban, Men and Illumination
are, respectively, the fraction of urban population, men and household reporting to have sources of illumination. Electricity TV
is the rate of households reporting to have current electricity and TV in the household. HDI stands for Human Development
Index, measured at the municipality. Income per capita is measured in $Reais. All the above are measured in the year 2000.
Contiguous Nbd measures the number of contiguous municipalities, sharing a border or a vertex with a nearby municipality. Nr
of Antennas measures the number of FM antennas placed in the municipality. Latitude and Longitude are coordinates of the
population centroid in the municipality. Max and Min Distance are, respectively, the farthest and the shortest distance from
the population centroid of the audited municipality to the population centroid of the contiguous neighbors. Dist. to State and
Federal capital measure the distances of municipalities from the state capital and Brası́lia, the federal capital. Surface is the
municipal area measured in square km. Electorate is the total electorate in the municipality, Residual votes is the fraction of
invalid votes, both measured in the electoral term previous to the audit, namely the year 2000 for municipal elections and 2002
for presidential elections. Munic. Council is the size of the municipal council measured as in 2000 electoral term. PSDB, PT,
PTB, PMDB, PP, PFL are indicators taking the value of one if its the party of the audited mayor.

44
 Table 2: Descriptive Statistics

(1) (2) (3) (4)

VARIABLES mean sd min max

Literacy rate 77.86 12.51 39.34 98.13

Pop. (2000) 25,414.61 46,006.36 795.00 461,534.00
Frac. Urban 0.59 0.23 0.04 1.00
Frac. Men 0.51 0.01 0.47 0.58
Frac. Illumination 0.22 0.06 0.03 0.32
Electricity TV 74.34 20.30 6.24 99.35
HDI 0.70 0.08 0.47 0.92
Income per capita 167.04 94.58 34.53 834.00
Contiguous Nbd 5.96 1.99 0.00 14.00
Nr Antennas 2.71 2.64 0.00 14.00
Dist. State capital 241.45 162.25 0.00 1,154.59
Dist. Federal capital 1,132.96 456.29 33.08 2,695.80
Surface 1,741.08 7,189.62 15.36 159,695.94
Electorate 16,340.64 28,722.38 979.00 313,221.00
Residual Votes 0.07 0.03 0.01 0.44
Munic. Council 11.01 2.87 9.00 21.00
Number Parties 2.79 1.04 1.00 9.00
Vote Share 55.22 11.67 23.89 100.00
PSDB 0.21 0.41 0.00 1.00
PT 0.04 0.20 0.00 1.00
PTB 0.06 0.24 0.00 1.00
PMDB 0.23 0.42 0.00 1.00
PP 0.12 0.32 0.00 1.00
PFL 0.17 0.37 0.00 1.00

Obs 1089 1089 1089 1089

Notes: Summary statistics for all municipalities under analysis. Literacy rate is the fraction of literate population, Frac. Urban,
Men and Illumination TV are, respectively, the fraction of urban population, men in the population and households that report
both having electricity and TV in the household. HDI is the human development index, income per capita is that of the household
in the municipality. Contiguous Nbd is the number of contiguous neighbors for each municipality, Nr of Antennas is the average
number of antennas in the municipality. Distance to State and Federal Capital are measured in kilometres. Electorate is the
number of people eligible to vote. Residual Vote is the fraction of invalid votes over total of valid votes. Council, Number of
Parties and Vote Share are, respectively, the size of the municipal council, the number of parties in the municipality and the vote
share of the winning party in year 2000. The political variables are indicators taking the value of one if the local government,
in the electoral term of 2000 (pre-audit) was headed by a mayor from that party and zero otherwise.

45
 Figure 5: Summary Example

Notes: Example of a summary in the municipality Curiuva in the state of Paraná in lottery 11. For each ministry, the irregularities
found are listed. Lei no 9.452/97 and Lei no 8.666/93 indicate, respectively, the law on Public Procurement and that on the
announcement of the received funds (from the federal authorities).

46
 Figure 6: Simulation of Procurement Process

Notes: An example of a simulation of a procurement process. The call for bids was for cleaning products. All bids submitted
intended to supply food. Not only is the layout of the bids the same, but they also coincide also on the type of the orthographic
mistakes. The only slight change the auditors point out is the font used across bids.

47
 Figure 7: Simulation/Fraud in Procurement Process

Notes: A second example of fraud in a procurement process. Three firms “participated” in the call for bids. The winner of the
contract was declared to be the firm with the unique identification number (CNPJ) 07.981.189/0001-90. However, none of the
firms were located in the domiciles indicated in the documents.

48
 Figure 8: Corruption Measures

Panel A

Panel B

Notes: Panel A depicts the correlation of the count of major irregularities and the first principal component of the following
variables: number of pages, number of lines, number of images and total amount of irregularities found in the municipality.
Panel B depicts the distribution of the first principal component of all measures jointly.

49
 Figure 9: Explained Variation

Notes: Explained variance principal component. The first principal component retains 80 percent of the explained variation.

50
 Figure 10: Numbers Attributed per Lottery

Notes: Each graph plots the numbers attributed to municipalities for a specific lottery. The distribution of numbers mimics a
uniform distribution consistent with a fair lottery.

51
 Figure 11: Numbers Selected per Lottery

Notes: Each graph plots the fraction of numbers selected over the total that number was attributed in that lottery. The
distribution of numbers within each lottery mimics a uniform distribution consistent with a fair lottery.

52
 Figure 12: Attributed and Selected Lottery Numbers

Notes: Fraction recording the number of times that the number is selected over the number of times that the number is allocated
for lotteries 3-6 and 7-31.

