the effect was more pronounced in some population subgroups than others. These analyses were conducted using both bivariate and multivariate models. The multivariate logistic regression models adjusted for socio-demographic, psychopathology, substance use, and treatment variables and whether or not the participants’ symptoms met the diagnostic thresholds for mood disorder diagnoses. However, there are limitations in the use of regression adjustment, in particular model dependence when the groups are different on the observed characteristics.40;41 Therefore, in addition to regression adjustment, we utilized the propensity score method of inverse probability of treatment weighting (IPTW)42;43 to adjust for differences between self-medicating and non-self-medicating participants. In this technique, first propensity scores (probability of self-medication) are computed using a logistic regression model. These scores reflect each participant's likelihood of self-medicating with alcohol given their socio-demographic and clinical characteristics. Next, data are weighted by their inverse probabilities of being in their observed group (self-medicating vs. non-self-medicating).41
