Mining the Criminal Data of Rio de Janeiro: Analyzing the Impact of the Pacifying Police Units Deployment

Abstract

Rio de Janeiro is recognized as a violent city, but the pacifying police program was established with the goal to modify this situation. This paper shows that after six years and many pacified communities, the program has had an impact on the decrease of violent deaths with effects both before and after the deployment of the pacifying police units (UPPs). We also evidence a spatial dynamic where the neighboring areas to the UPPs present a more pronounced decrease. Those results raise new issues in the analysis of public safety.

1 Introduction

The city of Rio de Janeiro is widely recognized by its touristic attractions, wonderful landscapes, beaches, and vibrant cultural life. In contrast to this image that pleases and attracts everyone, the city appears in the news in episodes of violence, usually related to drug trafficking. Nowadays, with large international events, such as the World Cup in 2014 and the Olympic Games in 2016, there is much talk about the violence and safety issues of sports delegations and tourists. To change this scenario, the state government implemented a pacifying program that creates Pacifying Police Units (UPPs) in communities dominated by criminal factions, expelling them and enabling the entry of public services and other investments [7]. The first UPP was deployed in December 2008. In 2014 there were 37 in the city of Rio de Janeiro, covering over 200 communities and an estimated population of 562,691 inhabitants. The expansion of the program raised several criticisms questioning the effectiveness of the UPPs in reducing criminality. In 2015, the Public Safety Institute (ISP), the organism responsible for the official criminality statistics in the state of Rio de Janeiro, published a report which showed a decrease in criminality within the UPPs [6]. However, it did not examine the areas neighboring areas that received UPPs. A program of such nature may influence criminality in the surrounding areas as well (hereafter called non-UPP areas), and thus deserves a broader evaluation.

In this work we used the open criminality statistics data from ISP (www.ispdados.rj.gov.br) and the demographic data from the census conducted by Instituto Brasileiro de Geografia e Estatística (IBGE) in 2000 and 2010. We needed the IBGE data to estimate the population of each Police Department (DP) area.

The goal of this work is to evaluate the impact of UPPs on the occurrence of violent deaths in the city of Rio de Janeiro. We conducted the analysis in two stages. First, we made a descriptive analysis of yearly series by region. Second, we used difference-in-differences (DID) models [4] with the monthly series in the DP areas with UPPs and in the adjacent DP areas.

The remainder of this paper is divided in four sections. In Sect. 2, we describe the models, variables, and modeling strategy adopted. Then, in Sect. 3 we describe and discuss the obtained results. Finally, Sect. 4 summarizes the contributions and points to future work.

2 Empirical Strategy

Public safety policy in Rio de Janeiro can be analyzed for two different periods: before UPPs and after UPPs. To evaluate the effect of the Pacifying Program, we consider the smallest area of data aggregation available, the DP areas. Those areas can be classified into two groups: areas which had received at least one UPP within the study period, and areas which had not. This way, a DP that received one or more UPPs can be analyzed in its period pre- and post-UPP.

The group of DPs that had received UPPs can also be compared, for the same period, to the group of DPs that did not undergo the intervention. Figure 1(b) shows the DP areas that had received UPPs. We focus on the city of Rio de Janeiro because, during the period of study, only DPs from that city received UPP interventions.

In the deployment process of UPPs, the criminal groups were expelled and the areas were kept densely occupied by police forces. Such an occupation reduces several crimes, especially those related to violent deaths. Thus, the variable adopted in this work, named homicides, corresponds to the aggregation of data on three types of crimes related to violent deaths and available for public access: murders, manslaughter, and robbery followed by death. This variable was transformed into rate by 100,000 inhabitants¹, named homicide rate (homic_r). We chose this indicator because of its relevance, the loss of lives. It is also less susceptible to subnotifications and, because it is aggregated, it eliminates problems due to subtle differences in classification. Even if the same kind of fact is at one time classified as murder and at another as manslaughter, they will both be included in the homicide aggregated variable, making the differences in the classification irrelevant for this indicator [3].

When planning an experiment, a series of measures are taken, such as randomization and restrictions in the choice of groups that will receive the treatment or not. This allows us to compare statistics calculated for each group to evidence the effects of the treatment. In a simplified way, for each group that receives a treatment, there is another one, similar and independent, that remains the same, without undergoing any treatment. This way, after a period of time, the difference between the two can be attributed to effects of the treatment. In the context of the Pacifying Program, the location of UPPs is not a planned experiment, but a deliberate choice of the designated authorities. Therefore, we need to take some measures to identify suitable control groups.