53
 Table 3: Direct Effect of Audit on Electoral Outcomes in the Audited Municipality

Local Elections Presidential Elections

Reelection ∆ VS ∆ VS
P anelA (1) (2) (3) (4) (5) (6)

Audit -0.094*** -0.099*** -1.952* -2.309** -4.135** -4.731***

(0.035) (0.036) (1.060) (1.099) (1.698) (1.667)

R-squared 0.066 0.116 0.091 0.145 0.290 0.334

State FE X X X X X X
Party FE X X X X X X
Controls X X X
P anelB

Audit -0.111*** -0.121** -2.609** -1.331 -4.183*** 0.006

(0.036) (0.058) (1.111) (1.979) (1.562) (2.420)
Audit * Corrupt. Index 0.013 0.249 -1.662**
(0.019) (0.536) (0.731)
Audit * 2o Tercile 0.054 1.949 -3.364
(0.081) (2.577) (3.592)
Audit * 3o Tercile 0.055 -0.225 -8.807**
(0.082) (2.587) (3.488)
Corrupt. Index -0.041** -1.012** 0.391
(0.018) (0.483) (0.609)
20 Tercile -0.108* -4.267** 0.216
(0.064) (2.108) (2.445)
30 Tercile -0.146** -3.224 1.352
(0.068) (2.196) (2.877)

Observations 973 973 973 973 387 387

R-squared 0.123 0.067 0.150 0.070 0.480 0.482
State FE X X X X X X
Party FE X X X X X X
Controls X X X X X X
Mean Corrupt. Index 0.0556 0.0556 0.0556 0.0556 -0.0121 -0.0121
F-test joint significance 5.120 2.369 3.383 0.480 4.365 4.475
Notes: State fixed effects and population measures in the year 2000 are included in all regressions. Municipal elections, columns
(1) - (4), also include term fixed effects. Controls are: Income per capita, fraction of urban population, fraction of male population
and literacy rate measured in 2000 and surface area, latitude and longitude, distance to state capital, Human Development Index,
and electricity in the household. Panel B in columns (1), (3) and (5) show an interaction with the corruption level in linear form.
Columns (2), (4) and (6) allow for a more flexible form in which terciles of the corruption index are interacted with all regressors.
Standard errors clustered at the municipality level in parentheses. *** Significant at the 1 percent level, ** Significant at the 5
percent level, * Significant at the 10 percent level.

54
 Table 4: Electoral Outcomes in neighboring Municipalities

Local Elections Presidential Elections

∆ VS ∆ VS
P anelA (1) (2) (3) (4)

Audit 0.313 -0.135 -3.308* -3.491*

(1.679) (1.669) (1.858) (1.872)

R-squared 0.087 0.177 0.322 0.345

State FE X X X X
Party FE X X X X
Controls X X
P anelB

Audit -0.645 5.195 -3.134* 2.120

(1.663) (3.273) (1.708) (2.175)
Audit * Corrupt. Index -2.130** -2.306***
(0.917) (0.783)
Audit * 20 Tercile -8.297* -4.858
(4.849) (3.844)
Audit * 30 Tercile -9.084** -9.847***
(4.351) (3.218)
Corrupt. Index 0.913 0.196
(0.664) (0.862)
20 Tercile 4.973 1.402
(4.217) (2.777)
30 Tercile 4.814 -1.432
(3.393) (3.080)

Observations 465 465 871 871

R-squared 0.139 0.140 0.490 0.499
State FE X X X X
Party FE X X X X
Controls X X X X
Mean Corrupt. Index -0.00490 -0.00490 -0.00515 -0.00515
F-test joint significance 1.950 1.667 4.468 3.553
Notes: State fixed effects and population measure in the year 2000 are included in all regressions. The municipal elections
column (1) - (2) also include term fixed effects. Panel B, columns (1) and (3) show an interaction with the corruption level in
linear form. Columns (2) and (4) allow for a more flexible form in which terciles of the corruption index are interacted with
all regressors. Controls include: income per capita, fraction of the urban population, fraction of the male population, and the
literacy rate, and surface area, latitude and longitude, distance to state capital, Human Development Index, and electricity in
the household. Standard errors clustered at the mesoregion area. *** Significant at the 1 percent level, ** Significant at the 5
percent level, * Significant at the 10 percent level.

55
 Table 5: Parties in coalition

Dependent Variable: Change in the Vote Share of the main coalition party
(1) (2) (3) (4) (5) (6) (7)

Audit 0.334 0.345 0.298 0.272 0.425 0.339 2.255

(1.472) (1.303) (1.319) (1.379) (1.330) (1.338) (2.243)
Audit * Corrupt. Index -0.926 -0.767 -1.073
(0.828) (0.778) (0.763)
Audit * 20 Tercile -1.145
(3.014)
Audit * 30 Tercile -4.436
(3.536)
Corrupt. Index -0.162 -0.006 0.092
(0.442) (0.404) (0.430)
20 Tercile -2.157
(2.394)
30 Tercile -0.519
(2.598)

Observations 498 498 498 489 489 489 489

R-squared 0.339 0.484 0.502 0.440 0.481 0.502 0.460
State FE X X X X X X X
Party FE X X X X X
Controls X X X
Notes: The effect of the audit on the electoral outcomes of the main party in the coalition. State fixed effects, party fixed effect,
and population measured in the year 2000, are included as controls in all regressions. Controls include income per capita and
population, the fraction of the urban population, the fraction of the male population and the literacy rate, surface area, latitude
and longitude, distance to state capital, Human Development Index, and electricity in the household. Standard errors clustered
at the municipality level in parenthesis. *** Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant
at the 10 percent level.

