Law Enforcement, Social Control and Organized Crime: Evidence from Local Government Dismissals in Italy

Abstract

Local governments suspected of Mafia infiltration can be dismissed in Italy through an administrative act not increasing formal deterrence but potentially signaling improved law enforcement among local communities. This paper finds that dismissals are associated to a persistent fall of petty crimes (e.g. thefts) but have little consequences on offenses more closely related to the activity of organized crime, as homicide, extortion, drug-trafficking or usury. Petty crimes are estimated to fall by around 10%, on average, a result that seems driven by the perception of enhanced deterrence (through media pressure, the signaling role of the policy, and other forms of social control) rather than induced by organized crime itself.

1 Introduction

The presence of organized crime is detrimental for local economic outcomes as growth and productivity, the allocation of public spending, corruption, let alone public safety and law offenses (Pinotti 2015; Barone and Narciso 2015). And yet there is limited empirical evidence on the effectiveness of policies aimed at fighting organized crime, especially in the territories where its presence is pervasive.¹ In this paper we estimate the consequences for crime of an Italian policy imposing the dismissal of local administrations suspected of Mafia infiltration (a measure we label local government dismissals—LGDs). Introduced in the early 1990s as an unconventional policy tool to combat the sharp increase in the activity of organized crime in the South of Italy, LGDs have since then been implemented in dozens of cases (Priolo 2004).

Are LGDs effective in lowering crime? The measure implies the dismissed administration is replaced by an external commission which governs the municipality until new elections, typically after around 2 years. Hence, it induces a sharp (though possibly temporary) increase in law enforcement in territories where the presence of the state is perceived as extremely weak. Being largely an administrative act, however, LGDs do not imply the strengthening of formal deterrence, as increased police deployment or financial resources allocated to public order. In addressing the above question we therefore adopt a broad perspective looking at alternative potential channels. The first is that, by breaking its ties with local politicians, dismissals have a direct effect on the presence of mafia and its local criminal activity. Alternatively, LGDs might more generally affect local crime by inducing the perception of greater enforcement if—as posited by influential theories since Sah (1991)—perceptions matter for deterrence.

Quantifying the consequences of LGDs on crime is challenging given that Mafia-infiltrated municipalities may feature peculiar patterns of criminal offenses. For example, if dismissals occur as a consequence of high levels of crime then simple comparisons with other municipalities would likely over-emphasize their effectiveness. Absent suitable instruments, our analysis leverages on newly available time series of municipality level data on a detailed list of criminal offenses. This allows us to exploit within-municipality variation, comparing changes in crime rates around the intervention net of time-invariant differences in crime between municipalities. Because dismissals might still occur as a consequence of spikes in crime, we will extend the analysis to the immediate neighborhood of the dismissed municipalities, often excluding the latter from the sample. Identification arises from contrasting changes in crime in (neighbors of) a dismissed municipality with contemporaneous changes in (neighbors of) municipalities that are not dismissed in the same year. Finally, we exploit the fact that reverse causality should be more of a concern for mafia-related crime (homicide, threat, extortion, arson, drug-trafficking, usury, etc.) and examine them separately from petty crime, including sexual offenses and thefts. Preliminary inspections of the data reassuringly show that in treated areas the patterns of petty crime are fairly stable prior to the intervention and do not anticipate the dismissal. They do drop in its aftermath, though.

We find that LGDs are associated with a significant and persistent fall in minor (petty) crimes, whose incidence is estimated to decrease by 9% on impact, and by 12.4% in the years following the dismissal. The fall in crime is detected irrespective of the dismissed municipality being included in the sample, and is robust to several sensitivity checks. Its geographical scope seems limited though, as dismissals do not affect crime rates beyond the municipality’s immediate neighbors. Our findings imply social gains from lower petty crimes ranging between 100 and 230 k Euro per dismissal, depending on the specification and discount rate. By contrast, dismissals do not prove effective in reducing mafia-related (power and enterprise syndicate) offenses, implying little direct effects of the policy on the activity of organized crime.

Such heterogeneous responses suggest the fall in minor offenses should not be ascribed to a ‘lay-low’strategy, whereas local crime syndicates command a generalized reduction of criminal activities following the intervention. With formal enforcement unlikely to be a relevant determinant either, we tentatively explore the role of perceived deterrence, as described for example by Lochner (2007). With LGDs, individual perceptions may change as a consequence of direct social learning (e.g. if friends, relatives, or affiliated to the same criminal organizations are caught), of indirect social learning (e.g. if some case determines a higher media pressure), or following changes in the surrounding environment (e.g. a greater urban order). Our results are consistent with forms of perception-based deterrence spurred by the signaling value of the policy or media attention. Moreover, LGDs look more effective where the endowment of civic capital is higher, consistently with social control theories (Keizer et al. 2008).

