This paper investigates the influence of peers on criminal behavior, using original data I collected by interviewing homeless people in Milan. Information on friends' names was elicited, which allows to map each respondent’s network. Each individual was also asked to report his criminal status prior to becoming homeless. To estimate the causal effects of network size and of the share of criminal friends on (subsequent) criminal behavior, I rely on two instruments. The first is the share of rainy days since the individual has become homeless: rainfall fosters concentration of homeless individuals in sheltered places and increases the probability of meetings. The second instrument is the fraction of inmates released by Milan’s authorities during one’s period as homeless, which affects the supply of criminal potential friends. I find that the probability of arrest decreases by 16 percentage points with the network size, but it increases by 20 percentage points with the share of criminal friends in the group.