56
 Table 6: Heterogeneous Treatment (Presidential Elections)

Dependent Variable: Change in the Vote Share

P anelA (1) (2) (3) (4)

1 year Before Elections -4.970** -5.534** -4.387* -4.890**

(2.246) (2.312) (2.302) (2.395)
2 year Before Elections -3.743* -3.675* -3.725* -3.581*
(1.909) (1.919) (1.920) (1.938)
3 year Before Elections -3.084 -3.386 -4.066* -4.391*
(2.095) (2.131) (2.360) (2.408)
1 year Before Elections * Corrupt. Index -1.873** -1.737**
(0.864) (0.845)
2 year Before Elections * Corrupt. Index -1.209 -1.225
(1.022) (1.047)
3 year Before Elections * Corrupt. Index -2.997 -3.166
(2.034) (2.114)

Observations 387 387 387 387

R-squared 0.458 0.473 0.466 0.481
State FE X X X X
Party FE X X X X
Controls X X
P anelB

1 year Before Elections -3.124 -3.026 -2.956 -2.714

(2.661) (2.868) (2.611) (2.852)
2 year Before Elections -3.298 -3.823 -4.098 -4.350
(2.563) (2.820) (2.474) (2.658)
3 year Before Elections -1.190 -1.070 -2.391 -2.192
(2.378) (2.555) (2.775) (2.994)
1 year Before Elections * Corrupt. Index -1.932** -2.081**
(0.844) (0.856)
2 year Before Elections * Corrupt. Index -2.851** -3.040**
(1.252) (1.433)
3 year Before Elections * Corrupt. Index -1.751 -1.983
(2.095) (2.246)

Observations 871 871 871 871

R-squared 0.459 0.476 0.476 0.491
State FE X X X X
Party FE X X X X
Controls X X
Notes: Panel A and B show results from the audited municipality and neighbors, respectively. State fixed effects and population
measure in the year 2000 are included in all regressions. Controls include: income per capita, fraction of the urban population,
fraction of the male population, and the literacy rate, and surface area, latitude and longitude, distance to state capital,
Human Development Index, and electricity in the household. Standard errors clustered at the municipality level for the audited
municipality and at the mesoregion area in the geographical spillovers.*** Significant at the 1 percent level, ** Significant at
the 5 percent level, * Significant at the 10 percent level.
57
 Table 7: Robustness (Change of the control group)

Outcome Party Reelected Changes in Vote Share

(1) (2) (3) (4) (5) (6)

Audit -0.128* -0.130* -0.148* -4.779* -4.638* -5.296**

(0.076) (0.074) (0.077) (2.525) (2.499) (2.571)
Audit * Corrupt. Index 0.002 -0.235
(0.046) (1.538)
Corrupt. Index -0.032 -1.014
(0.041) (1.367)

Observations 302 302 302 302 302 302

R-squared 0.140 0.220 0.224 0.161 0.230 0.235
State FE X X X X X X
Party FE X X X X X X
Controls X X X X

Outcome: Changes in Party’s Vote Share

(1) (2) (3) (4) (5) (6)

Audit -3.567* -4.095* -4.104* -3.387 -3.955* -3.956*

(2.151) (2.132) (2.188) (2.077) (2.083) (2.133)
Audit * Corrupt. Index -3.398*** -3.384*** -3.300***
(1.040) (1.038) (1.106)
Corrupt. Index 0.224 0.148 0.082
(0.623) (0.631) (0.656)

Observations 558 558 558 558 558 558

R-squared 0.122 0.123 0.135 0.137 0.139 0.150
State FE X X X X X X
Party FE X X X X
Controls X X
Notes: Municipal election are shown in the top panel. The bottom panel shows results from presidential elections. State fixed
effects and population measure in the year 2000 are included in all regressions. In municipal elections, term fixed effects are
also included. The dependent variable in the top panel is, respectively, the change in the reelection status in columns (1) - (3)
and the difference in vote shares in columns (4) - (6). State fixed effects and population, measured in year 2000, are included
as controls. Results for presidential elections are presented in the bottom panel, where the treatment group is a municipality
audited in one of the two main political parties and the control group is composed of municipalities audited in any other party.
Controls include: income per capita, fraction of the urban population, fraction of the male population, and the literacy rate,
and surface area, latitude and longitude, distance to state capital, Human Development Index, and electricity in the household.
Standard errors clustered at the municipality level, in parentheses.*** Significant at the 1 percent level, ** Significant at the 5
percent level, * Significant at the 10 percent level.

58
 Table 8: Excluding Lotteries 3-7 (Audited Municipality)

Local Elections Presidential Elections

Reelection ∆ VS ∆ VS
P anelA (1) (2) (3) (4) (5) (6)

Audit -0.097** -0.106*** -2.097* -2.644** -4.743** -5.380***

(0.040) (0.041) (1.249) (1.292) (1.934) (1.918)

R-squared 0.088 0.147 0.098 0.154 0.283 0.329

State FE X X X X X X
Party FE X X X X X X
Controls X X X
P anelB

Audit -0.112*** -0.117*** -2.165* -2.875** -4.125** -4.348**

(0.039) (0.041) (1.245) (1.298) (1.788) (1.809)
Audit * Corrupt. Index 0.016 0.015 0.555 0.418 -1.706** -1.593**
(0.019) (0.020) (0.542) (0.539) (0.796) (0.788)
Corrupt. Index -0.047*** -0.043** -1.077** -0.975** 0.132 0.027
(0.018) (0.018) (0.500) (0.493) (0.594) (0.637)

Observations 848 848 848 848 332 332

R-squared 0.082 0.155 0.063 0.159 0.458 0.476
State FE X X X X X X
Party FE X X X X X X
Controls X X X
Mean Corrupt. Index 0.0965 0.0965 0.0965 0.0965 0.0424 0.0424
F-test joint significance 5.436 4.727 2.554 2.871 4.217 4.268
Notes: Municipal elections are shown in the top panel. They also include state fixed effects and population as of the year
2000. The bottom panel shows results from presidential elections. In the municipal elections, columns (1)-(4), term fixed effects
are included. Controls include: income per capita, fraction of the urban population, fraction of the male population, and the
literacy rate, and surface area, latitude and longitude, distance to state capital, Human Development Index, and electricity
in the household. Standard errors clustered at the municipality level in parentheses.*** Significant at the 1 percent level, **
Significant at the 5 percent level, * Significant at the 10 percent level.

