University of Southern California Dissertations and Theses

Essays on political economy and corruption

(USC Thesis Other)

Essays on political economy and corruption

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS ON POLITICAL ECONOMY AND CORRUPTION
by
Karim Fajury
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2023
Copyright 2023 Karim Fajury
Acknowledgments
I am grateful to my main advisors, Je Weaver and Vittorio Bassi, who constantly challenged
and supported my ideas. Our regular meetings helped me become a better economist. Both of
them were true mentors. I learned from them how to navigate the intricacies and frustrations of
academic research, but they also taught me how to be kind and organized. It was a privilege to be
advised by them and to work with them as a teaching and research assistant. I also want to thank
Je Nugent, who was always excited to talk about my research ideas and gave me condence to
materialize them. The Thursdays applied-micro reading group meetings were also fundamental for
my research and were a highlight of my weeks.
I was fortunate to make long-lasting friendships and to solidify previous ones thanks to the PhD.
Juan Espinosa, Daniel Angel, Thomas Ash, Monira Al Rakhis, Amy Mahler, Sam Boysel, Rajat
Kochhar, Ruozi Song, Nicolas Roig, Xiongfei Li, Clement Boulle, Taraq Kahn, Islamul Haque,
Mateo Arbelaez, Carlos Molina, among many others, made my time at USC joyful and helped
me grow. I also want to thank professors Matt Kahn, Gerardo Munck, Monica Morlacco, Pablo
Kurlat, Dan Bennett, Simone Schaner, David Zeke, Joel David, Michael Magill, Augustin Bergeron,
John Strauss and Sandra Rozo, for teaching me in classrooms, in seminars, or through thoughtful
conversations.
I am especially grateful to my family. My parents, Isabel Arana and Emilio Fajury, are the
smartest, most creative and loving individuals I have ever{and probably will ever{meet. Their un-
conditional support and company are gifts I treasure deeply. The appreciation I have for everything
they have done for me is indescribable. Thank you. My brother, Raschid Fajury, was the rst {and
best{ professor I've had. For a long time, he was the only person who explained things in a way I
could understand. I am beyond grateful to him for his patience. His example made my life easier.
It has been thrilling to see him, and Alejandra Arbelaez, form their own family. Talia Fajury, and
ii
her future brother or sister, are fortunate to have you as parents.
My grandparents, Tamer Arana and Hendy Hissami, and my aunt, Nylse Fajury, died while I
was doing this PhD. This work is dedicated to their memory.
iii
Table of Contents
Acknowledgments ............................................................................. ii
List of Tables.................................................................................. vii
List of Figures................................................................................. ix
Abstract....................................................................................... xi
1 Chapter 1 Clientelism and Poverty: Voter Buying in Colombia ............................ 1
1 Introduction.................................................................................. 1
2 Data and Context ........................................................................... 5
2.1 Elections and Voter Buying ........................................................ 5
2.1.1 Data ....................................................................... 7
2.2 The Sisben and Familias en Acci on ................................................ 8
2.2.1 Data ....................................................................... 9
2.3 Voter Turnout ....................................................................... 10
2.3.1 Data ....................................................................... 11
2.4 Other Data: Municipal Characteristics ............................................ 12
3 Empirical Strategy .......................................................................... 12
3.1 Specication Tests................................................................... 15
4 Results........................................................................................ 16
4.1 Reduced Form Estimation .......................................................... 16
4.2 \First Stage" RD Estimation....................................................... 18
4.3 Fuzzy RD Estimation ............................................................... 19
4.4 Why voters bought are not migrants .............................................. 20
iv
4.5 Mechanisms .......................................................................... 21
4.6 Heterogeneity ........................................................................ 23
5 Conclusion ................................................................................... 26
Figures ............................................................................................. 28
Tables .............................................................................................. 35
Appendix........................................................................................... 42
2 Chapter 2 Campaign Spending Limits and Irregular Political Behavior .................... 50
1 Introduction.................................................................................. 50
2 Context ....................................................................................... 54
2.1 Elections in Colombia and Campaign Spending .................................. 54
2.2 Politics and Crime in Colombia .................................................... 55
3 Empirical Strategy .......................................................................... 56
4 Data .......................................................................................... 58
4.1 Municipal Outcomes: Crime, Violence, Finances and Service Provision ........ 58
4.2 Electoral and Political Variables ................................................... 59
4.3 Vote Buying Data ................................................................... 60
5 Threats to Identication .................................................................... 62
6 Results........................................................................................ 63
6.1 Crime and Violence ................................................................. 63
6.2 Electoral Outcomes and Vote Buying.............................................. 65
6.3 Service Provision and Municipality's Finances .................................... 66
6.4 Robustness ........................................................................... 67
7 Conclusion ................................................................................... 67
Figures ............................................................................................. 68
Tables .............................................................................................. 82
Appendix........................................................................................... 87
3 Chapter 3 A Digitization of Individual Voter Turnout in Colombia ........................ 90
1 Introduction.................................................................................. 90
2 Context ....................................................................................... 92
v
3 Data .......................................................................................... 93
4 Methodology ................................................................................. 94
5 An Application: Merging with Other Data................................................ 97
5.1 The Sisben ........................................................................... 98
5.2 Summary Statistics.................................................................. 98
6 Conclusion ................................................................................... 99
Figures ............................................................................................. 101
Tables .............................................................................................. 101
Appendix........................................................................................... 105
References ..................................................................................... 106
vi
List of Tables
1.1 Balance Table ................................................................................ 36
1.2 Balance Table (continued) .................................................................. 37
1.3 Reduced Form Estimation 2015 ............................................................ 38
1.4 First Stage Results: Multiple Bandwidths ................................................ 38
1.5 Fuzzy RD Results: Multiple Bandwidths ................................................. 39
1.6 Reduced Form RD: CCT on Turnout ..................................................... 39
1.7 Heterogeneity: Municipality's Population ................................................ 40
1.8 Heterogeneity: Municipality's Election Closeness 2011 .................................. 40
1.9 Heterogeneity: Municipality's Number Voting Stations per Sq. Km ................... 41
1.10 Heterogeneity: Neighboring Municipality's Number Voting Stations per Sq. Km ..... 41
1.11 Summary Statistics ......................................................................... 42
1.12 Transfer amounts for the FA program, 2015 .............................................. 43
1.13 Reduced Form RD Results: Multiple Bandwidths ....................................... 43
1.14 Reduced Form RD Results: Multiple Bandwidths and Controls ........................ 43
1.15 First Stage: Beneciaries of Familias en Accion .......................................... 44
1.16 First Stage Results: Multiple Bandwidths and Controls ................................ 44
1.17 Fuzzy RD: CCT Beneciaries and Vote Buying........................................... 45
1.18 Fuzzy RD Results: Multiple Bandwidths and Controls .................................. 45
1.19 Reduced Form RD Results: Proof of valid residence as Outcome ...................... 46
1.20 Reduced Form RD Results: Multiple Bandwidths after Proof of Residence ........... 46
1.21 Reduced Form RD Results: Multiple Bandwidths and Controls after Proof of Resi-
dence ......................................................................................... 47
1.22 Fuzzy RD Results: Multiple Bandwidths after Proof of Residence ..................... 47
vii
1.23 Fuzzy RD Results: Multiple Bandwidths and Controls after Proof of Residence ...... 48
1.24 Fuzzy RD Results CCT and Turnout: Multiple Bandwidths ........................... 48
2.1 Campaign Spending Limits, Colombian Pesos (2003-2015) .............................. 82
2.2 Balance Checks Across Thresholds ........................................................ 82
2.3 Political Outcomes .......................................................................... 83
2.4 Campaign Spending Limits and Voters-Bought imported per municipality.Pooled
Regression ................................................................................... 83
2.5 Crime Outcomes ............................................................................ 84
2.6 Violence and Con
ict Outcomes ........................................................... 84
2.7 Violence and Con
ict Outcomes in Non-Election Years .................................. 85
2.8 Municipality Finances ...................................................................... 85
2.9 Service Coverage ............................................................................ 86
2.10 Education ................................................................................... 86
2.11 Crime Outcomes Bandwidth 3000 ......................................................... 87
2.12 Crime Outcomes Bandwidth 5000 ......................................................... 87
2.13 Crime Outcomes Bandwidth 7000 ......................................................... 88
2.14 Violence and Con
ict Outcomes Bandwidth 3000 ....................................... 88
2.15 Violence and Con
ict Outcomes Bandwidth 5000 ....................................... 89
2.16 Violence and Con
ict Outcomes Bandwidth 7000 ....................................... 89
3.1 Turnout 2015. Colombia ................................................................... 102
3.2 Turnout 2015. Colombia Continued. ...................................................... 103
viii
List of Figures
1.1 GDP per capita and Corruption ........................................................... 28
1.2 Colombia's Electoral Transhumance as a % of municipality's population, 2015 ....... 29
1.3 Sisben normalized score distribution ...................................................... 30
1.4 McCrary Test ............................................................................... 31
1.5 Cattaneo, Jansson and Ma (2019) Density Test .......................................... 32
1.6 FA Eligibility and Voter Buying. Reduced Form Results ............................... 33
1.7 First Stage Results: Eligibility and Beneciaries of FA in 2015 ......................... 34
1.8 Example of Voter Turnout Records ....................................................... 49
2.1 Number of municipalities per campaign spending limit category (2003-2015) .......... 68
2.2 McCrary Test Around 25.000 Cuto ....................................................... 69
2.3 Cattaneo, Jansson and Ma Test Around 25.000 Cuto................................... 70
2.4 McCrary Test Around 50.000 Cuto ....................................................... 71
2.5 Cattaneo, Jansson and Ma Test Around 50.000 Cuto .................................. 72
2.6 Predetermined Outcomes ................................................................... 73
2.7 Political Outcomes .......................................................................... 74
2.8 Vote Buying ................................................................................. 75
2.9 Crime Outcomes ............................................................................. 76
2.10 Violence Outcomes .......................................................................... 77
2.11 Violence Outcomes in Non-Electoral Years ................................................ 78
2.12 Municipal Finances .......................................................................... 79
2.13 Service Coverage ............................................................................. 80
2.14 Education Outcomes ........................................................................ 81
ix
3.1 Example of Voter Turnout Records ....................................................... 101
3.2 Example of First Page of E-11 Form of Voter Turnout Records ........................ 104
3.3 Example of Jurors' Vote Report on E-11 Form .......................................... 104
3.4 Turnout Rates by Department: Code Generated and Ocial (2015) ................... 105
x
Abstract
This thesis gathers three papers which study how incentives and socioeconomic circumstances shape
electoral participation and corruption.
The rst paper is motivated by the fact that high income countries have lower levels of cor-
ruption, but it is still unclear whether higher incomes causally reduce the propensity to engage in
corruption. This paper studies the eect of income on corruption in elections. To do so, I digitize
novel information on vote buying from ocial records and exploit a nationwide cash transfer pro-
gram in Colombia. Using a discontinuity that determines eligibility for the transfer in a regression
discontinuity design, this paper nds that this increase in income reduced the likelihood of someone
having their vote bought by 18 to 27%. I show that this is explained by an income eect rather
than other mechanisms such as becoming more civically engaged. Back of the envelope calculations
indicate that around 30,000 votes were not sold as a result of the program. I also document how
other characteristics, such as closeness of elections, aect the political economy of vote buying, and,
therefore, the eectiveness of the cash transfers in reducing this form of clientelism.
The second paper studies the unintended consequences of political campaign spending limit
regulations in Colombia. To do so, I use discontinuities in electoral campaign spending limits
to estimate their causal eects on organized violence and vote buying. The campaign spending
limits are determined according to arbitrary cutos of potential voters for each municipality. Using
a pooled Regression Discontinuity Design, I nd that tighter campaign spending limits increase
organized violence and vote buying. The results suggest that tight campaign spending limits can
lead to undesirable outcomes, such as the use of fraudulent or illegal practices as a substitute to
legal nancial resources to attract voters for political gain.
The third paper describes the construction of a database {at the individual level{ of voter
turnout for Colombia. I use over a million pages of scanned ocial voting records and digitize the
xi
information using computer vision and optical character recognition techniques. I also provide an
example on how the data can be used and motivate future research.
xii
Chapter 1
Clientelism and Poverty: Voter Buying in Colombia
1 Introduction
Political corruption and clientelism{the exchange of goods or money for political support{are
widespread in developing countries and thought to be harmful for economic performance and gov-
ernment accountability (World Bank, 2017; Ferraz and Finan, 2011; J-PAL Governance Initiative,
2019). Richer countries have lower levels of corruption and vote buying (Figure 1.1). However, it
is not clear if these lower levels of corruption are due to increases in wealth or other factors that
may independently aect corruption and income (e.g. institutions, education, etc.).
In settings with low accountability and little law enforcement, individuals can benet from
exchanging their vote for cash or goods, expecting minor consequences for their individual behavior.
A theoretical literature claims that lower income individuals, who have a higher marginal utility
for private consumption, are more likely to engage in clientelistic practices and accept bribes for
their votes (Dixit and Londregan, 1996; Shleifer and Vishny, 1993). Nonetheless, there is still little
causal evidence of higher incomes reducing clientelism.
This paper addresses this gap in the literature by studying how exogenous increases in income
aect the likelihood of citizens to engage in clientelistic practices, such as vote buying. To do so,
I use a nationwide conditional cash transfer program in Colombia and newly digitized individual-
level data on vote buying. Exploiting discontinuities that determine the eligibility for the transfers,
in a regression discontinuity design I nd that the increase in income reduced the likelihood of
individuals having their vote bought.
As a measure of vote buying, I digitized data on all potential \voters bought" in Colombia for
a particular election. Voter buying is one of the most common forms of electoral fraud. It consists
1
in the mobilization of voters from their area of residence to another location to illegally vote for
a particular candidate. Under this practice, which is locally known as \electoral transhumance",
voters are paid with money or goods by political candidates or parties in exchange for their vote.
This is an illegal practice, as citizens who do not live in a certain location, but vote there, may
in
uence the electoral outcomes of an election in which they should not have a legitimate interest.
This type of vote buying is not exclusively a Colombian phenomenon. Several countries with
democratically elected governments punish the illegal registration of voters in dierent areas from
where they actually live. This practice has been found in a range of countries such as Brazil (Hidalgo
and Nichter, 2016), Per u (El Comercio, 2014), Botswana (AllAfrica, 2009), Mexico (MVSNoticias,
2018), Bolivia (IBCE, 2019), among many others.
It is rare to nd individual-level data that indicates clientelistic engagement at a national scale.
The literature on vote buying usually measures this variable using self-reported information (for
example, Vicente (2014), Blattman et al. (2019), Stokes et al. (2013a), Fergusson et al. (2020)).
Instead, this paper uses individual-level, non self-reported data, from Colombia to identify vote
buying. The country's electoral authority creates lists with the IDs of those voters for whom the
place of residence diers from the municipality where they registered to vote. To do so, they use
multiple administrative databases to examine if citizens who change the municipality where they
are registered to vote actually live, work, or study, in their new location. Using thousands of pages
of legislative decrees that include the lists of voters bought, I create a new database of all presumed
voters bought for Colombia's regional elections of 2015.
Using this voter buying data as a granular measure of clientelism, this paper tests for a causal
link between income and vote buying. I exploit a discontinuity determining eligibility for Colombia's
conditional cash transfer program, Familias en Acci on. Households are eligible for the cash transfer
program if they have a multidimensional poverty score below a pre-determined{and unknown to
them{cuto. Using a regression discontinuity design, I test how higher levels of income due to
the transfer aect vote-buying. There are no signs of potential manipulation of the poverty score
and, moreover, other observable characteristics are continuous across the cutos of eligibility for
the cash transfers, validating the regression discontinuity design.
I nd that beneciaries of the cash transfer program are between 18 to 27% less likely to engage
in voter buying than non-beneciaries. The estimates are robust across dierent specications and
2
denitions of optimal bandwidths around the eligibility cuto.
A reduction in voter buying caused by the conditional transfer program can be generated by
multiple channels. One potential explanation indicates that the subsidy decreases the marginal
utility of an additional unit of income, making it less desirable for a citizen to sell their vote.
However, the program could also make beneciaries more pro-social, making them less likely to sell
their vote. This paper shows that the beneciaries of the conditional cash transfers are not more
engaged with democracy for the elections of interest, in 2015. To test this, I digitize more than
a million pages of hand-written individual voter turnout data using computer vision and optical
character recognition techniques for the universe of voters in Colombia for the 2015 election. I nd
that beneciaries are not more likely to turn out to vote, suggesting no change in pro-sociality.
One potential concern about the vote-buying variable I use is that it can potentially capture an
individual's legitimate migration behavior instead of vote buying. Citizens
agged as voters-bought
can formally present documentation supporting that the location where they registered to vote is
where they actually live. This enables them to vote there in future elections. As a measure of
legitimate migration, I digitized additional data for citizens who presented valid documentation
and were allowed to vote in their new desired municipality. This variable is continuous across the
cutos of eligibility for the cash transfers, indicating that legitimate migration does not explain the
main results.
Certain municipal characteristics dictate the political economy of voter buying and can de-
termine the eectiveness of the conditional cash transfers in reducing this type of clientelistic
relationship (Hidalgo and Nichter, 2016). To explore this, I build a simple model following the
literature of voting behavior (Downs, 1957; DellaVigna et al., 2016). The model predicts that in
municipalities where people expect to be more pivotal, or where the value of deciding the elections
are higher, income increases are less eective in reducing vote buying. On the contrary, if someone
faces high costs of legally voting in their municipality of residence, the cash transfer program is
more eective in reducing voter buying. Finally, if there are higher costs of voting in a municipality
dierent from someone's own residence, the income transfers have a smaller eect in reducing voter
buying.
To test these predictions, I estimate heterogeneous eects of the regression discontinuity de-
pending on whether a citizen's municipality is more populated, electorally competitive, has more
3
voting stations or is surrounded by municipalities with a higher density of voting stations. The
results corroborate the predictions, but also highlight that the outcome of interest does measure
vote buying and not other variables, like migration.
This paper contributes to three main literatures. First, it builds on a growing body of work on
vote buying and its determinants. Politicians target individuals who are more reciprocal (Finan and
Schechter, 2012), socially connected (Duarte et al., 2019; Cruz, 2019), less informed (Cruz et al.,
2021; Vicente, 2014) or if their vote can be more easily monitored (Larreguy et al., 2016; Rueda,
2017). Other recent work has shown how the provision of interventions that reduce economic
vulnerability can reduce clientelistic relationships (Blattman et al., 2018; Bobonis et al., 2022).
The causal link between poverty and vote buying has been long theorized (Dixit and Londregan,
1996; Stokes, 2005) and discussed qualitatively (Auyero, 2001; Stokes, 2005; Nichter, 2018), but it
is dicult to show empirically. This paper contributes to the literature by demonstrating causally
that increases in income reduce vote buying. It does this by using novel, non-self reported measures
of vote-buying and studying the eects of income transfers at scale (Muralidharan and Niehaus,
2017). I also show that the transfers are less eective in reducing vote buying in municipalities
where voters are more likely to be pivotal and whose likelihood of deciding the elections is higher,
or where voters face lower costs to vote legally. These results highlight how vote-buying responds
to aspects of the institutional environment.
Second, this paper contributes to a broader research area on corruption and its determinants.
The literature identies how economic agents, such as politicians and government ocials, engage
in corruption when they have the incentives to do so (Ferraz and Finan, 2008; Avis et al., 2018;
Vannutelli, 2021; Olken, 2007; Garbiras-D az and Montenegro, 2022; Muralidharan et al., 2021).
This paper contributes to this literature by examining how the general public also engages in corrupt
behaviors. Particularly, this study uses large scale micro-data to provide evidence that individuals
respond to monetary incentives and their own socioeconomic situation to engage in corruption.
This goes in line with the experimental evidence shown in Bertrand et al. (2007), where individuals
who were rewarded with more money for obtaining a driver's license in India were more likely to
make extralegal payments to obtain it.
Third, this paper contributes to the literature of the eects of cash transfers on political be-
havior. Work by Manacorda et al. (2011); Conover et al. (2020); De La O (2013) suggests that
4
CCT programs in Uruguay, Colombia and Mexico increased the support for the beneciaries of
the programs towards the incumbent candidates or ruling parties. Frey (2019) shows the opposite
for Brazil. Unlike these studies, this paper studies vote buying, but also provides evidence on the
eect of a CCT on voter turnout at the individual level for a complete country, using digitized non
self-reported variables.
The structure of the paper is the following: The next section describes the context and the
data used. Section 3 documents the regression discontinuity design and its main specication tests.
Section 4 reports the main results of the paper, the mechanisms and the heterogeneity tests. Finally,
Section 5 concludes.
2 Data and Context
This paper uses the following data sets: 1) The lists, created by Colombia's electoral authority, of
voters bought for the 2015 regional elections; 2) The Sisben survey, which is essentially a census of
Colombia's poorest households and includes demographic, income and employment variables among
other information; 3) Transfers and beneciaries' information of the FA program; 4) Individual voter
turnout data for the 2015 elections, digitized from over a million pages of manually reported voter
records from the whole country; 5) Multiple municipal characteristics, such as the population's size,
the voting-age population, etc. Each data set is explained in detail in the following subsections
and, where necessary, some context is provided to understand the variables used.
2.1 Elections and Voter Buying
Every four years, since 2003, each one of Colombia's 1103 municipalities holds elections, in which
the mayor of the municipality is elected directly by popular vote. Elections for every municipality
happen during the same day, usually the last Sunday of October. Aside from the municipal mayor,
the departments' governors are also elected (a department is a higher administrative unit, composed
of multiple municipalities), alongside municipal and departmental councils. The presidential and
congress elections happen in a dierent date the year before the mayoral elections. In this paper,
the main election of interest are the 2015 regional elections. The analysis focuses on this year
because it provided the best and most complete data on vote buying and voters bought for the
5
whole country.
Voter buying, or \Electoral Transhumance", as is locally known, is one of the most pervasive
forms of electoral fraud experienced in Colombia. Under voter buying, individuals illegally register
and vote in a dierent municipality from where they live, in exchange of cash or in-kind transfers
paid by a politician or electoral candidate. This practice of corruption is illegal and punishable
with prison according to Colombia's penal code. The illegality of voter buying is explained by the
fact that citizens who don't live in a certain municipality may in
uence the electoral outcomes of
an election in which they shouldn't have a legitimate interest
1
. The \importation" of outsiders to
vote in a municipality is often organized by a local political organization, or \machinery", which is
in charge of targeting and mobilizing voters from one location to the other (Duque Daza, 2019).
News reports from Colombia indicate how the voter buying phenomenon is related to corruption
and clientelism. In most cases, people receive two transfers: one once they register their vote
in a dierent municipality from their location of residence, and a second transfer at the day of
the elections. The transportation costs, snacks, food and other expenditures are covered by the
corrupt political candidate. The qualitative evidence suggests that people who engaged in voter
buying received from 70,000 pesos (close to 25 US dollars in 2015, and equivalent to around 10% of
Colombia's monthly minimum wage at the time) to 450,000 pesos. People also claim to have received
groceries, help from the candidates to formalize their land ownership, bus tickets to celebrate a
town's carnival, cattle, among many other goods.
The magnitude of the presumed voters bought during 2015 was close to 1.5 million individuals of
the 34 million of people registered to vote. News articles and NGO reports indicate how pervasive
this corrupt practice is (El Heraldo, 2015c), El Heraldo (2015b), Radio Santa Fe (2015). For
example, some people in the department of Cauca, located in the south of Colombia's Pacic
coast, reported that buses and boats lled with people arrived from other municipalities directly
to the local registrar's oce to register their vote (La Campana, 2015). In a municipality, close
to the Caribbean coast, the number of registered voters exceeded the total population (which also
includes children, who are not allowed to vote until the age of 18) (El Heraldo, 2015a). In a small
1
Initially, the law 599 of 2000 punished only the brokers who organized the voters to register in a dierent
municipality from their place of residence. Committing this crime could lead to 4 to 9 years of prison and a monetary
ne. In 2017, the punishment was extended also to those individuals who register in a dierent municipality from
their place of residence. In this case citizens would be committing a \false testimony" crime and they can be punished
with 6 to 12 years of prison, according to Article 442 of Colombia's penal code.
6
rural town, located in the eastern part of the country, there were reports of abnormal quantities
of voter registrations; many of these individuals openly recognized that they were getting paid to
do it and couldn't answer basic questions about the characteristics of the municipality where they
were registering to vote, such as what was the name of the river that crossed the town (Montero,
2011). The MOE, a national NGO that supervises the elections and the democratic process in
Colombia, indicated that in 69 municipalities the quantity of registered voters was higher than the
overall population (Rubiano, 2015). These are just a few of the many other examples showing how
prevalent voter buying is in Colombia.
2.1.1 Data
Colombia's national electoral authorities, the Consejo Nacional Electoral (CNE) and the National
Registrar's Oce Registradur a Nacional del Estado Civil, are in charge of preventing the fraudulent
inscription of IDs on voting locations that dier from each citizen's municipality of residence. The
CNE cancels a citizen's change in their voting location if the new desired municipality doesn't
coincide with other administrative information that reports the actual residence of each individual.
If the CNE cancels a citizen's voting registration in a new voting location, the individual can still
vote in the previous municipality where they were previously registered to vote.
The administrative databases used by the CNE to prevent fraudulent changes in voting sites
are: the social security database, which includes information of all the beneciaries of the general
health care system; the national agency for overcoming extreme poverty's data (ANSPE); the
system of identication of potential beneciaries of social programs (SISBEN); the Victims' Unit,
which manages programs of reparation for the victims of Colombia's con
ict; electoral censuses
from previous elections; among other data. Furthermore, the CNE also organized visits in certain
municipalities to verify the location of residence of a subset of individuals who changed their voting
location.
The CNE, after merging this information, creates a list, for each municipality, with the citizens'
IDs for the people who registered to vote in a new location but whose residence data doesn't match
their new desired voting site. The data for voter buying used in this paper are the lists of voters
with their corresponding ID and their new municipality where they were intending to vote to, whose
change in voting location was denied by the CNE for the regional elections of 2015. I digitized this
7
information from the formal legal decrees produced by the CNE at the time of the election. This
is a unique data set because it provides objective non self-reported information on clientelism and
vote buying, at the individual level.
One potential concern about the data is that it can potentially capture an individual's legitimate
migration behavior instead of vote buying. I address this issue by digitizing information of the
citizens who, initially were
agged as voters-bought, but later presented documentation supporting
that the location where they registered to vote is where they actually lived. I discuss this with more
detail in section 4.4. The main analysis and results hold even after these \legitimate migrants" are
re-coded as not having their vote bought.
Figure 1.2 shows the per capita canceled voter registrations for presumed voter buying by
municipality. In the regional elections of 2015, the electoral authorities of the country canceled
close to 1.5 million voting registrations throughout the country (in 2015 there were around 46
million Colombians (DANE, 2020)). The map shows that the phenomenon was widespread across
municipalities.
2.2 The Sisben and Familias en Acci on
Familias en Acci on (FA) is a Conditional Cash Transfer (CCT) program. Its main objective consists
in aiding the most vulnerable and poorest households of the country, by giving them a monetary
transfer. To receive the subsidy, the families are required to: 1) take all kids younger than 7 years
old to health controls and 2) make sure all kids and adolescents in the household are going to
school. In 2014 around 2.7 million households received a subsidy from FA.
Only poor or vulnerable households are eligible to receive a FA transfer. To identify poverty,
Colombia's government has implemented the Sistema de Identicaci on para Potenciales Bene-
ciarios de los Programas Sociales (Sisben), which is essentially a census of the poor. The survey
developed in 2011, used in this paper, included close to 30 million individual responses. The Sisben
creates an index that measures multidimensional poverty, structured from multiple variables which
capture poverty and vulnerability and that are also hard to manipulate by the respondents. Based
on this index, the system assigns to each household a score that ranges (continuously) from 0 to
100, where lower values indicate more poverty. The algorithm used to calculate the index and the
score is unknown (and indecipherable) to the population.
8
The government identies the potential beneciaries of its social programs using the Sisben
score. For the FA program, three dierent cutos were determined according to the location of
each household. Families with scores below 30.56, 32.2 and 29.03 located in Colombia's 14 biggest
cities, other urban agglomerations, and the rural area, respectively, are eligible for the FA subsidy.
2
The amount transferred to each household depends on the number and the ages of the children
in the family and the school grade each child is attending. There are two main components that
determine the amount of the subsidy given: one based on health and the other on educational
outcomes.
A health subsidy is given to eligible households if at least one child is below the age of 7. This
transfer is given every two months conditional on every child below this age going to scheduled
health controls. This is a xed payment and its magnitude is of the same amount regardless of the
number of children below the age of 7 in a household. The educational component of the transfer
requires that all children in eligible households physically attend school. A subsidy is given for each
child that fullls this requirement; nonetheless, each household can get a subsidy for up to three
children. The transfer is given 5 times each year (bimonthly during the school calendar). Table
1.12 in the appendix shows the dierent amounts of each transfer depending on the component of
the program. As it can be seen, each municipality is grouped into four categories, which explain
slight variations in the amounts transferred. To gain some perspective, the monthly minimum wage
in 2015 was equal to 644,350 Colombian pesos, therefore the monthly value of a health transfer
was close to 5% of the minimum wage. Overall, the average monthly transfer is around 50,000 to
60,000 pesos, which corresponds to 20 dollars per month.
3
2.2.1 Data
The Sisben survey of 2011 is used in this paper, given that it is the closest one to the elections of
2015. The interviews for this census of the poor were collected between 2009 and 2010. The data
for beneciaries and the payments each household received from the FA program used in this paper
2
For more details on the criteria of eligibility, the determination of the program's cutos and the areas determined
for each cuto see (DPS, 2013a).
3
Colombia's minimum wage in 2015 was 86% of the median wage, which is relatively high with respect to OECD
countries (OECD, 2020).
9
range from the year 2012 to the year 2015. In this paper a household is considered a beneciary
of the FA program if they are coded as such by the administrative databases during the rst two
months of 2015.
The Sisben, the FA administrative data and the voter buying data include the IDs for each
individual, which is why it is possible to merge them. Table 1.11 shows summary statistics for
voters bought and non-bought voters identied in the Sisben. The table shows averages for dierent
variables for the population who, at the time of the Sisben survey, had a c edula, which is the ID that
certies that each individual is older than 18, and therefore could vote. The table doesn't show any
extreme dierences between voters bought and non-bought voters in demographic, employment,
education or income variables. It is important to note that, although the total number of voters
bought indicated by the lists developed by the CNE was close to 1.5 million people, it was only
possible to identify 382,289 individuals who were both at the Sisben and who had kids living in the
household in 2015, which is one of the key components for eligibility for the FA program.
Several reasons can explain non-matches between the voter buying data and the Sisben survey.
First, there is a temporal dierence between the Sisben surveys (2011) and the timing of the elections
(2015); people who turned 18 during this time and got their ID can't be matched. Second, the
lists of voters bought taken from the CNE's rulings are scanned les. It was necessary to use an
optical character recognition program to get the ID lists to a workable format; it is potentially
the case that some IDs were misread. Third, the Sisben is the census of the poor, there is still a
part of the population that is not surveyed on it. There are no reasons to think that mismatch is
selectively happening for the population close to the poverty score cutos that determine access to
the FA program; hence, the results of a Regression Discontinuity Design (described in the following
section) should not be biased.
2.3 Voter Turnout
In Colombia's electoral system, every citizen older than 18 years old who has a valid ID (locally
known as c edula de ciudadan a) can vote. Each person votes in a previously determined voting
location: once a citizen claims their ID (they can only do so once they turn 18) they are auto-
matically registered to vote in the closest polling site to their claimed address of residence within
their municipality. If a citizen changes their residency they can change their voting location before
10
the elections take place. For the 2015 elections citizens could change their voting location between
October 25th of 2014 and August 25th 2015 (Rubiano, 2015). After someone changes their voting
location, the electoral authorities proceed to evaluate whether the person's new location is legally
acceptable, in the case it is not, they are
agged as a voter bought, as was mentioned in the previous
sections.
Once someone is adequately registered and goes to vote in their corresponding voting location
the day of the elections, they are assigned to select their candidate of preference in a voting station
within the voting site. In 2015, there were 360 potential voters for each voting station. Here, each
person casts their vote anonymously{which is why it is not possible to know the political preferences
for each citizen{and are manually reported to have turned out to vote. This allows people to vote
only once. Voting jurors assigned to each voting station are responsible to ll out the forms of each
person who turns out to vote, and, more generally, assist in the overall electoral process for their
station during the election day.
2.3.1 Data
There is not a digitized database of individual voter turnout in Colombia. The lack of information
of voter turnout at the individual level happens because{as mentioned in the previous section{the
process to keep voting records is done manually at each voting station. Figure 1.8 in the Appendix
shows an example of how voter turnout is tracked during election days, this type of record is known
as a \form E-11". Each voting station initially has an E-11 form, which includes a list of all
registered voters' ID, with three empty contiguous columns. As the election day passes, for each
person who turns out to vote, the blank contiguous columns are signed, indicating that the person
with the corresponding ID number voted. If the columns next to an ID number are not lled, it
means the person did not vote.
This paper uses more than a million pages of scanned voting records for the 2015 elections,
provided by the Registradur a Nacional del Estado Civil and, using computer vision and optical
character recognition techniques, identies if an individual turned out to vote or not for the universe
of Colombian voters. The data of individual voter turnout is used to test multiple mechanisms. The
cleaned data includes the ID of each citizen registered to vote, their voting site and whether the
person turned out to vote or not. For the 2015 elections, 27 million IDs were identied (some of them
11
are subject to error) as potential voters. 58.8% of these individuals are recognized as voters, which
is in line with the ocial 59.4% of voter participation reported by the electoral authorities. An
upcoming article, separated from this paper, explains the data cleaning process and precision of the
extracted information, along with other descriptive information at the municipal and departmental
level.
2.4 Other Data: Municipal Characteristics
The heterogeneity tests in the results section explore multiple characteristics at the municipal level.
Variables of municipality size in squared kilometers are taken from the Municipal Panel created by
the CEDE. The ocially registered population to vote by municipality is taken from the Misi on
de Observaci on Electoral -MOE. Other electoral data, such as the number of votes received by
candidate in multiple elections is taken from Pach on and S anchez (2014b).
3 Empirical Strategy
To identify a causal relationship between income and vote selling or clientelism is challenging.
Higher incomes may lead individuals to sell their votes less. However, less vote buying and lower
levels of corruption may also lead to higher levels of income. An OLS estimation, regressing vote
buying on income would not identify a causal eect: the clientelistic behavior itself could have an
eect on income, which generates reverse causality issues. Income is also correlated with other
variables, like education or employment status, which in
uence the engagement of citizens in vote
buying, hence leading to omitted variable problems.
Ideally, one would want a national randomized experiment to test whether a set of randomly
selected individuals receive an income shock (treatments), while the rest of the population does
not receive any income transfers (controls). Additionally, vote-selling should be measured for both
the treated and control populations. The causal eect of an income transfer on vote selling would
consist in comparing this outcome in the treated and control groups. This ideal experiment has
not yet been eectively developed in the previous literature because of two reasons: 1) Only about
5-6% of the population sells their vote. Although this is a high number, to precisely identify a
treatment eect a high number of observations would be needed in an experimental setting, which
12
lead to prohibitively high expenses. 2) Properly measuring vote-selling in an experimental setting
is not straightforward. Previous studies usually rely on self reported measures of vote-selling, which
are not trustworthy, given the illegal features of this practice.
The criteria of eligibility for the FA program allows for a causal identication of the income
eects on vote buying. The size of the program and the potential beneciary population provides a
suciently large sample size to evaluate the eects of an income transfer on vote buying. Moreover,
the digitized data on voters bought from the CNE also provides a large sample of potential vote-
sellers.
A Regression Discontinuity design (RD) is used to causally identify (and quantify) the eects
of higher incomes on vote buying. Households are eligible for the FA program if their Sisben's
multidimensional poverty score is below 30.56, 32.2 and 29.03 in Colombia's 14 biggest cities, other
urban agglomerations, and the rural area, respectively (these three dierent subdivisions will be
noted as areas). To obtain a causal estimate of the program, this paper compares those individuals
whose score is close to the cutos of eligibility: those below the cuto are potential beneciaries
of the program (treated); while individuals above the cuto are non-beneciaries of the program
(controls). The validity of this specication relies on the assumption that individual characteristics
do not vary discontinuously across the cutos of eligibility.
A sharp RD and a Fuzzy RD around the standardized Sisben score are estimated to identify
the causal eect of the income shock on voter buying. The following sharp RD, reduced form
specication, is estimated to capture the eect of the FA program on vote buying:
VoterBought
ia
=
0
+
1
f(score
ia
) +
2
Above
ia
+

