Crime Victimization and Political Outcomes

Crime victimization has clear psychological effects on victims, such as increasing their levels of anger, fear and sadness (Greenberg and Ruback Reference Greenberg and Ruback2012). It can also have important political and electoral consequences.

In the case of Latin American countries, there is evidence showing that voters sanction incumbents for local homicides depending upon whether they consume information (Marshall Reference Marshall2015). Conversely, analyses using survey data indicate that crime victimization does not affect voters’ electoral decisions; however, perceptions of high levels of insecurity do impact respondents’ political choices (Perez Reference Perez2015). The discrepancies between these studies might be explained by the fact that incumbents can escape electoral punishment under particular circumstances (Kronick Reference Kronick2014).Footnote ³ There is also mixed evidence in the studies that explore the impact of crime on political participation (Bateson Reference Bateson2012; Trelles and Carreras Reference Trelles and Carreras2012; Berens and Dallendorfer Reference Berens and Dallendörfer2017; Ley Reference Ley2017). There is much more agreement about how crime victimization and perceptions about violence undermine support for (and the legitimacy of) democracy. Multiple studies have found evidence of this negative correlation (Carreras Reference Carreras2013; Fernandez and Kuenzi Reference Fernandez and Kuenzi2010; Malone Reference Malone2010; Merolla, Mezini and Zechmeister Reference Merolla, Mezini and Zechmeister2013; Liebertz Reference Liebertz2015).

The literature has paid less attention to how crime can modify victims’ policy preferences. Voters’ willingness to accept non-democratic measures, such as repression, can have critical consequences for the quality of democracy. Support for (or opposition to) iron-fist policies can inform politicians about citizens’ level of tolerance of human rights abuses by the state. Indeed, voters’ policy preferences can shape the adoption of policies (Lupu and Pontusson Reference Lupu and Pontusson2011) and impact parties’ ideological positions (Adams et al. Reference Adams2004). Additionally, a rise in crime might increase the electoral chances of parties associated with iron-fist measures to reduce crime such as right-wing or populist parties. Consequently, understanding the factors that influence citizens’ policy preferences regarding crime is particularly important.Footnote ⁴

It is challenging to address this research question for five main methodological reasons. First, being a crime victim is not a random event. Particular social circumstances can be associated with crime victimization, which can generate a serial victimization problem. This means that previous crime victims might be more likely to be crime victims again. Consequently, when using survey data it is hard to know if victimization is a unique or recurrent event in a respondent’s life (Bateson Reference Bateson2012). This problem can introduce biases, since the previous treatment status can affect the outcome (for example, serial victims might get used to crime).

Secondly, there might be a reverse causality problem. People who want strong-arm policies might be more likely to report a crime as a way to increase crime statistics and push for the implementation of those policies. Also, most of the literature based on respondents’ fear of crime has not adequately addressed relevant endogeneity concerns. For example, because political preferences can influence voters’ perceptions of insecurity, the literature might be overstating the political impact of these perceptions.

Thirdly, in any observational study the presence of hidden biases is a significant issue. Victims and non-victims can differ across multiple unobserved characteristics. This is particularly true when we use a national sample and compare individuals from different cities and, therefore, from diverse socioeconomic contexts. Fourthly, and related to the previous issue, neighborhood effects are a crucial concern (Bateson Reference Bateson2012). Some sectors or areas within a city might be more or less secure, which affects the probability of being a crime victim. This point is particularly salient when analyzing data from multiple countries or from diverse cities or states within a country. Crime has a very local nature, and neighborhood characteristics are hard to adjust for. Finally, when using survey data, the treatment (crime victimization) and covariates (respondents’ characteristics) are measured at the same time, which can lead to potential post-treatment biases. For instance, when trying to explain individuals’ policy preferences, the use of specifications that adjust for things such as income or presidential approval can be very problematic, since these covariates might be affected by exposure to crime.