59
 Table 9: Excluding Lotteries 3-7 (Geographical Spillovers)

Local Elections Presidential Elections

∆ VS ∆ VS
P anelA (1) (2) (3) (4)

Audit -0.582 -0.315 -4.653** -5.044**

(1.805) (1.847) (2.225) (2.183)

R-squared 0.076 0.213 0.310 0.343

State FE X X X X
Party FE X X X X
Controls X X
P anelB

Audit -0.879 -0.729 -4.082** -4.118**

(1.750) (1.840) (1.968) (2.003)
Audit * Corrupt. Index -2.105** -2.120** -2.417*** -2.582***
(0.863) (0.847) (0.833) (0.856)
Corrupt. Index 1.487** 1.205 -0.313 -0.178
(0.679) (0.737) (0.987) (1.132)

Observations 383 383 746 746

R-squared 0.089 0.225 0.474 0.495
State FE X X X X
Party FE X X X X
Controls X X
Mean Corrupt. Index 0.0112 0.0112 0.0346 0.0346
F-test joint significance 2.455 2.183 5.476 5.689
Notes: Municipal election are shown in the top panel. The bottom panel shows results from presidential elections. State
fixed effects are included and population as of the year 2000. In the municipal elections, columns (1)-(4), term fixed effects
are included. Controls include: income per capita, fraction of the urban population, fraction of the male population, and the
literacy rate, and surface area, latitude and longitude, distance to state capital, Human Development Index, and electricity in
the household. Standard errors clustered at the mesoregion area. *** Significant at the 1 percent level, ** Significant at the 5
percent level, * Significant at the 10 percent level.

60
 Table 10: Placebo Audited Municipality

Local Elections Presidential Elections

Reelection ∆ VS ∆ VS
P anelA (1) (2) (3) (4) (5) (6)

Audit -0.025 -0.014 -0.362 -0.221 0.538 0.713

(0.047) (0.049) (1.658) (1.720) (1.368) (1.445)

Observations 540 540 540 540 325 325

R-squared 0.064 0.130 0.131 0.219 0.812 0.818
State FE X X X X X X
Party FE X X X
Controls X X X
Notes: All regressions include state fixed effects and population as of the year 2000. Controls include: income per capita, fraction
of the urban population, fraction of the male population, and the literacy rate, and surface area, latitude and longitude, distance
to state capital, Human Development Index, and electricity in the household. Standard errors clustered at the municipality level
in parenthesis. *** Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant at the 10 percent level.

Table 11: Placebos Geographical Spillovers

Local Elections Presidential Elections

∆ VS ∆ VS
(1) (2) (3) (4)

Audit -0.171 0.168 -1.028 -1.353

(2.372) (2.678) (1.246) (1.185)

Observations 255 255 828 828

R-squared 0.185 0.232 0.817 0.823
State FE X X X X
Party FE X X
Controls X X
Notes: In the municipal regressions, columns (1) and (2), term fixed effects are included. All regressions include state fixed effects
and population as of the year 2000. Controls include: income per capita, fraction of the urban population, fraction of the male
population, and the literacy rate, and surface area, latitude and longitude, distance to state capital, Human Development Index,
and electricity in the household. Standard errors clustered at the municipality level and at the mesoregion level respectively for
audited municipality and neighboring municipalities in parenthesis. *** Significant at the 1 percent level, ** Significant at the
5 percent level, * Significant at the 10 percent level.

61
 Figure 13: Coefficient Distribution for Placebo Test

Panel A: Audited Municipal Panel B: Audited Presidential

Panel C: neighbors Municipal Panel C: neighbors Presidential

Density of predicted β0 from the placebo regressions. Each vertical line represents the esti-
mated value of the true coefficient.

62
 Table 12: Campaign Expenditure

Opposition Parties Incumbent Parties

(1) (2) (3) (4) (5) (6) (7) (8)
Panel A
Audit 0.229** 0.242** 0.208* -0.174 -0.137 -0.179 -0.232* -0.347
(0.114) (0.116) (0.116) (0.213) (0.127) (0.130) (0.128) (0.253)
Audit * Corrupt. Index -0.010 0.029
(0.059) (0.066)
Corrupt. Index -0.041 -0.027
(0.051) (0.056)
Audit * 20 Tercile 0.844*** 0.237
(0.281) (0.313)
Audit * 30 Tercile 0.272 0.114
(0.285) (0.309)
20 Tercile -0.313 -0.100
(0.225) (0.248)
30 Tercile -0.213 -0.099
(0.230) (0.264)

Observations 1,400 1,400 1,400 1,400 817 817 817 817

R-squared 0.310 0.344 0.363 0.368 0.418 0.438 0.477 0.478
Panel B
Opposition Parties Party of the Audited Neighbour