3
Above
ia
f(score
ia
) +
a
+
ia
(1.1)
Where i stands for individual and a represents each area that determine the eligibility cuto
value for the FA program. VoterBought
ia
is a dummy variable equal to 1 if an individual old enough
to vote is
agged as a voter bought by the electoral authority and zero otherwise. The score
ia
is
the standardized multidimensional poverty score assigned to each individual by the Sisben.
a
are
area xed eects. f() is a functional form that allows a non-parametric and continuous relationship
13
between the running variable and the outcome of interest. The Above
ia
variable indicates whether
an individual is above the thresholds of eligibility for the FA program. Therefore, this variable
takes values of zero for each eligible individual and one for non-eligible individuals according to
their Sisben score. More formally, Above
ia
is dened as:
Above
ia
=
8
>
>
>
>
>
>
<
>
>
>
>
>
>
:
1(score
ia
> 30:56) if
a
=\14 cities":
1(score
ia
> 32:2) if
a
=\Other urban agglomerations":
1(score
ia
> 29:03) if
a
=\Rural Area":
(1.2)
In addition to the reduced form specication 1.1, a fuzzy regression discontinuity is also esti-
mated. In this case, being an eective beneciary of the FA program (denoted byFABeneficiary
ia
)
is instrumented with the eligibility variable dened in 1.2. The rst stage is estimated as follows:
FABeneficiary
ia
=
0
+
1
f(score
ia
) +
2
Above
ia
+

3
FAeligible
ia
f(score
ia
) +
a
+
ia
(1.3)
Using the predicted beneciary status estimated in 1.3, the fuzzy RD specication is:
VoterBought
ia
=
0
+
1
f(score
ia
) +
2
\
FABeneficiary
ia
+

3
\
FABeneficiary
ia
f(score
ia
) +
a
+
ia
(1.4)
The coecients of interest are
2
and
2
, estimated in regressions 1.1 and 1.4. For robust-
ness purposes, all specications are estimated using four dierent bandwidths around the cutos:
2:5;5;7:5;10 points.
The baseline specications are estimated using a triangular kernels, results using uniform kernels
are also reported as robustness checks. The main specications are estimated using a linear local
polynomial unless otherwise noted. Furthermore, when noted, socioeconomic and demographic
controls are also included for robustness and precision purposes. Standard errors are clustered at
the household level, given that the FA transfers are assigned according to the household Sisben
score. The Sisben score is normalized to 0 in each area, such that individuals eligible for the FA
14
program have a normalized negative Sisben score.
3.1 Specication Tests
The Regression Discontinuity used in this paper is properly identied if the population does not
manipulate their multidimensional poverty scores around the cutos of eligibility and if other
variables, especially those that may be determinants for voter buying or clientelism, are continuous
around these cutos.
Figure 1.3 shows the distribution of households with respect to the normalized Sisben score.
Households below a value of zero are eligible for the FA program. The distribution is smooth
at 0, suggesting that there are no clear signs of manipulation. This is expected, given that the
algorithm to calculate the multidimensional poverty score implemented in this wave of the Sisben
was unknown (and indecipherable) to the population.
4
If there was manipulation, one would
expect a larger density just below 0, as people below this value are eligible for the FA subsidies.
The histogram is clearly not indicating such behavior. A more formal exploration of possible
manipulation is explored in Figures 1.4 and 1.5, which respectively show the McCrary (2008) and
the Cattaneo et al. (2019) tests for density dierences at the eligibility cuto. Both tests show that
there is a smooth distribution of the population across the cuto of eligibility for the FA program.
Identication could also be threatened if variables potentially correlated with vote buying and
clientelism are signicantly dierent across the two sides of the cuto. Using the reduced form RD
displayed in specication 1.1 and replacing the outcome variable with socioeconomic and demo-
graphic characteristics, it is possible to test if there are no other discontinuities at the cuto that
could be driving the results.
Tables 1.1 and 1.2 show the estimated coecients for being FA eligible (
2
in equation 1.1) on
each of the outcome variables reported in the rst column. The coecients are estimated using
four dierent bandwidths. When a narrower bandwidth is considered most variables don't show
clear dierences across the cutos of eligibility. The proportion of males and certain household
characteristics (number of rooms, and whether there is an oven or water heater in the household)
are signicant but the magnitude of the coecients is small. However, when the bandwidths are
4
Note that in previous versions of the Sisben there were clear signs of manipulation, as studied by Camacho and
Conover (2011).
15
wider there are more signicant dierences across the cuto, although most of them are of small
magnitude. Therefore, the preferred specications will be the ones estimated in the narrower band-
widths. The imbalances are a result of the large number of observations taken into consideration;
once the eects are analyzed for each of the variables, it is clear that what might be highlighted as
a signicant dierence, is not economically relevant.
For robustness and transparency purposes, the results are estimated using the previously men-
tioned four bandwidths and in most specications demographic controls are also included. The main
results are also estimated using the optimal MSE bandwidths calculated according to Calonico et al.
(2014).
4 Results
This section shows the main results. As detailed in the previous sections, to identify a causal link
between higher incomes and voter buying, a regression discontinuity is used around the cuto of
eligibility for the FA cash transfer program. First, results of the reduced form in specication
1.1 are shown. Then, the fuzzy RD estimates are described, including both details on the rst
stage (equation 1.3) and the second stage (equation 1.4). Following that, an alternative measure
of voter buying is used as an outcome. To build it, I use information of citizens who, after being

