I. INTRODUCTION

This paper discusses economic voting from a broad, multidimensional perspective, which considers other areas of performance aside from the economy in presidential approval, as is the traditional approach. Our focus here is on public security in its relationship with the themes of electoral accountability and clarity of responsibility ( ^{Silva and Whitten 2017} ), which are part of this theory.

The first reason for expanding the scope of economic voting relates to Latin America’s current scenario. The 2018 Latinobarómetro reported widespread pessimism in the region. Dissatisfaction with democracy grew from 51 percent in 2008 to 71 per cent in 2018, and the reasons are widely known. For 35 percent of Latin Americans, the economy is the main issue, especially unemployment and low wages, and for 19 percent, crime is something to consider. The study concluded that a decrease in income, job instability, violence, and corruption are the main issues that draw the attention of the Latin American population ( ^{Latinobarómetro 2018} ).

It is no mere coincidence that the three main countries of the region – Colombia, Mexico, and Brazil – elected the respective new heads of state in 2018: Iván Duque, Andrés Manuel López Obrador, and Jair Bolsonaro. They all share the following attributes: a) an anti-establishment attitude, b) anti-corruption rhetoric, and c) public security as priority number one or two.

Given the importance of security in the public opinion ( ^{Carreras 2013} ) and the recent elections in Latin America ( ^{Perez 2015} ), analysis of their impact on presidential approval is urgent.

The second reason refers to the limited number of studies relating public security and voter behavior. There are hundreds of studies on the economy’s impact on presidential elections; meanwhile, research on public security’s impact on government approval is restricted. Political science in Latin America should foment this debate along with discussions on retrospective voting.

This article aims to analyze the impact that feelings of insecurity and experiences of crime victimization have on presidential approval in Latin America. The first dimension focuses on voter subjectivity and considers their feelings of insecurity in their daily life (subjective dimension). The second dimension examines the concrete experience of crime victimization (objective dimension).

Using a multilevel model, we considered the different institutional arrangements for public security administration and their relation to presidential approval. We refer to the constitutional assignment of responsibility, which can be exclusive to the national government or shared. This subject is also tied to an important discussion within the scope of economic voting theory regarding clarity of responsibility ( ^{Powell and Whitten 1993} ; ^{Anderson 2000} ; ^{Ratto 2013} ; ^{Silva and Whitten 2017} ). Our research hypotheses propose that this institutional variable affects both clarity of responsibility and presidential approval and controls or maximizes the effects that feelings of insecurity and crime victimization have on presidential approval.

To answer our research question, we analyzed AmericasBarometer data from 2016/2017 (LAPOP 2017) and 2018/2019 (LAPOP 2019), which include specific questions regarding security and government approval.

The results confirmed the relevance of the two metrics of security in presidential approval in Latin America, wherein feelings of insecurity had a more significant effect. Citizens who feel unsafe, as well as those who claimed that they had been a victim of a crime in recent months, tend to evaluate the government more negatively. As for the relationship between context and the individual dimension of public security, we observed that when the responsibility for this area of public policy is exclusive to the national government, the effect of feelings of insecurity and crime victimization on presidential approval tends to increase.

This paper is divided into four parts aside from this introduction. In the next section, we present a theoretical review to support our variables and hypotheses for testing. We explain our hypotheses and how we processed the variables in the subsequent section. Finally, we disclose and discuss our results and conclusions.

II. ECONOMIC VOTING THEORY: RATINGS AND RESPONSIBILITY

III. DATA, METHODOLOGY, AND RESEARCH HYPOTHESES

To find the answers to our research questions, we used data from the 2016/2017 and 2018/2019 rounds of the Americas Barometer (LAPOP, 2017; LAPOP 2019), as they included specific questions about public security and government approval.

In Latin America, aggregated data on victimization and fear of crime are provided by opinion surveys such as Latinobarómetro and Americas Barometer (LAPOP). However, Lapop and Latinobarómetro feature different methodologies, as well as distinct ways of asking about fear of crime and victimization, which makes a general comparative analysis more difficult ( ^{Dammert and Tobar 2018} ). Despite the limitations of LAPOP in questions regarding fear of crime, it is still the most complete and broadly used survey among researchers, with time series and questions specifically tailored for the most violent countries, allowing for more depth in the analysis of violence and political aspects in Latin American nations.

As previously mentioned, our dependent variable is presidential approval, as measured by the following question in the LAPOP survey: “Speaking in general terms about the current administration, how would you rate the performance of President [name]? Very good, good, neither good nor bad, bad, very bad?” Due to the technical difficulties involved in adapting multilevel models to categorical dependent variables with multiple responses, we chose to dichotomize this variable, grouping the alternatives “very good” and “good” as “approve” and the others as “disapprove”.

