Hedges’ g was chosen as the ES because it adjusts obtained ES by sample size (Hedges, 1981). ES and their variances were entered into hierarchical linear modeling [HLM (Raudenbush and Bryk, 2002)] using the HLM 6.04 software package (Raudenbush et al., 2000). ES were inversely weighted by their sampling variances so that ES with larger samples and lower sampling error received greater weight than smaller samples with greater sampling error. Using full maximum likelihood estimation, the random intercept model was first fitted to calculate the overall mean ES, test whether it was significantly different from 0, and assess the heterogeneity of the ES distribution. Following significant heterogeneity, moderators were set to fixed effects and assessed by the t-statistic in level 2 analyses (least squares estimates of fixed effects, with robust standard errors). Moderator analyses for difference in years of education, duration of AUD, lifetime consumption, number of days abstinent, and treatment-seeking status were performed within the subsets of studies reporting those variables. In addition, mean ES were calculated for male and female samples for each study reporting data from men and women separately. Moderator analyses were not performed within male and female subsets due to power limitations.