agged as voters-bought, presented documentation certifying that where they registered to vote was
legitimate. This variable is used to avoid the potential concern that the measure of vote buying
might capture a legitimate migration across municipalities by an individual. Additionally, there is a
subsection which explores the possible mechanisms driving the results. Finally, the last subsection
analyzes multiple heterogeneous eects of interest.
4.1 Reduced Form Estimation
Table 1.3 shows the reduced form specication results. The coecient of interest indicates the
dierence in the likelihood of being a voter bought between individuals who are non-eligible for the
FA program with respect to eligible people. Therefore, the RD estimates measures the causal eect
of being eligible for the FA program on the likelihood of someone selling their vote. The results
indicate that individuals who are eligible for the cash transfer subsidy are less likely to engage in
16
voter buying. In particular, citizens who are barely below the threshold of eligibility, and therefore
eligible to the FA program, are 0.1-0.14 percentage points less likely to sell their vote. These results
indicate that the Conditional Cash Transfer program has an eect in reducing vote selling.
The coecients shown in 1.3 are all estimated using the optimal bandwidths calculated accord-
ing to Calonico et al. (2014). The results are signicant and robust when estimated using dierent
kernels (triangular and uniform) and using dierent local polynomials across the cutos (rst and
second order).
Specication 1.1 is displayed under dierent bandwidth selections. Table 1.13 shows the esti-
mated coecients for four dierent bandwidths. All specications are estimated using triangular
kernels and lineal local polynomials. The results are robust and statistically signicant across band-
widths. As mentioned in the previous section, the preferred estimates correspond to those with
narrower bandwidths, as the observable characteristics between the populations above and below
the eligibility cutos are statistically indistinguishable. It is clear that in narrower bandwidths
fewer observations are taken into consideration, which lead lo less statistical power.
The reduction in vote-selling caused by the FA program is also robust when socioeconomic and
demographic controls are included in the reduced form specication. Table 1.14 shows the results.
The coecients of interest are robust to the inclusion of controls, both in magnitude and statistical
signicance.
The results shown indicate that eligible individuals to the FA program are 0.1-0.2 percentage
points less likely to sell their vote with respect to non-eligible citizens. The variation in magnitudes
mostly comes from dierences in the bandwidths of choice in the RD, however, the results are
robust and statistically signicant across specications.
Figure 1.6 shows a plot for the estimated reduced form Regression Discontinuity with linear
polynomials. The graph indicates that around 3.75% of non-eligible individuals for the FA pro-
gram (those above the cuto) engage in voter buying. Being eligible to the FA program reduces
the likelihood in between 0.1-0.2 percentage points. This suggests that the FA program reduced
the likelihood of vote-selling in between 2.6 and 5.4% with respect to the average of non-eligible
individuals.
17
4.2 \First Stage" RD Estimation
An issue with the reduced form estimates is that they do not take into consideration the compliance
of the FA program and, therefore, don't measure the treatment eect with the actual beneciaries
of the program. This is why the reduced form estimates can be interpreted as an Intention To
Treat. To estimate the Average Treatment Eect of receiving the FA program on voter buying,
it is necessary to instrument the actual beneciary status with the exogenous assignment of the
program using the arbitrary cutos of eligibility, as described in Equations 1.3 and 1.4.
Table 1.4 shows the estimates of the rst stage described in equation 1.3 for multiple bandwidths
around the normalized Sisben cuto. The results show that voting-age individuals (who also live
in households with kids, and therefore who could potentially receive an FA subsidy) who are just
above the cutos of eligibility are 20 percentage points less likely to receive an FA subsidy. The
results are robust and are statistically signicant across specications (dierent kernels and local
polynomial orders).
Although one would expect a value for the coecients estimated in Table 1.4 closer to one,
several details of the FA program could in
uence this outcome. For example, indigenous citizens
and the victims of Colombia's internal con
ict are eligible for the program independently of their
score (Medell n and S anchez, 2015). Additionally, given that the FA program is a conditional cash
transfer program, households who are eligible must fulll the educational and health conditions
described in the data section. Moreover, the beneciaries considered in this paper are those voting-
age individuals who reside in a household where someone eectively received an FA subsidy in the
rst quarter of 2015. The results are close in magnitude and signicance to other studies that have
used the FA program, such as Conover et al. (2020) and Attanasio et al. (2021).
The rst stage RD plot is shown in Figure 1.7. The results are signicant and present little
variation when estimated using the optimal Calonico et al. (2014) bandwidths, dierent kernels,
and after controlling for socioeconomic and demographic controls, as shown in Appendix Tables
1.15 and 1.16.
18
4.3 Fuzzy RD Estimation
Results for the Fuzzy RD specication described in Equation 1.4 are shown in Table 1.5. The
results suggest that beneciary individuals of the FA program are between 1 and 0.5 percentage
points less likely to sell their vote. This variation depends on the bandwidth of choice. As in
previous sections, the preferred specications correspond to those in columns 1 and 2, because
the treated and control populations are more similar given their observable characteristics under
narrower bandwidths. The results indicate that the FA beneciaries are 18 to 27% less likely to
sell their vote when compared to the non-beneciaries of the program.
The estimates are all signicantly negative at conventional levels across multiple bandwidths
and dierent specications. Note that a negative coecient means that beneciaries are less likely
to sell their vote. Table 1.17 in the Appendix shows that the estimates are still signicant and
negative when dierent kernels and polynomials are used, with the exception of the uniform kernel
and quadratic polynomial, however the coecient of interest is close to signicance at the 10%
level. Table 1.18 in the Appendix shows the results for multiple bandwidths after socioeconomic and
demographic controls are included. The coecients and their signicance don't change signicantly
with respect to Table 1.5.
Although the magnitude of being FA eligible on vote-selling seems small (between 0.5 to 1
percentage points), the results are economically and politically relevant. A back of the envelope
calculation suggests that the transfers prevented a large number of citizens to engage in voter
buying. The number of households that get the subsidy is around 2.7 million, which means that
around 5 million adults benet from the transfers. Therefore, the FA program potentially reduces
the magnitude of voters bought in 30,000 people. This number comes from multiplying the local
treatment eects estimated in Table 1.17, with the number of adults who benet from the subsidies
(around 5 million people). Naturally, this is a suggestive interpretation of the results and likely
an overestimation of the eects of the overall eects of the program. The RD estimates a local
eect around the bandwidths of choice and, in this case, identies the likelihood of vote-selling at
the margin. It is clear that as the population of interest is further away from the cutos, other
variables in
uence whether an individual sells their vote or not.
19
4.4 Why voters bought are not migrants
One potential concern about the voter buying variable used in this paper is that it could capture
an individual's legitimate migration behavior . This could be the case if someone moves from one
municipality to another and registers to vote in their new municipality before any other adminis-
trative information is updated. In this case, a legitimate migrant would be
agged as a potential
voter bought by the electoral authority.
This subsection addresses the issue, using a modied variable of voter buying as an outcome
for the same RD estimations mentioned in the empirical strategy section. The results are robust
to the modication in the voter buying variable, indicating that regular migration is not driving
the eects highlighted in the previous sections.
To overcome the issue of migration, the electoral authority allows people to present what is
called a \recurso de reposici on". Recall that voters bought are identied as those citizens who
register to vote in a dierent municipality to where they were previously registered to vote, and
their administrative records (according to information from multiple administrative data-set) do not
indicate that the citizen actually resides in the new municipality where they registered to vote. The
electoral authority noties each citizen who has been
agged as a potential voter bought, and gives
them a period of ve business days to send (electronically or personally) information to certify that
the new municipality where they registered to is indeed their location of residence (CNE, 2015a).
Citizens can present certicates of work, health, Sisben, leases, taxes, study, displacements due to
violence, among others, to testify they reside in the new municipality they registered to vote (CNE,
2015b).
After a process of revision, the electoral authorities publish the list of citizens who, after pre-
senting a valid \recurso de reposici on", can vote in the new municipality where they registered. In
this section, the variable of voter buying is modied such that the citizens who present the valid
documentation to vote in a dierent municipality are considered regular non-bought voters. These
citizens' information was digitized from the resolutions emitted by the electoral authority.
5
Table 1.19 estimates the reduced form RD using the information of citizens who present valid
documentation of their change in residency location as an outcome. The results indicate that
5
The resolution was the CNE's 6170 of 2015 (CNE, 2015a). The document was digitized using Abbyy.
20
legitimate migration is not changing discontinuously across the cutos of eligibility for the cash
transfers. This provides evidence that the main results shown previously are not driven by individ-
uals legitimately and selectively migrating depending on whether they are potential beneciaries
of the income transfers.
Moreover, I use a modied version of the voters-bought variable, after considering the eective
proofs of residence, as an outcome for the estimated RDs. Table 1.20 shows the reduced form
described in specication 1.1, estimated for multiple bandwidths. The magnitudes and signicance
of the coecients vary slightly with respect to the results shown in the reduced form estimates
with the broader version of the voter buying variable. The coecients of interest still indicate
that potential beneciaries of the FA program are 0.1-0.18 percentage points less likely to sell their
vote with respect to potential non-beneciaries. The results suggest that even after correcting for
potential legitimate migration, vote-selling decreases due to the FA program.
A specication using the modied variable of voter buying and controlling for socioeconomic
and demographic characteristics is also estimated. Table 1.21 in the Appendix shows that the
results are more precisely estimated, as expected, and their magnitude and sign are similar to
the specication without controls. Moreover, Tables 1.22 and 1.23 indicate that the estimated
coecients don't signicantly change in the fuzzy RD specication with respect to the baseline
results shown previously, even after the voter buying variable is modied.
Overall, the results presented in this section indicate that legitimate migration is not driving
the main eects shown previously.
4.5 Mechanisms
This paper has provided evidence indicating that individuals are less likely to engage in vote buying
if they are beneciaries of a Conditional Cash Transfer program in Colombia. It is not entirely
clear, however, what is the channel that explains why individuals who receive the subsidy are less
likely to potentially sell their vote.
On one hand, as individuals receive higher (and stable) additional earnings, their marginal
utility from an extra unit of income decreases. Moreover, if there is a penalty to engage in vote
buying, then the beneciaries of the subsidy will be less inclined to receive a bribe or payment from
a politician (income channel). On the other hand, beneciaries of the conditional cash transfer
21
program could change their beliefs with respect to democracy and the functioning of The State.
After receiving the subsidy, citizens could see that the government works and, therefore, would be
less willing to contribute to the erosion of democracy by selling their vote (state building channel).
This section shows that the state-building channel is not explaining the main eect driving
the results shown in the previous sections, indicating that the income eect provides a compelling
interpretation of the main results.
The state-building channel implies that beneciaries of the FA program would be more appre-
ciative of democracy and the State. As a measure of democratic appreciation and trust in the
State, this paper uses novel individual voter turnout information, which was digitized from exten-
sive voting records, as described in the Data section. Note that the voter turnout information is
identied at the individual level, hence, it can be merged to the rest of the administrative data
using each citizens' national identication number. Therefore, the state-building channel is tested
by estimating the same specications described in the Empirical Strategy section, but using voter
turnout as the dependent variable.
Table 1.6 shows the reduced form results obtained from estimating equation 1.1 using voter
turnout as an outcome. The coecients estimated for multiple bandwidths indicate that the FA
program does not have an eect on increasing voter turnout at any conventional levels of signi-
cance. If anything, the sign suggests that the non-eligible population is slightly more likely to vote,
however the eect is indistinguishable from zero. Table 1.24, in the Appendix shows the estimated
coecients for the Fuzzy RD specication. The intuition and signicance coincides to the reduced
form estimates. Beneciaries of the FA program are not more likely than non-beneciaries to turn
out to vote.
Although the income-channel is not directly estimated, the evidence described in this section
indicates that the state-building channel is not the force driving the results shown in the previous
sections. Beneciaries of the Conditional Cash Transfer program are not more likely to turn out
to vote. This suggests that the subsidy is not necessarily making citizens value more democracy
and the State. However, the results indicate that individuals who receive the transfer are less likely
to engage in vote buying, potentially because their marginal utility of a bribe or payment by a
politician no longer compensates the costs associated to selling their vote.
22
4.6 Heterogeneity
This paper has shown that individuals who receive a transfer from the government are less likely to
engage in vote buying. However, these eects may depend on other aspects of the political economy
environments. This section presents a simple framework to highlight other aspects that may aect
voter-buying.
Assume that individuals face three choices with respect to voting: 1) Voting legally in their
own municipality; 2) Voting in another municipality, in exchange for a transfer from a politician;
3) Not voting at all.
Following a common model of voter turnout, such as in DellaVigna et al. (2016), a citizen who
votes legally in their own municipality gets the following utility:
u
i
=p
o
V
o
+g
o
c
o
(1.5)
Wherep
o
is the likelihood of being a pivotal voter in an individual's municipality of residenceo.
V
o
is the value of deciding the election at an individual's municipality of residence. g
o
is the \warm
glow of voting": This consists in the feeling of civic duty, value for community, etc. The term c
o
is
the cost of voting in the citizen's own municipality, which encapsulates the costs of transportation
to the voting station or/and the average opportunity cost of voting for an individual who resides
in municipality o.
If an individual is a voter bought, their utility is:
u
i
=c
od
+
1
y
i
b (1.6)
Equation 1.6 indicates that voters bought do not get any satisfaction from being pivotal and
do not get any warm glow from voting. b is the payment or bribe they receive from the politician
(namely, the price for their vote). This payment is multiplied by the inverse of each person's income
levely
i
to capture the negative marginal utility of income: an additional unit of income reduces the
likelihood of individuals to sell their vote, which corresponds to what is documented in the previous
sections of this paper.
6
c
od
is the cost from of transportation or/and the average opportunity cost
6
The term
1
y
i
b is an extreme simplication of more sophisticated models of vote buying that take into consideration
how marginally poorer individuals are more likely to sell their votes than wealthier people. For example, Stokes et al.
23
to move from an individual's municipality of residence o to vote in a destination municipality d.
There is not an additional penalty for selling the vote, however it can be considered whether as an
additional cost c
od
or as a new parameter with negative sign.
Additionally, assume that individuals receive a utility of 0 if they choose not to vote and don't
sell their vote.
An individual who resides in municipality o prefers to sell their vote if:
7
b [p
o
V
o
+g
o
+c
od
c
o
]y
i
(1.7)
The following hypothesis, motivated from the simple framework mentioned above, are used to
analyze the heterogeneous eects of the cash transfers on voter buying:
Prediction 1: Pivotality. If someone is more likely to be pivotal in their municipality of resi-
dence, the income transfer has a smaller eect in reducing voter buying.
To test this, the reduced form RD is estimated on two samples, one for all municipalities
where the voting-age population is above the municipal median for all the country, and
another for all municipalities with below the median voting-age population. People who live
in less populated municipalities have a higher likelihood to be pivotal. Therefore, the FA eect
on voter buying is expected to be smaller in municipalities below the voting-age population
median. Table 1.7 shows the estimated coecients of the reduced form RD. As expected, FA
eligible individuals from municipalities below the voting-age population median are less likely
to be vote-sellers. Moreover, the estimated coecients for the samples above and below the
median are dierent from each other at the 10% level of signicance.
Prediction 2: Value of Deciding the Election. If someone has a higher value of deciding the
election in their municipality of residence, the income transfer has a smaller eect in reducing
voter buying.
To measure the value of deciding an election, an index of how competitive the previous
(2013b) assumes that citizens assign a2 (0; 1) weight in the utility of voting to their ideological political dimensions
and 1 to material benets they get from politicians. Furthermore, they argue, increases with income. For
simplicity purposes, the utility of voters-bought in this paper just takes into explicit consideration how the weight to
material benets is decreasing in the level of each person's income
7
For the rest of this analysis it will be assumed that bc
od
yi.
24
mayoral elections is calculated for each municipality.
8
The index ranges from 0 to 1, with
smaller values representing more competitive elections. The FA eect on voter buying is
expected to be smaller in municipalities with more competitive elections. Table 1.8 shows
that in municipalities below the median with respect to the index of competitiveness the FA
program has a smaller eect in reducing voter buying than in those above the median, as
expected. Furthermore, the two coecients are signicantly dierent from each other at the
5% level of signicance.
Prediction 3: Cost of voting in a citizen's municipality of residence. If the costs to vote
in someone's municipality of residence are high, the income transfer has a larger eect in
reducing voter buying.
The number of voting stations per square kilometer in every municipality is used as a proxy
to measure the costs of voting within that municipality. If there are less voting stations,
then the voting costs increase, which according to the hypothesis, leads to more vote buying.
Table 1.9 shows the reduced form RD estimated for the sample of individuals who reside in
municipalities below and above the voting stations' municipal median. The cash transfer has
a larger eect in reducing vote-selling for the sample of individuals who live in municipalities
with higher costs to vote. It must be noted that the analysis is suggestive of the direction of
the eect; however, the estimated coecients, for the samples above and below the median,
are not dierent from each other at conventional levels of signicance.
Prediction 4: Cost of voting in a dierent municipality from a citizen's residence. If the
costs to vote in a dierent municipality from someone's location of residence are high, the
income transfer has a smaller eect reducing voter buying.
The average number of voting stations per square kilometer in all the neighboring munic-
ipalities from someone's municipality of residence is used as a proxy for c
od
. Neighboring
municipalities are dened as all the municipalities that share a border with municipality o.
If there are plenty voting stations in someone's neighboring municipalities, the costs to be
8
The index is calculated using the updated electoral data from Pach on and S anchez (2014b) for the mayoral
elections of 2011. Each municipality's level of competitiveness is dened as Compo =
winnerosecondo
winnero+secondo
. Where
winnero is the number of votes received by the winner candidate in municipality o. secondo is the number of votes
received by the second most voted candidate in the municipality.
25
a voter bought are lower and the likelihood to sell the vote increases. Therefore, the cash
transfer has a higher eect in reducing voter buying for citizens living in locations where
neighboring municipalities have a higher density of voting stations. Table 1.10 suggests that
the FA program reduces the likelihood of an individual selling their vote if they live in lo-
cations where neighboring municipalities have a higher density of voting stations (above the
national municipal's median). The results are suggestive of the direction of the eect; the
estimated coecients, for the samples above and below the median, are not dierent from
each other at conventional levels of signicance.
Motivated by a simple framework, this sections showed that the FA program has a dierent eect
on the likelihood of someone selling their vote depending on multiple municipal characteristics. In
particular, municipalities where people expect to be more pivotal or where the value of deciding
the elections are higher, the FA program has lower impacts in vote buying. Furthermore, if the
costs to legally vote in someone's municipality of residence are higher, the FA program has a bigger
eect in reducing voter buying. Finally, if the costs associated to vote in a municipality dierent
from someone's own residence are higher, and therefore if there are higher costs in being a voter
bought,the FA program has a smaller eect in reducing voter buying.
5 Conclusion
There is general consensus that high income countries have lower levels of corruption and electoral
irregularities. However, there is limited evidence showing that higher levels of income causally
reduce individuals' engagement in clientelism, in particular vote buying. The existing literature
focuses more on how politicians target individuals who are more reciprocal (Finan and Schechter,
2012), socially connected (Duarte et al., 2019) or easily monitored (Larreguy et al., 2016).
This paper documents that a Conditional Cash Transfer program leads to a substantial reduction
in voter buying, a particular form of clientelism. Under voter buying, individuals illegally register
and vote in municipalities where they don't live, usually in exchange for money, sponsored by local
politicians. By digitizing data from Colombia's national electoral authority on the individuals who
were identied as voters bought during the elections of 2015, this study shows that the beneciaries
of the conditional cash transfer program are signicantly less likely to engage in voter buying.
26
The estimates documented in this paper suggest that receiving the subsidy reduced the individual
probability of voter buying by 18 and 27%. Moreover, several municipal characteristics in
uence
the political economy of voter buying (Hidalgo and Nichter, 2016), which aect the eectiveness of
the cash transfer program on voter buying.
The beneciaries of the income transfers were not more likely to turn out to vote during the
elections of interest. This result suggests that the subsidies did not increase the sense of civic
responsibility or trust in the State, but rather that the reduction in vote buying was caused by the
pure income transfer. As a measure of democratic responsibility, this study uses voter turnout data
at the individual level. To build this variable, more than a million pages of handwritten documents
were digitized using computer vision and optical character recognition techniques.
Unlike most studies of vote buying and clientelism, this paper uses non self-reported individual
measures of voter buying for a whole country. Such variables are rare in the literature, as most
studies use individual self reported data or aggregate measures of corruption. Although this study
focuses solely on the eect of income transfers on voter buying, further research should explore
other factors that reduce vote buying.
27
Figures
Figure 1.1: GDP per capita and Corruption
Notes: This gure plots the logarithm of per capita GDP (taken from the World Development Indicators) in 2018 and the
perceived level of corruption index calculated by Transparency International for 2019.
28
Figure 1.2: Colombia's Electoral Transhumance as a % of municipality's population, 2015
Notes: This map shows voters bought per capita by municipality. Darker colors indicate more pontential voter imports. The
data for electoral transhumance is calculated using CNE's lists of electoral transhumance. The population data is taken from
the DANE. All data is for the year 2015.
29
Figure 1.3: Sisben normalized score distribution
0 .005 .01 .015 .02
Density
−40 −20 0 20 40 60
score
Notes: This histogram shows the distribution of households with respect to the normalized Sisben score. Those below a value
of zero are eligible for the FA program.
30
Figure 1.4: McCrary Test
0 .005 .01 .015 .02
−50 0 50 100
Notes: This graph shows the McCrary (2008) test of density dierences at household level across the normalized Sisben cuto
of 0.
31
Figure 1.5: Cattaneo, Jansson and Ma (2019) Density Test
.016 .0165 .017 .0175 .018
−10 −5 0 5 10
Score
point estimate 95% C.I.
Notes: This graph shows the Cattaneo et al. (2019) test of density dierences at household level across the normalized Sisben
normalized cuto of 0.
32
Figure 1.6: FA Eligibility and Voter Buying. Reduced Form Results
Notes: This gure plots the relationship between the likelihood of being a voter-bought and the normalized Sisben score.
Observations to the right of 0 indicate that individuals are above the cutos of eligibility for the cash transfer program.
33
Figure 1.7: First Stage Results: Eligibility and Beneciaries of FA in 2015
Notes: This gure plots the relationship between the likelihood of being beneciary of the cash transfer program and the
normalized Sisben score. Observations to the right of 0 indicate that individuals are above the cutos of eligibility for the cash
transfer program.
34
Tables
35
Table 1.1: Balance Table
Bandwidths
2.5 5 7.5 10
Age 1.10e-05 0.153*** 0.174*** 0.162***
(0.0776) (0.0549) (0.0449) (0.0390)
HH. Members 0.0167 0.0528*** 0.0411*** 0.0459***
(0.0181) (0.0128) (0.0106) (0.00915)
Married -0.000145 -0.00175 -0.00470*** -0.00376***
(0.00237) (0.00168) (0.00137) (0.00119)
Male -0.00833*** -0.00948*** -0.00619*** -0.00489***
(0.00136) (0.000963) (0.000788) (0.000684)
Power 0.000338 0.000620 -0.000261 -0.00160***
(0.00112) (0.000791) (0.000648) (0.000564)
Sewer -0.00257 -0.00333* -0.00339** -0.00531***
(0.00247) (0.00175) (0.00144) (0.00125)
Gas -0.00614** -0.00347* -0.00355** -0.00451***
(0.00266) (0.00188) (0.00154) (0.00135)
Telephone 3.13e-05 0.00200 0.00201 0.00229*
(0.00233) (0.00165) (0.00136) (0.00118)
Trash -0.00137 -0.000200 0.000364 6.28e-05
(0.00211) (0.00150) (0.00123) (0.00107)
Aqueduct -0.000632 -0.000283 -0.000507 -0.00156
(0.00233) (0.00164) (0.00134) (0.00117)
Rooms 0.0155* 0.0276*** 0.0266*** 0.0320***
(0.00817) (0.00575) (0.00469) (0.00407)
Toilets -0.000855 0.000207 -0.00244** -0.00748***
(0.00170) (0.00119) (0.000958) (0.000825)
Shower 0.000318 -0.00159 -0.00272* -0.00477***
(0.00278) (0.00196) (0.00161) (0.00140)
Fridge -0.00437 0.00328 0.00334* 0.00507***
(0.00305) (0.00216) (0.00177) (0.00153)
Washer 0.000809 0.00395*** 0.00297** 0.00356***
(0.00216) (0.00153) (0.00125) (0.00109)
TV -0.00219 0.00113 -0.000208 -0.000619
(0.00254) (0.00180) (0.00147) (0.00128)
WaterHeater -0.00253** -0.00143* -0.00121** -0.00113**
(0.00106) (0.000752) (0.000618) (0.000538)
Oven -0.00151* -0.000862 -0.000789* -0.000403
(0.000807) (0.000569) (0.000464) (0.000403)
Observations 1.265e+06 2.520e+06 3.744e+06 4.933e+06
* p < 0.10, ** p < 0.05, ***p < 0.01. Each variable in the column is the outcome for a reduced form RD, following specication
(1). Coecients measure the eect of being below the FA cuto. Standard errors are clustered at household level
36
Table 1.2: Balance Table (continued)
Bandwidths
2.5 5 7.5 10
Strata 0.00454 -3.81e-05 -0.00100 -0.00568***
(0.00402) (0.00284) (0.00233) (0.00202)
No Education -0.000700 0.000148 0.000269 0.000357
(0.00143) (0.00101) (0.000826) (0.000717)
Primary Educ. -0.000343 0.00422*** 0.00554*** 0.00698***
(0.00213) (0.00150) (0.00123) (0.00107)
Secondary Educ. 0.00127 -0.00322** -0.00491*** -0.00668***
(0.00204) (0.00144) (0.00118) (0.00103)
Technical Educ. 2.60e-07 -0.000411 -0.000391 -0.000345
(0.000482) (0.000340) (0.000278) (0.000241)
Bachelor -0.000293 -0.000678* -0.000442 -0.000271
(0.000543) (0.000384) (0.000315) (0.000274)
Graduate 6.28e-05 -5.93e-05 -5.99e-05 -4.32e-05
(0.000128) (8.98e-05) (7.34e-05) (6.36e-05)
Inactive 0.00125 0.00257** 0.00310*** 0.00346***
(0.00151) (0.00107) (0.000874) (0.000759)
Employed -0.00239 -0.00509*** -0.00509*** -0.00577***
(0.00171) (0.00121) (0.000990) (0.000859)
Unemployed 0.000953 0.000786 0.000539 0.000328
(0.000900) (0.000638) (0.000522) (0.000453)
Income -3,250 -10,957** -12,358*** -12,697***
(4,995) (4,698) (4,364) (4,275)
Observations 1.265e+06 2.520e+06 3.744e+06 4.933e+06
* p < 0.10, ** p < 0.05, ***p < 0.01. Each variable in the column is the outcome for a reduced form RD, following specication
(1). Coecients measure the eect of being below the FA cuto. Standard errors are clustered at household level
37
Table 1.3: Reduced Form Estimation 2015
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate 0.0012

