Do arrests (and killings) deter violent extremism? A comparative analysis

ABSTRACT

There is an ever-growing body of evidence that suggests that there exists a significant degree of overlap between violent extremism (VE) and ordinary crime, both at the conceptual level and in terms of patterns and predictors. Countries differ considerably in their approaches to countering violent extremism (CVE). Yet, at least in the west, one common feature is the criminal justice system, whose role is essentially the same for VE as it is for other forms of crime. Despite this, there is little quantitative research on policing and criminal justice system effects on VE. Among the few studies that do exist, most focus on single countries, and examine long observation periods. Our analysis compares two key democratic countries that have received less attention, Canada and Sweden, and finds evidence of heterogeneous effects and patterns concerning how arrests impact the risk of future VE. This suggests that studies focusing on single contexts may have limited generalizability and that current wisdom concerning deterrence-backlash effects is more limited than previously thought.

KEYWORDS:

• Counter violent extremism
• terrorism
• criminal justice system
• arrests
• series hazard models

Introduction

Since 9/11, combatting violent extremism (VE) has increasingly become a central feature of western countries’ national priorities. However, policies have often been developed and implemented absent of any substantial grounding in scientific research, because we have lacked a sufficient evidence base. As such, policies have to a large degree been based more on local culture, social identity, experiences, history, and politics, which explains the apparently different approaches of western countries (Hardy, 2018). However, there is one aspect of countering violent extremism (CVE) which is seemingly shared between almost all western countries and that pertains to the role of the criminal justice system, and the reliance on the criminal justice model for dealing with VE. That is, in virtually all countries, the police and security forces are responsible for investigating and apprehending offenders and potential offenders, prosecutors are responsible for laying charges, courts and judges are responsible for determining guilt and sentencing, and penal systems are responsible for supervising sentences (Hasisi, Perry & Wolfowicz, 2020).

The overlaps in how crime and VE are handled could be quite justifiable given the overlaps that exist between the phenomena. Such overlaps have been observed with respect to the risk and protective factors for offending (e.g. Wolfowicz et al., 2021), spatial and temporal patterns of offending (e.g. Hasisi, Perry, Ilan & Wolfowicz, 2020), patterns and predictors of recidivism (e.g. Hasisi, Carmel, Weisburd & Wolfowicz, 2020), and responses to situational prevention (e.g. Hasisi, Perry & Wolfowicz, 2020; Perry et al., 2017). With regard to the latter observations in particular, scholars have deduced that whilst VE offenders’ motivations and objectives may differ from those of ordinary criminals, they still broadly operate according to rational choice and decision-making process and as such, deterrence-based strategies are likely to be effective (Clarke & Newman, 2006; Newman & Hsu, 2012; Perry & Hasisi, 2015).

The question of whether violent extremism, with an earlier focus on terrorism, can be deterred represents one of the oldest and most extensive inquiries, with a large number of papers have been produced by scholars from various fields such as Political Science and International Relations (Knopf, 2010; Wilner, 2011). However, this inquiry has yielded mixed findings and, overwhelmingly, quantitative research has remained ‘near non-existent’ (Wenger & Wilner, 2012, p. 316). In recent years, Criminology has rekindled its focus on deterrence, and this has been applied to the study of violent extremism as well. Yet this has also produced mixed results. To date, there still exists a dearth of high-quality evaluations of deterrence-oriented interventions, in particular as they pertain to the role of the criminal justice system (Freilich et al., 2024; Sydes et al., 2023; Wolfowicz et al., 2023). The current study aims to address these above-mentioned literature gaps by providing a quantitative analysis of the impact of a criminal justice intervention, specifically arrest (and killing), on the occurrence of VE. Thus, the current study aims at assessing whether deterrence constitutes an area of overlap between VE and crime through the lens of criminological theories and methodologies. To do so, we conduct a comparative analysis between Canada and Sweden, two Western nations with notable similarities. Notably, both countries adopt a holistic, multi-level government approach to CVE (Andersson, 2022; O’Halloran, 2021). Additionally, Canada experiences relatively infrequent VE compared to the US, and Sweden relatively less compared to larger Western European countries such as nearby Germany. And both countries’ clearance rates for VE appear to be remarkably low. These shared characteristics, explored in greater detail in subsequent sections, make Canada and Sweden an insightful comparative framework for a deterrence study.

