As our data has five predetermined follow-up times, we used discrete-time survival analysis to examine the odds of relapse at subsequent time points and to examine predictors of relapse. This method is well-suited to longitudinal data with repeated measures of a discrete time outcome and allows for censoring of missing data at follow-up points. The initial model in a set of nested models included only the time effect represented by four dummy variables (one each for years 5, 7, 9 and 11). The time-only model estimates the overall population profile of the risk across time and indicates when events are more likely to occur (Xie et al., 2003). Abstinence versus non-problem use at 1 year was our main variable of interest so it was added to the second model (which also included the time effect) to see whether the hazard functions differed between these two groups. In the third model we added demographic, treatment and extra treatment variables in that order, to determine a final model with a best fit in predicting relapse. After finding our best fit model, we
