This study examines prison peer effects in an adult prison population in the United States using a unique dataset assembled from the administrative databases of the Pennsylvania Department of Corrections. The members of a first-time prison release cohort were identified and matched to each of the cellmates with whom they shared a double cell. These data were then linked to arrest history data from the Pennsylvania State Police.

Criminological theories of social influence expect unobserved and difficult to quantify factors, such as criminality, to affect criminal behavior both independently and through intermediate decisions, including the choice to maintain prison peer associations. Those theories, therefore, implicitly assume the presence of essential heterogeneity, which helps to account for the response heterogeneity observed in studies of social influence. This study introduces the concept of essential heterogeneity to criminology and is the first to apply a method to address it, local instrumental variables, to estimate causal social interaction effects.

The analyses presented in this study demonstrate that there is considerable response heterogeneity in prison peer effects. That response heterogeneity is attributable to essential heterogeneity, as implicitly expected by criminological learning theories. However, the null average effects estimated do not accord with the predictions of criminological learning theories, including differential association, balance, and prisonization theories, each of which expects peers who are, on average, more criminally experienced to exert criminogenic effects.

The presence of essential heterogeneity indicates that estimating average prison peer effects does little to adequately characterize the relationship between social interactions with cellmates and releasee reoffending behaviors. Within the null average prison peer effect estimates lies tremendous variation in marginal prison peer effects. Some marginal prison peer effects are significantly criminogenic, while others are significantly crimino-suppressive. That substantial variation in the measured effect of prison peers on reoffending persists despite rigorous analysis and the inclusion of robust theoretically relevant controls suggests that future work should focus on creating constructs more appropriate to the task of determining who is harmed and who is helped as a result of interactions with prison peers.