In this work, the treatment group comprises the DPs that received at least one UPP. The candidate control groups (which comprise DPs that had not received a UPP) were aggregated in four groups and shown in Fig. 1(a), and which are characterized as follows:

• Capital: aggregation of the DPs in the city of Rio de Janeiro that did not receive any UPP;

• Baixada: aggregation of the DPs in the cities of Itaguaí, Seropédica, Paracambi, Japeri, Queimados, Nova Iguaçu, Mesquita, Belford Roxo, São João de Meriti, Nilópolis, Duque de Caxias, and Magé;

• Greater Niterói: aggregation of DPs in the cities of Niterói, São Gonçalo, Itaboraí, Tanguá, Guapimirim, and Maricá;

• Interior: aggregation of DPs in the remaining cities (those that do not belong to the previous groups).

The aggregation of the subregions Capital, Baixada, and Greater Niterói, comprises the Metropolitan Region of Rio de Janeiro.

Fig. 1.

Subregions of the state of Rio de Janeiro.

The basic idea is to compare the treatment group with each control group to identify the effect of the Pacifying Program. In a certain way, the distance between those control groups observe a question of distancing from the areas with UPPs and could be affected in a different way in the case of crime migration.

Considering that the evaluation of the impact of UPPs is not a planned experiment, some steps need to be taken: first, a parallel trends model was used to compare the treatment group with each candidate control group in the period pre-UPP to verify whether the monthly homicide rates of each DP in each group are similar, i.e., whether there is evidence that these series are parallel. If we can state that a series of homicides in a certain candidate control group has a parallel trend to the series of the treatment group in the period of pre-UPP, so there are evidences that the subregions, for this dimension, do not differ in the pattern of occurrences, and therefore that control group can be deemed suitable for further analysis. This supports the idea that the differences in these two groups in the periods pre- and post-UPPs are effects of the UPPs. The second step was the use of panels of difference-in-differences (DID) models to evaluate the homicide rates in the UPP areas and the control groups. For this analysis we used linear models for data in panels [1] implemented in R [5]. The model of parallel trends between treatment and candidate control groups pre-UPP has the Eq. 1.

$$\begin{aligned} h_{it}= \gamma _{i}+ \sum _{t=1}^{n}\beta _{t} monthYear_{t} +\sum _{t=1}^{n} \delta _{t}(monthYear\cdot UPP_{it}) + e_{it}, \end{aligned}$$

(1)

where $h_{it}$ represents the homicide rate in the DP i at time $t, \gamma _{i}$ the fixed effect of the DPs, $\beta _{t}$ the general effect at time t, $monthYear_{t}$ a month-year time dummy, $\delta _{t}$ the effect of the interaction between time and the DPs that will receive a UPP in the future, $monthYear\cdot UPP_{it}$ the dummy for time and UPP presence, and $e_{it}$ the random error of DP i at time t. We want to assess the null hypothesis $H_{0}$ that $\delta _{t}=0$, implying that there is no interaction effect. At this stage time is limited to all months from 2003 to 2008, the period pre-UPP.

To verify the effect of UPPs, we used two variables: $UPP_{it}$, which is a dummy variable indicating the existence of a UPP in DP i at time t, and $cover_{it}$, which is the population coverage of the UPP. This variable is the ratio between the population in UPPs in DP i at time t and the total population of DP i at time t.² To verify the effect of UPPs, we used Eq. 2:

$$\begin{aligned} h_{it}= \gamma _{i}+ \lambda UPP_{it}+ \delta .cover_{it}+ \sum _{t=1}^{n}\beta _{t}.monthYear_{t} + e_{it}, \end{aligned}$$

(2)

where $h_{it}$ represents the dependent variable homicide rate in DP i at time $t,\gamma _{i}$ the fixed effect of DPs, $\beta _{t}$ the general effect at time t, $monthYear_{t}$ a month-year time dummy, $\lambda$ the effect of UPPs, $\delta$ the effect of coverage, and $e_{it}$ the random error of DP i at time t. At this stage time refers to the entire period of study, including all months from 2003 to 2014.

