Abstract
Research on offending has long noted the prevalence of co-offending, and researchers have argued that an important component of the decision to co-offend is the risk of arrest. Following this, the current research examines the group hazard, or the risk of arrest associated with co-offending, using National Incident-Based Reporting System (NIBRS) data. Results indicate that there is a hazard for robbery and homicide, whereas there is a negative relationship between co-offending and arrest for assault. Furthermore, arrest risk also varies significantly according to the group size and the demographic composition of groups. Taken together, the results suggest that the relationship between co-offending and arrest is complex, and heterogeneity in the relationship may be an important explanation for the contradictory results found in previous research.
Research on offending has long noted the prevalence of co-offending (e.g., Shaw & McKay, 1931), and a number of researchers have investigated the different motivations for offending with accomplices. Most prominently, Weerman (2003) argued that co-offending is a form of social exchange, wherein co-offending occurs when it is profitable enough or when the rewards exceed the costs. One important aspect of this exchange, and a primary cost associated with co-offending, is the risk of arrest. In this context, nearly all studies of co-offending, especially those that use official data sources based on arrest, acknowledge the potential “group hazard hypothesis,” which is the possibility that co-offenders may be more likely than solo offenders to be apprehended by law enforcement (e.g., Carrington, 2009; van Mastrigt & Farrington, 2009). Yet, despite a clear need for more insight into the circumstances that increase or decrease the risks of arrest when co-offending, there has been little research on the nature of this group hazard. Moreover, the research that has been conducted has yielded mixed findings. On one hand, some research has supported the group hazard (e.g., Erickson, 1973). Other research, however, has found results that do not support the group hazard (e.g., Feyerherm, 1980). As a result of these mixed findings, and despite the fact that the group hazard hypothesis was originally proposed several decades ago, Weerman (2014, p. 5181) recently argued that “whether co-offending exceedingly results in apprehension or not is still an open empirical question.” The current research argues that the primary reason for prior mixed findings is that there is significant heterogeneity in the risk of arrest for co-offenders according to offender and offense characteristics. Put simply, co-offenders are not monolithic in their offending behavior; consequently, neither is their risk of arrest. Following this, this research uses data from the National Incident-Based Reporting System (NIBRS) to examine differences in arrest risk for violent co-offending groups, focusing on differences (a) between solo offenders and co-offenders and (b) between group differences in arrest by co-offender characteristics and offense type.
Literature Review
Several scholars have argued that offending in groups is risky behavior, increasing the likelihood of betrayal and apprehension (e.g., McCarthy, Hagan, & Cohen, 1998). This risk is part of what Erickson (1973) called the “group hazard hypothesis” or the hypothesis that “violating the law in groups increases the likelihood of official detection and reaction (e.g., apprehension, arrest, court appearances, and so on)” (p. 128). Essentially, the group hazard can be separated into two processes. First, the hypothesis is one of opportunity: The involvement of more offenders increases the odds that at least one offender is caught. Second, once one offender is apprehended, the risk of betrayal increases the risk of everyone else being caught.
Early research on the group hazard was primarily based on comparisons of official records and self-report surveys, and generally supported the hypothesis. Erickson (1971), for example, reviewed 11 studies using official records and found that, on average, about 85% of offenses were committed by groups. In contrast, he found that only about 65% of self-report cases involved group offending. The difference between these estimates was interpreted as support for the group hazard (see also Hindelang, 1971). Hindelang (1976) also found support for the group hazard in a survey of sheriffs’ deputies, wherein respondents indicated that they were more likely to arrest a person who was verbally abusive than to arrest someone who was not. They also indicated that they believed juveniles in groups were more likely to be verbally abusive and that young offenders primarily offended in groups. Consequently, they were more suspicious of young groups than of other people.
Feyerherm (1980) argued, however, that there were several problems with these studies. First, when comparing official data and self-report data, researchers were comparing substantially different behavior. That is, official data tended to measure X, whereas self-report data tended to measure Y. Feyerherm also took issue with Hindelang’s (1976) analysis and argued that the survey had a low response rate (about 49%), that it was limited to a single jurisdiction, and that the measurement of officer behavior was not very precise. After Feyerherm’s review and analysis of the group hazard hypothesis, research on the subject decreased significantly. Morash (1984) conducted a survey examining the association between peer group characteristics and arrest and found that committing a high proportion of offenses with a group of peers increased the chances of arrest. Brownfield, Sorenson, and Thompson (2001), however, found no support for the group hazard hypothesis using a measure of gang membership.
At the same time, the rational choice or instrumental perspective on co-offending has argued that offenders may choose to co-offend because they anticipate that offending with others will be more profitable and less risky than committing the crime alone (e.g., Lantz & Ruback, 2017a; Weerman, 2003). That is, for some offenders, co-offending may actually be the result of a rational decision that includes the desire to minimize the risk of being caught for the offense (e.g., McCarthy et al., 1998; McGloin & Nguyen, 2012). Only a few studies have examined co-offending networks and cost avoidance (Bouchard & Nguyen, 2010; Bouchard & Ouellet, 2011; Kazemian & LeBlanc, 2007; Malm et al., 2017; McCarthy & Hagan, 2001; Tremblay & Morselli, 2000), but several of these studies have found that offenders may insulate themselves from detection within criminal networks (Baker & Faulkner, 1993; Krebs, 2002; Williams, 2001). Bouchard and Ouellet (2011), for example, found that larger co-offending networks decreased the risks associated with participation in the drug trade; as Bouchard and Nguyen (2010) pointed out, findings like these directly contradict the “conventional wisdom of keeping it small.”