0.0014

0.0010

0.0013

(0.0006) (0.0007) (0.0005) (0.0007)
Observations 10139565 10139565 10139565 10139565
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 2 1 2
Kernel Triangular Triangular Uniform Uniform
Bandwidth 8 12 7 10
E.No.Obs. 2764246 4047580 2571680 3540485
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.4: First Stage Results: Multiple Bandwidths
(1) (2) (3) (4)
Beneciary Beneciary Beneciary Beneciary
RD Estimate -0.1998

-0.2071

-0.2108

-0.2135

(0.0039) (0.0027) (0.0022) (0.0019)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Notes: This table reports the rst-stage RD specication of the eects of eligibility for the cash transfers on being a beneciary
of the program. The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility
for the transfers. Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
38
Table 1.5: Fuzzy RD Results: Multiple Bandwidths
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate -0.0103

-0.0077

-0.0059

-0.0047

(0.0053) (0.0036) (0.0029) (0.0025)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Notes: This table reports the fuzzy RD specication of the eects of the cash transfers on voter buying. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.6: Reduced Form RD: CCT on Turnout
(1) (2) (3) (4)
Turnout Turnout Turnout Turnout
RD Estimate 0.0010 0.0024 0.0013 0.0014
(0.0024) (0.0017) (0.0014) (0.0012)
Observations 10853567 10853567 10853567 10853567
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 905074 1802042 2677172 3529110
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter turnout.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
39
Table 1.7: Heterogeneity: Municipality's Population
(1) (2)
Above Median Below Median
RD Estimate 0.0021

-0.0015
(0.0008) (0.0019)
Observations 9051480 1088085
Household Clusters Yes Yes
Polynomial of Order 1 1
Kernel Triangular Triangular
Bandwidth 5 5
E.No.Obs. 1549543 233279
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The RD is estimated in the set of municipalities with population above and below the national median. The running variable
for each regression is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.8: Heterogeneity: Municipality's Election Closeness 2011
(1) (2)
Above Median Below Median
RD Estimate 0.0031

0.0003
(0.0010) (0.0010)
Observations 4912146 5227419
Household Clusters Yes Yes
Polynomial of Order 1 1
Kernel Triangular Triangular
Bandwidth 5 5
E.No.Obs. 866121 916701
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The RD is estimated in the set of municipalities with margins of victory between the winner and second most voted candidate
above and below the national median for the 2011 elections. The running variable for each regression is the normalized Sisben
multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are clustered at the household
level. * p < 0.10, ** p < 0.05, ***p < 0.01.
40
Table 1.9: Heterogeneity: Municipality's Number Voting Stations per Sq. Km
(1) (2)
Above Median Below Median
RD Estimate 0.0011 0.0031

(0.0008) (0.0014)
Observations 7595747 2543818
Household Clusters Yes Yes
Polynomial of Order 1 1
Kernel Triangular Triangular
Bandwidth 5 5
E.No.Obs. 1272317 510505
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The RD is estimated in the set of municipalities with a density of voting stations per square kilometer above and below the
national median. The running variable for each regression is the normalized Sisben multidimensional poverty score around the
cutos of eligibility for the transfers. Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p <
0.01.
Table 1.10: Heterogeneity: Neighboring Municipality's Number Voting Stations per Sq. Km
(1) (2)
Above Median Below Median
RD Estimate 0.0020

0.0013
(0.0009) (0.0011)
Observations 5482810 4637804
Household Clusters Yes Yes
Polynomial of Order 1 1
Kernel Triangular Triangular
Bandwidth 5 5
E.No.Obs. 933062 846053
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The RD is estimated in the set of municipalities for which its neighboring municipalities have a density of voting stations
per square kilometer above and below the national median. The running variable for each regression is the normalized Sisben
multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are clustered at the household
level. * p < 0.10, ** p < 0.05, ***p < 0.01.
41
Appendix
Table 1.11: Summary Statistics
Voter Bought Non-Bought Voter
Age in 2015 41.00 43.71
Household Members 5.04 5.07
Male 0.43 0.45
Education
None 0.07 0.08
Primary 0.43 0.42
Secondary 0.45 0.45
Technical 0.02 0.02
Bachelor 0.03 0.03
Graduate 0.00 0.00
Employment
No Activity 0.09 0.10
Work 0.50 0.50
Searching Job 0.05 0.05
Studying 0.04 0.04
Housekeeping 0.32 0.30
Rentist 0.00 0.00
Retired 0.01 0.01
Invalid 0.00 0.00
Income and Poverty
Score 37.47 39.10
Income 191426.82 202519.13
Observations 382289 9757276
42
Table 1.12: Transfer amounts for the FA program, 2015
Municipality Group Health($) Education($)
Kids Aged 0 7 Kindergarten Grade 1 5 Grade 6 8 Grade 9 10 Grade 11
1 63,525 0 0 26,475 31,775 47,650
2 63,525 21,175 10,600 26,475 31,775 47,650
3 63,525 21,175 15,900 31,775 37,050 52,950
4 74,100 21,175 15,900 37,050 42,350 58,225
All values are in Colombian pesos. One USD in 2015 corresponds to 2,743 Colombian pesos. Health amounts are given every
two months. Education transfers are given bimonthly during the school year (10 months). For more information on the transfer
amounts and their timing see Medell n and S anchez (2015), DPS (2013a) and (DPS, 2013b).
Table 1.13: Reduced Form RD Results: Multiple Bandwidths
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate 0.0023

0.0017

0.0013

0.0011

(0.0010) (0.0007) (0.0006) (0.0005)
Observations 10139565 10139565 10139565 10139565
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 895115 1782822 2654388 3500467
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.14: Reduced Form RD Results: Multiple Bandwidths and Controls
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate 0.0023

0.0017

0.0014

0.0012

(0.0010) (0.0007) (0.0006) (0.0005)
Observations 10139565 10139565 10139565 10139565
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 895115 1782822 2654388 3500467
Demogr. Controls Yes Yes Yes Yes
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on voter buying.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
43
Table 1.15: First Stage: Beneciaries of Familias en Accion
(1) (2) (3) (4)
Beneciary Beneciary Beneciary Beneciary
RD Estimate -0.2066

-0.2013

-0.2081

-0.2023

(0.0028) (0.0036) (0.0028) (0.0037)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 2 1 2
Kernel Triangular Triangular Uniform Uniform
Bandwidth 5 6 4 5
E.No.Obs. 1584576 1981626 1279015 1667544
Notes: This table reports the rst-stage RD specication of the eects of eligibility for the cash transfers on being a beneciary
of the program. The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility
for the transfers. Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.16: First Stage Results: Multiple Bandwidths and Controls
(1) (2) (3) (4)
Beneciary Beneciary Beneciary Beneciary
RD Estimate -0.2005

-0.2083

-0.2120

-0.2151

(0.0037) (0.0026) (0.0022) (0.0019)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Demogr. Controls Yes Yes Yes Yes
Notes: This table reports the rst-stage RD specication of the eects of eligibility for the cash transfers on being a beneciary
of the program. The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility
for the transfers. Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
44
Table 1.17: Fuzzy RD: CCT Beneciaries and Vote Buying
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate -0.0059

-0.0061

-0.0059

-0.0053
(0.0029) (0.0033) (0.0030) (0.0033)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 2 1 2
Kernel Triangular Triangular Uniform Uniform
Bandwidth 7 12 6 10
E.No.Obs. 2421589 3890442 1868534 3368053
Notes: This table reports the fuzzy RD specication of the eects of the cash transfers on voter buying. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.18: Fuzzy RD Results: Multiple Bandwidths and Controls
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate -0.0105

-0.0081

-0.0063

-0.0053

(0.0052) (0.0035) (0.0028) (0.0024)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Demogr. Controls Yes Yes Yes Yes
Notes: This table reports the fuzzy RD specication of the eects of the cash transfers on voter buying. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
45
Table 1.19: Reduced Form RD Results: Proof of valid residence as Outcome
(1) (2) (3) (4)
Valid Chage Valid Chage Valid Chage Valid Chage
RD Estimate 0.0063 0.0036 0.0026 0.0018
(0.0074) (0.0052) (0.0043) (0.0037)
Observations 382289 382289 382289 382289
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 33146 66201 98506 130850
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on whether
individuals present valid documentation indicating that they migrated and are not voters bought. The running variable is
the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.20: Reduced Form RD Results: Multiple Bandwidths after Proof of Residence
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate 0.0018

0.0014

0.0011

0.0009

(0.0010) (0.0007) (0.0005) (0.0005)
Observations 10139565 10139565 10139565 10139565
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 895115 1782822 2654388 3500467
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on an adjusted
variable of voter buying. The new variable individuals who present a valid proof of residence are recoded as non-bought-voters.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
46
Table 1.21: Reduced Form RD Results: Multiple Bandwidths and Controls after Proof of Residence
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate 0.0019

0.0014

0.0011

0.0010

(0.0009) (0.0007) (0.0005) (0.0005)
Observations 10139565 10139565 10139565 10139565
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 895115 1782822 2654388 3500467
Demogr. Controls Yes Yes Yes Yes
Notes: This table reports the reduced form RD specication of the eects of eligibility for the cash transfers on an adjusted
variable of voter buying. The new variable individuals who present a valid proof of residence are recoded as non-bought-voters.
The running variable is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers.
Standard errors are clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.22: Fuzzy RD Results: Multiple Bandwidths after Proof of Residence
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate -0.0081 -0.0065

-0.0050

-0.0041

(0.0050) (0.0034) (0.0027) (0.0023)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Notes: This table reports the fuzzy RD specication for the eects of cash transfers on an adjusted variable of voter buying.
The new variable individuals who present a valid proof of residence are recoded as non-bought-voters. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
47
Table 1.23: Fuzzy RD Results: Multiple Bandwidths and Controls after Proof of Residence
(1) (2) (3) (4)
Voter Bought Voter Bought Voter Bought Voter Bought
RD Estimate -0.0083

