We apply a novel approach to estimate the effects of exposure to peers with different attributes by using the predetermined leave-own-out attributes of all classmates in randomly assigned classes. This strategy allows a behavioural interpretation of the peer effect over and above the pure mechanical channel. We find that being exposed to peer groups with attributes conducive to academic achievements, induced by random variations in the shares of classmates with college-educated parents, increases exam scores. We show that estimates based on the commonly used leave-own-out measures are highly sensitive to sample selection bias arising from non-random tracking in the sample. We show that estimates based on the commonly used leave-one-out measures are highly sensitive to non-random tracking in the sample.