| Abstract: | There is a long-standing debate in criminology about the relative impact of static versus dynamic factors on criminal behavior. Researchers interested in estimating the impact of dynamic factors like prior offending or association with delinquent peers on criminal offending must control for static factors like intelligence, family background, or self-control, which could plausibly be correlated with criminal offending and the dynamic factor itself. Unfortunately, as a practical matter, it is not possible to observe all of these static factors. Statisticians and econometricians have shown that it is possible to identify the collective effect of static factors even though they cannot be observed. To achieve this objective, however, it is necessary to account for stable, unobserved individual characteristics through the use of "fixed-effect" or "random-effect" estimation. Criminologists often use random-effect estimators in these situations. We describe some of the assumptions that are necessary to develop valid inferences when time-varying covariates are used. Then, we use simulation evidence and an empirical application to show that bias can result when they are violated. |