A large body of research studies the origins of Mafia-type organizations, their functioning and internal organization (Buonanno et al. 2015; Dimico et al. 2017; Bandiera 2003; Mastrobuoni 2015; Campaniello et al. 2016). Fewer works attempted to assess their economic consequences (Pinotti 2015; Barone and Narciso 2015; Daniele and Marani 2011; Mirenda et al. 2019), and very little is known as to the effectiveness of policies attempting to increase law enforcement in areas where organized crime is pervasive. We contribute suggesting that enforcement policies can successfully reduce the incidence of (minor) crime in Mafia-dominated regions, and trace the effect to perception-based mechanisms, whose effectiveness is not much investigated relative to formal enforcement. Other recent work focused on LGDs but evaluated their consequences on public spending (Acconcia et al. 2014; Galletta 2017; Cataldo and Mastrorocco 2019) and the quality of local politicians (Daniele and Geys 2015), rather than crime rates. More generally, this work contributes to the empirical literature evaluating the effects of increasing enforcement on local crime (Machin and Marie 2011; Di Tella and Schargrodsky 2004), although the specific policy considered does not imply an increase in actual deterrence (police deployment or additional transfers). By studying the consequences of LGD beyond the borders of the dismissed municipality, it also relates to empirical works on the cross-border effects of law enforcement (Gonzalez-Navarro 2013; Dube et al. 2013; Bronars and Lott 1998).

The paper is organized as follows. Section 2 presents the institutional setting. Section 3 describes the dataset, the variables construction and presents some descriptive statistics. Section 4 explains the empirical strategy, while Sect. 5 discusses the main results, presents the robustness checks, and provides an attempt to size the social benefits from the implementation of the policy. Section 6 discusses the possible mechanisms. Section 7 concludes and derives policy implications.

2 Institutional Setting

In 1991, in response to a steer increase in the activity of organized crime in the South of Italy, the Italian Parliament passed a law establishing the possibility to dismiss the political body of a municipality (i.e. city council, major and governing council) in case of suspected or actual connections with criminal organizations.² Since its introduction in 1991, about 300 municipalities have been dismissed (Fig. 1). Almost all the dismissals (278) took place in the three regions in Southern Italy were Mafia-type organizations were historically born and are still pervasive in the social, political and economic context (Campania, Calabria and Sicily).³ More than 3 millions citizens experienced at least one dismissal of the municipality where they reside because of Mafia infiltration: this figure corresponds to almost 6.5% of the overall Italian population, and almost 19% of the population of Southern Italy.

Fig. 1

Number of Local Government Dismissals by year. The bar chart indicates the number of Local Government Dismissals (LGD) occurred in each year, since the first implementation of the policy in 1991. The black bars highlight the years exploited in the empirical analysis (2004–2011). The year of dismissal refers to the year of the official publication of the Presidential Decree Source: based on The dataset of dismissed municipalities—AvvisoPubblico. Last update: May 21, 2019

In his analysis of dismissals occurred from 1991 to 2006, Mete (2009) observed that these usually started following some resounding event (e.g. political intimidation during the elections, attacks involving the major or members of the city council), corruption scandals (e.g. in auctions for public provisions or in the management of public services), or as incidental follow-ups from other investigations on the activity of criminal organizations. Such information is collected by the police and transferred to the Prefects (the local representative of the Ministry of Interior) who decides whether to start an investigation activity.⁴

The policy intervention is then implemented in two distinct steps. In a first phase, a team of external commissioners is appointed and sent to the municipality to proceed with formal investigations. At this stage, the existence of suspects of connections between Mafia-type organizations and the local government becomes publicly known. On the ground of the evidence collected, the decision to dismiss the local government is taken by the central government, and formalized by issuing an ad hoc Decree.⁵ We label this first step as the investigation and dismissal phase. In the second phase, three external commissioners appointed by the central government replace the dismissed local government and administrate the municipality for a minimum period of 18 months to a maximum of 24 (i.e. the compulsory administration phase), after which new elections take place. The compulsory administration aims at re-establishing the normal conditions of legality and public safety, the provision of public services, licensing, sectoral regulation and the collection of taxes. The presence of the police force is not increased during the compulsory administration (Mete 2009; Cavaliere 2004).⁶

The implementation of this policy tool does not hinge upon a constant supervision by the police forces on all governing bodies in territories heavily influenced by the presence of organize crime. The police forces and the judicial authority are rather subject to stringent budget constraints which lead them to concentrate the investigation activities in cases in which the evidence for the dismissal are particularly resounding or when evidence of organized crime influence in the local governments is incidentally found. In the sample considered in the empirical analysis, the majority of the LGDs occurred because evidence of links between the local government and organized crime were incidentally found in investigations related public tenders run by the municipality. While in Italy, for public security reasons, it is not possible to obtain information on the deployment of police forces on the territory, all the works describing the establishment and the effects of LGDs show that starting from the mid 2000’s, LGD has become a purely administrative act, thus not followed by an increase of control in the territory established by additional police forces (Mete 2009; Cavaliere 2004). This was also confirmed by conversations with former commissioners, judicial authorities, and police officials.

3 Data and Descriptive Statistics

3.1 Data and Variables Definition

We exploit the Investigation System Database (ISD), a newly available dataset containing information on a detailed list of offenses committed in all Italian municipalities on yearly basis, from 2004 to 2011. The ISD draws directly from the IT system used for investigation activities by the police, and it is collected and managed by the Italian Ministry of Interior.⁷ The number of offenses in each municipality includes those directly discovered by the police in its day-by-day investigation activity, as well as those reported to the police by citizens, or by the judiciary authority starting an investigation.