Through criminological theories and methods, the current study carries out an evidence-based analysis of the impact of criminal justice interventions, namely arrest and killing, on VE trends. To do so, we draw on data from the Global Terrorism Database (GTD), as well as from Statistics Canada and Europol’s TE-SAT reports, and conduct several analyses on the relationship between likelihood of arrest and subsequent VE offending. The likelihood of arrest is a key deterrent factor which is also a prerequisite for the application, or applicability of other deterrent variables. The analyses also serve to address the question as to whether a strong or weak response to VE is preferred, or whether it has the potential for creating iatrogenic effects.

Current evidence on deterrence of VE

Deterrence theory has established foundations for how we understand crime and crime prevention and has overwhelmingly driven crime control policies and law enforcement strategies. Deterrence theory perspectives posit that criminals, much like non-criminals, rationally choose to commit an offense, after weighing the costs and benefits (Cornish & Clarke, 2016). Recent findings on rationality in VE bring further evidence of the applicability of criminological frameworks, both theoretical and methodological, to understanding and preventing it. Although VE offenders have often been deemed irrational, research has found that they too are driven by the anticipation of and a desire to maximize profits and rewards whilst minimizing risks (Perry & Hasisi, 2015). Despite displaying limited or bounded rationality, VE offenders use cognitive strategies that are similar to the ones used by other criminal and non-criminal individuals (Perry & Hasisi, 2015). Evidence for this can be found in how VE reacts to situational prevention measures, demonstrating responses akin to those seen in crime (Clarke & Newman, 2006; Dugan et al., 2005; Dugan & Chenoweth, 2012; Gupta, 2008; LaFree et al., 2009; Newman & Hsu, 2012; Perry & Hasisi, 2015).

From this point of departure, deterrence theory holds that would-be offenders weigh the ‘perceived probability of successful completion of the act’ against the certainty and severity of punishment (Nagin, 2013, p. 208). The presumption that crime will be sensitive to deterrence relies on the idea that an increase in police productivity will lead to an increase in the probability that an offender will be apprehended (Pratt et al., 2017). Whilst most research has found that severity of punishment has little or no deterrent effect, the pooled effects for certainty of apprehension and punishment show an appreciable deterrent effect (Bun et al., 2020; Pratt et al., 2017).

This finding raises two main implications. First, on a policing level, crime prevention will depend on the effectiveness of the police in apprehending offenders (Nagin, 2013). Second, on a methodological level, arrests are the most reliable variable to assess the impact of the certainty of punishment on crime levels. Indeed, arrests constitute the initial step of legal apprehension and thus paint a more accurate and all-encapsulating picture of the effect of apprehension on attack rates. Additionally, arrests represent a punishment in and of themselves (Pratt et al., 2017).

In the case of VE, there exists an issue in which a substantial proportion of arrestees may be released without charge. For instance, the UK's Home Office statistics shows that in 2018, of the 250 individuals arrested under the Terrorism Act (2000) 57% were released without charge (HMG, 2018). Using charge, conviction or sentencing rates as apprehension variables misrepresents the effects of the certainty of punishment on offending rates. So too, arrest or case clearance rates have traditionally been the preferred measures to assess the effect of the certainty of apprehension on crime levels (Levitt, 1998).

Unfortunately, to date, we do not know whether VE responds to such factors in the same way that crime does. Rather, deterrence research on VE has focused on the impact of different CVE strategies and policies. For example, studies have examined how metal detectors, police and military expenditures, military raids, targeted assassinations, and house demolitions impact VE offending. The collective evidence suggests that ‘harsh’ approaches can lead to substantial backlash effects, in which they are associated with increases, rather than reductions, in VE offending, and that ‘soft’ approaches are more likely to be associated with reductions in the risk or likelihood of VE (Argomaniz & Vidal-Diez, 2015; Dugan et al., 2005; Hsu & McDowall, 2020).

As is evident, these types of studies are quite limited in what they can tell us about deterrence, at least as it pertains to the traditional deterrence model framework and theory. It is difficult to compare these results with those from traditional criminological studies that examine the effects of the likelihood of arrest or punishment for example, by examining arrest or clearance rates (Argomaniz & Vidal-Diez, 2015; Hsu & McDowall, 2020; LaFree et al., 2009; Sageman, 2014). Additionally, the statistical models used in these studies rely on dichotomous variables used to operationalize the implementation of a policy that is qualitatively categorized as being either ‘hard’ or ‘soft’, and this leads to an oversimplification of their significance (Hsu & McDowall, 2020). These approaches appear to resemble more closely those taken by studies attempting to examine the effects of specific policies, such as the death penalty, which are known to suffer from several issues, especially as they pertain to identification. This is not to be critical of the work that has been done until this point, but rather to point out the limitations in what can be deduced from the results and what they mean (Charles & Durlauf, 2013).