Crime Policy Preferences

The problems associated with crime are highly visible in the largest country in the region, Brazil, where the homicide rate in 2006 was 29.2 per 100,000 inhabitants, making it the third most violent country in Latin America after El Salvador and Venezuela (Carreras Reference Carreras2013). These statistics have not improved in recent years, and ‘no country in the world has more cities plagued by violent crime than Brazil’ (Rapoza Reference Rapoza2016). Populist candidates have exploited this social context of insecurity and violence by promising to bring ‘authority’ back when fighting crime; this pattern was evident in the 2016 local elections (Winter Reference Winter2016). However, implementing iron-fist strategies comes at a cost. The Brazilian military police have perpetrated human right abuses such as extrajudicial and summary executions (Huguet and Szabó de Carvalho Reference Huguet and Szabó de Carvalho2008). More examples of police misconduct in Brazil include unwarranted searches, beatings and torture (Arias Reference Arias2006; Magaloni, Franco and Melo Reference Magaloni, Franco and Melo2015).

Support for these specific measures have important political consequences, because they define the boundaries that cannot be transgressed in an attempt to increase security. Moreover, state repression can affect citizens’ human rights and erode democratic institutions. The inviolability of citizens’ bodily integrity is a basic principle in contemporary democracies that can be undermined by the implementation of iron-fist policies (Fuentes Reference Fuentes2005). In multiple countries in Latin America the state is the main actor involved in human rights violations due to the implementation of military strategiesFootnote ⁵ to fight crime (Cruz Reference Cruz2010).

Where crime rates are high, it is important to understand whether victimization makes citizens more or less likely to support these policy approaches. What explains support for tougher crime-fighting measures? Prior research suggests two main explanations for citizens’ attitudes toward these particular policies. The first relies on voters’ ideological and/or party identification (long-term factors). The second focuses on how specific circumstances, for example a change in media coverage, can shape voters’ policy preferences (short-term factors).Footnote ⁶

Regarding the first explanation (long-term factors), right-wing voters are more likely to care more about crime than left-wing voters (Mayer and Tiberj Reference Mayer and Tiberj2004). In a similar vein, right-wing authoritarianism can predict support for punitive measures (Gerber and Jackson Reference Gerber and Jackson2016). Furthermore, the policies that emphasize punitive sanctions tend to be associated with conservative rather than liberal politicians. For example, former Republican US president Ronald Reagan summarized his views about how to fight crime by declaring that ‘here in the richest nation in the world, where more crime is committed than in any other nation, we are told that the answer to this problem is to reduce our poverty. This isn’t the answer […] [The] government’s function is to protect society from the criminal, not the other way around’ (Beckett Reference Beckett1999, 48). Moreover, there is evidence in the United States that the proportion of Republican legislators is correlated with imprisonment rates at the state level (Beckett and Western Reference Beckett and Western2001).

The link between ideology and crime policies is also evident in Latin America. Right-wing candidates in Honduras, Mexico and Peru have promoted strong-arm policies to combat crime (Cohen and Smith Reference Cohen and Smith2016). In El Salvador, the conservative party ARENA attempted to boost its support in a context of high crime rates by implementing iron-fist policies, such as diluting due process guarantees (Holland Reference Holland2013). In Brazil this pattern is also clear, as in the case of the right-leaning former governor of the state of Rio de Janeiro, Marcello Alencar. Alencar decided to provide semi-automatic weapons to the police and to implement a ‘bravery bonus’ to officers who engage in violent confrontations (Magaloni, Franco and Melo Reference Magaloni, Franco and Melo2015). In summary, right-wing politicians can be linked to these kinds of measures to combat crime. Right-wing citizens, similarly, are more likely to support tougher measures to reduce crime and to focus less on social policies.

Regarding the second explanation (short-term factors), citizens’ preferences can also be affected by particular circumstances, such that support for strong-arm policies might not be a static policy preference. For example, the literature has focused on how the media can influence voter preferences. There is extensive research showing that the way the media frame an issue can change how individuals think about that topic (Kinder Reference Kinder1998; McCombs and Shaw Reference McCombs and Shaw1972). In particular, certain news coverage of crime can impact individuals’ attitudes toward crime control policies (Howitt Reference Howitt1998; Krause Reference Krause2014). Less attention, however, has been paid to the consequences of direct exposure to crime. In this article, I show that crime victimization can have substantive and meaningful effects on victims’ policy preferences. Specifically, crime exposure can increase support for iron-fist policies such as state repression.Footnote ⁷