To verify possible effects before and after the inauguration of a UPP, we used a leads-lags model, represented by Eq. 3:

$$\begin{aligned} h_{it} = \gamma _{i} + \sum _{d=0}^{m}\lambda _{\pm d} UPP_{i,t\pm d}+ \sum _{d=0}^{m}\delta _{\pm d} cover_{i,t \pm d} +\sum _{t=1}^{n}\beta _{t} monthYear_{t} + e_{it} \end{aligned}$$

(3)

with pre-deployment lags of $-12$ and $-6$ months and post-deployment leads of 6, 12 and 24 months. In the model, the sum of the $m \, negative lags$ ($\lambda _{-1},\ldots ,\lambda _{-m}$ and $\delta _{-1},\ldots ,\delta _{-m}$) is the post-treatment effect, and the sum of the $q positive \,leads$ ($\lambda _{+1},\ldots ,\lambda _{+q}$ e $\delta _{+1},\ldots ,\delta _{+q}$) is the pre-treatment effect.

Another relevant hypothesis is that an effect of a DP with UPP may affect the neighboring DPs. To assess this, we used the Eq. 4:

$$\begin{aligned} h_{it}= \gamma _{i}+ \alpha vupp_{it}+ \sum _{t=1}^{n}\beta _{t} monthYear_{t} + e_{it}, \end{aligned}$$

(4)

where $h_{it}$ represents the homicide rate in DP i at time $t, \gamma _{i}$ the fixed effect of the DPs, $\beta _{t}$ the general effect at time t, $monthYear_{t}$ a month-year time dummy, $\alpha$ the effect in the non-UPP DP areas adjacent to DP areas that received UPPs, and the variable $vupp_{it}$ a dummy indicating that DP i at time t is adjacent to a DP with UPP at time t.

Table 1. Effects of UPPs on the homicides - results for the estimated models.

Based on the aforementioned equations, we developed 7 modeling strategies to access the effects of the variables. The estimated results, i.e. the variables’ effects, are shown in Table 1. Models 1, 2 and 3, were developed with Eq. 2. The first model tested only Presence of UPP, the second only the UPP coverage and the third both. Models 4, 5, 6 and 7 were developed with Eq. 3. The model 4 tested only effects of UPP, the 5 only effects of cover, the 6 used only significants variables from model 7, which is the estimated results of the complete model.

3 Results

The UPP program began at the end of 2008 with one deployed UPP, and grew rapidly. The year of the biggest impact was 2012, when the program doubled the number of involved communities.

In Figs. 2(a) and (b), we observe that homicides decreased throughout the state in the period from 2003 and 2011, and after that period they relapsed into increase. We observe that in the period from 2008 to 2012, corresponding to the beginning of the UPP program, there was a great decrease of homicides in all regions. In the following period, from 2012 to 2014, there is a relapse into increase in almost all regions; we highlight Baixada, with the largest increase ($39.15\%$). In the UPPs region, the increase was $5.58\%$, which calls attention for its contrast with the negative rate in the areas without UPPs (Capital) in the same city. As the regions affected are those of DPs with UPP, we can still investigate whether the increase in homicides occurred within the UPPs or outside the UPPs. The UPP balance report published by ISP [6] shows that the UPPs had a decrease of $31\%$ in homicides during the same period. We observe in the DP areas with UPP, in the same period, an increase of $7.72\%$ and, in the areas without UPP, a decrease of $5.02\%$. The difference between the rates suggests that violence has increased in the UPP margins, i.e., in the DP areas with UPP, but outside the boundaries of the UPPs themselves. This fact shows a change in the criminality dynamic. Considering the conclusions by [2, 3], there was a strong positive correlation between the presence of criminal factions and violence in the community surroundings. The results of our work, however, evidence that in the UPP margins the rates relapsed into increase. In the DP areas without UPP, the rates are still decreasing, although less strongly.

Fig. 2.

Evolution of the homicide rates in the subregions of Rio de Janeiro.

In the verification of the impact of UPPs with the DID model, the first step was to evaluate the possibility of using the aforementioned regions as control, and for this we used the parallel trends model described. As can be seen in Table 1, in the last row, the p-values (pre-UPP) are the p-values of the F test for the coupled significance of the interactions. In the case of Greater Niterói and Interior, we have evidence that, in the pre-UPP period, the homicide trends in those regions are parallel to the UPP area and we assume that those regions are acceptable as control to the UPP areas.