In searching for evidence that the deterrence model may have relevance to the issue of VE, there are few studies that measure arrests in a way that at least approaches the methods used in criminological studies. In one such study, Kaplan et al. (2006), taking Israel as a case study, find that preventive arrests decreased the number of suicide-bombing attacks. Their shot-noise model shows that a marginal increase in arrests in a given month leads to a decrease in the total number of future suicide bombings by 0.24 times. Similarly, Hussain’s thesis study (2010), which uses Pakistan as a case study, finds that each additional arrest is associated with a 5.0% decrease in the number of terrorism incidents in the subsequent period (Hussain, 2010, p. 138).

As Hussain (2010) explains, ‘the impact of terrorist arrests on terrorism has not been studied with enough details to make any generalizations’ (Hussain, 2010, p. 111). However, there have been some subsequent additions. Sandler et al. (2011) estimate that as the number of Interpol arrests increased, the number of transnational terrorist attacks decreased. Wolfowicz et al. (2023) conducted a panel data study on 28 EU member states and found that increased arrests were consistently associated with reductions in terrorism. However, as the small number of studies cited here indicates, the evidence-base is still quite limited, and is mostly limited to terrorism rather than a broader category of VE. In their study, Wolfowicz and Salama (2024) extend the evaluation of the effects of arrests on other forms of collective violence, finding that while arrests have deterrent effects, harsher forms of punishment display evidence of backlash effects.

The current study

Following from the above, the current study’s aims are to quantitatively investigate whether VE is deterred by arrests and to assess this dynamic in a way that is in line with theoretical and methodological frameworks employed in the study of crime. In particular, this means that we are interested in assessing how arrest rates, or the number of arrests impact the subsequent risk of VE. We are also interested in the effects of another potentially deterrent outcome, namely the elimination of offenders in the commission of an offense or in the course of authorities’ response to it. In line with deterrence theory, we hypothesize that increased arrests and killings of offenders will be associated with reductions in the risk of VE.

To date, a universally acknowledged or official definition of terrorism and VE does not exist (Easson & Schmid, 2011; Freilich et al., 2024; LaFree & Schwarzenbach, 2021). Therefore, the study will use as a framework a broad definition of VE, as outlined by the National Strategy on Countering Radicalization to Violence: ‘[v]iolent extremism is a term describing the belief and actions of people who support or use violence to achieve extreme ideological, religious or political goals’ (Canada Center, 2018). The focus of the paper is on incidents of VE, which not only capture terrorism incidents but also hate crimes and other related manifestations of ideological, religious, or political violence.

As noted above, most studies focus on only one specific country, and the types of policies that are used as proxies are often not comparable, beyond the qualitative labels assigned to them such as ‘hard’ vs. ‘soft’ approaches. The absence of comparable interventions, and direct country comparisons, serves to limit the more global implications that can be derived from findings. Another issue is that comparing results from these studies may ignore the different timelines being researched. For example, Argomaniz & Vidal-Diez (2015) examine the Spanish context from 1977 to 2010 and LaFree et al. (2009) examine the case of Northern Ireland from 1969 to 2002, both finding evidence of backlash effects.

There are some exceptions, however. Carson (2019) examined how targeted killings of terrorists impacted terrorism in three countries: Pakistan, Yemen and Somalia. The study found that there were differential effects between the countries. Hodwitz and King’s (2023) analysis compared the effects of the imposition of terrorism acts in Canada, the US, and the UK from 1997 to 2016. The analysis revealed that the effects of seemingly comparable national policies were different across countries, being associated with reductions in terrorism in Canada, and the US but increases in the UK. While these studies demonstrate the importance of country comparisons, they are still limited by being relegated to examining dichotomous, dummy variables to represent interventions.