Audit -0.393* -0.424* -0.288 -0.475 -0.346 -0.376 -0.303 -0.312

(0.221) (0.229) (0.219) (0.329) (0.267) (0.276) (0.289) (0.457)
Audit * Corrupt. Index 0.087 0.009
(0.089) (0.106)
Corrupt. Index -0.081 0.045
(0.062) (0.087)
Audit * 20 Tercile 0.496 0.106
(0.497) (0.551)
Audit * 30 Tercile 0.043 -0.208
(0.457) (0.521)
20 Tercile -0.759* -0.265
(0.387) (0.550)
30 Tercile -0.190 0.190
(0.400) (0.544)

Observations 766 766 766 766 412 412 412 412

R-squared 0.366 0.402 0.455 0.459 0.370 0.387 0.439 0.441
State FE X X X X X X X X
Party FE X X X X X X X X
Term X X X X X X X X
Party x Term FE X X X X X X
Controls X X X X
Notes: Panels A and B show results for the audited and neighboring municipalities, respectively. Column (1) - (4) depict results
for the opposition parties, columns (5) - (8) show results regarding incumbent parties. Controls include: income per capita,
fraction of the urban population, fraction of the male population, and the literacy rate, and surface area, latitude and longitude,
distance to state capital, Human Development Index, and electricity in the household. Standard errors are clustered at the
municipality level in the case of the audited municipality and at the mesoregion area in the case of geographical spillovers. ***
Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant at the 10 percent level.
 Table 13: Vote-buying

Audited Municipality Informed Neighbour

(1) (2) (3) (4) (5) (6)

Audit 0.014 -0.060 -0.087 0.087 -0.025 -0.025

(0.076) (0.155) (0.155) (0.060) (0.126) (0.127)

Observations 125 125 125 405 405 405

R-squared 0.000 0.122 0.173 0.005 0.062 0.100
State FE X X X X
Population X X X X
Controls X X
Notes: The dependent variable is an indicator taking value 1 if the individual replies yes to the question: ”Have you known of
someone in the last elections who was pressured or received something to change his vote in a certain way?”. Survey weights
are included in all regressions. In the last column, individual controls such as age, religion, sex, use of email, home ownership,
TV ownership, expenditure in health, family income, and health status are included. Population is the municipal population.
Robust standard errors in parenthesis. *** Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant
at the 10 percent level.

64
10 Appendix
10.1 More on the audit program
Although the rules for exclusion of municipalities in the lottery change over time, 85 percent
of the municipalities have been eligible for selection during my sample period. In Table A1,
for each lottery round, the date the lottery took place, the date of publication of lottery
rules, the date of publication of audit outcomes and replacement rules are shown. In Figure
14, the population cutoffs for each lottery are plotted. The population estimates are based
on the IBGE yearly projections.69
Another difference between the lottery rounds is the stratification. In earlier rounds,
up to lottery 8, the stratification was done at the state level. In later rounds, states with
a smaller number of municipalities are bundled together (see Table A2). For example in
lottery 9, states Acre, Amapá, Amazonas, Rondônia, and Roraima were bundled together
and one municipality was to be selected. In lottery 10 Acre, Amapá, Rondônia, and Roraima
were bundled together and one municipality was to be selected from the state of Amazonas
separately.

69
Except from lottery 32, which had two stages, the rest of the lotteries only had one stage. Each stage
implied two overlapping population brackets. Within each population bracket, 30 municipalities were audited.

65
 Figure 14: Population Cutoffs for Each Lottery

Notes: Population brackets for each lottery expressed in 1000 of inhabitants. In grey, minimum population cutoff and in orange
maximum population cutoff. The population brackets, except the first rounds, have been pretty constant over time.

66
 Table A1: Population Cutoffs

Lottery Draw Date Date Official Date Pub. Replacement

3 18/06/2003 11/06/2003 03/09/2003 past 3 lotteries

4 30/07/2003 24/07/2003 17/10/2003 past 3 lotteries
5 03/09/2003 28/08/2003 03/09/2003 past 3 lotteries
6 15/10/2003 09/10/2003 12/02/2004 past 3 lotteries
7 12/11/2003 10/11/2003 27/04/2004 past 6 lotteries
8 30/03/2004 24/03/2004 11/08/2004 past 6 lotteries
9 29/04/2004 26/04/2004 27/09/2004 past 6 lotteries
10 26/05/2004 17/05/2004 27/09/2004 past 6 lotteries
11 30/06/2004 08/06/2004 12/01/2005 past 6 lotteries
12 11/08/2004 26/07/2004 22/02/2005 past 6 lotteries
13 27/10/2004 23/09/2004 15/04/2005 past 6 lotteries
14 17/11/2004 16/11/2004 24/06/2005 past 12 lotteries
15 14/04/2005 12/04/2005 30/09/2005 past 12 lotteries
16 09/06/2005 03/06/2005 22/11/2005 past 12 lotteries
17 16/08/2005 08/08/2005 22/02/2006 past 12 lotteries
18 27/09/2005 14/09/2005 30/05/2006 past 12 lotteries
19 07/11/2005 27/10/2005 22/08/2006 past 12 lotteries
20 23/03/2006 15/03/2006 29/12/2006 past 12 lotteries
21 02/06/2006 24/05/2006 06/03/2007 past 12 lotteries
22 19/07/2006 12/07/2006 09/07/2007 past 12 lotteries
23 09/05/2007 24/04/2007 25/01/2008 past 12 lotteries
24 24/07/2007 18/07/2007 01/04/2008 past 12 lotteries
25 09/10/2007 1/10/2007 06/06/2008 past 12 lotteries
26 30/04/2008 17/04/2008 12/12/2008 past 6 lotteries
27 29/10/2008 22/10/2008 27/04/2009 past 12 lotteries
28 12/05/2009 07/05/2009 11/11/2009 past 3 lotteries
29 17/08/2009 11/08/2009 19/03/2010 past 3 lotteries
30 05/10/2009 28/09/2009 09/07/2010 past 3 lotteries
31 01/03/2010 23/02/2010 06/10/2010 past 3 lotteries
32 10/05/2010 30/04/2010 21/01/2011 past 3 lotteries
33 26/07/2010 20/07/2010 20/01/2011 past 3 lotteries
34 15/08/2011 05/08/2011 07/03/2012 past 3 lotteries
35 03/10/2011 27/09/2011 24/04/2012 past 3 lotteries
36 23/07/2012 11/07/2012 17/01/2013 past 3 lotteries
37 08/10/2012 04/10/2012 04/04/2013 past 3 lotteries
38 04/03/2013 18/02/2013 18/09/2013 past 3 lotteries
Notes: Lottery stands for the round of the lottery. Draw Date is the date in which the lottery took place for that specific round.
Date Official is the date the rules of the lottery were published. Date Pub. is the date in which the audit reports were published.
Replacement indicates which lotteries were eligible for selection among those in the published population brackets.