The dependent variables ($C_{it}^{j}$) are constructed as the sum of the offenses reported in each municipality (i) and year (t), for each offense category (j). We draw from the social science literature and exploit the offense categories based on the distinction among offenses linked to the power syndicate, enterprise syndicate of criminal organizations, and petty crimes (Block 1980; Mete 2009). In fact, the standard distinction between violent, property and economic-related crimes might not be suitable to study local crime dynamics in the areas with a pervasive and persistent presence of organized crime (Lupo 1993). Besides, the use of this categorization is new in the economic literature and offers interesting insights to the analysis.⁸Power syndicate encompasses offenses which appear to be strongly linked to the power of organized crime associations to control the territory. For example, this category includes murderers, attacks, assaults, extortions, arsons. Enterprise syndicate includes offenses connected with the business of organized criminal associations, such as drug related crimes, prostitution, usury and smuggling. These are also offenses not necessarily linked to the territory where Mafia type organizations are historically settled. The sum of power and enterprise syndicate crimes constitutes what we define to as Mafia-related crimes. The Petty crimes category collects criminal acts that are not under the monopolistic control of organized crime and could be also linked to criminal behavior outside organized crime (e.g. thefts, auto thefts, sexual offenses).

Additional socio-economic and demographic information is obtained from several sources: the Demographic Database (ISTAT Demos) from the Italian National Bureau of Statistics, which includes information on the population resident in Italian municipalities on yearly basis; the Public Finance Database (PFDB), containing information on personal taxable income at the municipality level; the Public Finance Reports (PFR) collected by the Italian National Bureau of Statistics (ISTAT) and the Ministry of Interior, which contain information on dismissals. These data are linked to the ISD by year and municipality identifier so to build a panel dataset with time-variant measures of criminal activity, population and income dynamics at the municipality level for 8 years (2004–2011). Finally, we collect additional information on the dismissals taken from the official Decrees (whenever available), and data from Google Trends to proxy for the media attention received by each dismissal (see Appendix A).

3.2 Descriptive Statistics

We focus on the 55 dismissals of local governments occurred between 2004 and 2011 located in Southern Italy (Sicily, Campania and Calabria) and Lazio (see Fig. 1).⁹ We identify 204 neighboring municipalities, those sharing at least one border with each dismissed municipality. All units are observed for 8 years (so that $N=1632$).¹⁰

Table 1 Descriptive statistics for neighboring and dismissed municipalities Sources: ISD (2004–2011), ISTAT-DEMOS, Ministry of Economics and Finance, Public Finance Database

Table 1 contains descriptive statistics on the dependent variables ($C_{it}^{j}$) and the set of the control variables, distinguishing between the subsamples of neighboring and dismissed municipalities. For descriptive purposes we also compute the crime rates (i.e. crime counts by 1000 resident population), which are easier to interpret. In both samples, petty crimes account for a substantial part of the local crime dynamics: we observe on average 12 petty crimes every 1000 inhabitants, 9 enterprise syndicate crimes, and 1 power syndicate crime. These figures are quite similar across neighboring and dismissed municipalities, thus confirming the local crime dynamics do not differ substantially among the two groups: we further investigate this aspect in the next section. The set of control variables includes the share of male aged 15–24, the share of male aged 25–29, the average personal taxable income and the share of households below a poverty threshold. Figure 2 shows the distribution of the crime counts, by offense categories: following a count data generating process, the variables are highly left skewed with a non-negligible mass of zeros (Osgood 2000).

Fig. 2

Distribution of crime counts by offense categories. Observations above the 99th percentile of each crime count distribution are excluded from the plot Source: based on ISD (2004–2011)

4 Empirical Strategy

4.1 Empirical Specification

One obvious concern when assessing the crime consequences of LGDs is that dismissed municipalities might differ from other municipalities in some fundamental and unobserved way. Put simply, the intervention might be a consequence of high levels of, or spikes in, Mafia-related, violent crimes which would not be observed in comparison groups. The empirical approach attempts to account for this major threat to identification in several complementary ways.

First, estimates will absorb time-invariant differences in crime with municipality-fixed-effects exploiting within-municipality variation in crimes over time. Second, the analysis allows the impact of LGDs to potentially spread beyond the borders of the dismissed municipality, and extend to the sample of neighbors (around 4 municipalities per dismissal, on average). In fact, to limit identification concerns our core results will be obtained excluding the dismissed from the sample. Hence, in the baseline specification identification arises from contrasting changes in crime in (neighbors of) a dismissed municipality with contemporaneous changes in (neighbors of) municipalities that are not dismissed in the same year. The results will be subject to standard robustness and falsification tests. Third, the analysis will separately focus on three typologies of crime: power syndicate crimes, enterprise syndicate crimes, and petty crimes. The presumption is that while dismissals might be triggered by spikes in Mafia-related crimes (as murders, extortions, arsons, usury, drug related crimes, etc. which fall in the first two categories) they are much less likely to be correlated to the local dynamic of petty crimes (theft and sexual offenses). Indirect supportive evidence to this hypothesis will be discussed in the next subsection.