Following from Hodwitz and King (2023), we sought to focus on Canada, which experienced only a fraction of the number of incidents that occurred on the US and the UK, and to find a more appropriate comparison. The identified comparison was Sweden. In line with calls for more comparative research, and the limitations of current studies examining only single countries, our analysis examines Canada and Sweden. We chose to compare these two countries for several reasons. First, our study is part of a larger project in which several research institutions are involved in comparative research between Canada and Sweden across a broad range of VE related topics and issues. Our analysis is therefore situated within this broader research agenda and context.

More substantively however, these two Western democracies share parallels in their CVE approach, placing a growing emphasis on the ‘soft side’ of counter-terrorism through a comprehensive ‘whole-society approach’ (Andersson, 2022; O’Halloran, 2021). This approach encompasses multi-level and inter-governmental policies and involves active collaboration with civil society and social welfare actors to address not only the immediate threat but also its perceived root causes (Silva & Deflem, 2020).

Furthermore, both Canada and Sweden’s CVE policies have been influenced by events that have occurred outside of their borders and many policy ideas have been transferred from abroad (Mattsson, 2019; O’Halloran, 2021). These countries are unique in that despite having suffered a comparatively smaller number of VE incidents, compared to their larger neighbors, the rate of incidents per capita is comparable. For example, whilst Canada certainly suffers fewer incidents than its larger US neighbor, the rate for both countries is 0.1. Similarly, although Sweden has suffered far fewer incidents than its larger neighbors Germany, its rate is actually higher at 0.9.

Additionally, compared to other countries where clearance rates for VE appear to be even higher than other types of offending (see Wolfowicz et al., 2023), clearance rates for VE in Canada and Sweden appear to be remarkably low, at least based on data from the GTD (see below). Generally speaking, Canada is considered to have a very high clearance rate, however, especially in the case of left-wing/environmentalist VE, which has included several bombing attacks, few arrests have been carried out. In the case of Sweden, clearance rates may be better than has been claimed. However, in the case of hate crime, Sweden has notoriously low rates. This appears to overlap with data in the GTD (see below) which indicates that especially in the case of arson attacks against refugee facilities, which represent a large proportion of all recorded events, clearance rates are quite low.

These unique characteristics may provide for the opportunity to assess how differences in clearance rates, a key factor in deterrence, may impact the risk of VE. While deterrence theory posits that higher clearance rates, as indicators of increased risk, ought to decrease offending, local contextual factors, such as the quality of arrests, can give rise to significant variability.

Methods

Data preparation

In this study we relied on data from the Global Terrorism Database (GTD). As opposed to some previous studies, we sought to focus on a period in which the specific VE threats were more comparable and therefore we limited our scope to the post 9/11 period of 2001-2020. The GTD is the most frequently relied upon data source in terrorism and VE research as it includes over 200,000 incidents from across the world with a relatively high level of detail (Johnson & Ackerman, 2023). Despite its name, the GTD also includes many incidents that are sub-terroristic forms of violent extremism. That is, the GTD tends to be highly inclusive, often including incidents that may be considered insignificant by others (LaFree & Dugan, 2007; Lehrke & Schomaker, 2016). It has been found that the GTD often (although not always) includes a significantly larger number of incidents than countries’ official statistics (e.g. Ščurek, 2021).

In the case of the current study, there is evidence in support of this. For example, for Sweden, several incidents in the mid 2000s suspected to have been carried out by Global Intifada (a left-wing extremist group) are included in the GTD data, including arson attacks on grocery stores in February 2009. This despite the incidents not being included in terrorism data reported by Sweden to Europol (Wolfowicz et al., 2023). So too, in Canada, various incidents, such as the bombing of a bank in Ottawa on 18 May 2010 by Anarchists, are included in the GTD data but not in official reports of terrorism. With several other similar examples constituting a large proportion of the data we believe that, at least in the cases of Canada and Sweden and the post-2001 period, the data more accurately refers to VE than the more limiting categorization of terrorism.

As described below in the analytic strategy, the dependent variable in this study is the occurrence of an incident, its date, and the number of days between events. There is a total of 174 events in the data, with N = 78 for Canada over a period of t = 6467 days, and N = 96 for Sweden over a period of t = 5285 days, with the number of days being the period between the first and last events. Overall, the average time between events is M = 68.33 (SD = 131.17), whereas for Canada it was 83.99 (142.40) and 55.63 (120.59) for Sweden, with the differences being marginally significant (t(170) = 1.414, p = .08).