-0.0068

-0.0054

-0.0046

(0.0050) (0.0034) (0.0027) (0.0023)
Observations 9572161 9572161 9572161 9572161
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 822490 1639144 2443829 3227179
Demogr. Controls Yes Yes Yes Yes
Notes: This table reports the fuzzy RD specication for the eects of cash transfers on an adjusted variable of voter buying.
The new variable individuals who present a valid proof of residence are recoded as non-bought-voters. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
Table 1.24: Fuzzy RD Results CCT and Turnout: Multiple Bandwidths
(1) (2) (3) (4)
Turnout Turnout Turnout Turnout
RD Estimate -0.0040 -0.0112 -0.0058 -0.0052
(0.0150) (0.0102) (0.0082) (0.0070)
Observations 6770676 6770676 6770676 6770676
Household Clusters Yes Yes Yes Yes
Polynomial of Order 1 1 1 1
Kernel Triangular Triangular Triangular Triangular
Bandwidth 2.5 5 7.5 10
E.No.Obs. 586757 1169096 1741949 2301374
Notes: This table reports the fuzzy RD specication of the eects the cash transfers on voter turnout. The running variable
is the normalized Sisben multidimensional poverty score around the cutos of eligibility for the transfers. Standard errors are
clustered at the household level. * p < 0.10, ** p < 0.05, ***p < 0.01.
48
Figure 1.8: Example of Voter Turnout Records
This is an example of the electoral form E-11. It reports the ID number of every potential voter in a voting station. The typed
numbers are the ID numbers for each registered citizen. If the spaces in the contiguous columns to a citizen's ID are left blank,
it means the citizen did not vote. If there is a signature next to a typed number it means the person turned out to vote.
49
Chapter 2
Campaign Spending Limits and Irregular Political Behavior
1 Introduction
During the past fty years we have seen a major expansion of democracy around the world: only
31.5% of countries were democratic in 1960, while in 2010 the number increased to 64.1% (Acemoglu
et al., 2019). Therefore, analyzing the factors that make democracies function is of great relevance.
In particular, political campaigns are one of the hallmarks of modern democracies and the regulation
on their expenditures and nancing is a topic of heated debate.
On one hand, higher campaign expenditures can signal the electorate which candidates have
more skills and capabilities to govern, because politicians were able to collect money and attract
contributors in the rst place. On the other hand, the involvement of money in politics also has
a negative connotation, as the process of selecting politicians according to their access to nancial
resources can be seen as inherently antidemocratic. This perception can be justied by candidates
potentially being more interested in satisfying the needs of their contributors rather than those of
the general public. This is why an increasing number of developed and developing countries have
established regulations for limiting campaign spending (Speck, 2013).
This paper studies the causal eect of campaign spending limits on violence, vote buying, crime
and the performance of politicians once they get elected. Although many countries implement some
sort of campaign spending limit regulation, there is still not much evidence on the causal eects this
policy has on irregular political behavior. The lack of causal evidence is mainly explained because
campaign spending limits are usually not xed exogenously.
I overcome this issue by using a Regression Discontinuity design, in which I exploit the fact
that, in Colombia, candidates running for mayors for a municipality are subject to dierent cam-
50
paign spending limits, which are dened according to the number of potential voters within the
municipality. In particular, candidates running for municipalities with less than 25,000 potential
voters are allowed to spend up to a certain cap, those between 25,001 and 50,000 potential voters
are allowed to spend up to a higher limit, and so on.
Given that the spending limits jump in the arbitrarily dened cutos and that no other gov-
ernment program or policy varies according to these thresholds, I can estimate the causal eect of
having higher campaign spending limits by comparing the municipalities with a number of potential
voters just above the threshold with those just below.
The main outcomes of interest are the eects of campaign spending limits on violence, crime
and vote buying. Limiting the legal contributions candidates are allowed to receive can enhance the
generation of links between politicians and criminal groups. When candidates' formal contributions
are more constrained, they have bigger incentives to associate with criminal groups, which can
provide illegal funds and coerce the opposition. Moreover, when the campaign spending limits are
low, it may be more attractive for criminal groups to lobby, given that they don't have to compete
for the candidate's support with more powerful lobbyists. This outcome has not been analyzed in
the literature and is a relevant aspect to consider, especially for weak democracies where the state
and criminal groups coexist, as in the case of Colombia. Moreover, the spending limits can also
generate incentives for politicians to substitute legal means to reach the electorate with illegal ones,
such as buying votes.
Regarding violence, the municipalities constrained with lower campaign spending limits show
higher levels of forced displacements, homicides and kidnappings executed by an armed group.
Although the signicance for homicides and kidnappings varies with dierent bandwidth choices,
the increase in displacements is relatively robust. In my preferred specication, when a municipality
is just below the potential voter cuto, the level of forced displacement is 88% higher than average
of a municipality just above. Interestingly, there are no statistically signicant dierences across
the cutos in other general variables of crime not necessarily associated to criminal groups' actions.
Overall, these results suggest that there is more participation of organized criminal groups when
politicians are subject to lower campaign spending limits.
I also test for the eects of campaign spending limits on political competition and elected
candidates' performance. The main objective of limiting campaign spending usually is to promote
51
the entry of more candidates and to foster competition in elections. It is also of great interest to
evaluate if the campaign spending regulation, by aecting the electoral competition, also aects
who wins the elections and how the elected politician performs while being in oce.
This paper contributes to several areas of research. First, this study is related to the papers
that analyze the causal eect of campaign spending limits on political entry and competition.
Avis et al. (2019) show that lower campaign spending limits lead to the participation of more
candidates running for mayor in Brazil's municipalities. Moreover, they nd that elected mayors
in higher campaign spending municipalities are able to get more funding for their government's
budget. Their paper identies a causal eect of lower campaign spending through a regression
discontinuity: the authors take advantage of a change in Brazil's regulation, which generated a
jump across municipalities in the campaign spending limit mayoral candidates were allowed to
spend.
The most closely related paper to mine is Ruiz (2018), who shows that Colombian donor-funded
mayoral candidates assign more contracts to the rms that contributed to their campaigns with
respect to non donor-funded candidates. Ruiz relies on several approaches to test this, among
which he exploits the variation I use in the cutos of campaign spending limits, although he only
focuses on one election in 2011. Even though it is not the core of his paper, Ruiz also nds that the
spending limits don't have a signicant eect on the number of candidates running for elections in
2011. In contrast, my study uses a bigger sample, since I pool for several elections, and I focus on
the possible links between politics, money and organized violence.
In other papers, such as Fouirnaies (2018) and Milligan and Rekkas (2008), the authors study
changes in the parameters of the spending limit formulas used in Great Britain and Canada, re-
spectively, where the limits are dened according to the number of electors and other city charac-
teristics. In both articles, the evidence shows that higher spending limits are associated to lower
candidate participation and less electoral competition. Nevertheless, their methodology implies
that the changes in the parameters' values aren't endogenous to electoral competition or other
political phenomenons, which seems unlikely.
I also contribute to the literature on the link between criminals and politicians. The cohabitation
and collaboration between democratic institutions and criminal groups challenges the traditional
Max Weber's argument, in which the foundation of the State is identied with its monopoly of
52
legitimate violence (Weber, 1946). For example, Acemoglu et al. (2013) show that there was a
symbiosis between paramilitaries and politicians in Colombia, where the rst guaranteed the ascent
of their preferred candidates in exchange of future benets, such as better terms with the justice
in the process of disarmament.
The association between criminals and politicians is not only limited to the Colombian case.
Alesina et al. (2019) suggest that the Italian Maa and other local criminal groups used violence
to protect their interests and support their preferred candidates. Moreover, the timing and mag-
nitude of violence was used strategically depending on how competed the elections were in each
municipality. In my paper I show that restricting the access of legal contributions for the electoral
campaigns can foster the link between politicians and criminal groups. In particular, and unlike
most of the literature, my study suggests that politicians are not exogenously honest or dishonest,
but their association with crime is an endogenous outcome that can respond to regulatory and
economic conditions.
Vaishnav (2017) studies the permeation of crime into politics in India, where around one quarter
of the members of Parliament face pending criminal investigations. Vaishnav proposes several
mechanisms that explain this feature, arguing that criminality can signal to the electorate the
politician's skill of \getting things done" and allows candidates to exercise coercion against their
opponents. Additionally, according to the author, the lack of information from voters can also
explain the success of criminals in politics. George et al. (2018) explores this last mechanism by
using a randomized experiment in which people receive information of the candidates' criminal
record by a mobile phone text message. The authors nd that the electorate votes less for criminal
candidates when they receive the information of the candidates' criminal record. Additionally,
Prakash et al. (2019) nd that elected politicians with pending criminal investigations have a
negative causal eect on their constituency's economic performance.
Following this Introduction, Section 2 discusses the context of Colombian elections and how
crime and politics have cohabited in the country, Section 3 describes the empirical strategy, Section
4 presents the data, Section 5 approaches the threats to my identication strategy, Section 6 shows
the results and Section 7 concludes.
53
2 Context
2.1 Elections in Colombia and Campaign Spending
Since 1988 each of Colombia's 1101 municipalities elect its own mayor through democratic elections.
After the year 2003, elections take place every four years, while before that they were held every
two years (1988-1994) and every three years (1994-2003). Candidates for mayoral elections can be
aliated to a political party, or can participate independently. Mayors have an important amount
of power over the municipality's resources, although transfers from the National Government are
usually tied to specic expenditures, mayors have discretion on around 20% of the total municipal
budget (Mart nez, 2016).
From 1994 it was determined by law (Law 30 of 1994), that private contributions to political
campaigns couldn't surpass certain limits, which are determined by the National Electorate Com-
mission (CNE, by its initials in Spanish). The law also included the guidelines for greater electoral
and political transparency, by creating a regulation on the reports of campaign expenditures and
their sources of income.
Therefore, since 1994, the private contributions for a mayoral electoral campaign are limited
according to the number of potential voters located in the municipality where the candidate is
seeking oce. A potential voter is dened as a Colombian citizen aged 18 or older, who has a
registered ID to vote. It's important to emphasize that the variation in the campaign limits comes
from the number of potential voters and not from the total population of a municipality. Most of
the Government's transfers are assigned to each municipality according to its total population, not
on its potential voter count. This will add strength to the empirical strategy shown below; to the
best of my knowledge, the only policy that varies according to the potential voter limits dened by
the CNE corresponds to the campaign spending limits.
Table 2.1 shows the spending limits in Colombian Pesos established for the mayoral elections
held since 2003 (2003, 2007, 2011, 2015). It is worth noting that 1 US Dollar was worth 2.875
Colombian pesos in 2003. This means that the campaign spending limit for municipalities with less
than 25.000 potential voters in 2003 was around USD 12.018; while the limit for the municipalities
with between 25.001 and 50.000 potential voters was close to USD 23.280. As it can be seen from
54
the table, the relative dierences in spending limits across thresholds are kept through all the
elections held since 2003.
Although the spending limit regulation started in 1994, I will focus only on the elections held
since 2003. The main reason for doing this is explained by the fact that previous to this year there
was no limit on the 25.000 potential voter cuto. This meant that the lowest cuto was xed
in 50.000 potential voters. The distribution of municipalities in Colombia is mostly concentrated
below the 25.000 electoral potential, as shown in Figure 2.1. Consequently, an analysis exploiting
the discontinuity in spending limits across thresholds previous to 2003 (as will be explained in
the empirical strategy), would be limited by the low quantity of observations available, as too few
municipalities had more than 25.000 potential voters. Furthermore, there is more data availability
for the outcomes analyzed since the start of the XXI century, making even more convenient to focus
on the period after 2003.
2.2 Politics and Crime in Colombia
The Law 30 of 1994 was designed with three main objectives: 1) Increasing transparency in elections;
2) Generating more pluralism in the political landscape; and 3) To avoid the interference of illegal
actors and criminal groups into politics (MOE, 2016a). With respect to this last point, it is worth
noting that during the late 1980s and 1990s the country's politics were corrupted by the inclusion
of dirty money, coming from the booming nacotrac business. At the time, Colombia faced several
scandals in which politicians and their electoral campaigns were nanced with the money of drug
cartels. The level of permeation of illegality into politics reached its highest point during the
presidential campaign of 1994, in which the elected president, Ernesto Samper, was accused of
receiving resources from the Cali Cartel, one of the most powerful criminal groups at the time, in
what is known as the \Proceso 8000".
The good intentions of the 1994 law were not enough for obstructing the interference of illegal
groups into politics. Starting from 2006, the media revealed major links between various politicians
and paramilitary groups in the country, which was known as the \parapol tica" scandal. The
paramilitaries were illegal groups of right-wing militias, whose main objective was to ght the
left-winged guerrilla groups that proliferated in the country. As most of the other major criminal
groups in the country, the paramilitaries nanced their activities through narcotrac, extortion
55
and kidnapping.
One of the most striking revelations of the \parapol tica" scandal consisted in a number of
regional pacts of collaboration between paramilitary leaders, local politicians and businessmen.
The most important pact, called \pacto de Ralito", signed by paramilitaries, mayors, governors,
senators and other important political gures, was written with the objective of \refounding our
nation, to sign a new social contract"
1
. After years of investigations, almost 100 mayors, 30
governors and more than 100 members of Colombia's Congress had faced investigations for being
connected to the paramilitaries (CNMH, 2018).
Although the link between politicians and criminals (cartels, paramilitaries, among others)
doesn't have an explicit foundation on the campaign spending limits, it is interesting to analyze
how these limits generate incentives on politicians to establish relations with criminals for reaching
power. As stated by `Ernesto B aez', the political commander of one paramilitary group, in his
declarations to the Supreme Court: \for power, politicians used to get alliances even with the
devil"
2
.
3 Empirical Strategy
Estimating the eects of campaign spending is a challenging task: politicians may decide to spend
more on their campaigns in municipalities where turnout is low, or in places where they have
chances to win but are faced with a competitive candidate. Therefore, an OLS estimation (for
example of electoral outcomes on campaign spending) would not be appropriate due to issues of
reverse causality, given that campaign spending would not be exogenous.
Furthermore, analyzing the impacts of campaign spending limits is also a daunting enterprise,
as countries tend to assign the limits using formulas that vary continuously on the population and
other variables (as in the cases of Canada (Milligan and Rekkas, 2008) and the UK (Fouirnaies,
2018)). This makes the determination of the spending limits endogenous to other important regional
characteristics that may have an eect on electoral outcomes.
The ideal method for studying the eects of campaign spending limits would be to randomize
across municipalities which ones get treated (bounded by a spending limit) and which ones are
1
The full text can be found at: https://www.semana.com/on-line/articulo/texto-del-acuerdo-ralito/83002-3
2
https://verdadabierta.com/la-historia-detras-del-del-pacto-de-ralito/
56
controls (without spending limit). On top of that, for analyzing dierent levels of campaign spend-
ing, one would also like to assign dierent (binding) spending limits to the treated municipalities
and examine the eects of more or less spending. Because this ideal experiment is impossible to
develop, due to administrative and ethical reasons, the best alternative is to use an exogenous
variation in the determination of campaign spending limits, exploring the eects of these between
the municipalities that are marginally subject to dierent spending limits.
The campaign spending regulation from Colombia oers a good setting for studying the general
eects of campaign spending limits. As it was mentioned in the Context, the spending limits vary
according to arbitrarily set cutos that depend on each municipality's total number of potential
voters. Table 2.1 shows the spending limits and their variation through the years. For example,
all the municipalities with less that 25.000 potential voters have the same spending limit, while
those between 25.001 and 50.000 potential voters have a campaign spending limit almost twice as
big. The municipalities that have between 50.001 and 100.000 potential voters have a campaign
spending limit almost three times as big as those between 25.001 and 50.000, and so on.
These big jumps in the spending limits across the arbitrarily dened potential voter thresholds
oer a natural setting for analyzing the eects of campaign spending limits. Using the exogenous
variation in the cutos that identify which municipalities get dierent campaign spending limits, I
will compare municipalities marginally below the potential voter threshold with those marginally
above to estimate the impact of the campaign spending limits on several outcomes. As shown in
Figure 2.1, most municipalities are concentrated below 100.000 potential voters; thus, I will focus
on the rst two thresholds (25.000 and 50.000 potential voters). I will use the following regression
discontinuity design
3
:
y
mt
=
0
+
1
below
mt
+
2
X
d=1