Taken together, research to this point has been largely inconclusive. Yet, it is important to better understand the nature of the relationship between co-offending and arrest for at least three reasons. First, given that prior research has suggested that co-offending is the result of social exchange processes wherein offenders choose to co-offend when the benefits of accomplices outweigh the risks of accomplices and that one of the most significant risks is that of arrest, it is crucial that we better understand the conditions that increase or decrease that risk. Second, a significant proportion of offending involves co-offending and, as a result, almost all empirical research on the phenomenon includes some sort of statement acknowledging the potential group hazard and the possibility that co-offenders may be more likely than other offenders to be counted using official measures. This acknowledgment is important, but it is not enough. Instead, it is more important that we begin to understand exactly how the group hazard might impact research on offending generally, and co-offending specifically. Finally, and relatedly, prior research has largely considered the group hazard as representing a stable risk for all co-offenders; it is likely, however, that this hazard is quite variable, and to fully understand the risk of arrest and co-offending, it is important to understand which co-offender and offense characteristics are more or less associated with this risk.
Heterogeneity in Co-Offender Group Arrest Risk
The current research suggests that one explanation for the mixed findings of prior research on the group hazard is that there is significant heterogeneity in the relationship. There is, more specifically, significant variation in the likelihood of arrest by offense type (Federal Bureau of Investigation [FBI], 2016) and by offender characteristics (e.g., Ouellet, Boivin, Leclerc, & Morselli, 2013). Following this, it is reasonable to expect that there may also be significant variation in the relationship between co-offending and arrest based on (a) offense type and (b) offender group characteristics.
Heterogeneity by offense type
Regarding the former, some types of crimes are riskier than others, and some types of crime more often involve co-offenders (Blumstein, Cohen, Piquero, & Visher, 2010; Carrington, 2009). Violent crimes, in general, are associated with an increased likelihood of arrest. More importantly, studies of groups and arrest have found different results using different offense-specific samples. Tillyer and Tillyer (2015), for example, examined robbery and found that offenses committed by groups were more likely than offenses committed by solo offenders to result in arrest. Cunningham and Vandiver (2016) focused on stranger kidnappings and found that incidents committed by groups had fewer arrests than incidents committed by solo offenders. Even Erickson’s (1973) research on the group hazard suggested variation by offense type, although not explicitly. In his analysis, he compared the frequency of group violations and solo violations for 14 different offenses, as well as for the sample as a whole (see also Erickson & Jensen, 1977). The group hazard was supported in eight of the 15 comparisons; the other seven offenses were either not significantly different or found results contrary to what the group hazard would suggest. Although he concluded that these results provided support for the group hazard, the results are perhaps better described as indicative of heterogeneity in the relationship. Feyerherm (1980) also found variation in the risk of arrest by offense type. On one hand, he found that drug offenses, and those offenses he called “force offenses,” which were primarily fighting behaviors, did not have a group hazard (see also Bouchard & Ouellet, 2011). On the other hand, he found consistent support for the group hazard only for theft offenses. He concluded that “the group hazard hypothesis appears to offer greater explanatory value for certain types of offenses than others” (p. 67).
Another reason to expect heterogeneity in arrest risk by offense is due to the group hazard itself. One of the most significant risks associated with co-offending is accomplice betrayal (Nguyen & McGloin, 2013). A co-offender can “divulge his or her identity to authorities or other potentially harmful people (e.g., victims); in other words, an offender who ‘defects’ may . . . increase the latter’s likelihood of detection, injury, or arrest” (McCarthy et al., 1998, pp. 155-156). Following this, it stands to reason that the risk of betrayal might also vary by offense type. More specifically, certain crimes present higher risks such that there are greater penalties if one is caught; homicide and robbery, for example, are often punished more severely than assault. Co-offending partnerships are often built on trust (Tremblay, 1993), but offenders will no longer be trusted if they give information to the police. Therefore, in some cases, co-offenders will protect their partners to preserve this trust and their own reputation (McCarthy et al., 1998). But, this protection is probably more likely when the costs are lower; when the punishment for an offense is more severe, it follows that the risk of co-offender betrayal may also be higher.
Heterogeneity by co-offender characteristics
Second, there is reason to expect that the relationship between co-offending and arrest may vary according to the gender, race, and age composition of groups. Returning to Tillyer and Tillyer’s (2015) analysis of co-offending and arrest, it should be noted that the authors decided to exclude offender demographics from most of their analysis. Moreover, when they did include these demographics in supplementary analysis, some key findings were no longer statistically significant. Most notably, the effect of two offenders on arrest was reduced to nonsignificance. In essence, the inclusion of offender demographic control variables washed out the significant relationship between co-offending in small groups and arrest. This pattern of results suggests that the relationship between co-offending and arrest likely varies significantly according to group demographic composition.
