Researchers typically examine peer effects by defining the peer group broadly (all classmates, schoolmates, neighbors) because of the lack of friendship information in many data sources as well as to enable the use of plausibly exogenous variation in peer group composition across cohorts in the same school. This paper estimates the effects of friend's health behaviors on own health behaviors for adolescents. A causal effect of friend's health behaviors is identified by comparing similar individuals who have the same friendship opportunities because they attend the same school and make the same friendship choices, under the assumption that the friendship choice reveals information about an individual's unobservables. We combine this identification strategy with a cross-cohort, within school design so that the model is identified based on across grade differences in the clustering of health behaviors within specific friendship options. This strategy allows us to separate the effect of friends behavior on own behavior from the effect of friends observables attributes on behavior, a key aspect of the reflection problem. We use a partial equilibrium model of friendship formation in order to derive the conditions under which our identification strategy will provide consistent estimates, and the key assumption required for our strategy to be feasible is supported by the empirical patterns of across cohort variation that we observe in our data. Our results suggest that friendship network effects are important in determining adolescent tobacco and alcohol use, but are over-estimated in specifications that do not fully take into account the endogeneity of friendship selection by 15-25%.