[영문]This paper models heterogeneity in the relationship between exposure to information at school or in the media and cannabis use and heavy drinking, using latent class techniques applied to data on French teenagers collected in 1993. Teenagers cluster in five classes which differ in their tastes for drunkenness and cannabis, and in the correlations between information exposure and cannabis use or heavy drinking. Teenager heterogeneity and habit-formation or precociousness effects limit the effectiveness of general information policies. Improving the impact of prevention requires that interventions be better targeted and personalised. We show how economic theory, latent class techniques and existing psychometric questionnaires can be used to build simple statistical tools for targeting prevention policies.