What causal mechanism explains the increase in support for more repressive measures? Consistent evidence shows that crime can affect victims’ democratic values and support for the rule of law (Merolla, Mezini and Zechmeister Reference Merolla, Mezini and Zechmeister2013; Carreras Reference Carreras2013). Crime can also undermine the legitimacy of the political system (Cruz Reference Cruz2010) and increase support for radical change (Seligson and Azpuru Reference Seligson and Azpuru2000). In fact, fear of crime has been connected with support for regimes that reduce civil liberties (Perez Reference Pérez2003). Additionally, there is evidence of a correlation between democratic preferences and support for policies that protect citizens’ due process rights (Seligson Reference Seligson2003). Consequently, a lower attachment to democratic values may explain why crime victims might accept the erosion of some basic rights in favor of radical measures to combat delinquency in their countries.

Civil liberties are directly linked to democratic values and the rule of law, and less support for democracy due to crime victimization might increase victims’ willingness to sacrifice these rights. State repression is not just another strategy to reduce crime. It implies a disposition to tolerate the dilution of procedural rights. Therefore, I expect voters to first need to have lower democratic values (because of the undermining effects of crime on the legitimacy of the political system) in order to then accept repression as a valid strategy.Footnote ⁸

In summary, I hypothesize that crime victimization has a substantive and significant effect on victims’ policy preferences: in particular, that crime exposure should increase support for iron-fist policies such as state repression. I expect that this change is explained by a lesser degree of support for democratic values, which makes victims more tolerant of certain strategies.

The study of crime victimization has been dominated by a sanctioning argument, the most common prediction of which is that victims will punish incumbent candidates. In this article, however, I focus on the prospective dimension of voters’ decisions by paying attention to the policies they most care about after crime victimization: in particular, support of state repression.

Research Design

Random assignment is the best strategy for establishing the causal effect of a particular intervention, because treatment assignment is independent of potential outcomes (Morgan and Winship Reference Morgan and Winship2014), and in expectation, observed and unobserved covariates should have similar distributions between treatment and control groups (Bowers Reference Bowers2011). However, randomization is not always feasible for ethical or practical reasons. The alternative strategy for studying a phenomenon that cannot be randomized, such as crime victimization, is a well-designed observational study structured to resemble a simple randomized experiment (Rosenbaum Reference Rosenbaum2010), which uses elements from the design-based approach to improve the study design (Keele Reference Keele2015). These include focusing on endogeneity (Imbens Reference Imbens2010), not including final outcome data (Rubin Reference Rubin2008) and not relying on statistical modeling (Keele Reference Keele2015).

What makes an observational study good? Following some of the recommendations provided by Rosenbaum (Reference Rosenbaum2010, Reference Rosenbaum2011), I apply four criteria. First, the treatment should be well defined. This means that we know when it starts, and therefore what the pre-treatment and post-treatment covariates are. Secondly, even though there is no random assignment, the intervention should seem haphazard or not obviously related to potential outcomes. Thirdly, treated and control groups should be comparable: in other words, the distributions of observed covariates should be similar across both groups. Fourthly, the design should use strategies to reduce sensitivity to unobserved biases, such as decreasing unit heterogeneity. I apply these four criteria in the design of this observational study.

Regarding the first recommendation, the main problem when working with survey data is the lack of pre-treatment covariates, since adjusting for post-treatment characteristics can introduce biases (Rosenbaum Reference Rosenbaum1984). Therefore, I use panel data from Brazil collected between 2002 and 2006 (Baker, Ames and Renno Reference Baker, Ames and Renno2006; Baker et al. Reference Baker2015) to adjust only on covariates captured in waves before respondents were victimized by crime. The survey questionnaire asked a standard battery of questions about political preferences, demographics, media exposure, crime victimization, feeling thermometers and social networks.Footnote ⁹ The panel structure allows me to include pre-treatment measures of the outcomes, the oldest and most basic tool for reducing the ambiguity of the effect of a treatment in an observational study (Rosenbaum Reference Rosenbaum2015).

Secondly, though crime victimization is not randomly assigned, it is possible to exploit certain aspects of the study design to make this situation more haphazard. In particular, I only select respondents who in wave t were not affected by crime. Then, if in wave t + 1 they were crime victims, they are incorporated into the treated group, and if they keep being non-victims they go into the control. Consequently, I exclude by design citizens who are serial victims of crime. This strategy also contributes mitigating the reverse causality problem since in the first wave I only focus on respondents that did not report a crime.

The third recommendation emphasizes the need to compare similar groups of exposed and unexposed individuals. I construct these groups by using an optimal matching algorithm that finds the largest representative pair-matched sample that is balanced by design (Zubizarreta and Kilcioglu Reference Zubizarreta and Kilcioglu2016; Visconti and Zubizarreta Reference Visconti and Zubizarreta2018). I explain the details of this technique later.

The fourth strategy focuses on decreasing sensitivity to hidden biases by reducing the heterogeneity of the sample. As Rosenbaum shows, reducing unit heterogeneity implies that larger unobserved biases will be needed to explain away a particular effect (Rosenbaum Reference Rosenbaum2005). Good examples of this strategy can be found in studies based on identical twins.Footnote ¹⁰ Consequently, in an observational study it is preferable to focus on more homogeneous and comparable subsets (Keele Reference Keele2015) or on natural blocks (for example, neighborhoods), since unmeasured covariates should be more similar between treated and control groups (Pimentel et al. Reference Pimentel2015). The use of national surveys does not help achieve this goal, because they increase the heterogeneity of the sample. Consequently, I exploit the design of the panel data since it focuses only on two mid-sized cities in Brazil: Juiz de Fora in the state of Minas Gerias and Caxias do Sul in Rio Grande do Sul (Baker, Ames and Renno Reference Baker, Ames and Renno2006). Both cities have similar characteristics, such as the size of the electorate, their educational and income levels, and racial composition.Footnote ¹¹ According to the unmatched sample, they also have similar crime rates in wave t + 1: 15 per cent of respondents were crime victims in Juiz da Fora, and 14 per cent in Caixas do Sul. Additionally, the data provides neighborhood indicators, which allows me to balance the sample by respondents’ geographic location.

How does one go about building a group of affected and unaffected citizens who are balanced in their observed characteristics? One alternative is matching, which attempts to generate treated and control groups with similar covariate distributions (Ho et al. Reference Ho2007; Stuart Reference Stuart2010). However, traditional matching techniques, such as propensity score and Mahalanobis distance, do not guarantee covariate balance and on some occasions can even make balance worse across observed covariates (Sekhon Reference Sekhon2009). These methods often involve a process of manually iterating the model until covariate balance is obtained (Hainmueller Reference Hainmueller2011). Moreover, a possible concern when using any type of matching technique is that it requires some level of pruning to obtain balance, which may cause the matched sample to be different from the unmatched sample.

To address these limitations, I use the designmatch package developed by Zubizarreta and Kilcioglu (Reference Zubizarreta and Kilcioglu2016), which allows me to find the largest representative sample that achieves covariate balance. This algorithm maximizes the size of the sample that: (1) meets the balance requirements defined beforehand and (2) is similar to a target sample also defined beforehand (in this case the unmatched sample). Point 1 addresses the limitations of traditional matching techniques because the algorithm directly balances the original covariates without needing to estimate a propensity score. Point 2 ensures that the samples before and after matching are similar on observed covariates, making pruning less of a concern.

I use mean balance constraints for forty-seven covariates. The algorithm matches individuals such that the treated and control matched groups cannot differ in their means by more than 0.1 standard deviation from the unmatched sample. As a consequence, the standardized differences between the matched treated and control groups cannot be larger than 0.1 x 2 standard deviation. In other words, the standardized differences between the matched groups cannot be larger than twice the standardized differences between the matched sample (that is, both matched groups) and the unmatched sample (see Zubizarreta and Kilcioglu (Reference Zubizarreta and Kilcioglu2016) and Visconti and Zubizarreta (Reference Visconti and Zubizarreta2018) for more details).

All of the mean balanced covariates are ordinal or binary; thus, adjusting their means is a meaningful decision.Footnote ¹² I also use fine balance for neighborhood, which implies that both groups will have the same frequency for this covariate but without restricting who is paired with whom (Rosenbaum, Ross and Silber Reference Rosenbaum, Ross and Silber2007; Zubizarreta Reference Zubizarreta2012). Therefore, I am adjusting for a total of forty-eight different observed covariates.Footnote ¹³

In the matching procedure I include pre-treatment covariates that can affect both the treatment assignment and the outcome (Stuart Reference Stuart2010), such as age, education, gender, ideology, job in the formal sector, media consumption, partisanship, policy preferences, political knowledge, race and religion (full list is provided in Figure 1 and in the Appendix).

Figure 1. Balancing the means of the observed covariates (i.e., mean balance) and building a representative matched sample. Dots represent the standardized differences between the matched treated and control group (balance requirement), and asterisks represent the standardized differences between the matched and unmatched sample (representative requirement)

Note: PMDB: Brazilian Democratic Movement Party, PFL: Liberal Front Party, PSDB: Brazilian Social Democratic Party, and PT: Workers’ Party

The treatment is a binary indicator for being a witness or victim of crimeFootnote ¹⁴ in wave t + 1 (only among a group of respondents who were not witnesses or victims of crime in wave t). The question used to construct the treated and control groups is the following: ‘Have you been a witness or a victim of crime in the past 12 months? This includes crimes such as assault, robbery, or aggression.’ Unfortunately, the question does not differentiate between different types of crimes.

The main outcome of interest is a binary indicator of support for the following statement: ‘The best way to reduce crime is with repression and an iron fist.’Footnote ¹⁵ I use a binary indicator of support for democracyFootnote ¹⁶ to explore the causal mechanism. To estimate the effect of crime victimization I use a linear regression with cluster standard errors at the neighborhood level:

(1)$$Y_{{it{\plus}1}} {\equals}\alpha {\plus}\beta _{1} T_{{it{\plus}1}} {\plus}\beta _{2} P_{{it}} {\plus}\beta _{3} X_{{it}} {\plus}\sigma _{n} {\plus}{\epsilon}_{i} $$

Y is a binary indicator that represents the outcome of interest in wave t + 1. T depicts the treatment (crime victimization in wave t + 1), P describes a pre-treatment measure of the outcome from wave t, and X corresponds to a set of pre-treatment covariates that might explain policy preferences (education and age). σ_n represents neighborhood fixed effects. I also provide the unadjusted estimates to increase transparency (Lin Reference Lin2013); this means no controls or fixed effects. Moreover, in the Appendix I use a one-sided Wilcoxon signed rank test statistic as another method of inference since it is less dependent on distributional assumptions, and allows us to conduct the amplification of a sensitivity analysis for hidden biases (Rosenbaum and Silber Reference Rosenbaum and Silber2009).

External Validity: Results Survey Data

Are these results a consequence of a particularity of the sample composition? Or of the year in which the survey was conducted? Is this pattern only present in Brazil? In an attempt to answer these questions, I use data from the Latin American Public Opinion Project (LAPOP) to study the correlation between crime victimization and policy preferences in eighteen Latin American countries in 2012.Footnote ²⁰ Since there is an evident trade-off between internal and external validity, this second study is less robust than the first because it is harder to reduce sensitivity to hidden biases without panel data. Nevertheless, it does help us check if similar results are obtained when we study all Latin American countries.

The main dependent variable is support for strong-arm policies.Footnote ²¹ I also test the effect of crime on the mechanism of interest: support for democracy.Footnote ²² To estimate the effect of crime victimization, I use a linear regression with standard errors clustered at the neighborhood level, and only include ‘placebo’ covariates as controls. Covariates should not be affected by crime victimization, because that can introduce post-treatment biases. Therefore, I use the following four controls: age, education, gender and ethnicity. I also include country fixed effects in the estimation. I do not use matching in this section to avoid concerns about pruning observations, since the main goal of this analysis is to improve external validity (see the Appendix for more detail).

(2)$$Y_{i} {\equals}\alpha {\plus}\beta _{1} T_{i} {\plus}\beta _{2} P_{i} {\plus}\sigma _{c} {\plus}{\epsilon}_{i}$$

Y is a binary indicator that represents the outcome of interest. T depicts the treatment (crime victimization), P describes the set of four ‘placebo’ covariates. σ_c represents country fixed effects. Table 3 displays the main results.

Table 3. External validity

*p<0.1; **p<0.05; ***p<0.01

The findings are similar to the results obtained using panel data. Crime victimization increases support for strong-arm policies for reducing crime, which might be explained by a lower support for democracy. This analysis allows us to increase the external validity of the results obtained in the two cities in Brazil.