Data for the main independent variables are also derived from the GTD, albeit from two variables that have rarely been used in research, but which are of specific relevance to the current inquiry. The first variable measures the number of individuals arrested for a specific incident and the second the number of individuals killed for a specific incident (offenders killed). However, this data is incomplete. For example, with respect to the aforementioned bank bombing in Ottawa, three suspects were arrested and charged a few weeks after the attack (Boesveld, 2010). As such, we carried out manual searches of open-source media, using the dates and descriptions of each incident, in order to try to identify evidence of additional arrests that were not captured by the GTD. For Canada, we also cross-referenced incidents with the newly created Canadian Incident Database (CIDB).¹ In total, we found nine instances, with five in Canada and four in Sweden, where information concerning arrests not coded in the GTD were readily identifiable. In two of these cases, the GTD event descriptions themselves included information concerning arrests despite the arrest variable having been coded as zero. The second variable is the number of offenders killed during the commission of the offense or because of authorities’ response to it.

Given that there is a debate as to whether raw count variables are suitable for use as independent variables in various types of regression models (see Wolfowicz et al., 2023), we carried out different models in which we treated these variables as both raw counts, as well as inverse hyperbolic sine (IHS) transformed variables. As discussed by Wolfowicz et al. (2023), this is a widely used approach in various literatures, including in VE research, to normalize count data in the presence of many zeroes, which prohibits using a standard log transformation. In total, there was an average of .42 (.81) arrests and .03 (.18) killings per event, meaning that the majority of events were not cleared by arrest or by incapacitation of the offender(s). However, for Canada, there was an average of .58 (.86) arrests per event, and for Sweden an average of .29 (.74), with the differences between the countries being statistically significant (t(172) = 2.532, p = .01).

During the data preparation process, we also spent extensive time inspecting and cleaning the GTD data for duplicates and other errors. A small number of relevant errors were found and corrected for, such as a duplicate event for a vehicular attack that appeared in Sweden as occurring on both the 10th and 11th of June 2017.

Analytic strategy

Following previous VE research, we employed a series hazard model approach to investigate the effects of arrests and deaths on VE. As is widely accepted, any government policy is unlikely to reduce VE to zero, and at some point, an incident will occur. The series hazard approach takes this into consideration and models the dependent variable as the occurrence of an incident whilst accounting for the amount of time between incidents. As a result, this approach takes advantage of a greater amount of information, and thereby explains more of the variation, than summaries of annual events or other levels of aggregation of the data (Dugan, 2011; Dugan & Yang, 2011).

Following similar procedures implemented in previous studies on terrorism and VE, the analysis controls the incident success density, as well as the monthly count time trend. Additionally, monthly and yearly dummy variables are entered into the model as fixed effects (Carson, 2014). This approach differs slightly from those who include various control variables, such as GDP or (un)employment rates (Hodwitz & King, 2023). Given that such factors remain stable over the course of a year, they can give rise to issues of multicollinearity, and hence the use of a fixed effects approach can account for the impact of both these and other unobserved time-context specific confounders. We also coded for the three different generations of GTD data coding that occurred during the observation period, which partially serves to control for differences in data comprehensiveness over the different periods (Hodwitz & King, 2023).

In some cases, the number of days between events is the same, even though the time period pertains to a different pair of events. This leads to a case of ‘ties’ in the dependent variable, which is accounted for by implementing the model using a partial maximum likelihood method. Another issue is that on some occasions there is more than one event on a single day, and often they even may be unrelated to each other. As a result, the two (or more events) are automatically scored as having ‘zero’ time elapsed between them, whereas in reality the time span between them is some number larger than this. The time of day on which an incident occurs can further complicate the issue. For example, an event occurring at 12:01 am will have a ‘1’ day difference from an event have occurred only two minutes earlier, whereas a ‘zero’ would be assigned to a subsequent event occurring at 11:59pm on the same day. In order to address this issue, all zeroes were recoded to be the number of days after the day of the event until the next event for all events that occur on the same day (see Carson, 2014).

