LAW-AND-ORDER POLITICS AND PUBLIC SECURITY MISMANAGEMENT

The focus of this article is politicians who during the electoral campaign highlight their professional background in law enforcement or in the military. These are law-and-order candidates. They use this signaling to show voters that they are committed to acting on public security and that if elected they will use all the power they have to champion policies in that policy area.Footnote ⁴ While for most candidates these powers are circumscribed by the laws and regulations of their political office, a subgroup of law-and-order candidates has an additional political asset that allows them to go beyond the formal capacities of ordinary law-and-order politicians. These are the politicians embedded in the public security bureaucracy: law-and-order candidates who work or have worked in the police.

Differently from the bureaucratic embeddedness stemming from the social relationships between bureaucrats and civil society (Grossman and Slough Reference Grossman and Slough2022; Pepinsky, Pierskalla, and Sacks Reference Pepinsky, Pierskalla and Sacks2017), the relationship between the politician and the bureaucracy is what defines embeddedness here. The professional experience of certain politicians allows them to extract hard-to-obtain knowledge of the bureaucracy’s procedures, and establish lasting relationships with its bureaucrats. When they become politicians, their embeddedness becomes an asset, as this privileged information gives them knowledge of the bureaucracy’s internal hierarchies, decision-making processes, resource constraints, and how much its norms and rules can be bent or broken with minimal risk of sanction or detection. Either through persuasion, coercion, or affection from former colleagues, politicians can use their embeddedness to influence the bureaucracy toward their political ends. Politicians will activate their embeddedness in the bureaucracy they once belonged to when favoring a specific portion of the electorate increases the chances of winning re-election. Politicians will employ favoritism when voters’ support depends on their honoring their commitment to act on the policy area they come from. Embedded politicians who have not forged these programmatic linkages will not have the same incentives to use their connections to capture policymaking because voters do not expect them to do so. For bureaucrats, having an embedded politician from their agency serving office can be advantageous, resulting in more resources, career opportunities, and job security.

Using embeddedness in the police to favor a segment of the electorate of voters is not a winning strategy for some politicians because favoritism distorts public security. While bureaucratic embeddedness may improve efficiency in policymaking (Evans Reference Evans2005), politicians’ embeddedness misallocates resources, generating a worse overall policy outcome. In public security, where policing is the crucial service, rearranging resources from one place to another will displace crime to neglected areas (Blattman et al. Reference Blattman, Green, Ortega and Tobón2021). It may also disturb the equilibrium among criminals, generating turf wars (Trejo and Ley Reference Trejo and Ley2020; Vargas Reference Vargas2016). These side effects compound given that those who support law-and-order politicians are relatively wealthier citizens or businesses (Gonzalez and Mayka Reference Gonzalez and Mayka2023; Lacey and Soskice Reference Lacey and Soskice2015; Moncada Reference Moncada2016). Young, poor men are those most likely to commit crimes and suffer violent deaths (Reza, Mercy, and Krug Reference Reza, Mercy and Krug2001), but they are not part of the law-and-order constituency. The favoring of law-and-order voters leads to a greater focus on the protection of property in the wealthier district, but also to less policing where individuals vulnerable to homicidal violence live. For these reasons, the number of citizens who suffer from favoritism might be larger than the number of voters who benefit from it. When running for office in races that require a majority of voters, playing favorites may backfire. However, in elections where officials are selected through proportional representation, catering to a fraction of all voters can deliver a victory. Politicians under these electoral rules, then, can concentrate their efforts on narrower demands and successfully win re-election (Myerson Reference Myerson1993).Footnote ⁵

Since favoritism puts particular interests above public governance, it depends on deficient bureaucratic accountability and transparency. Private information is a key factor in models of regulatory capture because if the government knows that the regulator and firms are in collusion against the public, it can prevent capture (Dal Bó Reference Dal Bó2006). But thanks to attributes that are common to police forces from many countries, changes in policing—for example, the location of police patrols, security cameras, outposts, and checkpoints—can be implemented without detection. The police notoriously receive insufficient oversight (Eckhouse Reference Eckhouse2022; González Reference González2020; Willis Reference Willis2014; Reference Willis2015), are seldom held accountable for transgressions (Brinks Reference Brinks2007; Cabral and Lazzarini Reference Cabral and Lazzarini2015), and are corrupt, unprepared, and brutal (Flom and Post Reference Flom and Post2016; Hinton and Newburn Reference Hinton and Newburn2008). Still, that does not make it easy for politicians to influence the police. Bureaucracies actively try to block politicians from meddling with their decisions (Flom Reference Flom2020; González Reference González2020; Pepinsky, Pierskalla, and Sacks Reference Pepinsky, Pierskalla and Sacks2017). Yet a former police officer is not an ordinary politician. The police force and their strong esprit-de-corps (Hinton Reference Hinton2005, 78) are open to their former colleagues, who may still be part of the police inner circle. In sum, the embedded politician is already inserted in the policing black box, allowing them to misallocate resources without the public and other members of the government noticing.

Police reform has proved difficult, but local measures can overcome the institutional drift from other federal spheres, remedying the lack of transparency in public security. Given the threat of capture by embedded law-and-order politicians, local institutions give other actors incentives to pursue ends that are not aligned with the individual goals of the candidate. Since these institutions produce information (González Reference González2019; Moncada Reference Moncada2009), they reduce the informational asymmetry between local actors and the police (while not necessarily improving public security delivery). Consequently, they increase the likelihood that any effort to capture the police will be detected and stopped.

In the next subsection, I will explain how law-and-order politics in Brazil is conducive to favoritism. First, I will present a brief overview of public security in Brazilian municipalities, and the scope of action of municipal councilors. Next, I will carefully validate the measure of law-and-order politicians I use.

Law-and-Order Politics in Brazilian Municipalities

Brazil is an unfortunate but ideal place to study politics and security. Levels of violence in the country are comparable to those of areas under civil conflict (Magaloni, Franco-Vivanco, and Melo Reference Magaloni, Franco-Vivanco and Melo2020) and the issue was ranked high in voters’ concerns during the 2018 elections.Footnote ⁶ The Brazilian police force is one of the most lethal police forces in the world (Bueno Reference Bueno, de Lima, Ratton and Azevedo2014b) and the criminal justice system investigates police killings less thoroughly than homicides committed by civilians (Willis Reference Willis2015). Brazilian police have also been noted for concentrating abuses in poorer areas (Caldeira Reference Caldeira2000; Magaloni, Franco-Vivanco, and Melo Reference Magaloni, Franco-Vivanco and Melo2020).

The context of Brazilian law enforcement allows for a test of how the relational capital of some law-and-order candidates affects public security outcomes. In Brazilian federalism, policing is mostly the responsibility of state governments, which control the investigative plainclothes branch (polícia civil), and uniformed patrolling police (polícia militar—henceforth only police).Footnote ⁷ The police are four times more numerous, thus more visible, than their investigative counterpart, and, following state constitutions, are the official guarantors of order. State government authorities dictate statewide strategies and coordinate resources, then police chiefs (“coronéis”) oversee regional police divisions (called “battalions”), and then at the municipality level, local police chiefs decide how to police neighborhoods given the number of police officers the state authority has decided to allocate to the municipality. The formal separation between patrolling and investigative branches leads to rivalry and competition (Medeiros Reference Medeiros2004). Military law-and-order candidates have no formal connections to law enforcement.

This article will analyze how the election of law-and-order municipal councilors affects homicides. Councilors are the most numerous elected officers in Brazil, but also the least powerful. Council candidates are often facing hundreds of adversaries in open-list proportional representation elections (see Table A.1 in the Supplementary Material). In contrast to mayoral candidates who on average compete among fewer than three candidates in first-past-the-post elections, council candidates can appeal to a select group of the municipality and still win a seat, allowing these candidates to “cultivate favored minorities” (Myerson Reference Myerson1993).Footnote ⁸ Councilors are elected when their ranking within a list of candidates (usually from a coalition of many parties) is above or equal to the number of seats allocated to that list. The ranking is defined by the number of votes candidates receive, encouraging council candidates to run individualistic campaigns. Local legislatures in Brazil vary in size according to the municipal population and 90% have a minimum of nine councilors (Mignozzetti, Cepaluni, and Freire Reference Mignozzetti, Cepaluni and FreireForthcoming).Footnote ⁹

Councilors have limited scope for influencing municipal policy. They can propose municipal laws and legislate local taxes, but cannot individually propose budgetary amendments. Councilors’ staffs lack professional expertise, and their legislative activity often centers on public honoring ceremonies (Silva Reference Silva2014). Clientelistic exchanges between councilors and voters are common during and after elections (Kerbauy Reference Kerbauy2005; Lopez and Almeida Reference Lopez and Almeida2017). As explained later, municipal governments can take more responsibility for public security, and Municipal Guards are part of a decentralizing push and are entirely locally managed (Arvate and Souza Reference Arvate and Souza2022). Councilors can support the creation of municipal guards, but ultimately it is the executive that can initiate their establishment. The largest asset a law-and-order candidate may have to affect local public security is relational. Police officers are allowed to run for office and if they win, they must take leave without pay from the police for the duration of their term. Military personnel can only run for office after retirement.Footnote ¹⁰

Law-and-Order Candidates and Platforms in Brazil

To classify law-and-order candidates, I use the ballot names candidates choose to communicate with voters. The Brazilian political system is a low-information environment, but using ballot names to convey signals to voters is common (Boas Reference Boas2014). Candidates often use the opportunity to tell voters they are, for example, part of the clergy, school teachers, or physicians. Many make reference to the police or military forces by including titles such as captain, colonel, corporal, detective, lieutenant, private, or sheriff in their ballot names. Candidates who use these titles are the law-and-order candidates in this article.Footnote ¹¹ Following this criterion, I was able to identify 7,888 law-and-order council candidates in the 2004–12 period. Of the 5,570 municipalities in Brazil, 2,491 have had a law-and-order candidate at some point and 594 had at least one elected.

But who among these law-and-order candidates are embedded in the police? To separate among subtypes of law-and-order candidates, I use the information they listed as an occupation. Among these, I classify three different subtypes of law-and-order candidates. First, there are 2,286 law-and-order candidates who reported being employed by the police. They correspond to the embedded law-and-order politician because they use a law-and-order ballot name and have worked in the police. Henceforth, police law-and-order candidates correspond to the embedded politician. Second, there are 752 investigative police officers (reported as polícia civil employees), and third, 2,561 nonpolice candidates with varying backgrounds that include former military (but not listed being public state employees, since these could still be police officers). Almost 10,000 law enforcement officers or military did not run as law-and-order candidates; that is, they listed their occupations as law enforcement or military but did not use a law-and-order ballot name as a candidate. By avoiding the use of law-and-order ballot names, they are revealing that they are pursuing an electoral strategy that is different from law-and-order politics. These will serve as a contrasting group to law-and-order candidates since according to the argument this article raises, their election should not affect violence. Additionally, those in this group from the police are embedded politicians, but according to the argument I presented, they do not use this asset for their electoral ends.

President Bolsonaro’s political trajectory helps illustrate how the classification works. In 1988, when he ran for Rio de Janeiro municipal council as “Capitão Bolsonaro,” he was a law-and-order candidate. He was not an embedded politician, since his rank as captain comes from his army service. He dropped the “Captain” for all other elections, and would no longer be classified as a law-and-order candidate. Another helpful example would be the mayor of New York, Eric Adams. He did not run for office in 2021 using his rank as a retired police officer, so would not be a law-and-order candidate here. However, he is an embedded politician, having served in the police for many years before running for office. In races such as New York’s, candidates may have more opportunities to advertise their platforms, finding other ways to show voters their commitment to public security. Even if these candidates would not enter the data here as law-and-order candidates, this type of misclassification is not a concern for identification due to the research design employed in the article.

Other ways in which candidates in Brazil communicate their law-and-order platform help validate the operationalization of law-and-order candidacy employed here. First, using publicly available campaign platforms of mayoral candidates, I check if law-and-order candidates use more security-related words.Footnote ¹² For each candidate, I count the number of public security keywords these candidates used.Footnote ¹³ Of around 30,000 mayoral candidates in 2012 and 2016 combined, 132 used law-and-order ballot names and 220 candidates stated past law enforcement or military occupation but did not advertise it through their name.

The text analysis of campaign platforms reveals that law-and-order candidates (of all subtypes) use words related to public security more often than candidates with a past in law enforcement without a law-and-order ballot name. The estimations in Table 1 confirm that ballot-name candidates employ public security words more often.Footnote ¹⁴ Compared with all other candidates, who are the baseline group in columns 1–3, law-and-order ballot names are associated with twice as many public security words. Importantly, there is a considerable and large difference between ballot and nonballot name candidates, as estimations that only include these candidates (columns 4–6) indicate.

Table 1. Unique Public Security Words Used in Campaign Manifestos

Note: Robust SEs are reported in parentheses. *** $ p<0.01 $ ; ** $ p<0.05 $ ; * $ p<0.10 $ .

Next, I check how elected candidates use social media to claim credit for achievements in public security. In order to make a credible commitment to law and order, candidates also need to show voters they are pursuing measures in the area. Nowadays, it is common for politicians in Brazil to advertise their purported accomplishments through social media (Bueno Reference Bueno2021). Law-and-order candidates may use the Internet to report that they have secured funds for CCTV systems, increased the ranks of the local police, protested for better wages and conditions for police officers, or participated in local security councils.Footnote ¹⁵ To systematically check how law-and-order politicians communicate with their public and contrast these candidates with those who are law enforcement professionals, but not law-and-order, I randomly selected 150 elected council candidates from the 2020 elections, half of whom are law-and-order candidates, and the other half, law enforcement without a law-and-order ballot name. I was able to retrieve 126 Instagram or Facebook personal accounts, and coded postings from the beginning of their term in January 2021 to February 2022. Credit claiming in public security is common among all of these candidates, but even more common among law-and-order candidates. I found that 41.3% of candidates that are not law-and-order have at least one social media posting showcasing achievements in public security. This number, however, climbs to 74.6% among law-and-order candidates.

Importantly, police officers and police chiefs figure frequently in the social media of law-and-order incumbents. Advertising a link with the police is a strategic choice that signals to voters that the candidate has the attention of those individuals who carry out the most important aspect of crime prevention. In the context of political favoritism, it might reveal a continued connection between police officers and their former work colleagues. To make a stronger test of police connections while in office, I randomly selected 160 former police candidates, either law-and-order or not. Law-and-order police candidates have their pictures taken with other police more often. Around 69% of law-and-order police incumbents posted pictures with patrolling police officers, 31.5 percentage points more than from accounts of police incumbents that did not use a ballot name. The two groups of police candidates are also substantially different when presenting connections to high-ranking police officers. Around 55.2% of law-and-order police candidates publicized meetings with police chiefs, such as battalion commanders (coroneis de batalhão), whereas 34.5% of ordinary police candidates had pictures with chiefs.Footnote ¹⁶

The empirical analysis only examines municipalities that have had a law-and-order candidate for the municipal council. These municipalities are not a random assortment of the median Brazilian municipality, as Table A.1 in the Supplementary Material attests. They have larger populations and are more violent, especially where a competitive candidate has run. Although this article does not consider the reasons behind why law-and-order candidates emerge, it is plausible that law-and-order platforms appeal more to voters as municipalities grow in population. Population size is related to graver public security concerns, and also have more public security resources (Post and Kuipers Reference Post and Kuipers2022), which police candidates can potentially capture. In municipalities with candidates that listed law enforcement as an occupation but refrained from using a law-and-order ballot name, the median population is 25% smaller than those with law-and-order candidates, suggesting that the decision to run under a security platform is more attractive in larger, more violent places.

Local Security Committees

The lack of comprehensive police reform harms public security at the local level. However, municipalities can remedy their lack of control over opaque state police by taking on some responsibilities in the area of public security. The National Fund for Public Security allows municipalities to request grants for security projects, but approvals are not automatic. They depend on local efforts to implement institutions that manage, oversee, and place local stakeholders in the policymaking process. But these local reforms are not guaranteed to improve public security. For example, Gonzalez and Mayka (Reference Gonzalez and Mayka2023) show that community security councils in São Paulo can increase demands for repression of marginalized groups. Despite the uncertain effects these institutions may generate on crime and violence, local institutions are a roadblock against police capture. They lock in local priorities in comprehensive plans and budgets, inserting priorities that may go against any attempt from law-and-order politicians to use public security for electoral gain. Importantly, these institutions establish formal channels through which local stakeholders oversee police activity, making other local actors attentive to the misallocation of security resources.

The local security committees (Conselhos Municipais de Segurança Pública) are one of these measures. They do not have formal powers over state police, but they can reduce the ability of the police to perform exceptional, violent tasks (Taylor Reference Taylor2011) unnoticed, and make any type of politicization of public security harder. According to the IBGE, in 2014, 71% of all established committees reported having met in the preceding year. These security committees are now widespread in Brazil, as the map in Figure A.1b in the Supplementary Material shows, and are increasing rapidly. A qualitative assessment of 25 committees from all regions of Brazil shows that around 90% of the committees include representatives from the local government, judiciary, and law enforcement agencies. On average, four different groups have reserved seats on the committee table.

After municipalities establish a committee, they are likely to have three other institutions, too. First, municipalities can write a multiyear public security goal-setting plan. These plans first diagnose local security problems, establish how local resources can be applied to help, and may recommend specific partnerships with state police where local resources are insufficient. Second, they put together a dedicated public security budget to finance new measures or those recommended in the multiyear plan. Finally, municipalities can help citizens create a nongovernmental security council, whose members often have a seat on the security committee’s table.Footnote ¹⁷ Figure A.2 in the Supplementary Material shows the relationship between committees and these other institutions.

However, municipalities do not create committees at random and other factors could explain differences in outcomes. For example, as Table A.3 in the Supplementary Material shows, municipalities without committees are where political favoritism can do the most harm, as on average they have a larger proportion of non-white residents. Additionally, these committees can be effective in preventing mismanagement caused by the election of law-and-order politicians, but they do not a priori remedy past distortions in the provision of public security, at least not in a manner that communities would take note of and, as a result, start supporting law-and-order platforms (evidence in Appendix A.5 of the Supplementary Material suggests that indeed is not the case).

Homicides in Brazil and Other Data

This subsection describes the data on homicides and shows that the group most likely to suffer from homicidal violence is the least likely to be the law-and-order constituency. The Ministry of Health collects mortality statistics through the Brazilian System of Death Registration (SIM/Datasus).Footnote ¹⁸ In contrast to crime data and police reports, every death is documented and the records include the International Statistical Classification of Diseases and Related Health Problems (ICD-10) classification. I include all deaths involving third-party aggression.Footnote ¹⁹ From 2000 to 2016, there have been 1.7-million deaths due to aggression. According to regulations, a coroner or an appointed physician must investigate deaths that involve aggression. Cause of death comes from a medical examination of the body, and through interviews with family members, witnesses, and the police. This decision sometimes contains errors (Cerqueira Reference Cerqueira2012). Death records also list the victim’s race.Footnote ²⁰ The National Treasury provides data on municipal spending on public security.Footnote ²¹ All demographic information, which includes Gini coefficients, population, and population according to race, among others, comes from the IBGE.Footnote ²² Municipal racial composition comes from census data and the proportions of white and non-white municipal residents used in the estimates (as well as any other covariate) are always pretreatment measurements.

THE REGRESSION DISCONTINUITY DESIGN

The research design aims to estimate the effect of electing a law-and-order candidate in municipalities that had a law-and-order candidate running for office. Next, it tests whether law-and-order candidates from the police are associated with more violence in comparison to other types of law-and-order candidates. Finally, it compares municipalities that had previously established a security committee with those that did not. The main estimations leverage a regression discontinuity design (RDD). RDDs can alleviate the problem of confounders and produce unbiased estimates under few, testable assumptions (Lee and Lemieux Reference Lee and Lemieux2010). Many factors drive public security outcomes at the district level, and without accounting for municipal demographics, such as social inequality, employment level, and other variables, the analysis would be confounded. For example, it is possible that municipalities only choose law-and-order candidates after a surge in homicides. RDDs exploit the fact that the treatment assignment, electing a law-and-order candidate, is discontinuous around a vote margin cutoff, but the potential outcomes of violence are continuous, making the observed outcomes in one group comparable to the unobserved potential outcome of units in the other. One caveat in RD designs is the local nature of estimated effects. Effects should be interpreted as causal within the subgroup of municipalities that had a law enforcement candidate elected or not elected by a small margin.

The estimand is the effect of a municipality electing a law-and-order council candidate. To estimate it, we need to define a score, which will depend on the competitiveness of the candidate in an open-list proportional representation system. In the Brazilian open list proportional representation system (OLPR), candidates are ranked solely according to the personal votes they receive from voters, without interference from party leaders. Following Boas, Hidalgo, and Richardson (Reference Boas, Hidalgo and Richardson2014), I define the score as the distance, in the percentage of valid votes, between the winning law-and-order candidate and most-voted losing candidate on a list.Footnote ²³ When the law-and-order candidate loses, the score is the distance between this candidate and the least-voted winning candidate on the list. Only municipalities that had a list that received a seat and had a law-and-order candidate will have a score and are thus included in estimations.Footnote ²⁴ To compare effects across groups of law-and-order candidates, I calculate a score based on each type of law-and-order candidate—police, investigative police, and others. As discussed in Boas and Hidalgo (Reference Boas and Hidalgo2011, 873–4) and shown in Table A.1 in the Supplementary Material, smaller vote share races tend to occur in larger municipalities, which is also the group more likely to have a law-and-order candidate.Footnote ²⁵

This type of design has two complications that are important to recognize and address. First, law-and-order candidates may be different from other candidates not only because they have law-and-order ballot names, but because they also differ in other characteristics. Thus, the causal effect the RDD uncovers is bundled with these different attributes. I show, however, that law-and-order candidates are not different from other candidates with respect to several observable characteristics. Because their differences may not be observable, I compare the effect of electing a law-and-order candidate with the effect of electing a law enforcement officer that does not use a law-and-order ballot name. These two sets of candidates share many characteristics, such as their professional past and social network. If these common traits, not the political platform, drive the results, tests with both groups should yield similar effects.

Second, if law-and-order candidates have specific attributes that make them more likely to appear in close elections, the local average treatment effect (LATE) might be biased (Marshall Reference Marshall2022). However, and contrary to single-member plurality races, the OLPR system makes it plausible that compensating differentials of law-and-order candidates do not strongly affect their vote shares, an assumption that, if met, is enough for identification (Marshall Reference Marshall2022). Voters are largely unaware of the competitiveness of candidates before the election, and thus unlikely to be strategic in their vote. There are no registered pre-election surveys and the multiple lists are populated by a large number of candidates, many of which are newcomers and whose electoral worth is unknown. The large district magnitude and the complexity of the D’Hondt rules that determine who wins a seat make it unlikely that voters know who is competitive or not. In sum, there is little reason to expect that attributes of law-and-order candidates correlate with them being in close contests. Robustness estimations in the Supplementary Material include models with municipalities with law-and-order candidates who are not in the office—a group with unproven electoral viability—and report results similar to the baseline models.

The basic regression discontinuity local-linear model that captures the causal effect of electing a law-and-order candidate is

(1) $$ \begin{array}{r}\Delta {Y}_{i_{t+1}}={\displaystyle \begin{array}{l}\alpha +{\beta}_1{LOC}_{i_t}+{\beta}_2{Margin}_{i_t}+{\beta}_3{LOC}_{i_t}\\ {} \times {Margin}_{i_t}+{\phi}_t+{\mu}_{i_t}\end{array}}\\ {}\forall \hskip0.3em i,t\hskip0.60em \mathrm{s}.\mathrm{t}.\,\left|{M}_{i_t}\right|<\epsilon, \end{array} $$

where $ {LOC}_{i_t} $ is a binary indicator equal to one if the law-and-order candidate won the election in t and $ {\beta}_1 $ measures its causal effect. $ {Margin}_{i_t} $ is score, $ {\phi}_t $ are the time fixed effects, and $ {\mu}_{i_t} $ is an error term. $ \left|{M}_{k{i}_t}\right| $ is the score for each municipality i with a list of candidates that won a seat and had a law-and-order candidate. It is possible for a municipality to have more than one law-and-order candidate. When all candidates lose, the margin of the municipality is the margin of the candidate closest to the cutoff. When more than one wins, the margin is also that of the candidate closest to the cutoff. If there are winning and losing candidates in the same municipality-year, I consider the winning candidates’ margin as the forcing variable. $ \epsilon $ is an arbitrarily small vote margin that defines the study group for each estimation.

The dependent variable $ \Delta {Y}_{i,{k}_{t+1}} $ measures the difference between the outcome in the period after the election and the outcome in the period before the election in municipality i. For example, for homicides, the dependent variable is the yearly homicide rate in the period after the elections minus the rate in the period before the election.Footnote ²⁶ With the exception of budget spending, the periods of analysis include all years in the study period except the election year.Footnote ²⁷

For the majority of the estimations, I deal with the issue of bandwidth selection using the nonparametric robust method devised in Calonico, Cattaneo, and Titiunik (Reference Calonico, Cattaneo and Titiunik2014). Particular to this article’s design, there is a risk that as bandwidths become larger, municipalities end up having more than one law-and-order candidate elected. This would not make results biased but would change the interpretation of the estimand. To also ensure that the estimations capture a study group close to the cutoff, I restrict bandwidth selection within municipalities whose law-and-order candidate won or lost by a 1.0% margin. Balance tests in the Supplementary Material confirm that there is no noticeable difference in the characteristics of municipalities and candidates between municipalities with winning or losing candidates, with the exception of their age (controlling for this covariate yields the same results). A density test in the Supplementary Material shows no sign of manipulation around the cutoff. It also shows that the density of the forcing variable is not centered on zero. For this reason, the nonparametric estimations select asymmetric bandwidths around the cutoff. Estimations with symmetric bandwidths in robustness tables, also presented in the Supplementary Material, show that these are almost invariably the same. These robustness tables also include estimations with a second-degree polynomial and a study group that only includes municipalities that do not have a sitting law-and-order candidate at the time of the election.

THE EFFECTS OF ELECTING A LAW-AND-ORDER CANDIDATE

The election of law enforcement candidates results in more homicides. Figure 1 plots several different estimations for homicides, from the smallest bandwidths at the left of each panel to increasingly larger bandwidths moving to the right. The effects are large: in the nonparametrically selected bandwidth (in red), approximately 14 more homicides per one hundred thousand inhabitants. The magnitude is in line with estimates from the Mexican drug war in Dell (Reference Dell2015). The median population in municipalities in the study group is almost 60,000, making this increase equivalent to ten murders per year in the median municipality. Appendix A.7 of the Supplementary Material shows that the results are robust to different polynomial specifications. Panel (b) presents the difference in homicide rates in a local randomization setup that more closely resembles an as-if random natural experiment (Dunning Reference Dunning2012). The difference of means between treatment and control groups reveals that the effect is significant even at only 10 observations closest to the cutoff. Panel (d) shows that most homicides are committed against non-white men, the group most vulnerable to violence but not the law-and-order constituency. Considering that in Brazil the correlation between skin color and income is high (Bueno and Dunning Reference Bueno and Dunning2017), when analyzing heterogeneity according to skin color, one is examining heterogeneity by social class, too. These deaths are mostly the result of firearm injuries, and the effects do not extend to any group of women, as Figures A.8 and A.9 in the Supplementary Material show.

Figure 1. The Effect on Homicide Rates of Electing a Law Enforcement Candidate

Note: Both plots estimate local linear models. Bars in panel (a) represent 95% robust CIs. Nonparametric bandwidth selection is in red. Table A.13 in the Supplementary Material reports estimates (first column) and different specifications. In panel (b), point estimates are differences of means. Bars are 95% CIs. “Treatment Group” and “Control Group” count the total number of observations in each group. Panel (c) illustrates the discontinuity with binned averages and local linear regression lines and panel (d) illustrates the difference in homicide rates, non-white and white men.

Police law-and-order candidates drive the results. Figure 2 shows that compared with municipalities that almost elected a police candidate, those that did face a considerable increase in murders. Those that elected an investigative police officer also see an increase, but it is not a well-powered test (nor robust; see Table A.15 in the Supplementary Material). Elected law-and-order candidates from the military or any other professional background have no effect. Tellingly, not having a law-and-order ballot name returns a well-estimated null effect, which corroborates that it is the campaign platform, not the profession of the candidate, that drives the results.

Figure 2. Contrasting the Effect of Electing Different Types of Law-and-Order Candidates and Law Enforcement Candidates

Note: Nonparametric RDDs. Bars represent 95% robust CIs. Tables A.14–A.17 in the Supplementary Material report estimates (first columns) and different specifications. Difference between Police and Not law-and-order is 24.7 ( $ t=3.24 $ ).

Municipalities can reserve part of their budgets for security and all law-and-order candidates can demonstrate their programmatic commitment in pressuring for more public security spending (Wenzelburger Reference Wenzelburger2015). Figure 3 shows that municipalities that elect a law-and-order candidate spend around R$23 more per capita per year,Footnote ²⁸ representing a doubling of the R$19.4 average. In contrast, municipalities that elect law enforcement agents who do not use a law-and-order ballot name do not experience an increase in spending.Footnote ²⁹ This result shows that the choice of using a ballot name is related to the actions these two groups take on an outcome that is at hand for all, independent of their connections to the police.

Figure 3. Heterogeneous Effects in Public Security Spending at the Municipal Level, Candidates from Police or from Armed Forces with or without the Use of Ballot Name to Signal Law-and-Order Platform

Note: Nonparametric RDDs. Bars represent 95% robust CIs. Tables A.18–A.22 in the Supplementary Material report estimates (first columns) and different specifications. Difference between Law-and-order and Not law-and-order (but law enforcement) is R$26.21 ( $ t=2.31 $ ).

Local Security Committees

Figure 4 shows the effects of the electing of law-and-order candidates conditioned on the municipality having or not having local security committees. Consistently, results show that places without committees are subject to a homicide increase following the election of a law-and-order candidate and especially police law-and-order candidate. In places with a committee, the effect is a zero.Footnote ³⁰ Although nonobservable attributes may account for the differences across these groups, controlling for observables yields the same results.

Figure 4. The Effects of Electing a Law-and-Order Candidate and the Presence of Local Security Institutions

Note: Nonparametric RD models. Bars represent 95% robust CIs. Larger and smaller municipalities are above or below 50,000 population, respectively. Controls include GDP per capita, Gini index, proportion of non-white population, and population. Tables A.23–A. 28 in the Supplementary Material report estimates (first columns) and different specifications.

The Embeddedness of Police Candidates

This subsection aims to find evidence of a transactional relationship between the police and police-politicians where the side that receives favors is the police. Recent literature has shown that partisans of winning candidates have a much higher probability of landing a public job when compared with allies of the losing side (Colonnelli, Prem, and Teso Reference Colonnelli, Prem and Teso2020). Likewise, police law-and-order candidates might also try to distribute jobs to their former work colleagues.Footnote ³¹ Especially, if these candidates want to use police resources for their benefit, using public jobs might be a way to reciprocate.

To classify political appointees who had been previously employed by the police in a given year, I first collect information about all individuals employed by the police in each state.Footnote ³² Next, I check how many of these individuals received wages from the municipal public sector after the election of the law-and-order candidate. The dependent variable represents the variation from before and after the election in the number of new appointments of former or on-temporary leave police employees through the 4-year span, rescaled to reflect appointments per one hundred thousand residents per year.

The results show that the election of a law-and-order police candidate causes more appointments of former police employees after the election. Over 10 past police employees per one hundred thousand residents start receiving wages from the local government.Footnote ³³ The election of law-and-order investigative police, however, appears to reduce the number of police employees in the administration (Figure 5). The relocation of police employees is unlikely to generate violence since the effects only represent 5% of the average number of police officers per capita (in 2014, there were 212 officers per one hundred thousand Brazilians; IBGE 2014).

Figure 5. Election of Law-and-Order Police Officers and Appointments of Police Employees to the Municipal Government

Note: Nonparametric RDDs. Bars represent 95% robust CIs. Tables A.29–A.33 in the Supplementary Material report estimates (first columns) and different specifications. Difference between Police and Not law-and-order is 11.39 ( $ t=0.64 $ ).

POLITICAL FAVORITISM AND THE LAW-AND-ORDER CONSTITUENCY

Violence has risen steadily in Brazil during the 2007–17 period. But as Figure 6 shows, the increase concentrates on non-white men—a group that corresponds to roughly one quarter of the population but suffered 71% of all homicides. They are also poorer than white men, as income and race are highly correlated in Brazil. If the primary goal of public security is to prevent murders, focusing resources on this group should be a priority.

Figure 6. Daily Homicides in Brazil by Gender and Race

Note: Yearly moving average.

Law-and-order candidates’ linkages, however, point to other groups. Matching polling station geolocation with fine-grained census tract data, it is possible to characterize neighborhoods precisely according to income and demographic profile.Footnote ³⁴ To classify rich and poor households, I define those with less than half of the minimum wage per resident as poor. I classify households with more than five minimum wages per resident as rich. The greater the proportion of households below or above these thresholds in the vicinity of a polling station, the poorer or richer the neighborhood.

Next, I rank polling station neighborhoods according to the support law-and-order candidates received in the municipality. I include in a Low support group all polling stations that are either in the lowest quantile or that returned zero votes for law-and-order candidates. Georeferenced polling stations (locais de votação) give us a reliable measure of neighborhood support for law enforcement candidates if voters go vote in polling stations that are closest to their homes. That is not an unreasonable assumption. First, voters can choose where they want to vote in their municipality, except when their preferred place to vote is full.Footnote ³⁵ Second, election day is always a Sunday, giving few voters a reason to choose a place that is close to their job, instead of their home. Finally, comparing the scarce demographic information the electoral authority provides for each polling station with detailed census tract data, I show that the two are exceedingly similar (Tables A.8 and A.9 in the Supplementary Material).

To analyze electoral support within municipalities that elected a law-and-order candidate, I employ a linear model for polling stations with municipality fixed-effects estimations.Footnote ³⁶ The dependent variable is the Low support group. Due to the high collinearity between variables, I run individual models instead of a saturated model to capture each association in detail, standardizing the variables to facilitate cross-comparisons. It is evident in Figure 7 that Low support areas have higher proportions of poor, young, and non-white voters relative to other areas in the municipality. In other words, Low support areas concentrate individuals vulnerable to violence.

Figure 7. Correlates of (Lack of) Support for Law-and-Order Candidates in São Paulo State

Note: Point estimates of individual standardized variables in fixed-effect models. Bars represent 95% robust CIs. Table B.6 in Appendix B of the Supplementary Material reports estimates.

Political Favoritism across Neighborhoods

Do neighborhoods that did not support law-and-order candidates (Low support neighborhoods) receive less police attention and suffer more from homicides? As a testament to the opacity of policing in Brazil, data on police resources and activities do not exist. I circumvent this issue using data from the state of São Paulo, the only state to consistently provide (since 2011) crime and homicide data with latitude and longitude coordinates, as well as the police reports of these events.Footnote ³⁷

Since there is no information about where and when police officers are present in a given neighborhood, I create a novel measurement of police activity. Using car robberies data,Footnote ³⁸ I leverage the fact that police reports include information on whether the robbers were caught in the act by the police (flagrantes in Portuguese). The police activity variable is the ratio of robberies caught in the act over total robberies in a neighborhood. The greater the proportion of robberies that are caught on the spot, the most likely that the police are present and intervening in that area:

$$ \begin{array}{rll}{\mathrm{Police}\ \mathrm{act}\mathrm{ivity}}_{rp}=\frac{{(\mathrm{total}\ \mathrm{caught}\ \mathrm{in}\ \mathrm{the}\ \mathrm{act})}_{rp}}{{(\mathrm{total}\ \mathrm{robberies})}_{rp}}\times 100,& & \end{array} $$

where r is a radius of either 250, 500, or 1,000 meters, centered around polling station p. I assume that the higher this ratio, the more police activity around the polling station.

Next, I build a violence score that is the sum of all homicides within a specified radius around the polling station. Figure 8 illustrates this variation in homicide score for two selected polling stations in Mogi das Cruzes, a municipality 60 kilometers from the state capital. The maps show the distribution of homicides in the city and how violence is unevenly distributed. The dots are the two polling stations and the circles are the area within a 1,000-meter radius of the polling station. Jundiapeba, a relatively poor neighborhood on the west side of town, has many more murders—the X’s in the maps—than Vila Oliveira, a wealthier district to the east.

Figure 8. Homicides Near Selected Polling Stations in the City of Mogi das Cruzes

Note: The eastern dot on each map represents Professor Camilo Faustino de Mello Public School in the Vila Oliveira neighborhood. The western dot is the Professor Paulo Ferrari Massaro Public School in the Jundiapeba neighborhood. Shaded areas are distances within a 1-kilometer radius of polling stations. X’s represent homicides.

For both the police activity and homicide scores, I calculate the difference between after and before the election of a law-and-order candidate. To estimate within-municipality variation, I employ a municipal fixed-effect model that uses polling stations as units of analysis. Fixed effects absorb factors that are common to the whole municipality that could affect public security, such as the type of mayor or overall level of delinquency. The fitted model is

(2) $$ \begin{array}{r}\Delta {V}_{p,m,2016}={V}_{p,2016}-{V}_{p,2012}=\alpha \\ {}+{\beta}_1{\mathrm{Low}\hskip2.22198pt \mathrm{LEC}\hskip2.22198pt \mathrm{Support}}_{p,2012}+\rho {X}_{p,2012}+{\phi}_m+{\mu}_{pt},\end{array} $$

where the variable $ \mathrm{Low}\hskip2.22198pt \mathrm{Support} $ is a binary indicator that is equal to one if the polling station did not support the elected law-and-order candidate. The coefficient $ {\beta}_1 $ is the quantity of interest, measuring the association between public security and support. The matrix X contains polling station-level controls that we have seen correlate with support and $ {\phi}_m $ is the fixed effect for municipality m. Although the model is not a difference-in-difference model (it is not possible to capture support for law-and-order candidates in the pretreatment, as few municipalities have had law enforcement candidates running two consecutive years), the analysis only includes municipalities that did not have a law enforcement incumbent in 2008. In this way, $ \beta $ captures associations in municipalities that have just elected a law enforcement candidate. Estimates cluster standard errors at the municipality level.

Table 2 confirms that communities that did not vote for law-and-order candidates are neglected. There is a consistent negative relationship between support and police activity, as columns 1–3 show. On average, 3.5%–4% of all car robbers are caught in the act, so the coefficients show that low-support areas are associated with a 10% lower police activity (note that these coefficients are consistent with all three radii).Footnote ³⁹ Columns 4–6 show the positive correlation between lack of support and homicides. Finally, favoritism only takes place when there is no formal oversight of public security. Results in Appendix A.5 of the Supplementary Material show that in municipalities with a local security committee and low support for law-and-order candidates, there is a considerable penalty for not supporting a winning law-and-order candidate, especially in municipalities lacking police oversight.

Table 2. Support for Law-and-Order Candidates, Police Activity, and Homicides

Note: Controls include the proportion of poor voters, proportion of rich voters, proportion of young men, proportion of non-white men, and total votes at the polling station. Standard errors (in parentheses) are clustered at the municipality level. Table B.5 in Appendix B of the Supplementary Material reports full results. * $ p<0.1 $ 0; ** $ p<0.05 $ ; *** $ p<0.01 $ .

ALTERNATIVE EXPLANATIONS

The many accounts of police confronting transnational drug organizations or state failure spiraling toward violence provide an alternative to political favoritism. It is plausible that law-and-order incumbents, especially those from the police, encourage their police colleagues to aggressively assault crime-dominated areas. In Brazil, there is a long history of the police using excessive force and carrying out killings of marginalized youth. Together with a popular perception that a “good bandit is a dead bandit,”Footnote ⁴⁰ this approach to crime could be electorally motivated. Furthermore, an action against crime may either take a wrong turn due to a lack of state capacity (Durán-Martínez Reference Durán-Martínez2015; Yashar Reference Yashar2018), or force criminals to directly confront the state (Calderón et al. Reference Calderón, Robles, Díaz-Cayeros and Magaloni2015; Lessing Reference Lessing2017), putting state agents in jeopardy. This, again, would not be surprising for Brazil, where the police suffer from high lethality rates (Bueno Reference Bueno, de Lima, Ratton and Azevedo2014b), and have recently been challenged by the rise of well-organized criminal groups (Feltran Reference Feltran2018; Willis Reference Willis2015).

Using death certificates to capture “legal interventions” (ICD-10 code Y35) and victims’ occupation codes to identify law enforcement agents, I am able to test for these competing explanations. It is worth noting that data on police killings are not completely reliable. For instance, some states have not reported any police killings over the entire period of analysis. To minimize this issue, I only include in the estimations municipalities from states that systematically recorded police killings in the previous period.Footnote ⁴¹ Figure 9 shows that there is no increase in the number of individuals killed by the police or in the number of law enforcement agents killed.

Figure 9. The Effect of Electing Law-and-Order Candidates on Killings of and by Law Enforcement Agents

Note: Nonparametric RDDs. Bars represent 95% robust CIs. Estimations only include municipalities from states that recorded killings in the past. Table A.34 in the Supplementary Material report estimates.