There are also clear gender, race, and age differences in co-offending and in arrest. Prior research on co-offending has demonstrated, for example, that females are more likely than males to co-offend (e.g., van Mastrigt & Farrington, 2009), that Black offenders are more likely than White offenders to co-offend (e.g., Felson, 2003), and that age is negatively associated with co-offending (Lantz & Ruback, 2017a). Past research has also demonstrated significant differences in arrest according to gender (e.g., Visher, 1983), race (e.g., Kochel, Wilson, & Mastrofski, 2011), and age (e.g., Steffensmeier, Allan, Harer, & Streifel, 1989). Taken together, it follows that we should expect significant variation in the relationship between co-offending and arrest by these characteristics as well.
Unfortunately, the literature on co-offender characteristics and arrest risk is very limited. Regarding gender, Ouellet et al. (2013) examined co-offending participations and found that men were more likely than women to be re-arrested (see also, Lantz & Wenger, 2019). Similarly, Fergusson, Swain-Campbell, and Horwood (2002) found that males were significantly more likely than females to be arrested. Finally, Morash (1984) found that male offenders in predominantly male peer groups were particularly likely to be arrested. Taken together, these studies suggest that co-offending groups involving more male offenders may be subject to a greater hazard than other groups.
Turning to race, scholars have long argued that Black offenders are more likely than White offenders to be arrested, controlling for criminal behavior (Kochel et al., 2011; Lantz & Wenger, 2019). There are a number of explanations for this relationship, ranging from disproportionate patrol practices (e.g., Mann, 1993) to suspect demeanor (Worden & Shepard, 1994), to racial bias (Anderson, 1990). Regardless of the mechanism, however, these studies suggest that co-offending groups involving more Black and non-White offenders may be subject to a more substantial group hazard than other groups. Brownfield et al. (2001) conducted the only analysis of the group hazard that has explicitly tested for differences by offender characteristics. They focused on offender race and found that Black offenders’ odds of arrest were 1.75 times higher than White offenders. They then used an interaction term to test whether the effect of the group hazard on arrest differed for Black and White offenders. The interaction term was not significant, suggesting there was no difference in the group hazard in their sample. 1
Finally, research suggests that the relationship between co-offending and arrest may vary significantly according to the age of co-offenders. The group hazard, as conceptualized by Erickson (1971), was intended to explain an increased occurrence of juvenile group violations in official data. Older offenders, he argued, more frequently commit crimes alone and work harder to reduce their risk of arrest by decreasing visibility and involving fewer co-offenders when they do co-offend. If this is indeed the case, then groups involving younger offenders should be subject to a greater group hazard than older co-offending groups. Yet, every explicit test of the group hazard has used juvenile samples or focused exclusively on youth. There is, of course, a significant problem inherent to this approach: One cannot test for significant differences between young offenders and older offenders without a meaningful comparison group. Regardless, over time, the group hazard has been used to refer to a potential bias in co-offending in general. Even if there is no clear reason that we might expect the group hazard to not apply to older offenders as well, there is currently no basis for discussing age differences in co-offending and arrest.
Some research—conducted outside of the group hazard theoretical lens—has supported the idea that age is related to co-offending and arrest. Ouellet et al. (2013) found that the probability of arrest was negatively associated with age, such that younger offenders were significantly more likely than older offenders to be arrested. Bouchard and Nguyen (2010) found that juvenile offenders who were embedded in larger adult networks were less likely than other offenders to be arrested. They suggested that juveniles in adult networks may be better protected from detection due to the presence of more experienced offenders. Similarly, Morselli, Tremblay, and McCarthy (2006) found that having a criminal mentor was negatively related to the number of days that an individual spent incarcerated, a finding that might mean that older, more experienced offenders may be better at avoiding arrest. Taken together, these studies suggest that co-offending groups involving more youthful offenders may be subject to more of a group hazard than other groups.
Current Study
The current study examines the relationship between co-offending and arrest within the context of the group hazard hypothesis. Although past research has yielded mixed support for the group hazard, this study suggests that one explanation for these mixed findings may be variation by offender and offense characteristics; the motivation behind the current research is to better understand the nature of this variation. To examine this variation, this analysis tests for significant differences in the slopes of the relationship between group size and arrest according to the gender, race, and age of co-offenders using a series of interaction terms for five different offenses samples: all violent offenses, assault, robbery, sexual assault, and homicide. When examining these interaction terms, I contend that it is worth paying special attention to the slope of these relationships and considering them in the context of the group hazard. Those groups with larger positive slopes may be considered as having a larger group hazard, those with smaller positive slopes as having a smaller hazard, and those with negative slopes as having no hazard.
Data and Method
The current study uses NIBRS data on physical assault, robbery, sexual assault, and homicide offenses during years 2003-2012. 2 To conduct the analysis, information from the offender, offense, victim, and arrestee segment are analyzed at the incident level. 3 Because a full sample of all incidents across 10 years would yield an unnecessarily large sample size and inflate statistical power, a random sample of 10% of incidents is selected for the final analysis of differences between solo offenders and co-offenders. 4 However, because the purpose of this research is to examine variation in arrest according to co-offender characteristics—some of which are comparatively rare—the full sample of co-offending groups is used for the analysis of differences between groups. The data are also limited to only violent offenses for two primary reasons. First, although useful, an examination of heterogeneity in arrest risk for all offenses in the NIBRS data would be beyond the scope of a single analysis. Because recent research has suggested that co-offending may be especially consequential in the production of violence (e.g., Lantz, 2018; Lantz & Kim, 2018; McGloin & Piquero, 2009), this research focuses on violent offenses as a starting point. Second, because the victim does not necessarily encounter the offender during property crimes, information regarding offender demographics—one of the primary foci of this research—are considerably more reliable for violent crimes.