The above described model can be expressed as in Equation 1, as being the estimation of the number of days until the next event (Y) as being a function of ${\lambda }_{0,}$an unspecified hazard baseline hazard function, the main independent factors (e.g. arrests and killings), as well as both contextual (e.g. success density) and time trend (e.g. fixed effects) variables measured at the time of each recorded event (Dugan et al., 2005). $h(Y)={\lambda }_0\exp ({\beta }_1\mathrm{Arrests}+{\beta }_2\mathrm{Killings}+{\beta }_3\mathrm{Context}+{\beta }_4\mathrm{Trends}$It is important to note that our macro-level approach, which relies on aggregated data for both arrests (and killings) and VE rates, is grounded in two fundamental mechanisms of deterrence. The first one is the indirect deterrence effect, which refers to the assumption that an increased likelihood of getting arrested or killed deters potential offenders from offending in the first place. The second one is the direct deterrence, or ‘incapacitation’, effect, which refers to the principle that when physically detained, or killed, offenders are unable to further offend (Nagin, 2013; Piquero & Blumstein, 2007). Given that at least some arrests will lead to conviction and sentencing, our model cannot disentangle the deterrence effect from the incapacitation effect, and our results reflect both mechanisms.

Results

The results of our models are presented in Table 1 and can be viewed as two sets of models. In models I-III we first examine the effects of arrests, then killings, and then both factors respectively, whilst controlling for all context and trend related factors. Models IV-VI repeat this order using the IHS transformed variables.

Table 1. Results of Series Hazard Models for the effects of arrests and killings on violent extremism in Canada and Sweden.

	I		II		III		IV		V		VI
	Canada	Sweden	Canada	Sweden	Canada	Sweden	Canada	Sweden	Canada	Sweden	Canada	Sweden
Arrests	1.262 (.279)	.640 (.162)*			1.248 (.267)	.634 (.164)*
Killings			1.814 (1.391)	1.364 (1.694)	1.713 (1.314)	.890 (1.219)
Arrests							1.31 (.403)	.467 (.182)*			1.308 (.400)	.462 (.183)*
Killings									1.965 (1.710)	1.422 (2.004)	1.935 (1.678)	.800 (1.270)
LR	54.33**	103.60***	53.74**	100.22***	54.80**	103.61***	53.94**	104.23***	53.74**	100.22***	54.50**	104.25***

Notes: Coefficients are Hazard Ratios with standard errors in brackets. * < .05. Context and time-trend controls are included in all models but not presented in the table.

Across all models, the only statistically significant findings pertain to the effects of arrests in Sweden. Here, in models I and III, the coefficient indicates that for each additional arrest per event, there is a 36% reduced hazard. Another way of appreciating this is to say that an additional arrest is associated with a hazard of only (approximately) 0.6% per day of experiencing an event, of a 99.4% chance of there not being an event. With each successive day the hazard grows (e.g. .994 at day two and .994 at day three and so forth). In models IV and VI, the coefficient is somewhat larger, representing an approximate hazard reduction of 54% for each additional arrest and a daily hazard of just above 0.46%.

While not statistically significant under any specifications, the direction of the effect for arrests in Canada is always positive, pointing toward a potential iatrogenic effect. Similarly, in the case of killings, the direction of the coefficient is always positive for Canada, and generally the same for Sweden. In an ad-hoc analysis (not shown here), we combined the counts for arrests and killings into a single variable. Using both the raw count of this variable as well as IHS transformation, the results followed those in Table 1. Here, only in the case of Sweden was the variable statistically significant, and of virtually identical magnitudes to those in the main analysis.²

Discussion

The current study sought to examine the idea that arrests (and killings) can deter VE, reducing the frequency and timing of events. While there are some terrorism studies that have found that arrests are associated with deterrent effects, they are few in number, with most studies examining policies, rather than counts of arrests, as proxies for increased risk of apprehension. As such, the studies that do exist have only covered a narrow set of contexts, generally focused on only one specific country, or more recently, a larger panel of countries (Wolfowicz et al., 2023). By comparing two contexts that have received considerably less attention, we sought to test not only whether any effects exist but also whether context matters. In doing so, we indeed found that there are some significant differences, with arrests reducing the hazard of VE in Sweden but not in Canada. And this despite there being a higher rate of arrests in the latter than in the former.

Both theoretical and empirical evidence can explain these differences. First, deterrence theory holds that the risk of apprehension acts as a deterrent in conjunction with the risk and severity of punishment. Our study was unable to examine such factors. It is also possible that Canada has a higher rate of release without charge for arrestees. In some countries, such as the UK, rates of non-prosecution for suspects in VE cases can be quite high. Wolfowicz et al. (2023) suggest that such arrests can offset deterrence and even lead to iatrogenic effects. In general, over policing, especially when directed at suspect communities in the context of VE, can contribute to stigmatization, which in turn can serve as a risk factor for the very extremism that authorities seek to reduce and prevent. As such, it is possible that the differential effects are an outcome of Sweden engaging in more effective arrests, even if fewer in number.