For the independent variables, we used both individual and national metrics. Among individuals, our main interest is related to security. As stated above, one of the main criticisms when studies along this line of research began to emerge in the United States was the lack of correspondence between actual crime rates and the feelings of insecurity demonstrated in opinion polls. As a way to subvert this criticism in reference to moral panic, we analyze the subjective dimension (feelings of insecurity) as well as the objective experiences of those who had effectively been victims of a crime (crime victimization). We have also adopted a contextual variable to represent homicide rates in each country, as we shall later observe. In this study, feelings of insecurity are measured by the question: “Considering the place or neighborhood where you live and the possibility of falling victim to assault or robbery, do you feel very safe, safe, unsafe, or very unsafe?” We used a 4-point scale and recoded it as a dichotomous scale by joining the two first points as “Safe” and the other two as “Unsafe”. Crime victimization is measured by the question: “Have you been a victim of any type of crime in the past 12 months? That is, have you been a victim of robbery, theft, assault, fraud, blackmail, extortion, violent threats, or any other type of crime in the past 12 months?” As the original metric already has a binomial distribution, we did not need to recode it.

For the contextual variables (level 2 = country), our main interest was to assess the effect of clarity of responsibility on government approval. In this sense, we searched for data on the institutional organization of public security in the countries under analysis.

The debate on clarity of responsibility is less common in the area of public security, which to a certain extent can be explained by the diversity of public actors involved in this field, the legacy of dictatorships and the impasses in these institutions during the transition between political regimes ¹ , as well as by the influence of penal populism and the structural factors involved in the increase of violence.

To deal with the issue of clarity of responsibility, we considered it relevant to approach the institutional organization of police in Latin American countries, seeing as how police work relates to people’s daily assessments of their safety. Citizen perception of justice in institutional procedures, in turn, could foster and increase trust in and legitimacy of police forces among citizens ( ^{Zanetic 2017} ).

In the Latin American context, lack of trust in police and in the criminal justice system would not be related only to their ineffectiveness, but also to the perception of arbitrariness and selectiveness in their operations ( ^{Dammert 2012} ).

Although the article is not dedicated to further understanding the policing models adopted by the countries or even how police work is evaluated in each one, we believe that discussing clarity of responsibility from the perspective of institutional organization and constitutional design ² can contribute to deep-ening connections between public security and political aspects, presidential approval in particular.

Therefore, using data taken from the National Constitution of each country and based on studies by the ^{PNUD and SENASP (2016)} , we developed a metric that divides the countries between those that assign responsibility for public security exclusively to the national executive branch and those that share the responsibility with other government entities. We assume here that the countries with exclusive responsibility have higher clarity of responsibility in this area; therefore, it is easier for voters to rate the administration’s performance. Finally, as a second metric for the context, we used homicide rates for the country. This can help measure the salience of the issue of public security in each country, and it allows us to observe crime in objective terms while keeping in mind criticisms of moral panic and penal populism. ^{Mendez (2018)} also used an objective metric at a macro level, the geographical distribution of crime, in his study about Mexico, finding statistically significant results.

Even though our main interest is specifically in these predictors, we cannot ignore other recurring variables in the literature about economic voting, especially those related to economic ratings. For this reason, we posed a question related to sociotropic retrospective voting: “Do you think the country’s economic situation is better than, the same as, or worse than it was 12 months ago?” We recoded the original question so that “0” represents “worse” and “2” represents “better.” We also included a similar metric at the individual level, which we obtained from the following question in LAPOP: “Do you think your economic situation is better than, the same as, or worse than it was 12 months ago?” We applied the same recodification.

As previously seen, studies show that political variables such as ideological identity and political party affinity can affect a voter’s rating of the government’s performance in the economy ( ^{Veiga and Ross 2016} ) and public security ( ^{Berens and Dallendörfer 2019} ). Therefore, they are the control variables. We also use sociodemographic information such as sex, age, education level, race, and residential region as control variables. We do not provide a detailed description of each of these control variables, but we have provided a complete list of them below ( Table 1 ) with details on the recoding procedures.

The appropriate statistical model for this data structure is the multilevel model for binary response variable, which determines the direct effects of individual and contextual explanatory variables. It also allows us to evaluate explanatory variables at a macro level as moderators of individual relations and observe whether there is variability in the impact of a predictor of interest in comparison to the others ( ^{Hox 2010} ).