There are, of course, limitations to the use of NIBRS data. Most obviously, NIBRS is not a nationally representative sample (Addington, 2008). Furthermore, the data are also official records, meaning that they are limited to only those incidents known to law enforcement. As such, those incidents that are not reported to the police are not included. Although there is no empirical research on differences in victim reporting for co-offending and solo offending incidents, these potential differences should be kept in mind when considering the results. That said, despite these limitations, NIBRS is the most detailed national level data on crime clearance, wherein incidents for which an arrest occurred can be separated from incidents for which an arrest did not occur (Roberts, 2009). The data also include detailed information on co-offending and demographic information for each recorded co-offender. Finally, these data also include a substantial number of cases, allowing for the examination of statistically rare, but theoretically significant groups, like robberies committed by all-female groups (Lantz, 2019). Taken together, these advantages situate NIBRS as an extremely useful data source for the examination of the group hazard.
Measures
The primary dependent measure of interest is arrest. Arrest is coded as a dichotomous measure indicating whether any arrest occurred as a result of the incident (1 = yes). Because this outcome is dichotomous, the analysis is conducted using logistic regression models. 5 For the first part of the analysis, the primary independent measure is a dummy measure indicating whether multiple offenders are involved in the incident (1 = yes). For the second part, a measure of the number of co-offenders involved (i.e., group size) in the incident is created. Three different interaction terms are then created between number of co-offenders and (a) the proportion of co-offenders who are male, compared with female; (b) the proportion of co-offenders who are White, compared to non-White; and (c) the mean age of the co-offenders. These interaction terms allow for an examination of heterogeneity in the relationship between co-offending and arrest according to the gender, race, and age composition of the group, respectively.
Several incident-level control measures are also included in the analyses. First, a number of offender demographic measures are created. For the comparison of solo and group offenders, gender is coded into three dichotomous measures: male, mixed gender, and female. 6 For the analysis of group differences, a measure of the proportion of co-offenders who are male, compared with female is coded. Group race is coded into similar measures: four dummy measures indicating that offenders are White, Black, other race, or mixed race and a proportion measure indicating the proportion of offenders who are White, relative to the proportion of offenders who are non-White. 7 The mean age of the co-offenders is computed to account for age.
At the offense level, a dichotomous measure of whether a weapon was used during the incident is included (1 = yes). 8 Victim injury is coded into a dichotomous measure, that is, coded as no injury (0) when the victim or victim(s) were not injured in any way and coded as injury (1) if any victim suffered any injury. Alcohol and drug use may also be related to offender behavior and the likelihood of arrest, and dummy measures for both are included. Brownfield et al. (2001) also argued that gang members may be particularly likely to be subject to a group hazard; the analysis controls for potential gang involvement by including a dichotomous measure indicating whether any of the offenses involved suspected gang activity.
Finally, several victim characteristics are included as control measures. First, a continuous measure indicating the number of victims is coded. Second, victim age is coded as the mean age of all victims. Third, several separate proportion measures are created to indicate the proportion of victims who were male (compared with female), Black, or other race (compared to White), Hispanic ethnicity (compared to non-Hispanic), and a resident of the location in which they were victimized. If the incident involved only a single victim, the respective gender, race, ethnicity, and residential status is coded. If multiple victims were involved in the incident, the proportion value for each variable is calculated. Finally, a dichotomous measure is created to capture the relationship between the victim(s) and the offender(s), indicating whether any victim is known to any of the offenders involved in the incident (1 = yes).
Results
The current research proceeds in two parts. First, differences in the likelihood of arrest between solo offenders and co-offenders are examined. Second, the sample is limited to groups, and the analyses test for heterogeneity in the relationship between group size and arrest by offense type and group characteristics. In total, roughly 12.8% of assaults, 37.4% of robberies, 7.9% of sexual assaults, and 23.7% of homicides involve co-offending, and—among co-offending incidents—approximately 38.7% of assaults, 27.7% of robberies, 26.7% of sexual assaults, and 75.5% of homicides are cleared by arrest. Additional sample descriptive statistics are presented in Table 1.
Descriptive Statistics, by Sample.
The arrest rate is consistent across samples, such that roughly 37% of incidents result in an arrest. Table 2 presents results from logistic regression models of arrest on co-offending for five different samples: all offenses, assault, robbery, sexual assault, and homicide. Because of the large sample size, odds ratios are presented as a measure of effect size, rather than coefficients.
Logistic Regression Models of Arrest on Co-Offending by Offense Type, NIBRS 2003-2012 Random Sample.
Note. OR = odds ratio; Sig. = significance; NIBRS = National Incident-Based Reporting System.
p < .05. **p < .01. ***p < .001.