Taking also into account the count data nature of the dependent variables ($C^{j}_{it}$), the reduced form estimating equation is:

$$\begin{aligned} E[C^{j}_{it}]=\exp ( \beta _1 Inst_{it}+ \beta _2 Post_{it}+ X^{'}_{it}\alpha + \lambda ln(Pop_{it}) + \gamma _i + \delta _y ) \end{aligned}$$

(1)

where $C^{j}_{it}$ is the count of crimes of category j in municipality i, year t. Years t are centered on the year of the dismissal ($t=D$).¹¹ The specification allows the impact of the policy on crime to vary over time: $Inst_{it}$, is an indicator equal to one if $t=[I;D]$, thus measuring the consequences of Inspection and dismissal; $Post_{it}$ is a dummy variable capturing the Post-dismissal effects (i.e. covering the period of, and subsequent to, the compulsory administration, $t\ge 1$). Vector $X'_{it}$ includes relevant time-variant characteristics (e.g. population characteristics and wealth in municipality i), while the (natural) logarithm of the population ($ln(Pop_{it})$) is included as the standard exposure variable in count-data models (Cameron and Trivedi 2013). Finally, $\gamma _{i}$ are municipality fixed effects and $\delta _y$ are calendar year fixed effects capturing aggregate shocks to crime.

While the analysis will provide a broad set of results on the crime consequences of dismissals, estimates exploiting changes in the patterns of petty crimes in the immediate proximity of the dismissed municipality are the least likely to reflect spurious correlation of crime rates with the intervention. Conditional on the identification assumption, the coefficients of interest can be interpreted as intention-to-treat (ITT) parameters, and Sect. 6 will discuss the relative importance of alternative potential underlying mechanisms. Finally, the analysis will consider alternative metrics of distance and testing for the displacement of criminal activities from the center to the periphery.

4.2 Preliminary Evidence

The empirical setting allows quantifying the consequences of LGDs on crime if crime rates are orthogonal to the dismissal. It is therefore important to provide supportive evidence of this assumption. To this purpose, in Fig. 3 we plotted the pre-intervention patterns of crime rates of dismissed municipalities (and their neighbors), distinguishing petty crimes from mafia-related crimes. Specifically, we grouped municipalities by dismissal year and computed the average crime rate of each group and year, focusing on the period before the inspection (($t=I$), which is potentially part of the treatment). For example, the yellow line in Panel A plots the patterns of petty crimes for municipalities dismissed in 2010, in the years before 2008. Similarly, the green line plots crimes for municipalities dismissed in 2009.¹² The pictures show that, when they overlap, the average patterns of crime rates before the interventions appear reasonably parallel, in particular starting in 2005.

Fig. 3

Petty and Mafia-related crime pre-trends. The figure shows the average crime rates (Petty crime in a and Mafia-related crime in b) for neighboring and dismissed municipalities, grouped by the year of dismissal, in the years before the policy intervention (i.e., $t < I$). Dismissals occurred on 2004 and 2005 are excluded Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

Appendix Fig. 6 deals with another potentially relevant source of confounds, showing that dismissals are not systematically related to a specific period of the year (seasonality) or to other relevant events, in particular the occurrence of local elections.

We also perform tests of spatial autocorrelation in the crime rates of the dismissed municipalities and its neighbors in the years before the LGD (see Appendix B for a detailed discussion of the approach and results). The null of spatial autocorrelation can be rejected for all dismissed municipalities in the case of petty crimes, and for most in the case of power and enterprise syndicate crimes. Our core findings are, however, robust to the exclusion of such municipalities.

5 Estimated Consequences of Dismissals on Local Crime

5.1 Baseline Results on Alternative Typologies of Crime

We estimate Eq. (1) using Quasi-Maximum Likelihood (QML) methods for panel fixed-effects Poisson estimators (Cameron and Trivedi 2013; Osgood 2000), while in the robustness checks we also test alternative specifications. As $Inst_{it}$ and $Post_{it}$ are dummy variables, the corresponding estimated parameters from Poisson regressions ($\beta _1$ and $\beta _2$) have a straightforward interpretation as semi-elasticities.¹³

Table 2 Baseline results: LGD and petty crimes Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

Panel A of Table 2 reports our baseline estimates of the consequences of dismissals starting from the case of petty crimes (thefts, sexual offenses). Identification exploits within municipality variation and distinguishes immediate from delayed effects. Results in column (1) focus on neighboring municipalities, where the LGD has a statistically significant crime reducing effect: in the inspection and dismissal years petty crimes decrease by 9% relative to the pre-policy period; after dismissal they are more than 12% lower. The following two columns highlight that these effects are determined by the response of thefts (column 2). The last two columns extend the analysis to dismissed municipalities. The results suggest no major differences in the response pattern of crime relative to the neighboring areas (despite less precision in the estimate, see column 4). As a result, the estimated fall in petty crime does not change pooling the two groups of municipalities (column 5).

Panel B further distinguishes the post dismissal case into the compulsory administration phase (Post-dismissal: short term, $t=[1;2]$), and the following years (Post-dismissal: medium term, $t \ge 3$). The estimates do not highlight any sizable difference between the two periods suggesting that the crime reducing effects from increased enforcement are not concentrated in the dismissal period or during the compulsory administration phase, but extend to the medium-term, when new elections take place.¹⁴

By contrast, government dismissals seem to have little consequences on the typical activities of organized crime. Replicating the previous analysis on mafia-related crimes (power and enterprise syndicate crimes), in fact, yields smaller and largely non statistically significant (albeit negative) estimated coefficients (Table 3). This is confirmed if restricting to those mafia-related crimes with little chance of going unreported (arsons and murders, see Appendix Table 8).¹⁵ This asymmetric response of different crime typologies is difficult to reconcile with the idea that organized crime syndicates can effectively determine all criminal activities in their territory, and are therefore the ultimate responsible of the observed decrease in,for example, the incidence of thefts following the dismissal.

Table 3 Baseline results: LGD and Mafia-related crimes Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

5.2 Robustness and Falsification Tests

Table 4 shows the results of several robustness checks for our baseline specification (i.e. column 1 in Table 2). The results in column 1 are obtained adding municipality-specific trends, thus exploiting deviations in crime rates from an underlying municipality-specific dynamic. This is a demanding specification given the short time span of the data, and indeed the estimated coefficients fall significantly relative to the baseline (a result that will be taken into account later in the analysis). Importantly, however, they remain statistically significant. The specification in column 2 accounts for potential spillover effects from crime in the dismissed municipality to the neighborhood. The results are not affected in this case. Columns 3 and 4 alter the underlying sample. Dropping duplicate observations, that is municipalities that happen to be neighbors of two dismissed at the same time, does not alter the core findings (column 3). In column 4, the sample is restricted to interventions occurred in the central years in the available time window (2005–2010) so to work with a similar number of observations before and after the dismissal. In this case, point estimates are slightly lower and less precisely estimated than in the baseline.

Table 4 Robustness and specification tests Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

Appendix Table 10 shows that the same results hold if the robustness checks are performed on the larger sample including the dismissed municipalities. It also reports robustness to changing estimation methods (again in both samples) and running a negative binomial, as opposed to the Poisson model. This check is meant to assess whether the Poisson standard errors are ‘too small’due to over dispersion in the data (i.e. the conditional variance is larger than the conditional mean).¹⁶ While the point estimates can not be directly compared to those in the main text (Cameron and Trivedi 2013), the results confirm that dismissals induce a significant and persistent fall in petty crime following the policy intervention.

We also performed a falsification test looking at whether government dismissals not due to mafia infiltration end up affecting crime. Because they still imply the presence of technocrats for a certain period, a comparison with such episodes could help disentangling the sources of the deterrence effects. Similarly to other papers studying the same policy intervention, we focused on municipalities dismissed due to the mayor death by natural causes, restricting to Southern Italian regions for the sake of comparability. There are several advantages in focusing on mayors’ deaths. First, it is reasonable to assume that (natural) deaths are exogenous to the crime dynamics of the municipality (i.e. the treatment can be considered as-good-as random). Second, the number of occurrences (65) is in line with the number of dismissals for Mafia reasons (55). Table 5 replicates our baseline regressions, showing no significant changes in the patterns of petty and mafia-related crimes around dismissals of this kind. This suggests that our baseline finding in Sect. 5.1 is associated to the detection of illegal activities, rather than the presence of commissioners per se. The next session will explore some potentially relevant underlying channels in greater extent.

Table 5 Falsification test: dismissals occurred because of the (natural) death of the mayor Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

5.3 Do LGDs Displace Crime?

Spill-over effects of law enforcement are not new in the literature. The positive impact of LGD on neighboring municipalities is in line, for example, with findings on compliance with tax payments following an increase in inspections (Rincke and Traxler 2011). However, highly mobile crimes can also be easily displaced if prevention efforts are not uniformly implemented over the territory (Gonzalez-Navarro 2013; Dube et al. 2013; Bronars and Lott 1998). If this is the case, the observed fall in crime in and around dismissed municipalities might not imply a net aggregate reduction if accompanied by an increase in crime in further away areas. The core analysis was therefore replicated looking at municipalities located at further and increasing distances from the dismissed. In almost all cases, neighboring municipalities fall within a radius of 20 km from the dismissed. Here, the sample is composed of those located within non-overlapping rings (annuli) with inner circle of radius d $=$ {15, 20, 25,..., 45} and larger circle of radius d+5.¹⁷

Fig. 4

Displacement effects in petty crime. The graphs show the estimated coefficients (blue dots) and the 90% confidence interval (dashed black lines) for the inspection and dismissal effect (a i.e. 1 from Eq. 1) and the post-dismissal effect (b i.e. 2 from Eq. 1). Estimates obtained using samples including municipalities at increasing distance from the dismissed one (i.e. annuli increasing their radius each time by 5 km) for petty crimes Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011) (color figure online)

The results are summarized in Fig. 4, which plots the Inspection and dismissal effect ($\widehat{\beta _1}$) in Panel A, and the Post-dismissal effect ($\widehat{\beta _2}$) in Panel B. The estimated coefficients are negative and marginally significant only in the case of the smallest circle (with 15 km radius), with point estimates close to those obtained for neighboring municipalities and reported in Table 2. In no other case we estimate positive and significant coefficients, suggesting that displacement effects are not present (or not strong enough to offset the crime reducing spillover from the intervention) outside the closest area around the treated unit.

5.4 Some Implications for Welfare

While tackling minor offenses is not the main motivation behind dismissals, it certainly contributes to their economic effectiveness. A simple measure allowing to quantify the gains, in terms of social benefits (SB), from the intervention can be obtained as in Machin and Marie (2011): $SB=[(\widehat{\beta }_{1} + (1+d)^{-1} \widehat{\beta }_{2}) \cdot \overline{K}_{pre}\cdot USC]$, where $\widehat{\beta }$ are the estimated policy impacts (Table 2, column 1), $\overline{K}_{pre}$ is the average number of recorded offenses in the pre-policy periods, USC indicates the unitary social cost of petty crime crime (e.g. a theft) and d is a discount rate of the post-dismissal benefits.¹⁸ The unitary social cost for petty crimes (USC) is taken from Detotto and Vannini (2010), who produced detailed estimates for each of the offenses available in the ISD data underlying the present analysis. Clearly, this calculation does not allow for a full welfare assessment of dismissals, which should factor in the potential social benefits from the improved quality of local administration (e.g. public expenditure, tax collection, law enforcement) and the social cost of the policy.¹⁹

We report the the calculation of social benefits in the Appendix Table 11 (figures are expressed in 2007 euros). Based on the baseline coefficients in Table 2 column 1 a dismissal implies an average reduction of about 140 petty crimes (essentially, thefts). The (monetary) cost of each such crime, USC, amounts to about 7650 euros. The social benefits of dismissals would amount, in this case, to between 231 k and 187 k euros. However, the more demanding specification in Table 4 column 1 (including municipality specific linear trends) yielded to smaller coefficients. Based on those estimates the gains from lower crime incidence following the intervention would amount to between 128 k and 102 k euros.²⁰

For reference, the above figures amount to between 0.73 and 0.58 times (based on the most conservative estimates) the social benefits from the implementation of a policy (the Street Crime Initiative, SCI) explicitly targeted to high-density street crime areas of England and Wales (see Machin and Marie 2011). While the estimated effect of the policy on crime rates is quite similar to that obtained here, the difference is explained by the larger size of the population (and number of crimes) targeted by the SCI relative to the LGDs.

6 Potential Channels

What explains the sizable reduction in petty crime following dismissals? As discussed in Sect. 2, the policy does not imply an increase in actual deterrence, for example through higher sanctions, transfers or police deployment. Several alternative channels remain, however, whose relative importance can be indirectly tested.

Since Sah (1991) and Lochner (2007) it is known that, even absent changes in actual or formal deterrence, criminal behavior can be influenced by the perceived certainty or severity of punishment (perceived deterrence). Absent direct measures, one can test for heterogeneity in the effects of LGDs along dimensions that are plausibly correlated with the perception of law enforcement and the presence of the state among the the local population. One such dimension is media pressure, an important determinant of perceptions according to modern criminology (Surette 2011; Jewkes 2010). Variation in the degree of media attention across different dismissals can be measured through the Google Trend Index, which is available at the monthly and municipality level.²¹ Another plausible correlate of perceived deterrence across dismissals is the occurrence of arrests during the inspection and dismissal phase (which are reported in the formal Decree issued upon the dismissal, see Sect. 3).

Figure 5 summarizes the results obtained running the baseline petty-crime regression on the two subsamples of municipality featuring, respectively, high and low intensity of each characteristic.²² The first panel clearly shows that the crime reducing effect of dismissals is statistically significant only when these attracted a lot of attention by the media (these coefficients are plotted as diamonds, with blue indicating the instantaneous effect ($\beta _1$), and red the subsequent effect ($\beta _2$)). While less clear-cut, the evidence on arrests is also consistent with perceptions-based deterrence channels.²³

Fig. 5

Potential channels: heterogeneous effects according to media attention, contextual arrests, civicness. The graphs show the coefficients and 90% confidence intervals estimated in the sub-samples of municipalities characterized by: a high/low level of media attention to the dismissal, b dismissals occurred with (high formal deterrence) or without contextual arrests (low formal deterrence), and c high/low level of civicness. Coefficients in a are sourced from Panel A of Appendix Table 12, columns 5 and 6; those in b come from Panel B, columns 3 and 4; those in c from Panel C, columns 1 and 2 Sources: based on ISD, ISTAT-DEMOS, Public Finance Database, Google Trends Index (color figure online)

The third panel tests whether the crime decreasing effects are higher in municipalities with higher civic capital endowment.²⁴ This should be the case if, as argued by many influential theories, the consequences of increased (actual or perceived) enforcement are stronger where institutions (in this case, informal institutions measured by civic capital) are of better quality (Keizer et al. 2008). The results strongly support this hypothesis.²⁵

Table 6 Heterogeneous effects according to the share of non-native residents in the municipality Sources: based on ISD, ISTAT-DEMOS, Public Finance Database, Census 2001

One last explorable explanation for the crime reducing effects of LGD is organized crime controlling, and determining the intensity of all type of offenses in their territory. Rather than a reaction to the increased law enforcement, the fall in the incidence of minor crime might just be a consequence of a Mafia-determined ‘lay-low’strategy. Section 4 already discussed that this hypothesis is somehow inconsistent with the observed patterns of Mafia-related (power or enterprise syndicates) crimes, which, unlike minor crimes, do not experience any reduction in or around dismissed municipalities. Further evidence is provided in Table 6 which distinguishes municipalities based on the presence of immigrants (i.e. the share of immigrants in the population being higher or lower than the median in the sample). The idea is that, for reasons including their weaker links to the local community and their higher geographical mobility, immigrants (and their crime decisions) should be less likely to be influenced by organized crime. The results support this presumption, as petty crime falls only in municipalities with higher share of immigrants. Interestingly, Mafia-related crimes fall following the dismissal in municipalities where immigrant concentration is lower.