The results obtained in the adjustments of the fixed effect models for DP, with the two control groups, are in Table 1. In the rows of this table we find the estimated parameters. The columns are divided in two groups, one for each control, G. Niterói and Interior. In those groups, each column refers to a model identified by a number, according to aforementioned.

In general, there is an agreement regarding the significant estimated parameters. For model 1, the estimate of the homicide rate decrease due to the presence of UPP is $-0.39$ homicides-month per 100,000 people when compared to G. Niteroi. When compared to the Interior region, the decrease was of $-1.06$ homicides-month per 100,000 people, i.e., more than twice as much.

In the tested models, our preferred model was model 6. In this model, the effects six months before the inauguration of the UPP (Presence of UPP $-6\,months$) are significant and indicate that there was already a decrease in homicide rates at that time. When comparing to G. Niterói, the estimated monthly rate was $-0.35$, and for Interior $-0.79$, which implies a decrease of $13\%$ and $34\%$, respectively, when compared to the 2014 rates. This effect can be explained by observing the process of UPP deployment.

In general, the tactical intervention begins one year before the deployment, and within six months the area is stabilized and with strong policing, which promotes the decrease of homicides. In practice, considering that in 2014 the average monthly rate in Interior was 2.34 homicides-month per 100,000 people, a decrease of 34% represents on average 34 homicides a month. In the case of Greater Niterói, the decrease of $12\%$ represents an average decrease of 9 homicides per month.

In the moment of deployment, the coverage (cover) has significant effects, with $-0.158$ and $-0.166$ for G. Niterói and Interior, respectively. The parameters estimated for Interior represent a decrease in the monthly rate of 0.166 for a coverage of 1 ($100\%$). Put in another way, for each $1\%$ of coverage, the decrease is of 0.00166 homicides-month per 100,000 people. This represents an average decrease of 0.07 homicides each month for each $1\%$ of coverage. If we used the average coverage of 2014, which was $21\%$, we would have an average decrease of 1.47 homicides each month. When compared to Greater Niterói, also considering the average coverage of 2014, the decrease would be, on average, 0.84 homicides-month per 100,000 people. The effect of coverage is significant and stronger two years after the inauguration of the UPP (UPP coverage ${+}24\,\text {months}$). At this moment the decrease for each percentage of coverage is of $0.14\%$ with respect to G. Niterói and of $0.17\%$ with respect to Interior.

In the models for the group of non-UPP DPs neighboring the DPs with UPP, model 4, we also used the parallel trends model to define the control groups. The best control groups were Baixada and Greater Niterói, p-values (pre-UPP) and 0.29 respectively. Using the model 4, the estimated effects were significant ($1\%$), with parameters $-1.097$ and $-0.8315$ when compared to Baixada and Niterói, respectively. This represents a decrease of $30\%$ in the rates, or 44 fewer homicides each month, when compared to Baixada.

It calls attention that the decrease effect due to the presence of a UPP, in the UPP areas (see Table 1 model 1), is lower (the estimated parameter is higher) than the same effect in the non-UPP areas neighboring UPPs, both with relation to Greater Niterói. This shows that the neighboring areas present a stronger decrease in homicide. Knowing that approximately 80% of the DPs without UPP in the city of Rio de Janeiro are neighbors to an area with UPP, this result corroborates that areas without UPP in the city of Rio de Janeiro have kept the decrease in yearly homicide rates between 2012 and 2014.

4 Conclusions

The presented results indicate that the DPs which received at least one UPP presented a decrease in homicide rates. The estimated effects are general, obtained by monthly series from 2003 to 2017, pre- and post-UPP periods, and therefore do not show a change in trends of the yearly series (see Table 1), but the estimates are certainly affected.

We also observe that the decrease in homicide is stronger in non-UPP areas neighboring the areas with UPP. The analysis of yearly series showed a new dynamic of homicide occurrences. After a period of decrease throughout the state, the rates increased between 2012 and 2014 in almost all regions. We highlight the city of Rio de Janeiro where homicide rates increased around the UPPs, i.e., in the DP areas that have UPPs, but outside the pacified communities (i.e., in its margins). In the remaining regions of the city the homicides have decreased: in the UPP areas, 30% [6], and in the areas without UPP, 5%. The new strategy of crime behavior can be attributed to the police occupation. With the territory reclaimed and densely policed, the intervention inhibits criminal action and moves the confrontations to the margins of UPPs. The fact that the rates in the remaining regions have increased is also an evidence of a change in criminality, which deserves attention in future work.