This method is applied to hierarchical data structures, and it is useful in research involving the relationship between individuals and their collectives, assuming that people interact and are influenced by the groups or social contexts they belong to. This forms a hierarchical system ( ^{Raudenbush and Bryk 2002} ). A multilevel analysis approaches this system and clarifies the underlying effects within a certain group (or cluster) and among different groups ( ^{Hox 2010} ; ^{Sommet and Morselli 2017} ).

For this purpose, the models are built so that the response variable is at the lowest level. In other words, we assume that a set of data exists with one single result or response variable that is measured at the first level and that it has explanatory variables at all other levels.

IV. RESULTS AND DISCUSSION

Before we present the results of our hypotheses, we think it is important to disclose the descriptive information for the more directly relevant dependent variables and predictors at the individual and national level. ³

To begin with the dependent variable, we noticed a low percentage of presidential approval, with an overall average of 34%, considering all the countries and years involved in the analysis.

The data for the two independent variables at the individual level indicate a very alarming scenario in the region considering that half of the citizens in Latin America (47.06 percent) stated that they feel unsafe, and approximately a quarter (24.85 percent) claimed they had been the victim of a crime in the past 12 months. The variation is also considerable: from 37.7 per cent (Honduras) to 62.7 per cent (Dominican Republic) in the first metric and from 17.7 per cent (Nicaragua) to 35.8 per cent (Peru) in the second. Although correspondence is not perfect, there is a strong connection between the relative positions of the countries in these two metrics. However, the relation is less clear when we include the homicide rate in this scenario. For example, we highlight the coherence of Chile’s case, which has the lowest rate per 100,000 inhabitants (3.5) and is among the countries with lower percentages of insecurity (38.7) and a percentage of victimization that is below the overall percentage for the region. We see the opposite in the case of Honduras, which has the lowest percentage of insecurity and the second highest homicide rate.

Finally, the variable that distinguishes countries based on constitutional responsibility presents a prevalence of federal leadership, as only five cases (Argentina, Bolivia, Brazil, Ecuador, and Mexico) have shared responsibility. Since N constitutes two rounds per country, the final value is 10.

To assess whether multilevel modeling is feasible, we estimate a two-level null model, which is an equation with only one intercept and group effect that allows us to estimate the president’s approval probabilities, without including the predictors. The fixed intercept is a general constant term, and is shared by all countries included in the sample, while the random effect is country-specific. The results in Table 2 indicate that the odds ratio of approval in an “average” country is .57. The odds ratio variance between countries is 1.77, and to find out if this difference is statistically significant, we conducted an odds ratio test to rule out the hypothesis that this variability is equal to 0. In this procedure, the null model is compared with its individual level equivalent, that is, without the random effects of the countries, whose result of the “log-lik” line presented the value of -2769.85, much higher than what would indicate the nonexistence of variability. The bottom of the table, which compares adjustments of the model with and without variation among the national units, also reinforces this position as it shows a significant difference at a very demanding level (0.000).

In the Caterpillar chart below, composed of the residuals and with bars representing the respective confidence intervals around each estimate, we note that there is considerable variation, since the effects are mostly below or above the confidence interval. In a model like the one used here, it is assumed that there is a hierarchical dataset, with an outcome or response variable that is measured at the lowest level, and explanatory variables at all existing levels ( ^{Hox 2010} ). Thus, after observing the existence of this effect at the country level, we sought to test the micro-level predictors, that is, we sought to identify individual-level factors linked to presidential approval.

As we confirmed the relevance of multilevel modeling, we present the results of six different models in Table 3 and Figures 1 and 2 . We decided to split the two main independent metrics at the individual level ⁴ , so the first model (Mod.1) includes the metric of crime victimization, the two economic rating variables, and the control variables as predictors. ⁵ The second model (Mod.2) replaces victimization with the subjective metric of feelings of insecurity, while the other predictors remain the same. The following two models (Mod.3 and Mod.4) follow the same logic but include two national-level variables. Finally, the last two models (Mod.5 and Mod.6) present cross-level interactions between the two main individual variables and the assignment of responsibility metric. ⁶ For all models in each row, the B coefficients are presented, followed by their respective values exponentiated between braces. ⁷

Figure 1 Variation of conditional modes for presidential approval, Latin America, 2016-2018. 

Figure 2 Predicted probabilities of presidential approval by feelings of insecurity, Latin America, 2016-2018. 