d
D
mtd
+
2
X
d=1

d
D
mtd
f
d
(dist
mtd
)
+
2
X
d=1

d
below
mt
D
mtd
f
d
(dist
mtd
) +
t
+
mt
(2.1)
Wherey
mt
will be my outcome of interest at municipalitym in yeart. Unless otherwise specied,
3
Dell and Querub n (2018) use a similar specication of a pooled RD.
57
the years considered will correspond with municipal election years, namely 2003, 2007, 2011 and
2015. below
mt
is an indicator variable equal to 1 if the municipality is below to its nearest potential
voter threshold in year t. D
mtd
is an indicator variable equal to 1 if threshold d is the closest
threshold to municipality m in year t; this allows me to compare municipalities just below the
25.000 threshold with those just above it and municipalities just below the 50.000 threshold to
those just above it. f
d
(dist
mtd
) is a polynomial that varies with the distance to the nearest voter
threshold.
t
is a set of year xed eects. The coecient of interest is, therefore,
1
.
For the baseline estimates I use the optimal bandwidths as calculated by Calonico et al. (2018),
uniform kernels and local linear regressions across the thresholds. The optimal bandwidths are a
function of the data, thus they will vary depending on the outcome variable. The standard errors
are clustered by municipality, as the municipalities located in the selected bandwidths across years
tend to persist through the years. Nonetheless, the results don't change signicantly when not
clustering at the municipality level.
In the baseline graphs I center the cutos around zero to have a simple pictorial representation
of the pooled estimations of equation 2.1. In the graphs, I partial out the cuto and year xed
eects, so I'm comparing for the same year only the municipalities below the threshold of 25.000
with those above the same threshold, and those below the cuto of 50.000 with those above.
4 Data
4.1 Municipal Outcomes: Crime, Violence, Finances and Service Provision
The main variables I will focus on related to crime correspond to the total number of homicides
and theft (which includes theft of residences, vehicles, robberies against people, etc.) in each
municipality in each year. Both variables are calculated since 2003 by Colombia's Ministry of
Defense, and I extracted them from the Municipal Panel constructed by the Center of Studies on
Economic Development (CEDE).
Theft and homicides, nonetheless, don't represent appropriately the features of organized crime
and may fail to capture the interference of organized criminal groups and their links with politics
and politicians. Because of this, I also include variables of number of homicides, kidnappings and
displaced people exclusively executed by any armed group since 2003 for each municipality. The
58
data was extracted from the Unidad de V ctimas (UV), which is the agency in charge of managing
the programs of assistance to Colombia's victims of the armed con
ict. Interestingly, there are
several inconsistencies between the data withdrawn by the Ministry of Defense and the UV, so the
results should be interpreted with caution.
4
Finally, for the variables of each municipality's nances and their service provision I also use
the CEDE's Municipal Panel information. The municipality's nances are originally calculated by
the National Planning Department and cover the full period of my analysis. The service provision
variables I got access to, though, don't have information for all my period of analysis. For the
variables of aqueduct, sewer, gas and trash coverage services, I use information gathered from 2008
to 2016; all of these are originally calculated by the Superintendence of Public Utilities.
I also consider education variables to capture dierences in how mayors may in
uence the service
provision in each municipality. I use the number of schools, number of teachers, and an average
score of each municipality's standardized tests, which students take during their last year of high
school. These are all also included in CEDE's Municipal Panel. The education variables used cover
the period from 2004 to 2013.
4.2 Electoral and Political Variables
Electoral outcomes for all municipal elections were taken from the most updated version of the
database compiled by Pach on and S anchez (2014a). The data's main source is Colombia's national
electoral authority, the Registradur a Nacional de la Naci on. The variables included correspond to
the names and party aliation of each candidate for each municipality for all mayoral elections,
and their vote shares. This allows me to calculate the number of candidates running for elections
in each municipality in every year.
The data in Pach on and S anchez (2014a) doesn't include the electoral census, namely the
number of potential voters, of each municipality. This is my running variable, which is crucial for my
analysis since its variations determine the level of campaign spending allowed in each municipality.
Therefore, I directly obtained the number of potential voters for each municipality for each mayoral
4
One could think that there is an over-report of crimes by the victims of Colombia's con
ict who may benet
from subsidies or support programs. There could be also an under-report of violence from the Ministry of Defense.
Nonetheless, for the purposes of this paper, there shouldn't be a systematic variation in either variable across the
potential voter cutos.
59
election year from Colombia's national electoral authority
5
.
The number of potential voters also allows me to calculate turnout for each election and each
municipality. Furthermore, I dene a measure of eective political concentration for each munici-
pality by calculating a Herndahl Index of concentration of vote shares:
PoliticalConcentration
mt
=
X
i2Imt
v
2
mti
(2.2)
Where i is a mayoral candidate and I
mt
is the set of all candidates running for mayor at
municipality m in year t. v
mti
is the vote share of each of these candidates. Unlike the plain
measure of the number of candidates, this index captures if the electorate is eectively voting more
for more candidates; if its value increases, then there is less competition and more concentration.
4.3 Vote Buying Data
Voter buying, or \Electoral Transhumance", as is locally known, is one of the most pervasive forms
of electoral fraud experienced in Colombia. Under voter buying, individuals illegally register and
vote in a dierent municipality from where they live, in exchange of cash or in-kind transfers paid
by a politician or electoral candidate. This practice of corruption is illegal and punishable with
prison according to Colombia's penal code. The illegality of voter buying is explained by the fact
that citizens who don't live in a certain municipality may in
uence the electoral outcomes of an
election in which they shouldn't have a legitimate interest
6
. The \importation" of outsiders to vote
in a municipality is often organized by a local political organization, or \machinery", which is in
charge of targeting and mobilizing voters from one location to the other (Duque Daza, 2019).
News reports from Colombia indicate how the voter buying phenomenon is related to corruption
and clientelism. In most cases, people receive two transfers: one once they register their vote
in a dierent municipality from their location of residence, and a second transfer at the day of
5
I would like to especially thank William Enrique Mu~ noz, who was in charge of the Registradur a's library and
very kindly helped me accessing the data.
6
Initially, the law 599 of 2000 punished only the brokers who organized the voters to register in a dierent
municipality from their place of residence. Committing this crime could lead to 4 to 9 years of prison and a monetary
ne. In 2017, the punishment was extended also to those individuals who register in a dierent municipality from
their place of residence. In this case citizens would be committing a \false testimony" crime and they can be punished
with 6 to 12 years of prison, according to Article 442 of Colombia's penal code.
60
the elections. The transportation costs, snacks, food and other expenditures are covered by the
corrupt political candidate. The qualitative evidence suggests that people who engaged in voter
buying received from 70,000 pesos (close to 25 US dollars in 2015, and equivalent to around 10% of
Colombia's monthly minimum wage at the time) to 450,000 pesos. People also claim to have received
groceries, help from the candidates to formalize their land ownership, bus tickets to celebrate a
town's carnival, cattle, among many other goods.
The magnitude of the presumed voters bought during 2015 was close to 1.5 million individuals
of the 34 million of people registered to vote.
Colombia's national electoral authorities, the Consejo Nacional Electoral (CNE) and the Na-
tional Registrar's Oce Registradur a Nacional del Estado Civil, are in charge of preventing the
fraudulent inscription of IDs on voting locations that dier from each citizen's municipality of res-
idence. The CNE cancels a citizen's change in their voting location if the new desired municipality
doesn't coincide with other administrative information that reports the actual residence of each
individual. If the CNE cancels a citizen's voting registration in a new voting location, the individual
can still vote in the previous municipality where they were previously registered to vote.
The administrative databases used by the CNE to prevent fraudulent changes in voting sites
are: the social security database, which includes information of all the beneciaries of the general
health care system; the national agency for overcoming extreme poverty's data (ANSPE); the
system of identication of potential beneciaries of social programs (SISBEN); the Victims' Unit,
which manages programs of reparation for the victims of Colombia's con
ict; electoral censuses
from previous elections; among other data. Furthermore, the CNE also organized visits in certain
municipalities to verify the location of residence of a subset of individuals who changed their voting
location.
The CNE, after merging this information, creates a list, for each municipality, with the citizens'
IDs for the people who registered to vote in a new location but whose residence data doesn't match
their new desired voting site. The data on voter buying used in this paper are the totals {aggregated
by municipality of destination{ of voters whose change in voting location was denied by the CNE
for the regional elections of 2011, 2015 and 2019. I digitized this information from the formal legal
decrees produced by the CNE at the time of the election.
61
5 Threats to Identication
The identication strategy used in this paper relies on several assumptions. Probably the most
important ones are: 1) the potential voter cutos are arbitrarily dened and 2) the potential
voter count is not manipulated with the objective of getting a municipality to get lower or higher
campaign spending limits. To evaluate the previous concerns, I test if there is a discontinuity in the
number of potential voters across the thresholds by using the McCrary (2008) test and a similar
test developed by Cattaneo et al. (2019).
Figure 2.2 and Figure 2.3 show the McCrary (2008) and the Cattaneo et al. (2019) tests,
respectively, for the distribution of potential voters across the 25.000 cuto. Figures 2.4 and 2.5
show the McCrary (2008) and the Cattaneo et al. (2019) tests, respectively, across the 50.000 cuto.
As it can be seen for both cutos and both tests, there seems to be no evidence of manipulation
on the number of potential voters across the thresholds.
The Cattaneo et al. (2019) tests yield a robust bias-adjusted p-value of 0.66 and 0.43 for
distributional dierences across the 25.000 and 50.000 potential voter cutos, establishing further
evidence of no manipulation. Despite the statistical evidence, around electoral times is common
the crime denominated \trashumancia", in which voters are registered in dierent municipalities
from those they live in, probably in support for a given candidate. I should analyze the importance
of this phenomenon and its eects on the overall manipulation across the thresholds, although the
previous tests show that there is not a concentration of this crime around the cutos.
Another concern that may arise with the empirical strategy I use is that there could be im-
balances in pre-determined covariates across the cutos. Although the previously shown tests of
no manipulation indicate that the potential voter thresholds are arbitrary, it is worth exploring if
there are dierences across pre-determined covariates across the thresholds. Table 2.2 shows the
results of estimating equation 2.1 on the following pre-determined outcomes: a dummy variable
indicating the presence of indigenous population in the XVIth century; the average altitude of the
municipality; the distance of each municipality to its department's capital
7
; and the distance to
Bogot a, Colombia's capital. All of these variables were taken from the CEDE's Municipal Panel.
The variables chosen should not have, on average, signicant variations across the cutos, as
7
A department is one of Colombia's political subdivisions and it is conformed by a group of municipalities.
62
they are completely pre-determined with respect to the application of the spending limit legisla-
tion. The results show that this is indeed the case. Not having signicant dierences in historical
indigenous presence is of interest, as this variable's association with the current institutional devel-
opment and economic performance might be relevant (Acemoglu et al., 2002). Moreover, the fact
that there is not a statistically signicant dierence in the distance to Bogot a or to the municipal-
ity's department capital is also worth emphasizing, as closer distances to bigger cities may have an
eect on economic prosperity, among other outcomes. Figure 2.6 displays graphically the results,
showing no discontinuity in the pre-determined variables.
6 Results
6.1 Crime and Violence
In the Context section, it was mentioned how organized criminals permeated into Colombia's pol-
itics. In this section I explore if there is a causal link between the campaign spending limits and
higher levels of organized crime in a municipality. The evidence shown in this section indicates
that lower campaign spending limits increase the number of forced displacement and increase the
number of kidnappings by an armed group. These results suggest that in lower campaign spending
limit municipalities criminal groups are more active, which could be a sign that there is a closer tie
between crime and politics in the municipalities just below the campaign spending cutos than in
those just above.
Table 2.5 shows the results of estimating 2.1 on the total number of homicides, calculated by
the Ministry of Defense, theft and homicides per capita. The evidence indicates that there is not
a statistically signicant dierence in crime outcomes between municipalities above and below the
campaign spending limit cutos. This is not surprising, as the total number of homicides and theft
do not necessarily re
ect with precision the operation of organized groups. Figure 2.9 illustrates
graphically the lack of variation in crime outcomes around the cutos.
To investigate the relation and intervention of armed groups and organized crime in politics and
their interaction with campaign spending, I use the variables extracted from the UV, which report
forced displacements, homicides and kidnappings developed systematically by one of Colombia's
groups involved in con
ict. These variables re
ect in a more precise way the presence of armed
63
groups and their involvement in elections. Table 2.6 and Figure 2.10 report the results on these
outcomes.
The results indicate that there is an increase in all the variables related to con
ict when a
municipality is just below the campaign spending limit cuto, suggesting that when the spending
limit is more binding, there is more organized violence. In particular, forced displacement is
88% higher when a municipality is below the cuto with respect to those that are just above the
cuto, and the eect is signicant at the 5% level. Moreover, kidnappings are also 63% higher in
municipalities just below the cutos; the result is signicant at the 10% level. Finally, homicides
related to con
ict are also higher in municipalities with more binding campaign spending limits,
and are close to signicance.
The evidence should be interpreted with caution. It is important to emphasize the hypothesis
that the only thing that varies at the cutos of potential voters is the level of campaign spending.
If this is true, the only factors that can explain why violence is higher in municipalities just below
the thresholds must be related to the levels of campaign spending.
One hypothesis to explain this is that as candidates get more constraints on their campaign
spending, their incentives to use other mechanisms to reach the electorate increase, even those
that involve engaging with criminal groups. In particular, getting alliances with armed groups,
among other illegal strategies, may be also worthwhile approaches for signaling to the voters that
a candidate has skills for \getting things done" and accomplishing their purposes, as Vaishnav
(2017) points out for India. Furthermore, using coercion guarantees a lower participation for the
opposition's supporters. Hence, when candidates are forced to have lower (legal) spending on their
campaigns, there may be a substitution eect in which they use other illegal strategies to get
elected. Additionally, when the campaign spending limits are low, it may be more attractive for
criminal groups to lobby, given that they don't have to compete for the candidate's support with
more powerful lobbyists.
To add strength to my previous results, Table 2.7 and Figure 2.11 display the estimates of
equation 2.1 but considering only non-election years and using the election-year number of potential
voters. If organized violence and the association with armed groups is used by candidates more
constrained by the campaign spending limits, then, there should be a reduced or insignicant
outcomes in my variables of violence when there is not a mayoral election. As suggested, my results
64
indicate that displacements and kidnappings are insignicant for non-electoral years, although
con
ict-related homicides are signicant at the 10% level in municipalities below the cutos, namely
those that have more binding campaign spending limits. Again, this is only suggestive, as I can't
reject that the coecients between electoral and non-electoral years are statistically dierent.
6.2 Electoral Outcomes and Vote Buying
In this section I estimate the causal eects of higher campaign spending limits on political com-
petition, turnout and vote buying. As mentioned in the Data section, I have two measures of
competition: a plain measure of the number of candidates running for mayor at each municipality
for each election and a Herndahl Index of eective concentration of votes.
The graphical evidence for my baseline estimation strategy is shown in Figure 2.7. The im-
age shows that, on average, the municipalities that are marginally below the cutos of campaign
spending limits have more candidates and more eective political competition than the municipal-
ities just above the thresholds. This result is in line with what Avis et al. (2019) nd for Brazil's
campaign spending legislation. Particularly, their theoretical model implies that when there are
higher campaign spending caps, a candidate must incur in more expenditures to attain the same
vote share, consequently disincentivizing a candidate's entry. Nevertheless, Table 2.3 shows that
the results are not signicant at conventional levels.
One possible implication of restricting campaign spending can be that voters are less informed
about candidates and elections. Therefore, one could think that municipalities with lower campaign
spending limits have lower turnout. The results suggest that this is not the case, as there is not a
statistically signicant dierence between municipalities above and below the cutos; if anything,
Table 2.3 indicates that lower spending is associated with slightly higher turnout.
On the other hand, Table 2.4 suggests that municipalities with tighter campaign spending limits
have higher levels of vote buying. The variable of vote buying used is the number of imported voters
from other municipalities known as voter buying. The results are robust to dierent specications
and to the transformation of the outcome variable to the hyperbolic sine. The results suggest that
municipalities with tighter campaign spending limits have, on average, 60% more voters-bought
that municipalities with looser limits. The results are complemented in Figure 2.8.
65
6.3 Service Provision and Municipality's Finances
One question that naturally arises is if the elected politicians dier in their policies once they get
to oce, given that they were subject to dierent campaign spending limits. The implications on
this aspect are ambiguous. Raising funds for campaigning can be used to reach uninformed voters
and it can also be interpreted by the electorate as a signal of a candidate's skills, if true then
limiting the spending limits may benet unskilled politicians. On the other hand, allowing higher
campaign spending can also reshape the candidates' policies in favor of special interest groups who
lobby, as shown in Ruiz (2018)
8
. Hence, the implications on welfare of the campaign spending
limits are uncertain, much more when the limits seem to enhance a link between criminal groups
and politicians, as shown in the previous section.
For measuring dierences in elected mayor's performance I estimate the baseline equation using
the outcomes in non-electoral years to capture the actual performance of the elected candidates,
and keeping constant the number of potential voters, to maintain the treatment condition of each
municipality according to the previous election. Each mayor's performance is captured by using:
1) variables of municipality's nances, namely the total income (coming from taxes, transfers, etc.)
and total expenditures (in public employees, public investment, etc.); 2) the coverage in aqueduct,
sewer, gas and trash; 3) education variables such as the total number of teachers in a municipality,
the average scores in the standardized test scores and the number of schools.
Table 2.8 and its corresponding graph (Figure 2.12), show that the municipalities just below
the cutos have higher income and higher expenditures (measured in millions of Colombian pesos)
at the 5% level of signicance. Moreover, the magnitude of this dierence is relevant, as income is
around 27% and expenditures are more than 30% higher than the average municipality above the
cuto. Nonetheless, even if these municipalities have higher income and expenditures, this fact is
not re
ected in a signicant increase in the service coverage variables nor in education variables
(Tables 2.9 and 2.10, and Figures 2.13 and 2.14).
8
An extensive theoretical literature has focused also on the aspects of lobbying, elections and welfare, for example:
(Baron, 1994), (Prat, 2002), (Coate, 2004)
66
6.4 Robustness
Tables 2.11 through 2.16 included in the Appendix show the estimation of equation 2.1 for band-
width choices of 3.000, 5.000 and 7.000 potential voters around the cutos.
The tables show that the results on general crime remain insignicant at conventional levels
for all the bandwidths considered. For the variables related to criminal group violence the results
indicate that forced displacements are consistently signicant for any bandwidth's choice, while the
signicance of homicides and kidnappings related to con
ict varies and tends to fade out as the
bandwidth gets wider, which is also natural, given that other confounding variables take place as
I expand the universe of municipalities involved in the RD with wider bandwidths.
7 Conclusion
In this paper I have shown the implications of campaign spending limits on electoral competition,
crime, organized violence and service provision. To achieve this, I use a Regression Discontinuity
design, by exploiting the fact that in Colombia the candidates' campaign spending from each
municipality is limited according to arbitrary potential voter cutos. This regulation implies that
the municipalities that have a higher number of potential voters than the cuto are allowed to spend
more than those below the threshold. I compare the variables of interest of those municipalities
that are just above and just below these arbitrary cutos.
I show that municipalities with lower campaign spending limits have more cases of organized
violence. In particular, forced displacement, which was one of the main forms of brutality in
Colombia's armed con
ict, is signicantly higher in municipalities with lower campaign spending
limits. Homicides and kidnappings associated to an armed group are also higher when campaign
spending is lower, although their signicance varies depending on the bandwidth's choice.
My main hypothesis with respect to this result, in addition to the qualitative evidence from
Colombia and the related literature from other countries, indicates that the link between politics
and crime is enhanced when there are low campaign spending limits because candidates overcome
the legal limitations to get into power with informal or illegal means. Furthermore, tighter spending
limits can lead criminal groups to capture and eectively nance politicians.
67
Figures
Figure 2.1: Number of municipalities per campaign spending limit category (2003-2015)
This gure plots the distribution of municipalities with respect to each of the potential-voter cutos that dictate changes in
campaign spending limits. The source of the data is Colombia's National Electoral Authority.
68
Figure 2.2: McCrary Test Around 25.000 Cuto
0 .00005 .0001 .00015
10000 20000 30000 40000
The data is pooled for all election years considered in the period (2003-2015).
69
Figure 2.3: Cattaneo, Jansson and Ma Test Around 25.000 Cuto
0 .00001 .00002 .00003
15000 20000 25000 30000 35000
potencial_elec
point estimate 95% C.I.
rddensity plot (p=2, q=3)
Results testing for dierences in the densities across the 25.000 potential voter threshold estimated using Cattaneo et al. (2019).
The data is pooled for all election years considered in the period (2003-2015).
70
Figure 2.4: McCrary Test Around 50.000 Cuto
0 .00002 .00004 .00006 .00008 .0001
30000 40000 50000 60000 70000
The data is pooled for all election years considered in the period (2003-2015).
71
Figure 2.5: Cattaneo, Jansson and Ma Test Around 50.000 Cuto
0 .00001 .00002 .00003
20000 40000 60000 80000 100000
potencial_elec
point estimate 95% C.I.
rddensity plot (p=2, q=3)
Results testing for dierences in the densities across the 50.000 potential voter threshold estimated using Cattaneo et al. (2019).
The data is pooled for all election years considered in the period (2003-2015).
72
Figure 2.6: Predetermined Outcomes
.3 .4 .5 .6 .7
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Indigenous Population
600 800 1000 1200 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Altitude
50 60 70 80 90 −5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Distance Dpt Capital
300 350 400 450
−4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Distance to Bogota
This gure reports the eect of campaign spending limits on pre-determined variables. Standardized cutos around zero and
estimations include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions are
calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local linear
regressions. Cuto and year xed eects have been partialed out from graph.
73
Figure 2.7: Political Outcomes
2.5 3 3.5 4 4.5 −10000 −5000 0 5000 10000
Electoral Potential
Sample average within bin Polynomial fit of order 1
No. Candidates
.4 .45 .5 .55
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Turnout
.08 .09 .1 .11 .12
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Concentration
This gure reports the eect of campaign spending limits on electoral outcomes. Standardized cutos around zero and estima-
tions include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions are calculated
using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions.
Cuto and year xed eects have been partialed out from graph.
74
Figure 2.8: Vote Buying
This gure reports the eect of campaign spending limits on voter buying. Standardized cutos around zero and estimations
include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions are calculated using
the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions.
Cuto and year xed eects have been partialed out from graph.
75
Figure 2.9: Crime Outcomes
20 25 30 35 40
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Homicides
.0005 .0006 .0007 .0008 .0009 −10000 −5000 0 5000 10000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Homicides per Capita
20 40 60 80 100
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Theft
This gure reports the eect of campaign spending limits on crime outcomes. Standardized cutos around zero and estimations
include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions are calculated using
the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions.
Cuto and year xed eects have been partialed out from graph.
76
Figure 2.10: Violence Outcomes
80 100 120 140
−4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Conflict Homicides
3 4 5 6 7 −10000 −5000 0 5000 10000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Kidnappings
200 400 600 800 1000 1200
−5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Displacements
This gure reports the eect of campaign spending limits on organized group violence outcomes. Standardized cutos around
zero and estimations include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions
are calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local
linear regressions. Cuto and year xed eects have been partialed out from graph.
77
Figure 2.11: Violence Outcomes in Non-Electoral Years
60 80 100 120 140 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Conflict Homicides
0 2 4 6 8
−4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Kidnappings
0 1000 2000 3000 4000 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Displacements
This gure reports the eect of campaign spending limits on organized group violence outcomes for non-electoral years. Stan-
dardized cutos around zero and estimations include year and cuto xed eects. Optimal bins determined following Calonico
et al. (2014). Regressions are calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use
uniform kernels and local linear regressions. Cuto and year xed eects have been partialed out from graph.
78
Figure 2.12: Municipal Finances
0 10000 20000 30000 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Homicides
0 10000 20000 30000 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Subversive Actions
This gure reports the eect of campaign spending limits on municipalities' nances outcomes. Standardized cutos around
zero and estimations include year and cuto xed eects. Optimal bins determined following Calonico et al. (2014). Regressions
are calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform kernels and local
linear regressions. Cuto and year xed eects have been partialed out from graph.
79
Figure 2.13: Service Coverage
50.00 60.00 70.00 80.00 90.00 −5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Aqueduct
30.00 40.00 50.00 60.00 70.00 80.00
−4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Sewer
70 80 90 100 110
−4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Gas
40 50 60 70 80 90 −4000 −2000 0 2000 4000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Trash
This gure reports the eect of campaign spending limits on service coverage outcomes during non electoral-years. Standardized
cutos around zero and estimations include year and cuto xed eects. Optimal bins determined following Calonico et al.
(2014). Regressions are calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform
kernels and local linear regressions. Cuto and year xed eects have been partialed out from graph.
80
Figure 2.14: Education Outcomes
300 400 500 600 −5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
No. Teachers
45 46 47 48 49 50 −10000 −5000 0 5000 10000
Electoral Potential
Sample average within bin Polynomial fit of order 1
Standardized Test Scores
7 8 9 10 11 12 −5000 0 5000
Electoral Potential
Sample average within bin Polynomial fit of order 1
No. Schools
This gure reports the eect of campaign spending limits on education outcomes during non electoral-years. Standardized
cutos around zero and estimations include year and cuto xed eects. Optimal bins determined following Calonico et al.
(2014). Regressions are calculated using the optimal bandwidths as dened by Calonico et al. (2018). Estimations use uniform
kernels and local linear regressions. Cuto and year xed eects have been partialed out from graph.
81
Tables
Table 2.1: Campaign Spending Limits, Colombian Pesos (2003-2015)
Potential Voters 2003 Elections 2007 Elections 2011 Elections 2015 Elections
5:000:001 3.449.505.411
[5:000:000; 1:000:001] 1.726.086.150
[1:000:000; 500:001] 803.214.378 989.000.000 1.318.000.000 1.617.936.803
[500:000; 250:001] 468.542.032 577.000.000 745.000.000 1.221.956.853
[250:000; 100:001] 401.604.684 494.000.000 659.000.000 1.080.672.789
[100:000; 50:001] 200.807.352 247.000.000 330.000.000 541.087.905
[50:000; 25:001] 66.932.444 82.000.000 110.000.000 180.362.635
< 25:000 34.552.420 43.000.000 58.000.000 94.690.384
The limits were dened by the following resolutions and decrees: Resolution 0127 of 2015, Resolution 0078 of 2011, Resolution
82 of 2007 and Decree 2207 of 2003.
Table 2.2: Balance Checks Across Thresholds
(1) (2) (3) (4)
Indigenous Pop Altitude Distance Dpt Capital Distance to Bogota
below -0.051 42.035 1.461 -21.319
(0.079) (220.410) (10.519) (37.691)
Observations 418 214 355 343
Cutos FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
This table reports the variation of pre-determined outcomes around the potential-voter cutos. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal
bandwidths as calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects
of cuto and year of election are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05,
***p < 0.01.
82
Table 2.3: Political Outcomes
(1) (2) (3)
No. Candidates Turnout Concentration
below 0.195 0.003 -0.011
(0.321) (0.025) (0.010)
Observations 566 393 392
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 4.585 .5700000000000001 .115
This table reports the eect of campaign spending limits on electoral outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal bandwidths as
calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and
year of election are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.4: Campaign Spending Limits and Voters-Bought imported per municipality.Pooled Re-
gression
Imported Voters asinh(Imported Voters)
(1) (2) (3) (4) (5) (6)
RD Estimate -349.7219

-346.3938

-323.9490

-0.5306

-0.5306

-0.4671

(107.9744) (108.4843) (119.2125) (0.2393) (0.2404) (0.2679)
Observations 2707 2707 2707 2707 2707 2707
Kernel Tri Tri Unif Tri Tri Unif
Bandwidth 3720 3666 2580 3862 3815 2776
E.No.Obs. 210 207 154 219 217 161
Std.Errors Robust Cl. Mpality Cl. Mpality Robust Cl. Mpality Cl. Mpality
This table reports the eect of campaign spending limits on vote buying. RD Estimate is an indicator variable equal to 1 for
municipalities with more potential voters than the closest cuto. Regressions are calculated using the optimal bandwidths as
calculated by Calonico et al. (2018). Estimations use local linear regressions. Fixed eects of cuto and year of election are
included. The robust standard errors are clustered as indicated in the table.* p < 0.10, ** p < 0.05, ***p < 0.01.
83
Table 2.5: Crime Outcomes
(1) (2) (3)
Homicides Theft Homicides p.c.
below -0.290 -3.199 -0.000
(3.869) (9.941) (0.000)
Observations 425 404 582
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 22.435 61.62000000000001 0
This table reports the eect of campaign spending limits on crime outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal bandwidths as
calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and
year of election are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.6: Violence and Con
ict Outcomes
(1) (2) (3)
Displacement Con
ict Homicides Kidnapping
below 451.579

20.834 0.952

(206.974) (12.677) (0.567)
Observations 489 339 657
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 511.605 46.76 1.505
This table reports the eect of campaign spending limits on organized group violence outcomes. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal
bandwidths as calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects
of cuto and year of election are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05,
***p < 0.01.
84
Table 2.7: Violence and Con
ict Outcomes in Non-Election Years
(1) (2) (3)
Displacement Con
ict Homicides Kidnapping
below 163.425 18.540

0.535
(155.657) (10.048) (0.602)
Observations 789 641 831
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 397.45 36.105 1.155
This table reports the eect of campaign spending limits on organized group violence outcomes. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal
bandwidths as calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects
of cuto and year are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.8: Municipality Finances
(1) (2)
Income Expenditures
below 7258.072

8963.870

(3188.607) (4019.494)
Observations 628 511
Cutos FE Yes Yes
Year FE Yes Yes
Mean Above Cuto 26351.34 26602.14
This table reports the eect of campaign spending limits on municipalities' income and expenditures. below is an indicator
variable equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the
optimal bandwidths as calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed
eects of cuto and year are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p
< 0.01.
85
Table 2.9: Service Coverage
(1) (2) (3) (4)
Aqueduct Sewer Gas Trash
below 2.519 2.617 -1.082 -1.971
(5.143) (6.109) (5.197) (6.790)
Observations 706 642 510 463
Cutos FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
Mean Above Cuto 67.84999999999999 56.82 92.875 59.915
This table reports the eect of campaign spending limits on service coverage. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal bandwidths as
calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and
year are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.10: Education
(1) (2) (3)
No. Teachers Standarized Test Scores No. Schools
below 39.776 0.079 1.317
(31.802) (0.433) (0.901)
Observations 705 1191 773
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 523.23 47.94 10.715
This table reports the eect of campaign spending limits on education outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the optimal bandwidths as
calculated by Calonico et al. (2018). Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and
year are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
86
Appendix
Table 2.11: Crime Outcomes Bandwidth 3000
(1) (2) (3)
Homicides Theft Homicides p.c.
below 4.942 8.968 0.000
(4.985) (13.724) (0.000)
Observations 229 229 229
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 24.44 60.685 0
This table reports the eect of campaign spending limits on crime outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth mentioned in
the title of the graph. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year are included.
The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.12: Crime Outcomes Bandwidth 5000
(1) (2) (3)
Homicides Theft Homicides p.c.
below -0.292 -6.407 -0.000
(3.953) (10.033) (0.000)
Observations 382 382 382
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 22.575 60.015 0
This table reports the eect of campaign spending limits on crime outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth mentioned in
the title of the graph. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year are included.
The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
87
Table 2.13: Crime Outcomes Bandwidth 7000
(1) (2) (3)
Homicides Theft Homicides p.c.
below -2.223 -2.779 -0.000
(3.477) (8.462) (0.000)
Observations 580 580 580
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 21.945 63.475 0
This table reports the eect of campaign spending limits on crime outcomes. below is an indicator variable equal to 1 for
municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth mentioned in
the title of the graph. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year are included.
The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.14: Violence and Con
ict Outcomes Bandwidth 3000
(1) (2) (3)
Displacement Con
ict Homicides Kidnapping
below 465.584

44.234

1.409
(253.402) (15.116) (0.967)
Observations 224 224 224
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 479.16 51.68 1.455
This table reports the eect of campaign spending limits on organized group violence outcomes. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth
mentioned in the title of the table. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year
are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
88
Table 2.15: Violence and Con
ict Outcomes Bandwidth 5000
(1) (2) (3)
Displacement Con
ict Homicides Kidnapping
below 522.563

22.190

1.038
(224.880) (11.972) (0.736)
Observations 374 374 374
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 470.925 48.015 1.305
This table reports the eect of campaign spending limits on organized group violence outcomes. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth
mentioned in the title of the table. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year
are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
Table 2.16: Violence and Con
ict Outcomes Bandwidth 7000
(1) (2) (3)
Displacement Con
ict Homicides Kidnapping
below 380.376