7 Concluding Remarks

This paper suggests that policies strengthening law enforcement mainly through perceptions-based deterrence or stimulating social control may have non-negligible effects on crime in territories characterized by the pervasive presence of criminal organizations. Our results show that the increase in law enforcement spurred by the dismissal determines a persistent reduction in petty crimes (e.g. thefts) in excess of 10%, on average. The consequences of the intervention are quantitatively similar in both the dismissed and neighboring municipalities, but no significant changes in the patterns of minor crime (in either direction) are observed at further distances. By contrast, dismissals do not have significant consequences on crimes that are more likely to be related to the activities of organized crime as murders, extortions, arsons, usury and drug related crimes.

While not the main goal of the policy, the estimated reduction of minor crimes implies non-negligible social benefits which can be monetized in about 130 to 100 k euro (in the more conservative specification). To reference this figure, we estimated it amounts to about two thirds the social benefits implied by a policy introduced in 2002 to counter petty crime in England and Wales (Machin and Marie 2011). This is of course also a consequence of LGDs being extremely cheap relative to alternative forms of deterrence (Mete 2009).

Administrative dismissals may impose other direct or indirect costs on local communities, and should obviously not be seen as panacea.²⁶ And yet, it seems important to be aware of the potential crime-reducing effect of policies strengthening law enforcement largely through perceptions-based deterrence or forms of social control (Sah 1991; Lochner 2007). Moreover, we document the existence of positive spillovers in enforcement, which are not offset by the displacement of criminal activities. This aspect is rarely taken into consideration in the design and implementation of policies aimed at combating crime.

Appendices

Appendix A: The Google Trends Index for LGD

LGDs are usually resounding (but local) events, only sometimes reported in the national newspapers. To build a measure that would proxy the attention devoted by the media to each LGD we resort to the Google Trend Index (GTI, https://www.google.it/trends/?hl=it). Recent papers have shown that the GTI is a good predictor of near-term values of economic indicators, such as automobile sales, unemployment claims, travel destination planning, consumer confidence, and media attention (Choi and Varian 2012; D’Amuri and Marcucci 2017; Kearney and Levine 2015). The GTI provides a time series index of the volume of queries users enter into Google in a given geographic area. The index is based on query share: the total query volume for the search term in question within a particular geographic region divided by the total number of queries in that region during the time period being examined. The maximum query share in the time period specified is normalized to be 100 (Choi and Varian 2012).

We collect the GTI for each dismissal defining a time window of 25 months centered in the month of the dismissal ($t\in [-12;+12]$), and restricting the geographical area to Italy, the keyword for the search being the name of the dismissed municipality.²⁷ Figure 7 shows an example of the GTI for a dismissal occurred in 2013: the GTI peaks exactly in the month of the dismissal ($t=0$). This is somehow confirming what previously found in existing works: the GTI is more appropriate to capture the short term attention of the media, in our case, the media attention around the dismissal (Choi and Varian 2012).

The GTI is provided from the year 2004: for this reason, it was not possible to retrieve the index for nine dismissals (occurred in 2004 and at the beginning of 2005) determining a reduction of 352 observations from the baseline sample (i.e. 35 neighboring municipalities per 8 years). Moreover, municipalities located on the seaside displayed seasonal trends in the search volumes plausibly due to touristic flows. To clean out monthly and yearly trends, as well as trends linked to observable characteristic of the municipality, such as geographical location and dimension, we calculate a clean version of the GTI, taking the residuals from OLS regressions of the raw GTI on dummies for the municipality characteristics (population dummies and seaside location dummy), calendar month and calendar year fixed effects. A comparison between the raw and the clean version of the GTI is shown in Panel A and Panel B, respectively, of Fig. 8. In the main analysis and in the reminder of the section we refer to clean version of the GTI, although the results do not change substantially using the raw index.

As a final step, we build for each dismissal an average index of media attention, averaging the monthly GTI over a given number of months centered in the dismissal. Given that the GTI is suitable to capture the media attention in the short term (i.e. around the dismissal, as shown in Figs. 7 and 8), the larger the time window the less accurate could be the measure. In our baseline empirical analysis (see Table 12) we use the average GTI in the month of the dismissal or over a window of 5 months centered in the dismissal.²⁸ Finally, we distinguish between LGDs with High or Low media attention defining two subsamples according to whether the average GTI is, respectively, higher or lower than the median value. In Fig. 9 we show the GTI separately for the subsamples of LGD with High or Low media attention: the group of high media attention LGD clearly shows a peak around the dismissal (indicated with the vertical line in $t=0$), while LGD with low media attention do not show peculiar trends around the dismissal.

Appendix B: A Test of Spatial Autocorrelation

Figure 10 plots the results of testing for the presence of spatial autocorrelation in crime rates between the dismissed municipalities and their neighbors. The test (the Local Moran’s I test, see Mendes and Mendes (2015)) is computed in the years before the intervention (i.e. $t\le -1$). For each dismissed municipality, indicated on the horizontal axis, the graph reports the P-values of the test for each of the three typologies of crime considered in the paper; the 10% level of acceptance is indicated by a horizontal line.

The two panels differ in the definition of neighboring municipalities. The left panel focuses on the whole sample of neighbors, as in the core exercises in the main test (Fig. 10, Panel a). In this case the presence of spatial autocorrelation can be rejected in all but one municipality (in the case of power syndicate crimes). Figure 10, Panel b refers to a more restrictive version of the test, focused on the sample of closest neighbors (i.e. municipalities having an Euclidean distance to the dismissed falling within a band of 0–5). The test is more restrictive as spatial correlation statistics are likely to decrease in sample size. In this case, the null of spatial autocorrelation can be rejected in all but few (9), and in any case in which petty crimes are considered.

This evidence supports the idea that the incidence of crimes in the sample of neighbors would likely be unrelated to the intervention, even if this was a response to the patterns of crime in the dismissed municipality.

Appendix C: Additional Figures and Tables

See Figs. 6, 7, 8, 9, 10 and Tables 7, 8, 9, 10, 11 and 12

Fig. 6

Fraction of Local Government Dismissals by month of dismissal (a) and by distance (in months) from the local elections (b). The histograms indicate the fraction of Local Government Dismissals (LGD) by calendar month in which the LGD took place (a), as derived from the formal Presidential Decree, and the fraction of LGDs by distance (in months) from the previous local elections (b); $N=55$ Source: based on The dataset of dismissed municipalities—AvvisoPubblico. Last update: May 21, 2019

Fig. 7

The Google Trends Index for LGD: an example. The graph plots the monthly GTI (vertical axis) by timing of the intervention (horizontal axis) derived from the municipality of Sedriano, dismissed in October 2013; the vertical line in $t=0$ indicates the month of the dismissal Sources: Google Trend Index (GTI, https://www.google.it/trends/?hl=it)

Fig. 8

The Google Trends Index for LGD: a visual inspection. The graph plot the monthly GTI (vertical axis) by timing of the intervention (horizontal axis) for the subsample of dismissed municipalities for which it was possible to compute the GTI; the vertical line in $t=0$ indicates the month of the dismissal. a Shows the raw index; b shows the residuals from OLS regressions of the raw index on municipality characteristics (i.e. five dummies identifying municipalities with: (i) less than 5000 inhabitants, (ii) 5001–10,000 inhabitants, (iii) 10,001–20,000 inhabitants, (iv) 20,001–60,000 inhabitants, (v) more than 60,001 inhabitants, a dummy indicating whether the municipality is on the seaside), calendar month and calendar year fixed effects Sources: Google Trend Index (GTI, https://www.google.it/trends/?hl=it) and ISTAT (2009)

Fig. 9

The Google Trends Index: LGD with high vs. low media attention. The graph plot the monthly GTI (vertical axis) by timing of the intervention (horizontal axis) for the subsample of dismissed municipalities for which it was possible to compute the GTI, and by distinguishing among the groups of dismissals that received High or Low media attention. The groups are defined according to whether the average GTI (calculated over the GTI residuals in a time window of 5 months centered in the dismissal, $t=0$) is, respectively, higher or lower than the median value observed in the data Sources: Google Trend Index (GTI, https://www.google.it/trends/?hl=it) and ISTAT (2009)

Fig. 10

Moran’s I test for spatial autocorrelation. The scatter plots depict the results of binary Moran’s I tests between the crime rates of each dismissed municipality and those of the corresponding neighboring municipalities: the vertical axes indicates the p values of test; the horizontal axes the anonymized (dismissed) municipality ID; the red horizontal line indicates the 10% level of acceptance in the p values. a Performs the test including all neighboring municipalities (i.e. those having an Euclidean distance to the dismissed municipality falling within a band of 0–10); b restricts to the closest neighbors (i.e. municipalities having an Euclidean distance to the dismissed lower than 5) Source: based on ISD (2004–2011)

Table 7 Offense categories definition

Table 8 Additional results: separate effects on arson and murders Sources: based on ISD, ISTAT-DEMOS, Ministry of Economics and Finance, Public Finance Database, Google Trends Index

Table 9 Petty and Mafia-related crimes in neighboring and dismissed municipalities vs. all Italian municipalities Sources: based on ISD, ISTAT-DEMOS

Table 10 Robustness and specification tests on dismissed and neighboring municipalities Sources: based on ISD, ISTAT-DEMOS, Ministry of Economics and Finance, Public Finance Database, Google Trends Index

Table 11 Welfare analysis: the social benefits from the LGD Sources: based on ISD, ISTAT-DEMOS, Public Finance Database (2004–2011)

Table 12 Heterogeneous effects according to media attention, arrests, civicness Sources: based on ISD, ISTAT-DEMOS, Ministry of Economics and Finance, Public Finance Database, Google Trends Index