The first two models evidence the relevance of this dimension of public policies in presidential approval in the Latin American context, thus confirming our first hypothesis. They also indicate that feelings of insecurity are slightly more important, reducing the probability of the president’s approval by 24%, versus a 13% reduction caused by victimization, which confirms our second hypothesis. Figures 2 and 3 present the predicted values for points 0 and 1 of these two predictors, displaying on the y-axis the percentage increase in approval probability. While an individual who feels insecure has a 28% chance of approving the president, another who feels secure has approximately a 34% chance, this being controlled for the effects of all other model predictors. Victim status, in turn, generates a probability of 29%, against 32% predicted for a non-victimized individual. Therefore, the results reinforce previous findings on the relevance of violence in the declining trust in political institutions in general, considering its effects on presidential approval. Accordingly, the results reinforce the abovementioned reports that public security affects presidential approval ( ^{Mendez 2018} ; Perez 2018) and that feelings of insecurity have a greater impact on presidential approval than crime victimization.

Figure 3 Predicted probabilities of presidential approval by victimization, Latin America, 2016-2018. 

This last aspect can also be observed in the importance that public security issues have assumed in elections, especially concerning changes in crime control and the emergence of penal populism ( ^{Garland 2008} ). It is important to highlight that the crime victim’s concrete experience also influences approval, which is significant as this variable considers crimes that occurred within the last 12 months and does not contemplate indirect experiences of crime victimization of family and friends.

Although this effect may seem smaller, we note that the equations include the two economic rating metrics which the literature on economic voting has already agreed upon. Mendez’s study (2018) demonstrated the relationship between the economic situation and the influence of violence on presidential approval in Mexico. The effects of the two variables in relation to the national economic situation were confirmed once again as they presented increases in the probability of presidential approval of 122% and 29%, respectively. Moreover, all the predictors were relevant.

As for our analysis of the relationship between the contextual dimension (clarity of responsibility and homicide rate) and government approval, models 3 and 4 indicate that the national homicide rate and assignment of responsibility do not directly affect approval, and their inclusion does not significantly alter the effects of the individual-level predictors.

The homicide rate may not have had an effect as a result of its high level of aggregation, as there are significant differences between regions, states, and municipalities of the same country, which might by considerably diluted in the national homicide rate. For example, from 2006 to 2016, while seven Brazilian states reduced their homicide rates, the homicide rate increased by 50 to 100 percent in other eight states. ⁸ Also, other variables related to crime, such as robberies and theft, may have greater influence.

As for the clarity of responsibility variable, we did expect some form of direct association between the variables.

Despite these results, we developed the last two models using cross-level interactions to test our third hypothesis, which suggests that assignment of responsibility for security policies could moderate or magnify the effects of crime victimization and feelings of insecurity. The results confirmed our hypothesis, as the national-level variable for clarity of responsibility increases the negative effect of feelings of insecurity by 15%. The interaction involving the victimization measure did not reach the required level of statistical significance.

The contextual variable in reference to clarity of responsibility magnifies the effects of feelings of insecurity and contribute to the decline in presidential approval. Therefore, clarity of responsibility for public policies alone does not affect presidential approval, but it does when it interacts with feelings of insecurity, as was expected given the literature and theoretical framework. It is a moderating variable.

V. CONCLUSIONS

This study affirms that feelings of insecurity and crime victimization have a significant and negative impact on presidential approval in Latin America. Feelings of insecurity have a more substantial effect on approval than crime victimization, thus confirming our first and second hypotheses and reinforcing ^{Perez’s (2015)} findings.

We also observed that when responsibility for public policies is exclusive to the national government, feelings of insecurity and crime victimization have a stronger effect on presidential approval than in contexts where this responsibility is shared. This confirms our third hypothesis and the theory on the effect of clarity of responsibility proposed by ^{Powell and Witten (1993)} .

For research in Latin America, this paper’s contribution is in its exploration of the consequences of one of the three main concerns of the Latin American population in their approval of the president – public security.

This study contributes to research on electoral behavior in two ways. First, it expands our understanding of retrospective economic voting from a multidimensional perspective for public security policies, an area which has been explored little. It considers the issue of public security in retrospective voting, which is much needed in Latin America. Its second contribution pertains to testing and confirming the clarity of responsibility theory in other areas other than the economy. This study demonstrated that institutional arrangements affected the relationship between feelings of insecurity and crime victimization on one side and presidential approval on the other.

For public security analyses, this paper determined that homicide rates, a contextual variable, did not have a direct or moderating effect on presidential approval. Finally, it emphasizes the relevance of feelings of insecurity as a stronger predictor of presidential approval than crime victimization. In other words, our findings do not refute criticisms of moral panic.

While this paper may provide answers, it also raises new questions. In the case of shared responsibility over public security, which entity is affected more by feelings of insecurity and crime victimization? What determines feelings of insecurity? We believe that there is a very strong agenda to be pushed for public security and electoral behavior research in Latin America today.