Self-selection bias poses a major threat to the validity of research findings in naturalistic, quasi-experimental, or single-group designs. A new method of addressing self-selection bias in naturalistic evaluations of prevention programs was implemented. The study, involving voluntary exposure to multicomponent interventions, was developed and applied to an evaluation of an alcohol abuse prevention program in which student participation in hall-based programs was conditioned by choice of where to live. A longitudinal mail survey of students in 1978 (N=274) and 1980 (N=197) assessed the impact of the alcohol abuse prevention program implemented in the intervening years. The effects of three interventions were compared: (1) alcohol education; (2) structured drinking environments; and (3) living group self-regulation activities. Choice of living environments and other variables were controlled. Outcome measures included alcohol consumption and alcohol-related problems. Results generally showed a lack of program impact. Aggregate levels of both drinking and problems differed substantially across living groups, and a strong selection rule was found predicting the probabilities of being in each living group. Problem drinkers did not avoid program exposure even though programs were offered in the living groups. The findings suggest that after correcting for possible self-selection bias, differences in alcohol consumption and problems must be accounted for primarily by sorting between living groups rather than by living group climate or other environmental factors. (JAC)