67
 Table A2: Division of Brazilian States

State Municipalities Microregion Mesoregion

Acre 22 2 5
Alagoas 102 3 13
Amapá 16 2 4
Amazonas 62 4 13
Bahia 417 7 32
Ceará 184 7 33
Espı́rito Santo 78 4 13
Goiás 246 5 18
Maranhão 217 5 21
Mato Grosso 141 5 22
Mato Grosso do Sul 78 4 11
Minas Gerais 853 12 66
Pará 143 6 22
Paraı́ba 223 4 23
Paraná 399 10 39
Pernambuco 185 5 19
Piauı́ 224 4 15
Rio De Janeiro 92 6 18
Rio Grande do Norte 167 4 19
Rio Grande do Sul 496 7 35
Rondônia 52 2 8
Roraima 15 2 4
Santa Catarina 293 6 20
São Paulo 645 15 63
Sergipe 75 3 13
Tocantins 139 2 8
Notes: Subdivision of states into municipalities, microregions and, mesoregions. Distrito Federal (Federal Districtis) not included.
According to the Brazilian Constitution the Federal District is not separable into municipalities.

68
10.2 Selection Probability
To further address the large t-statistics for the PT dummy in the presidential sample, I
use all documentation on the published lotteries in the Official Gazette and available on
the CGU website. To investigate if a mayor from PT is more or less likely to be selected
for an audit, I run a linear probability model of the audit probability on a dummy taking
value one if the municipality is headed by a PT mayor. In Panel A of Table A3 the full
sample of eligible municipalities for each lottery is included. In Panel B, I restrict the sample
to the municipalities that have a mayor either from PT or PSDB. The subsample of the
municipalities actually audited constitutes the presidential sample.

[Table A3 Here]

Findings suggest that having a mayor from PT does not differentially affect the likelihood
of selection. This is also the case when lottery fixed effects are added. In column (2), results
from within lottery variation are shown. In column (3), I add the probability of selection of
each municipality in the lottery. The latter is computed as the number of municipalities to
be selected over the total municipalities eligible in the state or group of states. In column
(4), both lottery fixed effects and the probability of selection are included in the regression.
In column (5), I also include group fixed effects. In column (6) I include state fixed effects
instead of group fixed effects. The coefficient on the PT changes only marginally indicating
that PT does not receive any preferential treatment.

69
 Table A3: Likelihood of an Audit

P anelA (1) (2) (3) (4) (5) (6)

PT -0.0005 -0.0005 -0.0008 -0.0007 -0.0005 -0.0005

(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Selection Prob 2.8693*** 2.9183*** 5.9840*** 1.5288***
(0.519) (0.524) (1.938) (0.589)

Observations 156,551 156,551 156,551 156,551 156,551 156,551

R-squared 0.000 0.001 0.001 0.001 0.003 0.002
Lottery FE X X X X
Probability X X X X
State FE X
Group FE X
P anelB

PT 0.0006 0.0010 0.0002 0.0006 0.0001 0.0003

(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Selection Prob 3.0889*** 3.1315*** 1.2639 1.0018
(1.072) (1.076) (5.218) (1.356)

Observations 35,618 35,618 35,618 35,618 35,618 35,618

R-squared 0.000 0.001 0.001 0.002 0.004 0.004
Lottery FE X X X X
Probability X X X X
State FE X
Group FE X
Notes: Likelihood of audit on party dummy. Selection Prob is the selection probability of the municipality in each lottery. ***
Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant at the 10 percent level.

70
10.3 Scraping and Coding Lotteries 3-7
Lotteries 3-7 are problematic to scrape and code because the information in these PDF is
not stored in a homogeneous format. This phenomenon does not only affect reports across
time but also those within the same state (and within lottery) where the same auditors are
investigating and compiling the report. The auditing procedure, however, seems to be similar
over time and across states.
In order to maximise the sample size, I include the total amount of irregularities found
by the auditors and create a count of severe irregularities. Provided that these documents
do not have a summary indicating the irregularities found by the auditors, I try to locate the
paragraph where the irregularity is described.70
In most of the cases, the paragraph can be located by searching words like “Constatação
da Fiscalização” (Audit Findings). In some cases, a phrase summarising the paragraph is
included before the irregularity description. Figure 15 provides an example. Usually the
summary phrase is followed by a word such as “Fato” (fact) and “Evidencia” (Evidence –
brought by the auditors).

