Estimating the effects of friends on health behaviors of adolescents.

This paper estimates the effects of friends' smoking and drinking on own behavior while controlling for correlated unobservables between friends. The effect of friends' behaviors is identified by comparing similar individuals who have similar friendship opportunities and make similar friendship choices, exploiting the idea that friendship choice reveals information about unobservables. We combine this identification strategy with an across-cohort within school design so that identification arises in our reduced form estimates from across-grade differences in the clustering of health behaviors. Finally, we use estimated information on correlated unobservables to examine longitudinal data on the onset of health behaviors, where the likelihood of reverse causality should be minimal. We find evidence that this strategy produces somewhat smaller (no more than 16% smaller) friendship effect estimates than the more standard school fixed effect models consistent with at most modest bias from correlated unobservables.