There are three important findings in Model 1. First, in the full sample, there is a significant negative relationship between co-offending and arrest (OR = 0.729, p < .001). Second, several of the offender characteristics are significantly related to arrest. Incidents involving males are significantly more likely than other incidents to result in arrest (OR = 1.107, p < .001). Age is also negatively associated with arrest (OR = 0.994, p < .001), such that an increase in the age of the offender(s) is associated with decreased likelihood of arrest. Third, there are significant differences in the likelihood of arrest by offense type. Compared with assaults, robbery (OR = 0.665, p < .001) and sexual assaults (OR = 0.592, p < .001) are significantly less likely to result in arrest. Homicides, however, are more likely than assaults to result in an arrest (OR = 3.967, p < .001).
Models 2 through 5 split the incidents into offense-specific samples, supporting the proposition that there is heterogeneity in the relationship by offense type. Most offenses in the data are assaults, and for these incidents, co-offending is significantly and negatively associated with arrest (OR = 0.702); assaults committed by groups are about 30% less likely than those committed by solo offenders to result in arrest. There is, however, a significant positive relationship between co-offending and arrest for robbery and homicide (p < .001). For robbery, co-offending is associated with a 9% increase in odds of arrest (OR = 1.093). For homicide, incidents committed by groups are associated with an 81% increase in arrest (OR = 1.811).
Next, the sample is limited to co-offending groups and between-group differences in the likelihood of arrest are examined; results are presented in Table 3.
Logistic Regression Models of Arrest on Co-Offending Group Size by Offense Type, NIBRS 2003-2012 Groups Only Sample.
Note. OR = odds ratio; Sig. = significance; NIBRS = National Incident-Based Reporting System.
p < .05. **p < .01. ***p < .001.
For the sample as a whole (Model 6), group size is positively associated with arrest (p < .001), net of controls. Moreover, co-offending incidents involving a higher proportion of male offenders, relative to females, are significantly associated with an increased likelihood of arrest (p < .001). Age, however, is negatively related to arrest (p < .001). Finally, co-offending incidents involving a higher proportion of White offenders are significantly more likely than other incidents to result in arrest (p < .001).
That said, the relationship between co-offending and arrest varies significantly by offense type. Group size is negatively associated with arrest for assault incidents (OR = 0.955, p < .001), such that each additional co-offender is associated with about a 4.5% decrease in the likelihood of an arrest occurring. Conversely, group size is significantly and positively related to the likelihood of arrest for robbery (OR = 1.297, p < .001) and homicide (OR = 1.290, p < .001), suggesting that, for these offenses, there may be a group hazard. For robbery, each additional co-offender is associated with a 30% increase in the likelihood of arrest, whereas, for homicide, each additional co-offender is associated with about a 29% increase in the likelihood of arrest.
Next, Table 4 presents results from three sets of interaction models examining differences in the slope of the relationship between group size and arrest according to the sex, race, and age composition of groups. Model Set 1 shows results for models including an interaction between group size and proportion male, Model Set 2 shows results for models including an interaction between group size and proportion White, and Model Set 3 shows results for models including an interaction between group size and age. Control variables were included in all analyses but are not shown for brevity. Instead, only the conditional effects of the two variables included in the interactions and the interaction effects themselves are shown.
Logistic Regression Models of Arrest on Co-Offending Group Size by Offense Type With Interaction Terms for Sex, Race, and Age (Condensed Models).
Note. Control variables were included in all analyses but are not shown for brevity. OR = odds ratio; Sig. = significance.
p < .05. **p < .01. ***p < .001.
The results from Model Set 1 indicate that the relationship between group size and arrest varies significantly according to gender composition all offense types except homicide. The predicted probabilities for these relationships are presented graphically in Figure 1.
Predicted probability of arrest, by gender and offense type: (a) predicted probability of arrest by gender, full sample, (b) predicted probability of arrest by gender, assault only, (c) predicted probability of arrest by gender, robbery only, and (d) predicted probability of arrest by gender, sexual assault only.
In the full sample (Figure 1a), the results indicate that male groups are more likely than other groups to be arrested regardless of group size but that this effect is particularly pronounced in larger groups. If these slopes are considered in the context of a “group hazard” as proposed, male groups appear to have a group hazard, mixed gender groups have no hazard, and female groups are actually less likely to be arrested when committing offenses in large groups. That said, the nature of the relationship varies significantly by offense type. For assaults (Figure 1b), the likelihood of arrest decreases as group size increases, regardless of gender. For robbery (Figure 1c), however, female group incidents are most likely to result in arrest, especially in small groups; male group incidents are least likely to result in arrest. Moreover, there is a group hazard for robbery regardless of gender: as group size increases, so does the likelihood of arrest. That said, this group hazard is largest for males; female group incidents, although they are most likely to result in an arrest overall, have the smallest hazard. Once groups reach a size of six, there are no longer significant differences in the likelihood of arrest by gender. The results for sexual assault (Figure 1d) indicate a bifurcation in the likelihood of arrest by gender and group size. In dyads, sexual assaults committed by males are more likely than those committed by other dyads to result in arrest. Female groups, however, have a significant group hazard: Large female group incidents are more likely than incidents committed by other groups to result in an arrest. 9
Next, results from Model Set 2 in Table 4 indicate that the relationship between group size and arrest varies significantly according to the racial composition of co-offending groups for the full sample (p < .001), assault (p < .001), robbery (p < .001), and homicide (p < .01). These interactions are presented graphically in Figure 2.