[Figure 15 Here]

I am able to systematically count the number of times a line starts with “Constatação da
Fiscalização” and how many times a line starts with “Fato”. Having two different ways of
locating the paragraph where the irregularity is included allows me to reduce the measurement
error in the variable. Both approaches point to a similar result, allowing me to get the total
amount of irregularities encountered by the auditors in the municipality.71
The second step is to get the irregularity count from these reports. For this purpose, I
start by extracting the first three sentences in the paragraph where the auditors describe the
irregularity. Next, I again follow two different approaches. First, I classify the text with the
same method in Section 3.1.1. Second, I use a Naive Bayes classification technique.
The classifier is based on Bayes’ Theorem with the assumption of independence among
70
This is done again, by regular expressions.
71
This variable may be prone to measurement error especially belonging to lotteries 3 and 4. The reason
being that in earlier lotteries the auditors were not only auditing every fund transferred, they were also
writing about all findings, both positive and negative. From lottery 5, auditors only point out irregularities
found. Namely, they only point out only irregularities, or negative facts, which, depending on the situation,
may or not be improved by the municipal management.

71
predictors.72 Regardless of the assumption, the classifier has a very accurate performance. I
am able to use the Naive Bayes because the training sample, automatised classified phrases
from the reports 8-38, is large enough and the out of sample prediction is similar in magnitude.
When working with the Naive Bayes classifier, the text pre-processing is different.73 First,
stopwords need to be removed from the text. Stopwords are common words in a language.
In English, the equivalent would be: “of”, “to” , “on”. I do so mainly using the packages
in Natural Language Toolkit (NLTK) for Portuguese. The library works well in English but
it has some limitation for other languages such as Portughuese. For this reason, I include
another list of stopwords, which I create myself, from the most commonly occurring words
among the PDF reposts. The classifier, as expected, provides results similar to the first
methodology. The performance of the Naive Bayes is in Table A4. The test on the classifier
indicates that it is a good one; both in-sample and out-of-sample forecasts have a good
precision, recall, and f1-score.

[Table A4 Here]

72
A Naive Bayes classifier assumes that the presence of a particular feature in a class is unrelated to the
presence of any other feature.
73
Some steps are similar such as accent removal and conversion of upper case letters to lower case letters.

72
 Table A4: Naive Bayes Classifier

In sample forecast

precision recall f1-score support

0.0 0.94 0.88 0.91 52936
1.0 0.82 0.91 0.86 32129
avg / total 0.90 0.89 0.89 85065

Out of sample sample forecast

0.0 0.87 0.94 0.90 17269

1.0 0.90 0.81 0.85 12504
avg / total 0.88 0.88 0.88 29773
Notes: In-sample forecast refers to the performance of the classifier in the training sample. Out-of-sample forecast refers to the
performance of the classifier in the test sample. In sample forecast refers to the performance of the classifier in the training
sample. Out of sample forecast refers to the performance of the classifier in the test sample. Precision (or positive predicted
value) and recall (or sensitivity) are, respectively. the fraction of retrieved instances that are relevant and the fraction of relevant
instances that are retrieved. Both precision and recall are based on an understanding and measure of relevance. F1-score (also
F-score or F-measure) is a measure of a test’s accuracy. It considers both the precision p and the recall r of the test to compute
the score. The f1-score can be interpreted as a weighted average of the precision and recall, where an f1-score reaches its best
value at 1 and its worst value at 0. Support stands for the number of samples that support the classification. In this particular
case, support counts the number of phrases that support the classification.

73
 Figure 15: (Slight) Differences in audit reports

Panel A: Example of an Audit Report with a summary phrase before the full
description of the irregularity.

Panel B: Example of an Audit Report without a summary phrase before the full
description of the irregularity.

74
10.4 Correlation with other measures in the literature
I perform a second test on the corruption measure using two independently coded measures of
corruption. The first one is the coding by Ferraz and Finan (2011), henceforth FF (available
on the AER webpage) and the data by Garfias and Timmons (2015), GT henceforth.74 The
second measure was coded similar in spirit to the measure by FF. I match these data and
use the subsample where the coders mainly agree. This is the subsample where the coders
attribute to the same municipality the same number of corrupt activities or they disagree by
1 corrupt activity. The correlation among both measures in this subsample is of 0.99. To
test the performance of my classifier, I regress it on the corruption measure by FF and GT
conditioning for state fixed effects. The rationale for including state fixed effects is due to
the fact that my measure depends heavily on the language used by the auditors which varies
across states (in different states there are different teams of auditors). I report the partial
correlation coefficient and the R-squared of the regression. I also present a binned scatterplot
of corruption index and the measure of corrupt activities in FF and the overlaid regression
line.75

74
I thank Claudio Ferraz, Frederico Finan, Francisco Garfias and, Jeffrey Timmons for sharing their data
with me.
75
These graphs were produced using binscatter, a user-written Stata command written by Michael Stepner,
with input from Jessica Laird and Laszlo Sandor.

75
 Table A5: Correlation with FF and GT

(1) (2) (3) (4)

Corrupt. Index Corrupt. Index Sd Major Irreg Sd Major Irreg

Corrupt FF 0.376*** 0.274***

(0.041) (0.025)
Corrupt GT 0.352*** 0.260***
(0.045) (0.033)

Observations 99 99 99 99
R-squared 0.751 0.725 0.730 0.704
Notes: Partial correlation of the corruption index, in columns (1) and (2), and the standardised measure of irregularity count,
in columns (3) and (4), with the measures of corrupt activities as in Ferraz and Finan (2011) and Garfias and Timmons (2015).

Figure 16: Correlation with FF and GT

Notes: Binned Scatterplot of the Corruption Index and Corruption Activities (as in FF
and GT).

