Bivariate associations with remission status at follow-up for demographics, alcohol and drug use, social measures and treatment were tested using multinomial logistic regression, with current AUD as the reference group. Variables from bivariate analyses with p-values of .25 or less in relation to the outcome category being evaluated were selected for initial inclusion in the multivariate model (Hosmer and Lemeshow, 2000). Multivariate models were developed using logistic regression, rather than multinomial regression, because bivariate models showed that few variables were associated with all outcome categories; therefore a multinomial model would not provide a good fit to the data. Separate logistic regression models were fit for high-risk drinking, remitted low-risk drinking, and abstinence, using persistent AUD as the reference group for each. This method approximates a multinomial model in cases such as this where covariates are significant in some categories but not in others (Hosmer and Lemeshow, 2000). Variables that were not associated with the outcome category and that did not contribute to model fit based on comparison of fit statistics were eliminated from the model if they were not necessary
