The participants in this study were not randomly assigned to the 2 study groups, although efforts were made to match the 2 samples as described above. To address possible sample selection bias, we used a modification of the propensity score approach proposed by Rosenbaum and Rubin.18- 19 We constructed a logistic regression model with a binary dependent variable indicating 1 = MHC and 0 = TAU. First, we entered in the model the basic variables of age, race, sex, and site. Second, we included a pool of all available potential explanatory variables: personal characteristics including ever married, education, lived with biological father until 15 years of age, father ever arrested, and father used illegal drugs; mental health history including age the individual first saw a mental health professional, age at first mental hospitalization, and ever having psychiatric hospitalization; current mental health factors including mutually exclusive diagnostic category, Insight and Treatment Attitudes Questionnaire, Colorado Symptom Index scores, mental health treatment in past 6 months, self-reported compliance with treatment and medication, and other types of medical treatment; child physical and sexual abuse and baseline violence; substance use and treatment such as alcohol and illegal drug use in past 30 days and received substance abuse treatment in past 6 months; criminal justice variables including age at first arrest, number of arrests since 15 years of age, number of pre–18-month incarceration days and arrests; annualized pre–18-month arrests; and target arrest, charge level, and most serious offense. Variables selected for the model by the stepwise procedure were marital status, Colorado Symptom Index, days using illegal drugs in last 30 days, diagnosis of depression, ever been hospitalized, received treatment for medical problem, violence at baseline interview, age at first arrest, and target arrest charge level (warrant, violation, misdemeanor, felony). These variables, along with those entered on the first stage (age, race, sex, and site), constitute the variables included in the propensity score model. The model has good characteristics, with a pseudo R2 (Nagelkerke) of 0.244 and an area under the curve of 0.750. We used this model to generate the propensity scores, and the resulting propensity score is included in all models comparing MHC and TAU samples, thus adjusting for selection bias.