76
10.5 Exogeneity Check (Geographical Spillovers)
Given the randomization in audits, both groups should be homogeneous. There should be
no major differences between the treatment and the control group. To assess this, I test the
validity of the exogeneity assumption by using predetermined municipal characteristics. I
run the following specification:

Auditpit−s = λ0 + λ1 xj + ηc + pop2000 + εpjt (7)

In this case, i is the audited municipality and j corresponds to the informed neighbor.
Auditpit−j is an indicator taking value one in case there was an audit in a nearby municipality
i at time s previous to elections and value zero in case the audit in town i occurred after
elections. xj is a characteristic of municipality j, measured in the year 2000. ηc state fixed
effects and pop2000 are included in each pairwise regression given that they are determinants
of the eligibility of the municipality. λ1 is expected to be zero in each regression. This is
indeed the case for most of the municipal characteristics; however, there are some differences
in the presidential sample. These differences vanish if instead of controlling for state fixed
effects and population, I control for the selection probability of the municipality.
If I follow the latter strategy, it requires for the sample to be restricted to the municipali-
ties that are eligible for selection in the same lottery as their (audited) neighbor. This reduces
the samples by hundreds of observations (municipal and presidential), thus the samples will
not coincide with those used for the main analysis.76 This test provides further credibility to
the identification strategy. I present results both with and without the vector of municipal
characteristics as controls. Similarly to the specifications used for the geographical spillovers,
I cluster the standard errors at the mesoregion area. These results are shown in Table A6.

[Table A6 Here]

76
Other possibilities would be to control for the probability of selection in the same year or the same
electoral term as the audited neighbor. In these cases, an average of the selection probability must be
attributed given that the municipality is eligible for selection various times within the same year. These
results are available upon request.

77
 Table A6: Exogeneity Checks neighbors

Municipal Presidential
Corrupt. Index -0.0272 -0.0168
(0.0321) (0.0247)
Literacy Rate -0.0030 0.0092
(0.0039) (0.0055)
Frac. Urban 0.2288 0.1540
(0.2599) (0.1767)
GDP per Capita 2.1871 -2.0610
(2.7150) (3.0867)
Income per Capita 1.2777 1.6207
(1.4676) (0.9219)*
Electricity TV -0.0016 0.0040
(0.0019) (0.0027)
HDI -0.9477 1.0830
(1.1132) (0.7097)
Income per capita -0.0005 0.0005
(0.0007) (0.0005)
Coverage -0.0491 0.1525
(0.0743) (0.0671)**
Latitude -0.0163 -0.0145
(0.0329) (0.0187)
Longitude -0.0604 -0.0268
(0.0596) (0.0171)
Dist. State capital 0.0003 0.0003
(0.0004) (0.0002)
Dist. Federal capital 0.0002 0.0003
(0.0003) (0.0002)
Surface 0.0000 0.0003
(0.0000) (0.0002)
Electorate -0.0000 -0.0000
(0.0000) (0.0000)
Residual Vote -0.9491 -1.9190
(1.3666) (0.9970)*
Munic. Council 0.0030 0.0016
(0.0092) (0.0050)
PT 0.0419 -0.1792
(0.0567) (0.0889)**
PSDB 0.0222
(0.1106)
PMDB -0.0775
(0.0761)
PP -0.0878
(0.1177)
PFL 0.0701
(0.0812)

Notes: Coefficients and standard errors reported in this table result from a regression in which the dependent variable is an
indicator taking value one in case the neighbor was audited before elections and zero after elections. The regressor is defined by
each row. In the regression are also included population and state fixed effects as controls. Standard errors are clustered at the
mesoregion area. *** Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant at the 10 percent level.

78
10.6 Exit of parties
To investigate for potential differential exit of parties, I use a linear probability model where
I regress the exit dummy on the audit dummy and other controls. The exit dummy takes
value one in case the party does not show up in upcoming elections (and zero otherwise).
The results are shown in Table A7.
Columns (1) - (4) present results for the audited municipality whereas those from (5) -
(8) are related to exit of the audited party in the case of the geographical spillovers. The
interaction of the treatment with the corruption index is shown in columns (3), (4) for audited
municipalities and (7), (8) for neighboring municipalities. Across columns, the coefficient on
the audit dummy is not statistically different from zero and this also applies to the interaction
term. These results indicate that there is no differential exit of parties.

Table A7: Party Exits

Audited Municipality Geographical Spillovers

(1) (2) (3) (4) (5) (6) (7) (8)

Audit -0.038 -0.030 -0.037 -0.031 -0.016 -0.014 -0.016 -0.016

(0.024) (0.024) (0.024) (0.024) (0.045) (0.046) (0.046) (0.047)
Audit * Corrupt. Index -0.006 -0.003 -0.010 -0.010
(0.012) (0.013) (0.018) (0.019)
Corrupt. Index 0.008 0.002 0.010 0.008
(0.010) (0.011) (0.017) (0.017)

Observations 1,834 1,834 1,828 1,828 898 898 898 898

R-squared 0.140 0.175 0.141 0.175 0.142 0.143 0.165 0.165
State FE X X X X X X X X
Party FE X X X X X X X X
Controls X X X X
Notes: All regressions include state fixed effects, time fixed effects and also population as of the year 2000. Columns (1) - (4)
show results in the audited municipality. Columns (5) - (8) show results from municipal elections in neighboring municipalities.
Controls include: income per capita, fraction of the urban population, fraction of the male population, and the literacy rate,
and surface area, latitude and longitude, distance to state capital, Human Development Index, and electricity in the household.
Standard errors are clustered at the municipality level in case of the audited municipality and at the mesoregion area in the
case of geographical spillovers. *** Significant at the 1 percent level, ** Significant at the 5 percent level, * Significant at the
10 percent level.

79