Revisiting the Use of Propensity Score Matching to Understand the Relationship between Gang Membership and Violent Victimization: A Cautionary Note

Abstract

Previous research has consistently reported that gang members are more likely to experience violent victimization compared to non‐gang members. Recently, however, a study challenged this conventional wisdom using the Gang Resistance Education and Training (GREAT) data. Employing propensity score matching (PSM), this study reported no significant differences in violent victimization between gang and non‐gang members. Upon closer examination of the GREAT data and the PSM process used in this study, we note several theoretical, methodological, and statistical concerns. We reanalyze the GREAT data using both negative binomial regression and PSM. We find that self‐reported gang members were significantly more likely to report subsequent violent victimization compared to non‐gang members. Although contrary to this previous study, our findings are consistent with the bulk of previous empirical research and widely held beliefs about the relationship between gang membership and violent victimization.

Keywords:

• gangs
• victimization
• violence
• methods
• propensity score matching

Notes

1. According to Hill, “matching on propensity scores induces dependence across the matched samples. Failure to address this aspect of the observational study design can result in standard errors that are overly conservative” (2008, p. 2058).

2. Only 18 students reported with more than 12 victimizations; these were re‐coded to 12.

3. Due to the small sample size (N = 68), t‐test results were also confirmed with standardized differences as proposed by Rosenbaum and Rubin (1985). No covariate was significant using the standardized difference formula: [( x ₁− x ₂)/(s ₁ + s ₂/2)] × 100. Rosenbaum and Rubin (1985) suggest that standardized differences over 20% are significant.

4. We conducted additional analyses using Gibson et al.’s measure of victimization, however we were still unable to replicate their findings. These analyses are not reported in this article, however are available from the authors upon request.