13.913 0.996
(188.934) (11.135) (0.606)
Observations 568 568 568
Cutos FE Yes Yes Yes
Year FE Yes Yes Yes
Mean Above Cuto 514.49 47.805 1.425
This table reports the eect of campaign spending limits on organized group violence outcomes. below is an indicator variable
equal to 1 for municipalities with less potential voters than the closest cuto. Regressions are calculated using the bandwidth
mentioned in the title of the table. Estimations use uniform kernels and local linear regressions. Fixed eects of cuto and year
are included. The robust standard errors are clustered by municipality. p < 0.10, ** p < 0.05, ***p < 0.01.
89
Chapter 3
A Digitization of Individual Voter Turnout in Colombia
1 Introduction
Voting is the foundation of democracy. Citizens vote and elect candidates who best represent their
preferences. They also vote to remove governments that fail to fulll their expectations. Voter
participation is an indicator of major importance to measure the health of democracies. In free so-
cieties, more people willing to vote is usually an indicator that citizens trust their local institutions.
Voter turnout is, therefore, a crucial variable for politicians, policymakers, and ordinary citizens.
Social scientists and policymakers still grapple to understand the causes of voter turnout. Data
from almost every country suggests that individuals vote for other reasons besides their probability
to be pivotal. Studies indicate that citizens choose to turn out to vote for multiple reasons. Empiri-
cal evidence suggests that people turn out to vote because: they developed a habit (Fujiwara et al.,
2016; Gerber et al., 2003), to tell others (DellaVigna et al., 2016; Gerber et al., 2008), altruistic
motives (Fowler and Kam, 2007), etc. These studies use turnout data either aggregated at an ad-
ministrative level (counties, voting locations, etc.), or a sample of individuals, or use representative
surveys in which citizens self-report if they voted in the past.
Studies of voter turnout that use representative surveys or experiments cannot capture nuances
of electoral participation or can be subject to measurement error and biases. On one hand, in
surveys individuals self-report their voting behavior and can lie about their actions (Karp and
Brockington, 2005). On the other hand, representative samples of voter turnout are also limited by
the scale of the experiment implemented by the researcher and its costs. Hence, there is an absence
of information and research in more remote regions {poorer and less populated{ of developing
countries. Because of this, researchers struggle to study the heterogeneity of voter turnout rates,
90
especially within developing countries, which lack digitized information for the universe of their
voters. Having digitized databases of voter turnout from ocial records at the most granular level
is important to deepen our understanding of politics and the political economy of development.
In this paper I describe the process to digitize the universe of voter turnout records from
Colombia for several elections. I also show summary statistics, after generating a database of voter
turnout at the individual level, for the regional elections of 2015. Voter-turnout data in the country
is ocially collected via hand-written records during election days. In this paper, I use millions
of pages with this information to create a database of individual voter turnout for the universe of
Colombian adults. To do so I use optical character recognition and computer vision techniques.
The generated data is of high quality and is not biased. Manually identied ID numbers for a
random set of polling stations, compared to the ones generated by code, coincide in around 90% of
the total observations. Furthermore, aggregated turnout rates estimated via the generated output
described in this paper are not signicantly dierent to the ocial turnout statics published by the
electoral organization. This paper highlights how the generated turnout data can be merged with
other large administrative databases. I show summary statistics from merging the Sisben with the
cross-section of the universe of citizens who voted for the regional elections of 2015.
This paper contributes to two strands of literature. First, it provides a novel database of voter
turnout records from administrative records for the universe of voters for a whole country. A
recent literature has used similar information for other countries to explore questions related to
electoral participation. Using information of voters from the United States, Cantoni and Pons
(2021) show that more strict laws that require identication to vote did not have an eect on
turnout. Evidence from Mexico and the United States using individual voter turnout records also
indicates that neighborhoods or the location of residency aect individuals' electoral participation
and political preferences (Cantoni and Pons, 2022; Finan et al., 2021). Other studies show how
changes in income, or participating in a particularly relevant election for the future of a country,
increase voter turnout in subsequent elections (Schafer et al., 2022; Kaplan et al., 2023).
This paper also contributes to a second strand of literature which develops methods for digi-
tization at scale using computational techniques. Recent studies have developed tools from deep
learning techniques to parse information with complex layouts in scanned documents (Shen et al.,
2021). Using these tools it is possible to extract historical information stored in documents with
91
unconventional layouts. For example, Shen et al. (2020) uses these tools to digitize data on bi-
ographies of prominent Japanese citizens from historical records. In this paper I do not use deep
learning techniques to digitize the information, but I follow the example of using computational
techniques to digitize large amounts of data at scale.
The rest of this paper is organized as follows. Section 2 describes the context and Colombian
elections, while section 3 discusses the data. Section 4 explains the methodology used to clean the
data, while section 5 provides an application. Section 6 concludes and discusses future avenues of
research.
2 Context
Colombia is a democracy where political candidates win an election if they get more votes than
their competitors. In each election, every citizen older than 18 years old who has a valid ID
(locally known as c edula de ciudadan a) can vote. Each person votes in a previously determined
voting location: once a citizen claims their ID (they can only do so once they turn 18) they are
automatically registered to vote in the closest polling location to their claimed address of residence
within their municipality. If a citizen changes their residency they can change their voting location
before the elections take place.
Once someone is adequately registered and goes to vote in their corresponding voting location
the day of the elections, they are assigned to select their candidate of preference in a polling station
(mesa de votaci on) within the voting site. In 2015, for example, there were 360 potential voters for
each polling station. Here, each person casts their vote anonymously{which is why it is not possible
to know the political preferences for each citizen{and are manually reported to have turned out
to vote. This allows people to vote only once. Voting jurors assigned to each polling station are
responsible to ll out the forms of each person who turns out to vote, and, more generally, assist
in the overall electoral process for their station during the election day.
The period of analysis {for which there is data availability{ described in this paper ranges from
2011 to 2019. Throughout this time span the country held three regional elections (2011, 2015,
2019); two legislative elections (2014 and 2018); two presidential elections (2014 and 2018); and a
national Plebiscite (2016).
92
In regional elections, citizens vote to select the mayors and councilors in each of the country's
1103 municipalities. Additionally, in these elections individuals also vote for governors and assembly
deputies for each of Colombia's 32 departments. In legislative elections, citizens vote for senators
and congressmen. The presidential elections described in this paper include a rst round and a
second round. A second round happens if neither candidate got the majority (more than 50%) of
the votes in the rst round. The second round was necessary in both elections of 2014 and 2018,
which means that there are two sets of turnout information for each presidential election. Finally,
in 2016 a national plebiscite was implemented, in which citizens voted in favor or against a national
peace agreement with the FARC guerrilla. A small majority (50.2%) voted against the agreement,
a result that was surprising locally and internationally (BBC, 2016).
Individual voter turnout information will be a fundamental tool for researchers to understand
results such as the Plebiscite's. Perhaps more importantly, the individual voter turnout data
will motivate research to answer questions about local politics in Colombia and their own local
heterogeneity. This has not been approached with sucient depth given data limitations and the
costs to collect information, especially in the most isolated regions of the country.
3 Data
There is not a digitized database of individual voter turnout in Colombia. The lack of information of
voter turnout at the individual level is explained because the voting records are reported manually
at each voting station.
Figure 3.1 shows an example of how voter turnout is tracked during election days, this type of
record is known as a \form E-11". Each polling station initially has an E-11 form, which includes a
typed list of all registered voters' ID numbers in that station, and three empty contiguous columns.
Throughout election day, voting jurors ll the blank columns next to an ID for each person who
turns out to vote. Jurors write the rst and second last names, and the names of each individual
who turns out to vote in the three blank columns, respectively. This procedure is implemented
to keep track of voters within each polling station and avoid citizens to vote more than once. At
the time elections close, the E-11 forms include all the information, at the individual level, of who
voted and who did not. If a citizen votes, their ID will have written information in the contiguous
93
columns. Otherwise, it implies a citizen did not vote.
Voting jurors sign in the rst and last pages of the E-11 form. Each juror can also vote in the
station they were assigned to perform their duty. The rst page of a typical E-11 form is shown in
Figure 3.2. Figure 3.3 shows how jurors report their vote. These dierent page layouts are taken
into consideration in the process that reads the IDs and identies whether an individual voted.
This paper uses information for the universe of E-11 les to identify individual voter turnout
for all the country for multiple elections. The scale of this project is substantial given the amount
of information. Colombia's population ranged between 44 to 50 million people between 2010 and
2020 (DANE, 2023). For example, close to 34 million people were registered to vote in the 2015
regional elections. There were 93,303 polling stations(MOE, 2016b). 360 citizens are assigned to
potentially vote in each voting station.
1
An E-11 form is lled in every polling station and each
form has 12 pages. This implies that there is a total of about 1.1 or 1.2 million pages just for
the 2015 elections. The number of voters per polling station varies depending on the type of the
election, but the total number of pages with information for each election is substantial for every
election.
In this paper I use the scans of every E-11 form for all the country, which were provided by
the Registradur a Nacional del Estado Civil, Colombia's electoral administrative organization. In
particular I focus on the elections of 2015, for which I cleaned the information for all the country.
4 Methodology
In this paper I use computer vision and optical character recognition (OCR) techniques to identify
if an individual turned out to vote, or not, for the universe of Colombian voters. Using the scans
of E-11 forms described in the previous sections I identify {at the individual level{ who turned out
to vote and who did not. This section describes the process to capture the information of interest
and provides statistics indicating the precision of the outcome.
There is an E-11 form {hence, a le{ for each polling station in a particular election. The
process to identify whether an individual voted or not in each polling station works following these
1
Not all polling stations have 360 voters. Stations in Corferias, Bogot a and other central municipal voting locations
include a higher number of voters. Voting locations with a total number of potential voters that is not a multiple of
360 also have a polling station with less than this number of potential voters.
94
steps:
1. Split each le {or E-11 form{ into single pages.
2. Open each page and parse the table with the ID numbers and contiguous columns. The
location of the table changes according to the page number. First pages in each form have a
dierent layout.
3. Identify the main contours of the table.
4. Parse the table's rows.
5. For each row, use OCR techniques to identify the typed numbers {the IDs{ in the rst column
of each half of the row. Note that each row in the table contains information of two individuals
and whether they voted or not (Figure 3.1).
6. Calculate the ratio of white to black pixels in the three contiguous columns to the ID column,
for each half of each row. If this ratio is smaller than a threshold value, indicating that there is
evidence of written information, then turnout is equal to one for that observation. Otherwise
{too many white pixels with respect to black pixels{, turnout is equal to zero.
7. Repeat the OCR ID identication and the white-to-black pixel ratio calculation for each row
in a table. Repeat the process for each page in a le. Repeat for every le.
The process is implemented using multiple loops. Each le is split into pages, and each table in
a page is split into rows. The code identies in each row whether a pair of individuals {with their
own ID numbers{ voted or not. The iterative nature of the code makes the process time consuming.
Each voting station {which include 360 potential voters{ takes between one or two minutes to read.
2
There are more than a million pages in each election and there are multiple elections. Because of
this, the code was parallelized and executed using high-performance computing via the Center of
Advanced Research Computing (CARC) at USC.
The cleaned data includes the ID of each citizen registered to vote and whether the person turned
out to vote or not. For the 2015 elections, 27 million IDs were identied (some of them are subject
2
Files with dierent image resolution explain the dierences in running time.
95
to error in the OCR process) as potential voters. 58.8% of these individuals are recognized as voters,
which is close to the ocial 59.4% of voter participation reported by the electoral authorities.
I test the quality of the generated turnout data in two ways. First, I digitize a random set of
E-11 les by hand and compare it with the output generated by the code. Second, I compare the
average turnout data at the department level for the elections of 2015 generated by the code with
the ocial statistics published by the national electoral agency. In both cases the generated data
looks of high quality and reliable.
I digitize by hand a set of fteen E-11 les randomly selected from across the country to
contrast how reliable the OCR process of ID identication was. I transcribe the ID numbers and
manually report if someone turned out to vote for 4,475 individuals. This information was also
used to determine the accuracy of the voter-turnout variable generated using the pixel-counting
ratio described previously.
Of the 4,475 individuals manually digitized (58.32% turned out to vote) 4075 were eectively
identied by the OCR process. Furthermore, voter turnout for the 4075 individuals was 58.28% in
the original data. The code identied a turnout rate of 58.36% for these 4075 individuals. For the
400 observations with an incorrect OCR outcome, turnout rates are 58.75%, which indicates that
the code does not perform better identifying turnout for individuals whose ID can (or can't) be
correctly extracted. This suggests that the correctly extracted data generated by the code is not
biased towards either voters or non-voters. If an ID is not correctly identied we can assume it was
because of random factors, not by any systematic causes which could be potential sources of bias
in the end outcome.
The results indicate that around 9% of observations were not correctly identied because the
OCR technology could not or did not properly read the typed ID numbers. Additionally, the pixel-
counting ratios identify adequately whether someone votes or not. Of the 4075 individuals correctly
identied, only in 7 cases the original turnout rates did not match the code's generated turnout
values. The code generated 5 false positives {individuals who did not turn out to vote, but whose
white to black pixel ratio led to count them as voters{ and two false negatives.
Overall, the results suggest that the code is relatively precise identifying the correct ID numbers.
The white-black pixel ratio criteria used to identify whether someone voted {or not{ is also accurate.
Table 3.4 in the Appendix shows the voter turnout rates generated by the code and the original
96
rates reported by Colombia's national electoral agency. The ocial turnout statistics published by
the electoral authority are, at most, disaggregated at the polling-station level. Once the elections
are over during election day, voting jurors count the total number of votes {and how many votes
each candidate or party got{ in their polling station. The electoral agency aggregates all the data
from every polling station to declare the winners and losers. Note, however, that the individual
level information is not processed or used by the electoral agency to calculate any outcome.
The turnout statistics at the department level indicate that the information on voter turnout
generated by the code (using individual inputs) is comparable to the average turnout published
by the electoral agency. The dierences range in most cases between -3 and 0 percentage points.
Although the dierences are small in every case, it is worth noting that in most departments the
code is generating slightly higher turnout rates than what the ocial statistics indicate.
5 An Application: Merging with Other Data
Having a data set of individual voter turnout is a resource for new and exciting avenues of research.
In particular, having the universe of voters for a particular election (or set of elections) can be
benecial to study the political implications for dierent public policies. In Colombia, most adults
have a c edula de ciudadan a identication number. This number essentially identies citizens
in every administrative exchange with the government and in many private sector transactions.
Furthermore, the voter turnout data described in the previous section includes this ID number.
Therefore, the data can potentially be merged with multiple other data sources, either public or
private.
In this section I describe how the voter turnout data can be merged with another large adminis-
trative data set, widely used in the country: the Sisben. In Fajury (2023), I show how a conditional
cash transfer program reduces vote buying and has no eect on citizens' voter turnout. To do so,
that study uses the voter turnout data described in this paper, the Sisben, and Familias en Acci on
data, which is Colombia's national conditional cash transfer program. In this section I describe the
Sisben and provide statistics of interest that can be studied using that data source as a motivation
for further studies.
97
5.1 The Sisben
Colombia's government implements the Sistema de Identicaci on para Potenciales Beneciarios
de los Programas Sociales (Sisben) {essentially a census of the poor{ to identify the potential
beneciaries of subsidies or other social programs. The Sisben survey of 2011 is used in this paper.
The interviews for this census of the poor were done between 2009 and 2010. The data includes
close to 30 million individual responses. The survey includes questions about their dwelling's
features (number of rooms and bathrooms,
oor material, public services availability, etc.) and
socioeconomic characteristics of the household (education, gender, age, etc.).
The Sisben also includes an index that measures multidimensional poverty, structured from
multiple variables which capture poverty and vulnerability and that are also hard to manipulate
by the respondents. Based on this index, the system assigns to each household a score that ranges
(continuously) from 0 to 100, where lower values indicate more poverty. The algorithm used to
calculate the index and the score is unknown (and indecipherable) to the population. This score
is used by the government to assign subsidies or to determine the eligibility of social programs
(Medell n and S anchez, 2015; DPS, 2013a,b).
5.2 Summary Statistics
Table 3.1 and 3.2 show summary statistics for multiple variables and voter turnout. Voter turnout
was measured using the code described in previous sections for the regional elections of 2015 for
all Colombia. This information was then merged using each individual's ID with the Sisben data.
The tables shows averages for dierent variables for the population who, at the time of the Sisben
survey, had a c edula, and could vote. Around 23 million individuals are included in the Sisben and
about 10.8 million matched with the turnout data. This further indicates that the turnout data
is of high quality. The statistics are relevant from a national political perspective. First, around
a third of all Colombian voters were successfully merged (in 2015 close to 33 million people were
registered to vote). Second, the population included in the Sisben are the relatively poorer citizens
in the country.
The dierences across voters and non voters suggest some interesting and surprising patterns:
First, the data does not suggest that there is a large wealth gap between voters and non-voters.
98
The dierent dwelling characteristics, as well as the Sisben score, don't indicate that wealthier
individuals vote more. If anything, the score indicates that slightly poorer individuals are more
likely to vote. This statistic is surprising and worth exploring, as it goes against evidence from other
contexts, particularly from developed countries (Schafer et al., 2022; Kasara and Suryanarayan,
2015).
Second, voter turnout for the matched Sisben sample is particularly high. The ocial statistics
of voter turnout for this election indicate that about 60% of registered citizens voted (MOE, 2016b).
Within the Sisben sample, which comprises the relatively poorer individuals of the country, voter
turnout is of 66%. This evidence suggests {as mentioned in the previous point{ that higher income
individuals did not turn out to vote more. It is worth noting that turnout is higher in regional
elections in Colombia. Nonetheless, it is surprising that turnout is not increasing by wealth. Future
work should study whether this result holds for presidential and legislative elections. Furthermore,
the granularity of the data allows to study if these patterns of wealth and lower turnout hold for
dierent municipalities. Heterogeneity across municipalities can contribute to explain the intuition
of the political economy that is generating such outcomes.
Third, older and married individuals vote more, however the dierence in age between voters
and non-voters {3 years{ is not particularly salient. Future research should study whether marriage
causally increases voter turnout in this context and what are the mechanisms of this phenomenon.
6 Conclusion
Voting is the foundation of democracy. The act of voting is of great relevance for political parties,
public ocials and policymakers. Through elections and individual participation, voters choose
the candidates whose platforms are closer to their preferences, but also punish governments that
do not fulll their expectations. Although voting plays such an important role for democracies,
ocial data on voter turnout at the individual level is not widely collected and digitized, especially
for developing countries. In Colombia, a middle-income country, the voter-turnout data is ocially
collected via hand-written vote records during election days. In this paper, I use such information
to create a database of individual voter turnout for the universe of Colombian adults. To do so, I
use optical character recognition and computer vision techniques.
99
The generated data is of high quality and is not subject to biases. Manually cleaned ID numbers
compared to the ones generated by code, for a random set of polling stations, coincide in around
90% of the total observations. Furthermore, aggregated turnout rates estimated via the generated
output described in this paper are not signicantly dierent to the ocial turnout statics published
by the electoral organization. Finally, this paper highlights how the generated turnout data can
be merged with other large administrative databases. I show summary statistics from merging the
Sisben with the cross-section of the universe of citizens who voted for the regional elections of 2015.
The generated data described in this paper will be an important input for future research. The
granularity of the data will allow researchers to dive deeper into the determinants of political par-
ticipation in developing countries. I also expect the data to be relevant for future economics and
political science studies that explore the most remote and poorest regions of the country, which
desperately need more research focused on their local contexts. Finally, I expect this paper to mo-
tivate researchers to use similar sources {handwritten voting records, or other scanned information,
such as education or health records{ from other countries and digitize them using computational
techniques. The explosion of these new technologies and computing power now allows researchers
to study a wide variety of public policies and programs at scale, especially in the developing world,
at a low cost.
100
Figures
Figure 3.1: Example of Voter Turnout Records
Electoral form E-11 for a particular voting station. It reports the ID number of every potential voter in a voting station. The
typed numbers are the ID numbers for each registered citizen. If the spaces in the contiguous columns to a citizen's ID are left
blank, it means the citizen did not vote. If there is a signature next to a typed number it means the person turned out to vote.
Tables
101
Table 3.1: Turnout 2015. Colombia
(1) (2) (3)
Variable Did Not Vote Voted Dierence
Age 43.465 46.767 3.302***
(16.668) (15.533) (0.010)
PeopleinHousehold 4.475 4.413 -0.062***
(2.255) (2.134) (0.001)
Married 0.505 0.605 0.101***
(0.500) (0.489) (0.000)
Male 0.491 0.454 -0.038***
(0.500) (0.498) (0.000)
Energy 0.952 0.945 -0.008***
(0.213) (0.229) (0.000)
Sewer 0.703 0.594 -0.110***
(0.457) (0.491) (0.000)
Gas 0.467 0.400 -0.068***
(0.499) (0.490) (0.000)
Telephone 0.359 0.269 -0.091***
(0.480) (0.443) (0.000)
Trash 0.775 0.668 -0.107***
(0.418) (0.471) (0.000)
Aqueduct 0.822 0.772 -0.050***
(0.383) (0.420) (0.000)
Rooms 2.870 2.869 -0.001
(1.256) (1.246) (0.001)
Toilets 0.993 0.966 -0.027***
(0.395) (0.432) (0.000)
Shower 0.713 0.607 -0.106***
(0.452) (0.488) (0.000)
Fridge 0.610 0.562 -0.048***
(0.488) (0.496) (0.000)
Washer 0.267 0.228 -0.038***
(0.442) (0.420) (0.000)
TV 0.809 0.769 -0.039***
(0.393) (0.421) (0.000)
WaterHeater 0.086 0.070 -0.016***
(0.280) (0.256) (0.000)
Microwave 0.057 0.046 -0.012***
(0.233) (0.209) (0.000)
Observations 3,679,215 7,174,352 10,853,567
Means and standard deviations (in parenthesis). Voters and Non-Voters are all those citizens identied in the electoral censuses
of this particular election and in the Sisben. The sample is restricted to citizens old enough to vote ( 18) who have a c edula.
*** denotes 1% signicance, ** denotes 5% signicance, * denotes 10% signicance.
102
Table 3.2: Turnout 2015. Colombia Continued.
(1) (2) (3)
Variable Did Not Vote Voted Dierence
EconStrata 1.573 1.489 -0.084***
(0.731) (0.727) (0.000)
None 0.083 0.084 0.001***
(0.276) (0.277) (0.000)
Primary 0.390 0.446 0.056***
(0.488) (0.497) (0.000)
Secondary 0.458 0.404 -0.054***
(0.498) (0.491) (0.000)
Technical 0.026 0.026 -0.000***
(0.160) (0.159) (0.000)
Bachelor 0.040 0.038 -0.002***
(0.196) (0.191) (0.000)
Graduate 0.003 0.002 -0.000***
(0.051) (0.049) (0.000)
GoesSchool 0.089 0.069 -0.020***
(0.284) (0.253) (0.000)
No Activity 0.122 0.101 -0.022***
(0.328) (0.301) (0.000)
Work 0.492 0.479 -0.013***
(0.500) (0.500) (0.000)
Searching Job 0.055 0.043 -0.012***
(0.228) (0.202) (0.000)
Studying 0.074 0.055 -0.019***
(0.261) (0.227) (0.000)
Housekeeping 0.236 0.301 0.066***
(0.424) (0.459) (0.000)
Rentist 0.003 0.003 -0.000***
(0.056) (0.055) (0.000)
Retired 0.016 0.018 0.001***
(0.127) (0.132) (0.000)
Invalid 0.003 0.001 -0.001***
(0.050) (0.035) (0.000)
Weeks SearchJob 15.690 15.328 -0.362***
(13.856) (13.746) (0.040)
Score 41.754 40.615 -1.138***
(19.070) (19.399) (0.012)
Observations 3,679,215 7,174,352 10,853,567
Means and standard deviations (in parenthesis). Voters and Non-Voters are all those citizens identied in the electoral censuses
of this particular election and in the Sisben. The sample is restricted to citizens old enough to vote ( 18) who have a c edula.
*** denotes 1% signicance, ** denotes 5% signicance, * denotes 10% signicance.
103
Figure 3.2: Example of First Page of E-11 Form of Voter Turnout Records
This image shows a typical rst page of the form E-11 for a particular voting station. The layout of the page is dierent than
the rest of pages in the form. Voting jurors sign and provide their ID number in the upper part of the page.
Figure 3.3: Example of Jurors' Vote Report on E-11 Form
This image shows a typical penultimate page of the form E-11 for a particular voting station. In the last pages of the E-11
form jurors hand-write their own ID number and sign next to it if they choose to vote.
104
Appendix
Figure 3.4: Turnout Rates by Department: Code Generated and Ocial (2015)
This table compares the turnout rates by department generated by code with the ocial turnout rates revealed by the electoral
authority and taken from (MOE, 2016b). The code data is aggregated by department after cleaning the universe of 2015
electoral data. The aggregates calculated by the electoral agency are computed after adding the reports of total vote counts
done by the jurors in each polling station on elections day.
105
References
Acemoglu, D., Johnson, S., and Robinson, J. A. (2002). Reversal of fortune: Geography and
institutions in the making of the modern world income distribution. The Quarterly Journal of
Economics, 117(4):1231{1294.
Acemoglu, D., Naidu, S., Restrepo, P., and Robinson, J. A. (2019). Democracy does cause growth.
Journal of Political Economy, 127(1):47{100.
Acemoglu, D., Robinson, J. A., and Santos, R. J. (2013). The monopoly of violence: Evidence from
Colombia. Journal of the European Economic Association, 11(suppl 1):5{44.
Alesina, A., Piccolo, S., and Pinotti, P. (2019). Organized crime, violence, and politics. The Review
of Economic Studies, 86(2):457{499.
AllAfrica (2009). Botswana: Parties in row over voter tracking.
https://allafrica.com/stories/200903250847.html (Accessed August 29, 2022).
Attanasio, O., Sosa, L. C., Medina, C., Meghir, C., and Posso-Su arez, C. M. (2021). Long term ef-
fects of cash transfer programs in Colombia. Working Paper 29056. National Bureau of Economic
Research.
Auyero, J. (2001). Poor people's politics: Peronist survival networks and the legacy of Evita. Duke
University Press.
Avis, E., Ferraz, C., and Finan, F. (2018). Do government audits reduce corruption? Estimating
the impacts of exposing corrupt politicians. Journal of Political Economy, 126(5):1912{1964.
Avis, E., Ferraz, C., Finan, F., and Varj~ ao, C. (2019). Money and politics: The eects of campaign
spending limits on political competition and incumbency advantage. Working Paper.
Baron, D. P. (1994). Electoral competition with informed and uninformed voters. American Political
Science Review, 88(1):33{47.
BBC (2016). Colombia referendum: Voters reject Farc peace deal.
https://www.bbs.com/news/world-latim-america-37537252 (Accessed March 15, 2023).
106
Bertrand, M., Djankov, S., Hanna, R., and Mullainathan, S. (2007). Obtaining a driver's license in
India: An experimental approach to studying corruption. The Quarterly Journal of Economics,
122(4):1639{1676.
Blattman, C., Emeriau, M., and Fiala, N. (2018). Do anti-poverty programs sway voters? Experi-
mental evidence from Uganda. Review of Economics and Statistics, 100(5):891{905.
Blattman, C., Larreguy, H., Marx, B., and Reid, O. R. (2019). Eat widely, vote wisely? Lessons
from a campaign against vote buying in Uganda. Technical report, National Bureau of Economic
Research.
Bobonis, G. J., Gertler, P., Gonzalez-Navarro, M., and Nichter, S. (2022). Vulnerability and
clientelism. American Economic Review, 112(11):3627{3659.
Calonico, S., Cattaneo, M. D., and Farrell, M. H. (2018). On the eect of bias estimation on
coverage accuracy in nonparametric inference. Journal of the American Statistical Association,
113(522):767{779.
Calonico, S., Cattaneo, M. D., and Titiunik, R. (2014). Robust nonparametric condence intervals
for regression-discontinuity designs. Econometrica, 82(6):2295{2326.
Camacho, A. and Conover, E. (2011). Manipulation of social program eligibility. American Eco-
nomic Journal: Economic Policy, 3(2):41{65.
Cantoni, E. and Pons, V. (2021). Strict ID laws don't stop voters: Evidence from a US nationwide
panel, 2008{2018. The Quarterly Journal of Economics, 136(4):2615{2660.
Cantoni, E. and Pons, V. (2022). Does context outweigh individual characteristics in driving voting
behavior? Evidence from Relocations within the United States. American Economic Review,
112(4):1226{72.
Cattaneo, M. D., Jansson, M., and Ma, X. (2019). Simple local polynomial density estimators.
Journal of the American Statistical Association, pages 1{7.
CNE (2015a). Resoluci on No. 5352 de 2015 (21 de Octubre). Technical report, Consejo Nacional
Electoral.
CNE (2015b). Resoluci on No. 6170 de 2015 (21 de Octubre). Technical report, Consejo Nacional
Electoral.
CNMH (2018). Paramilitarismo. Balance de la contribuci on del CNMH al esclarecimiento hist orico.
Centro Nacional de Memoria Hist orica.
107
Coate, S. (2004). Political competition with campaign contributions and informative advertising.
Journal of the European Economic Association, 2(5):772{804.
Conover, E., Z arate, R. A., Camacho, A., and Baez, J. E. (2020). Cash and ballots: Condi-
tional transfers, political participation, and voting behavior. Economic Development and Cultural
Change, 68(2):541{566.
Cruz, C. (2019). Social networks and the targeting of vote buying. Comparative Political Studies,
52(3):382{411.
Cruz, C., Keefer, P., and Labonne, J. (2021). Buying informed voters: New eects of information
on voters and candidates. The Economic Journal, 131(635):1105{1134.
DANE (2020). Proyecciones y retroproyecciones de poblaci on.
https://www.dane.gov.co/index.php/estadisticas-por-tema/demograa-y-
poblacion/proyecciones-de-poblacion (Accessed August 4, 2020).
DANE (2023). Proyecciones de Poblaci on. https://www.dane.gov.co/index.php/estadisticas-por-
tema/demograa-y-poblacion/proyecciones-de-poblacion (Accessed March 15, 2023).
De La O, A. L. (2013). Do conditional cash transfers aect electoral behavior? Evidence from a
randomized experiment in Mexico. American Journal of Political Science, 57(1):1{14.
Dell, M. and Querub n, P. (2018). Nation building through foreign intervention: Evidence from
discontinuities in military strategies. The Quarterly Journal of Economics, 133(2):701{764.
DellaVigna, S., List, J. A., Malmendier, U., and Rao, G. (2016). Voting to tell others. The Review
of Economic Studies, 84(1):143{181.
Dixit, A. and Londregan, J. (1996). The determinants of success of special interests in redistributive
politics. The Journal of Politics, 58(4):1132{1155.
Downs, A. (1957). An economic theory of democracy. New York, Harper.
DPS (2013a). Denici on de puntos de corte del Sisb en III para el Programa M as Familias en
Acci on. Technical report, Departamento para la Prosperidad Social.
DPS (2013b). Redise~ no del programa Familias en Acci on. Technical report, Departamento para la
Prosperidad Social.
Duarte, R., Finan, F., Larreguy, H., and Schechter, L. (2019). Brokering votes with information
spread via social networks. Working Paper.
108
Duque Daza, J. (2019). Fraudulent electoral democracy. Electoral transhumance as a strategy to
win local elections in Colombia. Estudios Pol ticos, (55):61{86.
El Comercio (2014). Realizan operativos para combatir el voto golondrino.
https://elcomercio.pe/lima/realizan-operativos-combatir-voto-golondrino-347607-noticia/
(Accessed August 29, 2022).
El Heraldo (2015a). Colos o sale del anonimato por el trasteo masivo de elec-
tores. https://www.elheraldo.co/politica/coloso-sale-del-anonimato-por-el-trasteo-masivo-de-
electores-217085 (Accessed August 4, 2020).
El Heraldo (2015b). La trashumancia, un negocio \redondo" en Sucre.
https://www.elheraldo.co/trashumancia-electoral-en-la-region-caribe/la-trashumancia-un-
negocio-redondo-en-sucre-218355 (Accessed August 3, 2020).
El Heraldo (2015c). >C omo opera la cadena de la trashumancia y la compra de vo-
tos? https://www.elheraldo.co/politica/como-opera-la-cadena-de-la-trashumancia-y-la-compra-
de-votos-218349 (Accessed August 3, 2020).
Fajury, K. (2023). Clientelism and Poverty: Voter buying in Colombia. Unpublished Manuscript.
Fergusson, L., Molina, C., and Robinson, J. A. (2020). The weak state trap. Technical report,
National Bureau of Economic Research.
Ferraz, C. and Finan, F. (2008). Exposing corrupt politicians: the eects of Brazil's publicly
released audits on electoral outcomes. The Quarterly Journal of Economics, 123(2):703{745.
Ferraz, C. and Finan, F. (2011). Electoral accountability and corruption: Evidence from the audits
of local governments. American Economic Review, 101(4):1274{1311.
Finan, F. and Schechter, L. (2012). Vote-buying and reciprocity. Econometrica, 80(2):863{881.
Finan, F., Seira, E., and Simpser, A. (2021). Voting with one's neighbors: Evidence from migration
within Mexico. Journal of Public Economics, 202:104495.
Fouirnaies, A. (2018). How do campaign spending limits aect electoral competition? Evidence
from Great Britain 1885-2010. Working Paper.
Fowler, J. H. and Kam, C. D. (2007). Beyond the self: Social identity, altruism, and political
participation. The Journal of Politics, 69(3):813{827.
109
Frey, A. (2019). Cash transfers, clientelism, and political enfranchisement: Evidence from Brazil.
Journal of Public Economics, 176:1{17.
Fujiwara, T., Meng, K., and Vogl, T. (2016). Habit formation in voting: Evidence from rainy
elections. American Economic Journal: Applied Economics, 8(4):160{188.
Garbiras-D az, N. and Montenegro, M. (2022). All eyes on them: A eld experiment on citizen
oversight and electoral integrity. American Economic Review, 112(8):2631{68.
George, S., Gupta, S., and Neggers, Y. (2018). Coordinating voters against criminal politicians:
Evidence from a mobile experiment in India. Working Paper.
Gerber, A. S., Green, D. P., and Larimer, C. W. (2008). Social pressure and voter turnout: Evidence
from a large-scale eld experiment. American Political Science review, 102(1):33{48.
Gerber, A. S., Green, D. P., and Shachar, R. (2003). Voting may be habit-forming: Evidence from
a randomized eld experiment. American Journal of Political Science, 47(3):540{550.
Hidalgo, F. D. and Nichter, S. (2016). Voter buying: Shaping the electorate through clientelism.
American Journal of Political Science, 60(2):436{455.
IBCE (2019). Piden investigar traslado de votantes de riberalta a pando.
https://ibce.org.bo/principales-noticias-bolivia/noticias-nacionales-detalle.php (Accessed
August 29, 2022).
J-PAL Governance Initiative (2019). Governance initiative review paper. Technical report, J-PAL.
Kaplan, E., Saltiel, F., and Urz ua, S. S. (2023). Voting for Democracy: Chile's Plebiscito and the
Electoral Participation of a Generation.
Karp, J. A. and Brockington, D. (2005). Social desirability and response validity: A comparative
analysis of overreporting voter turnout in ve countries. The Journal of Politics, 67(3):825{840.
Kasara, K. and Suryanarayan, P. (2015). When do the rich vote less than the poor and why?
Explaining turnout inequality across the world. American Journal of Political Science, 59(3):613{
627.
La Campana (2015). En buses y canoas se dio en el Cauca el trasteo de vo-
tos. https://www.periodicolacampana.com/en-buses-y-canoas-se-dio-en-el-cauca-el-trasteo-de-
votos/ (Accessed August 4, 2020).
110
Larreguy, H., Marshall, J., and Querubin, P. (2016). Parties, brokers, and voter mobilization:
How turnout buying depends upon the party's capacity to monitor brokers. American Political
Science Review, 110(1):160{179.
Manacorda, M., Miguel, E., and Vigorito, A. (2011). Government transfers and political support.
American Economic Journal: Applied Economics, 3(3):1{28.
Mart nez, L. R. (2016). Sources of revenue and government performance: evidence from Colombia.
Working Paper.
McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity design: A
density test. Journal of Econometrics, 142(2):698{714.
Medell n, N. and S anchez, F. (2015). >C omo funciona M as Familias en Acci on?: Mejores pr acticas
en la implementaci on de programas de transferencias monetarias condicionadas en Am erica
Latina y el Caribe. Nota T ecnica, (884).
Milligan, K. and Rekkas, M. (2008). Campaign spending limits, incumbent spending, and election
outcomes. Canadian Journal of Economics/Revue canadienne d' economique, 41(4):1351{1374.
MOE (2016a). Entre Reglas e Incertidumbres: Financiaci on de las campa~ nas pol ticas y la rendici on
de cuentas en Colombia. Misi on de Observaci on Electoral.
MOE (2016b). Resultados Electorales en Colombia: Elecciones de Autoridades Locales 2015. MOE.
Montero, D. (2011). Los turistas electorales de Cabuyaro: Un pueblo con trasteo de votantes. La
Silla Vacia https://lasillavacia.com/historia/los-turistas-electorales-de-cabuyaro-un-pueblo-con-
trasteo-de-votantes-24332 (Accessed August 4, 2020).
Muralidharan, K. and Niehaus, P. (2017). Experimentation at scale. Journal of Economic Per-
spectives, 31(4):103{24.
Muralidharan, K., Niehaus, P., Sukhtankar, S., and Weaver, J. (2021). Improving last-mile ser-
vice delivery using phone-based monitoring. American Economic Journal: Applied Economics,
13(2):52{82.
MVSNoticias (2018). Tepjf ordena al ine sancionar al pri por turismo electoral.
https://mvsnoticias.com/nacional/2018/4/26/tepjf-ordena-al-ine-sancionar-al-pri-por-turismo-
electoral-371757.html (Accessed August 29, 2022).
Nichter, S. (2018). Votes for survival: Relational clientelism in Latin America. Cambridge Univer-
sity Press.
111
OECD (2020). Minimum relative to average wages of full-time workers, OECD.Stat.
https://stats.oecd.org/Index.aspx?DataSetCode=MIN2AVE (Accessed August 10, 2020).
Olken, B. A. (2007). Monitoring corruption: Evidence from a eld experiment in Indonesia. Journal
of Political Economy, 115(2):200{249.
Pach on, M. and S anchez, F. (2014a). Base de datos sobre resultados electorales CEDE, 1958{2011.
Documentos CEDE.
Pach on, M. and S anchez, F. (2014b). Base de datos sobre resultados electorales CEDE 1958-2011.
Serie documentos Cede, 2014-29.
Prakash, N., Rockmore, M., and Uppal, Y. (2019). Do criminally accused politicians aect economic
outcomes? Evidence from India. Journal of Development Economics, (forthcoming).
Prat, A. (2002). Campaign advertising and voter welfare. The Review of Economic Studies,
69(4):999{1017.
Radio Santa Fe (2015). Medell n salen 12 buses con ciudadanos que votar an.
http://www.radiosantafe.com/2015/10/24/medellin-salen-12-buses-con-ciudadanos-que-
votaran/ (Accessed August 3, 2020).
Rubiano, D. A. (2015). Mapas y factores de riesgo electoral. Elecciones de autoridades locales
Colombia 2015, chapter Nivel de Censo Electoral frente al tama~ no de la poblaci on: Riesgos
desde la zonicaci on y el cambio de la residencia electoral. MOE.
Rueda, M. R. (2017). Small aggregates, big manipulation: Vote buying enforcement and collective
monitoring. American Journal of Political Science, 61(1):163{177.
Ruiz, N. (2018). The power of money. The consequences of electing a donor funded politician.
Working Paper.
Schafer, J., Cantoni, E., Bellettini, G., and Berti Ceroni, C. (2022). Making unequal democracy
work? The eects of income on voter turnout in Northern Italy. American Journal of Political
Science, 66(3):745{761.
Shen, Z., Zhang, K., and Dell, M. (2020). A large dataset of historical Japanese documents with
complex layouts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition Workshops, pages 548{549.
112
Shen, Z., Zhang, R., Dell, M., Lee, B. C. G., Carlson, J., and Li, W. (2021). LayoutParser: A
unied toolkit for deep learning based document image analysis. In Document Analysis and
Recognition{ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5{
10, 2021, Proceedings, Part I 16, pages 131{146. Springer.
Shleifer, A. and Vishny, R. W. (1993). Corruption. The Quarterly Journal of Economics,
108(3):599{617.
Speck, B. (2013). Money in politics: Sound political competition and trust in government. Technical
report, OECD. Background Paper.
Stokes, S. C. (2005). Perverse accountability: A formal model of machine politics with evidence
from Argentina. American Political Science Review, 99(3):315{325.
Stokes, S. C., Dunning, T., Nazareno, M., and Brusco, V. (2013a). Brokers, voters, and clientelism:
The puzzle of distributive politics. Cambridge University Press.
Stokes, S. C., Dunning, T., Nazareno, M., and Brusco, V. (2013b). Brokers, voters, and clientelism:
The puzzle of distributive politics, chapter 6. Cambridge University Press.
Vaishnav, M. (2017). When crime pays: Money and muscle in Indian politics. Yale University
Press.
Vannutelli, S. (2021). From lapdogs to watchdogs: Random auditor assignment and municipal
scal performance in italy. Working Paper.
Vicente, P. (2014). Is vote buying eective? Evidence from a eld experiment in West Africa. The
Economic Journal, 124(574):F356{F387.
Weber, M. (1946). From Max Weber: Essays in Sociology, chapter Politics as a Vocation. Oxford
University Press. Translated by H.H. Gerth and C. Wright Mills.
World Bank (2017). World development report 2017: Governance and the law. The World Bank.
113

