Partners in crime? Post-release recidivism among solo and co-offenders in Norway

ABSTRACT

Co-offending may increase offenders’ criminal capital in ways that impact their subsequent offending behaviour, and while highly theorized, the relationship between co-offending and reoffending has received less attention in empirical research. This study relies on Norwegian registry data to explore patterns of registered co- and solo offending before and after offenders’ first release from prison, by assessing differences in total, solo and co-reoffending between (1) co-offenders and solo offenders and (2) co-offenders embedded in different co-offending networks. The sample is based on 10 complete release cohorts, and co-offending networks are constructed from 22 years of administrative police data. Egocentric network analysis is used to obtain measures of degree centrality and tie strength. Results show that recidivism rates are higher among individuals with a co-offending network at release, and there is a consistent, positive relationship between degree centrality and reoffending. There is also a positive correlation between time spent in prison and the likelihood of co-offending after release, but there are no incidents of repeated co-offending (i.e. reoffending with co-offenders acquired before incarceration). The analysis hereby confirms several well-known patterns of co-offending in a new national context and highlights how incarceration can shape the nature and longevity of egocentric co-offending network ties.

KEYWORDS:

• Co-offending
• egocentric networks
• recidivism
• registry data
• Norway

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. The main reason for focusing on first-time releasees is that this is a theoretically interesting group of offenders who have all reached a point in their criminal career where variation in both co-offending and overall offending (i.e. in both the explanatory and dependent variables of the analyses) is plausible and they have experienced the intervention of incarceration – which might impact both existing co-offending ties and the overall trajectory of the criminal career – for the first time. This is only one of the many potential ways to go about exploring the questions at hand, and potential consequences of this particular sample criterion for the nature and generalizability of the results are addressed further in the discussion section.

2. Co-offending groups and co-offending networks are used interchangeably and refer to any constellation of two or more co-offenders (Carrington, 2014). These concepts differ from gangs, organized crime and criminal groups/organizations; see, e.g., Hashimi et al. (2016) for a review of terminology and, e.g., Pyrooz, Sweeten, and Piquero (2013) for an analysis of the association between gang embeddedness and desistance/persistence.

3. While most theoretical concepts in the influence perspective postulate a positive effect of co-offending on reoffending, the concepts of e.g., cost avoidance (Bouchard & Nguyen, 2010) and individual arrest risk (Ouellet & Bouchard, 2017) suggest a negative effect.

4. Note that some evidence suggests that group size and degree vary by crime type (see, e.g., McGloin and Piquero (2009) for violent vs. non-violent offences); however, any moderator effects between network measures and offence types remain outside the scope of this article.

5. As these data are available from 2001 onwards, the first release record represents the first-ever release record for offenders born in 1985 or later (the minimum age of criminal liability in Norway is 15 years). The sample also includes the first release records of offenders born between 1964 and 1985 who have only one record of a prison sentence in the data on criminal sanctions (Statistics Norway, 2018c), which are available from 1980 onwards.

6. As the logit transformation rests on an assumption of an error term distribution with a fixed variance of 3.29, logit estimates are affected by the degree of unobserved heterogeneity in the model (Mood, 2010). Comparing log odds ratios or odds ratios across samples, groups or over time is therefore problematic. Common criticisms of LPMs include the risks of inappropriate statistical tests due to heteroscedastic and non-normal residuals, predicted probabilities that are out of range and a misspecified functional form (cf. Hellevik, 2009; Long, 1997; Mood, 2010). The former critique is addressed by using heteroscedasticity-robust standard errors (Mood, 2010, p. 81), while the second is of limited practical importance (Hellevik, 2009; Long, 1997). It remains plausible that the recidivism rates are related to the network variables in a non-linear fashion; however, as long as the misspecification does not alter (more than marginally) the substantive conclusions, it remains reasonable to choose LPM over logistic regression (Mood, 2010, p. 78). To explore the relevance of this concern for the current analysis, I have re-estimated the models using a logit specification, and reassuringly, with only one minor exception (see note 12), the size, direction and p-values of all marginal effects are similar to those from the OLS models (see Appendix B, Table B1).

7. In calculating the average tie strength, the denominator equals the degree centrality, and offenders with larger networks would, therefore, have to repeatedly co-offend more times than those with smaller networks to obtain higher values on this variable. Additional descriptive statistics on the separate and joint distribution of the two network variables, as well as their relationship with recidivism, can be seen in Tables A1-A3 and Figure A1 in  Appendix A.

8. Western countries include the United States, Canada, Australia, New Zealand and European countries within the EU/EEA zone, and non-Western countries include the Central and South Americas, Africa, Asia (including Turkey) and the European countries outside the EU/EEA zone.

9. Second (or higher) degree polynomials for sentence length and time since the last co-offence were tested but not added as this did not significantly improve model fit.

10. If included in the same model (results not shown) the estimate sizes change somewhat, but the overall conclusions remain unchanged.

11. This finding stands in contrast to previous research, which tends to show that repeated co-offending is rare but still occurring (cf. Charette & Papachristos, 2017; McGloin et al., 2008). Plausible explanations for this puzzling finding are explored further in the discussion section.

12. These estimates are somewhat larger and reach statistical significance (p < 0.5) in the logit models (see Appendix B), while p < 0.1 for the LPMs. This is the only case in which the choice of an LPM over a logit model impacts the substantial conclusion of the results.

Additional information

Funding

This work was supported by The Research Council of Norway [grant number 202453].