Using Propensity Score Analysis to Explore Subgroup Differences in Intimate Partner Violence Research

Background and Purpose: The exploration of group differences is of substantive concern to many social work researchers, such as differences in intimate partner violence (IPV) perpetration/victimization prevalence between individuals within different subgroups. Recently, potential differences between members of same-sex couples and those in mixed-sex couples have drawn increasing attention. Unfortunately, prior research has failed to employ robust methods for handling endogeneity bias, non-equivalent group comparison, and disproportionate subgroups. Instead, many analysts rely solely on covariate control methods. Thus, it remains challenging for researchers to disentangle the “true effect”, if any, of group membership on violence perpetration/victimization from other potentially confounding differences between subgroups. We present propensity score analysis (PSA) as a possible method for handling some of the challenges that accompany research in this area. Our key aim is to heuristically apply propensity score matching and weighting to the investigation of differences between members of same-sex and mixed-sex couples with respect to the occurrence of various types of IPV. Applying these methods may generate improved counterfactuals, bolster internal validity, and illustrate how PSA can be used to strengthen answers to a variety of important research questions.

Methods: Data came from the International Dating Violence Study (2001-2006). Our illustrative example uses an analytic sample of 3,960 participants in mixed-sex relationships and 121 participants in same-sex relationships. We examine group differences in the occurrence of five forms of IPV perpetration/victimization. We estimate propensity scores via logistic regression with group membership as the binary outcome, and eight covariates as predictors. Propensity scores with a logit-function transformation were used to conduct nearest-neighbor, within-caliper matching. After examining the common support region of propensity scores, we set the caliper to .10 of a standard deviation. As a sensitivity analysis, we also used propensity score weighting procedures to estimate an “average treatment effect” (ATE), and an “average treatment effect for the treated” (ATT).

Results: Results indicated that both matching and weighting procedures effectively balanced subgroups across all eight covariates. Prior to the application of propensity score procedures, results indicated two significant differences between subgroups. Propensity score matching reduced the total analytical sample to 245 participants, thus diminishing the statistical power of the analysis. In this context, the significant coefficients were reduced to non-significance, but still remained notably large. Propensity score weighting procedures retained the full analytical sample (4,081), re-established statistical significance, and even increased the magnitude of the coefficients.

Conclusion and Implications: Although propensity score methods are only capable of handling observed, rather than unobserved, subgroup heterogeneity, our results suggest that PSA may improve the counterfactuals used for subgroup comparisons. Importantly, our findings varied based on the propensity score method used. We suspect that post-matching results were more an artifact of limited statistical power, and less an artifact of correcting for confoundedness. Our methodology and results provide a meaningful heuristic by which PSA might be applied in various areas of social work research, particularly the examination of subgroup differences in risk and protective factors, prevalence, and impacts of IPV.