Abstract (if available)

Abstract This thesis gathers three papers which study how incentives and socioeconomic circumstances shape electoral participation and corruption.
The first paper is motivated by the fact that high income countries have lower levels of corruption, but it is still unclear whether higher incomes causally reduce the propensity to engage in corruption. This paper studies the effect of income on corruption in elections. To do so, I digitize novel information on vote buying from official records and exploit a nationwide cash transfer program in Colombia. Using a discontinuity that determines eligibility for the transfer in a regression discontinuity design, this paper finds that this increase in income reduced the likelihood of someone having their vote bought by 18 to 27%. I show that this is explained by an income effect rather than other mechanisms such as becoming more civically engaged. Back of the envelope calculations indicate that around 30,000 votes were not sold as a result of the program. I also document how other characteristics, such as closeness of elections, affect the political economy of vote buying, and, therefore, the effectiveness of the cash transfers in reducing this form of clientelism.
The second paper studies the unintended consequences of political campaign spending limit regulations in Colombia. To do so, I use discontinuities in electoral campaign spending limits to estimate their causal effects on organized violence and vote buying. The campaign spending limits are determined according to arbitrary cutoffs of potential voters for each municipality. Using a pooled Regression Discontinuity Design, I find that tighter campaign spending limits increase organized violence and vote buying. The results suggest that tight campaign spending limits can lead to undesirable outcomes, such as the use of fraudulent or illegal practices as a substitute to legal financial resources to attract voters for political gain.
The third paper describes the construction of a database -at the individual level- of voter turnout for Colombia. I use over a million pages of scanned official voting records and digitize the information using computer vision and optical character recognition techniques. I also provide an example on how the data can be used and motivate future research.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on development economics and adolescent behavior

PDF

Social policy and presidential ideology in Latin America: the political economy of social spending and anti-poverty programs

PDF

Essays on political economy of privatization

PDF

Essays on applied microeconomics

PDF

Politics behind appointments: policy, patronage, and party systems in the Andes

PDF

The political economy of R&D collaboration: micro- and macro-level implications

PDF

Three essays on the microeconometric analysis of the labor market

PDF

Essays in the economics of education and conflict

PDF

Voter engagement in the 2020 presidential election: what we can learn from the voter engagement initiatives in California-based and national nonpartisan organizations to increase voter participation

PDF

The role of local actors in international development: how incentive types affect foreign aid effectiveness

PDF

Three essays on the political economy of federal spending

PDF

Essays on macroeconomics and income distribution

PDF

Creating destruction: the political economy of zombie firms

PDF

Essays on sovereign debt

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Essays in political economy and mechanism design

PDF

Essays in labor economics: demographic determinants of labor supply

PDF

Essays on mitigation, adaptation, and resilience to urban fiscal crises

PDF

Can election reforms improve democracy?

PDF

Politics is something we do together: identity and institutions in U.S. elections

Asset Metadata

Creator Fajury, Karim (filename)

Core Title Essays on political economy and corruption

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2023-05

Publication Date 04/11/2023

Defense Date 03/23/2023

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag campaign spending,cash transfer,Colombia,Corruption,OAI-PMH Harvest,Political Economy,vote buying,voter turnout

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Weaver, Jeffrey (committee chair), Bassi, Vittorio (committee member), Munck, Gerardo (committee member), Nugent, Jeff (committee member)

Creator Email fajury@usc.edu,ka.fajury312@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113004113

Unique identifier UC113004113

Identifier etd-FajuryKari-11597.pdf (filename)

Legacy Identifier etd-FajuryKari-11597

Document Type Dissertation

Format theses (aat)

Rights Fajury, Karim

Internet Media Type application/pdf

Type texts

Source 20230411-usctheses-batch-1019 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu