Drug courts provide a unique environment in which to understand how substance users respond to contingencies within a criminal justice environment. These programs rely explicitly on a system of rewards and sanctions to reduce substance use and offending. Because substance use is monitored through regular supervised urine testing, they offer an opportunity to understand actual behavioural responses to those contingencies. While these longitudinal data provide a very powerful tool for researchers, the data also have certain properties that need to be appropriately accounted for in statistical analyses. One issue is that substance use is clustered within individuals. A second issue is the fact that some participants are removed from the program earlier than others as a result of poor performance. This drop-out is non-ignorable and unless it is explicitly modelled may result in biased and inconsistent treatment effects. This paper will outline a Bayesian approach to modeling longitudinal patterns of substance use with clustered data and non-ignorable treatment drop-out. The analyses are described in the context of a randomized controlled trial testing the impact of intensive judicial supervision on drug court outcomes.