In saying this, we do not suggest that our own null findings for a deterrent effect of arrests in Canada should be taken to completely reject the hypothesis. Rather, they may simply point to issues of levels of measurement and comparison. As noted above, omitted variable bias could play a role, with such issues being prevalent in deterrence research more generally and especially in the case of VE (Wolfowicz et al., 2023).

Exploring further the differences between the Swedish and Canadian criminal justice systems may provide insight into the contextual nuances affecting the link between arrests and VE trends. Criminal justice system interventions such as arrests, charging, convictions, and sentence length are simultaneously implemented. They collectively constitute the broader deterrence efforts and therefore should be analyzed together. As such, developing a more comprehensive model to examine these variables together within a comparative framework may shed light on observed discrepancies while concurrently mitigating omitted variable bias. This represents an avenue for future evidence-based research within the specific context of Canada and Sweden, and beyond (Landes, 1978; Wolfowicz et al., 2023).

It is also possible that comparative criminology is ill suited to this type of analysis, and that it is necessary to increase the number of panels (e.g. countries, states, cities, etc.) to obtain a degree of variance that can serve for a more meaningful analysis. In saying this, researchers will need to carefully consider whether these panels are indeed comparable, and that modeling strategies are in fact appropriately geared towards capturing the dynamic of deterrence; both time series and regression approaches carry their own set of methodological issues (Charles & Durlauf, 2013).

While our study provides evidence that arrests can be effective in reducing VE, at least sometimes, they should not be relied upon as a single strategy. While classical deterrence strategies may be useful for reducing the risk of an incident, they ‘do little to reduce the impact that future attacks have’, as opposed to interventions that target ‘finances, communication, or access to material support’, which ‘may not diminish their attack frequency, but will reduce the impact of those attacks’ (Porter et al., 2011, p. 93). As such, a whole-system approach is needed to combat VE. While most western countries, such as Canada and Sweden are taking such an approach, increasing clearance rates, whilst simultaneously avoiding unnecessary deaths of offenders, appears to be an important part of the puzzle.

Conclusions

To date, most of the deterrent research in VE has modeled deterrence by using dummy variables to represent specific types of policies, often conceptualized as being ‘soft’ or ‘hard’ approaches. The evidence from these studies is mixed and limited. Given that the literature points to significant degrees of overlap between crime and VE at various levels of the phenomenon, especially when it comes to rational choice and decision making, it seems intuitive to believe that deterrent effects may be similar for both outcomes. Our results add to the growing body of evidence that points to such similarities, finding that increasing the risk of arrest can reduce offending; but this is not always the case, and we need to consider why.

Certainly, more evidence-based research is needed in this area, especially given the central role of the criminal justice system in handling cases of VE. More integrated analysis of how VE responds to the criminal justice system, the effects of policing interventions, and CVE intervention evaluation studies, constitute some of the most important, yet also under-researched issues in the field (Freilich et al., 2024; Sydes et al., 2023).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This research was funded by the Canadian Safety and Security Program of Defence Research and Development Canada, Contract No. W7714-217852/001/SV, the Swedish Civil Contingencies Agency (Myndigheten för samhällsskydd och beredskap), grant 2019–13780, and the Swedish Research Council (Vetenskapsrådet), grant 2023-00611.

Notes on contributors

Michael Wolfowicz

Michael Wolfowicz, PhD., is Senior Lecturer at the Institute of Criminology, Faculty of Law, Hebrew University of Jerusalem. His research focuses primarily on the application of traditional criminological frameworks, both theoretical and methodological, to the study of radicalization, extremism, and terrorism. His work in this area is reflected in published studies on terrorism and place, recidivism, policing, deterrence, risk and protective factors, as well as media effects.

Esther Salama

Esther Salama is a PhD candidate at the Institute of Criminology at the Hebrew University. Her work centers primarily on the application of deterrence-based theories to the study of terrorism. Her thesis focuses on examining the effectiveness of counter-terrorism and CVE interventions, using advanced quantitative methods.

Notes

1 www.cidb.ca

2 Raw variable, HR = .633, SE = .164, IHS transformed, HR = .465, SE = .184.