Control theory implies that social services will decrease levels of deviant behavior by strengthening the bond between the adolescent and society. Labeling theory implies the reverse, i.e., that the process of formal adjudication through the juvenile court will first stabilize and then increase levels of deviant behavior. Diversion programs were originally developed as an application of labeling theory, with the objectives of minimizing involvement with the juvenile court, referring adolescents to less stigmatizing social services, and ultimately reducing levels of deviant behavior. This study tested hypotheses that over time control theory variables are related to decreases in drug use and to each other; labeling theory variables regarding drug use are related to increases in drug use, increases in each other, and negatively to control theory variables; participation in a nonsystem juvenile justice diversion program is related to an increase in drug use, more deviant self-labels, and weaker societal bonds; and the effects of the diversion program differ across gender of respondent. Baseline and follow-up data were collected on a randomly selected comparison group of adolescents (n=100, 46 females) and a diversion program sample of adolescents (n=88, 46 females). Slope differences across samples were tested by using interaction terms. Where significant interaction terms were found, additional separate sample regression equations were performed. Changes in drug use over time were apparently a function of several complicated processes, including weaker social bonds and a more deviant self-concept. Nonsystem diversion apparently magnified these processes, especially for girls. Interactional theory was supported, in that changes in drug use resulted in a more deviant self-concept over time and in that variables derived from control and labeling theories influenced each other in complex ways over time. Studies of deviant behavior thus need to be multitheoretical, and longitudinal methodologies with nonrecursive modeling should be used. 2 tables and 30 references