Predicted probability of arrest, by race and offense type: (a) predicted probability of arrest by race, full sample, (b) predicted probability of arrest by race, assault only, (c) predicted probability of arrest by race, robbery only, (d) Predicted probability of arrest by race, homicide only.
Overall (Figure 2a), incidents involving White groups are significantly more likely than other incidents to result in arrest, regardless of group size. That said, minority groups have the greatest group hazard, such that the increase in likelihood of arrest by group size is largest for minority offenders. These results are also disaggregated by offense type. Again, for assault, co-offending in larger groups reduces arrest risk. There is, however, a significant group hazard for groups that commit robbery: As group size increases, the likelihood of arrest significantly increases regardless of race. White groups are more likely than other groups to be arrested, but the group hazard for minority groups is significantly larger than the group hazard for White groups, net of controls. For homicide, there are no significant differences in arrest by race for small groups (i.e., Size 2 and 3). But, as the size of the group increases, significant racial differences in the likelihood of arrest emerge. Minority groups, in particular, have the most significant group hazard, such that homicide incidents committed by large (i.e., Size 4 or greater), minority groups are significantly more likely than other incidents to result in arrest.
Finally, results from Model Set 3 in Table 4 indicate that the relationship between group size and arrest varies significantly according to the age of the group for the full sample (p < .001), assault (p < .001), robbery (p < .001), sexual assault (p < .001), and homicide (p < .01). These interaction terms are presented graphically in Figure 3.
Predicted probability of arrest, by age and offense type: (a) predicted probability of arrest by age, full sample, (b) predicted probability of arrest by age, assault only, (c) predicted probability of arrest by age, robbery only, (d) predicted probability of arrest by age, sexual assault only, and (e) predicted probability of arrest by age, homicide only.
The figure presents the relationship for three different ages: 26 years old, which is the average age of groups in the sample; 16 years old (roughly one standard deviation below the mean); and 36 years old (roughly one standard deviation above the mean). In the overall sample, younger groups are more likely than older groups to be arrested, regardless of group size. But, there is also a group hazard for those who co-offend in large groups of young offenders: For younger groups, the likelihood of arrest increases as group size increases. For older offenders, however, the relationship is the opposite: For older groups, the predicted probability of arrest occurring decreases significantly as group size increases. On the average, committing violent offenses with young groups increases risk of arrest, whereas committing offenses with older, and potentially more experienced offenders, appears to reduce the risk of arrest.
Again, co-offending in larger groups appears to reduce arrest risk for assaults (Figure 3b). Moreover, the relationship is stronger in older groups, suggesting that assault co-offending may be more protective for older offenders than for younger offenders. For robbery (Figure 3c), however, there is a group hazard for all age groups. Committing robbery in young groups, however, is especially risky: In these groups, as group size increases, the predicted probability of arrest increases from about .28 in co-offending partnerships to about .57 in groups of Size 6, an arrest risk that more than doubles in size.
The relationship between co-offending and arrest for sexual assault also varies significantly by age (Figure 3d). Co-offending in larger, older groups reduces the risk of arrest. There is, however, an increasing risk in younger groups as group size increases. More specifically, incidents committed by older dyads are more likely than incidents committed by other groups to result in arrest. There are no significant differences in the predicted probability of arrest by age for small groups of Sizes 3 and 4. But, for incidents committed by larger groups, those sexual assaults committed by younger groups are more likely than other incidents to result in arrest. Finally, there are significant differences in arrest for homicide incidents by age (Figure 3e). Overall, homicides committed by younger groups are significantly more likely than incidents committed by older groups to result in arrest. There is also a group hazard effect for all age groups. Again, however, this hazard is strongest in younger groups: The predicted probability of arrest increases from about .78 in dyads to .94 in larger groups of six, net of controls.
Sensitivity analyses
The focus of the present research is the examination of heterogeneity in the relationship between co-offending and arrest. To that end, the current study has examined heterogeneity by co-offender characteristics and by offense type. That said, it is possible that there is further heterogeneity within these offense types. Although a full explication of this potential heterogeneity is beyond the scope of a single analysis, sensitivity analyses were conducted including additional offense type variables for each offense type. Although these models are not presented to keep models parsimonious and comparable, the overall pattern of findings remained consistent, without exception. More specifically, additional measures were included in all assault models indicating whether assault involved intimidation, simple assault, or aggravated assault: co-offending (OR = 0.682, p < .001) and group size (OR = 0.959, p < .001) remained substantively similar, whereas the sex (OR = 1.025, p < .001), race (OR = 1.034, p < .001), and age (OR = 0.998, p < .001) interaction terms remained consistent as well. 10 Second, additional measures were included in all sexual assault models indicating whether the incident involved forcible rape, forcible sodomy, forcible sex with an object, or forcible fondling. Co-offending and group size remained nonsignificant, whereas the sex (OR = 0.760, p < .001), race (OR = 1.032, n.s.), and age (OR = 0.994, p < .001) interaction terms remained substantively similar. Finally, measures were included in all homicide models indicating whether the incident involved murder or manslaughter. Co-offending (OR = 1.786, p < .001) and group size (OR = 1.288, p < .001) remained positive and significant; the sex (OR = 1.127, n.s.), race (OR = 0.810, p < .01), and age (OR = 0.988, p < .01) interactions remain consistent as well. All robbery incidents are treated the same within the NIBRS data and thus could not be divided further.
Discussion
A significant amount of co-offending research has acknowledged the potential for increased risk of arrest, referred to as the group hazard, when offending with accomplices. Yet, the nature of this relationship is still an open empirical question (Weerman, 2014). The current study is an important step toward clarifying and resolving this question, conducting a thorough examination of the relationship between co-offending and arrest with a focus on between-group heterogeneity in the relationship. This research not only found some support for the group hazard but also found that there are co-offenses for which the hazard is less applicable. Following this, these results also offer an explanation for the inconsistent results in prior research: heterogeneity by offense type and group composition.
Erickson’s (1973) group hazard hypothesis posited that offending in groups increases the likelihood of arrest. The current research did not support this hypothesis in an overall comparison of solo and group offenders. Instead, the results indicated that, overall, co-offending was significantly and negatively related to the likelihood of arrest. There was, however, some support for the group hazard when the sample was disaggregated by offense type, supporting the proposition that there is substantial heterogeneity in the relationship between co-offending and arrest. More specifically, four main findings support this conclusion. First, there are clear differences in the likelihood of group arrest by offense type. There is, for example, evidence for a robust group hazard among robbery groups and homicide groups. The current research also found, however, that co-offending also reduces the likelihood of arrest in some circumstances. In particular, co-offenders who commit assault are significantly less likely than solo offenders who commit an assault to be arrested, regardless of the gender, race, or age composition of groups.
There are at least three possible explanations for these contrary relationships. First, if co-offending is a form of social exchange, as Weerman (2003) posits, then it may be that offenders are better at evaluating the cost of co-offending, at least in terms of cost avoidance, for assault offenses than they are at evaluating the utility of co-offenders for robbery and homicide offenses. In other words, offenders may inaccurately judge the value of co-offenders when committing these offenses. Research on bounded rationality has, in fact, emphasized that this decision-making process is flawed. Offender decision making, especially for robbery and homicide, may be imperfect or flawed by lack of information or poor information processing (e.g., Nagin & Paternoster, 1993). Cost avoidance, however, is only one component of this equation; it is also possible that offenders choose to co-offend for robbery and homicide offenses in spite of the risk of apprehension. That is, co-offending may offer benefits for these offenses that outweigh the risk of arrest. Some researchers have, for example, suggested that co-offending may offer social support (e.g., Lantz & Ruback, 2017a); increased peer approval and respect (e.g., Warr, 2002); future information about potential targets (e.g., Felson, 2003; Lantz & Ruback, 2017b); and increased enjoyment and excitement (e.g., Katz, 1988). Second, in the NIBRS data, assaults are much more likely than robberies to involve victims and offenders who know each other. Roughly, 78% of assaults involve victim(s) that know the offender in some way, whereas only about 15% of robbery offenses involve victim(s) who know the offender. Because the offender(s) are more likely to know each other in assault cases, compared with robbery, retaliation may be more likely and victim(s) may be less likely to cooperate with police (Felson & Lantz, 2016). As the size of the group increases, victim fear of retaliation may also increase.
Finally, these findings could be related to the potential severity of punishment for these offenses. The group hazard hypothesis posits that co-offenders should be more likely to be apprehended because once an offender is caught, he or she can be coerced to betray co-offenders. Following this, it is reasonable to assume that offenders are probably more likely to betray their offenders when the potential punishment is more severe. They may, for example, be offered incentives for their betrayal, like a less severe punishment. The punishment for homicide offenses is most severe, and the results indicate that there is a robust group hazard for these offenses. The punishment for robbery offenses is also more severe than assault offenses: According to the United States Sentencing Commission (USSC; 2016) Guidelines, the recommended term of imprisonment for an individual with no prior criminal history would be 0 to 6 months for simple assault, 15 to 21 months for aggravated assault, and 33 to 41 months for robbery. Robberies are also more likely than assault offenses to involve weapon use (59% compared with 15%), a common criterion for sentence enhancement. If co-offending resembles a prisoner’s dilemma (McCarthy et al., 1998), the severity of potential punishment may be an important predictor of co-offender behavior. It is easier to protect fellow co-offenders when the punishment is that which may accompany an assault offense, but the cost of such protection increases when the offense involves the severe punishment that is likely to accompany robbery or homicide offenses.
Second, the relationship between co-offending and arrest varies significantly by group gender. Overall, there is a general group hazard for male groups. But, this relationship also varies by offense type. For robbery offenses, the group hazard is largest for male groups. It is worth noting, however, that robbery is actually the only offense for which there is a gendered group hazard; although there is a group hazard for homicide offenses, the relationship does not vary significantly by gender. Third, there is significant heterogeneity in the relationship between groups and arrest by race. Although, overall, incidents committed by White groups are more likely than other incidents to result in an arrest, the current results suggest strong evidence for a substantial group hazard for minority co-offenders. That is, the likelihood of an arrest increases as the size of minority groups increases, whereas it does not increase in the same way as the size of White groups increases. Again, the nature of this relationship varies by offense type. There is a group hazard for all groups that commit robbery or homicide, but the hazard is more substantial for minority groups. Once more, assault reduces arrest risk, and there is no group hazard regardless of racial composition.
Finally, there is heterogeneity in the relationship between co-offending and arrest by the age of the group. In the overall sample, there is a significant group hazard for younger co-offenders. There is, however, no such hazard for older groups; instead, co-offending in adult groups is protective. Yet again, these relationships are heterogeneous by offense. For robbery, there is a group hazard for all groups, although the hazard is larger in younger groups; for assaults, however, co-offending is protective for all ages, but most protective for older groups. Taken together, the results indicate that offending in young groups is generally risky. Offending in older groups, however, is considerably less risky: Co-offending is protective for assault and sexual assault, and there is a smaller group hazard for older co-offenders that commit robbery or homicide.
Altogether, this research suggests support for the group hazard for robbery and homicide offenses, and for young or minority co-offenders. Although these findings are an important advance in our understanding of the relationship between co-offending and arrest, it should be noted that the data are limited in some respects. First, the current analysis is limited to a single offense observation, and it is possible that some offenders or groups are committing more crimes not recorded in the data. More criminal activity increases risk of detection, and these offenders may be more likely to be arrested; unfortunately, however, this possibility cannot be measured in the data. Second, the current analysis is limited in its ability to examine the mechanisms between co-offending and arrest. Although the group hazard hypothesis would suggest that co-offender betrayal may be acting as a mechanism, recent research by McGloin and Rowan (2015) has indicated that people are willing to engage in riskier behavior in larger accomplice groups. As a result, it is possible that certain co-offenders act differently while engaging in offending behavior with larger groups, a process that may also increase arrest risk. Future research should investigate these mechanisms in greater detail.
Relatedly, prior research has indicated that, to some extent, individuals may choose to co-offend to reduce arrest risk; these decisions may be impacted by a number of factors, including trust (Tremblay, 1993) and criminal experience (Lantz & Ruback, 2017a; Ouellet & Bouchard, 2017). Although these mechanisms cannot be examined using the data at hand, it is possible that they may play an important role. Those offenders with more criminal experience, for example, may be more effective at evaluating the risks associated with a potential co-offending partner, and thus both (a) less likely to co-offend and (b) less likely to be apprehended when they do co-offend. Following this, future research should consider examining the nature of these relationships in a longitudinal framework. Finally, NIBRS includes an enormous amount of data, but not all agencies submit data. As such, the sample may not be generalizable to the entire population (Addington, 2004). 11 Currently, 29.3% of the population is covered by NIBRS, or about 28% of crime in the United States (McCormack, Pattavina, & Tracy, 2017). As a result, it is possible that the arrest patterns observed in these data are less generalizable to urbanized areas than they are to less urban areas. That said, more generalizable data in which these extensive analyses could be conducted does not currently exist.
Conclusion
Many studies of co-offending have acknowledged the potential risk of arrest associated with group offending, and nearly all studies that rely on official arrest or conviction records recognize that differential arrest patterns between co-offenders and solo offenders could potentially bias the data at hand (e.g., Ouellet et al., 2013). To this point, however, the nature of this potential bias has been somewhat unclear. Based on the current analysis, if we assume that (a) those incidents that do result in an arrest would be those most likely to appear in official record data based on arrests or convictions and (b) those that do not result in an arrest would be those incidents most likely to be missing from these data sources, we can make several conclusions about how co-offending and arrest may impact such records. Because co-offenders are more likely to be arrested for robbery and homicide, official data including robbery and homicide offenses likely overestimate group offending, compared with solo offending. In contrast, official data on physical assaults likely underestimate group offending, relative to solo offending. Because there are also significant differences in the relationship between co-offending and arrest by offender characteristics, we might also conclude that there may be significant differences in these characteristics between official record data and other data sources. Based on these results, official records, relative to other types of data sources are likely to disproportionately include (a) male co-offenders who commit robbery, (b) female co-offenders who commit sexual assault, (c) minority co-offenders in general, except for in the case of assault, and (d) young offenders, in general.
Taken together, the results from this research suggest that there is significant heterogeneity in the risk of arrest associated with co-offending, and the group hazard hypothesis, as currently posited, cannot be broadly applied to the relationship between co-offending and arrest for all offense types and all co-offenders. That said, it may still be possible to strive toward the development of a general theory of co-offending and arrest. It is important, however, that any theoretical approach account for heterogeneity, rather than assuming homogeneity. To accomplish this, such a theory must account for diverse co-offenders and their varying motivations for co-offending, including the role of material and nonmaterial rewards (Weerman, 2003), individual experience (Carrington, 2009), trust (Tremblay, 1993), and larger criminal networks (Lantz & Hutchison, 2015). By accounting for this variation and specifically attending to heterogeneity in co-offending, this reformulated theory may be better able to account for the relationship between co-offending and arrest in future research.
Footnotes
Acknowledgements
The author wishes to thank Barry Ruback, Wayne Osgood, Holly Nguyen, Jeremy Staff, Scott Gest, and Marin Wenger for comments on this and earlier versions of this research.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported in part by the Bureau of Justice Statistics (2015-R2-CX-K032). The views presented represent those of the author and do not necessarily represent those of the Bureau of Justice Statistics.
