What you don’t see can’t hurt you? Panel data analysis and the dynamics of unobservable factors

Abstract

We investigate the consequences of using time-invariant individual effects in panel data models when the unobservables are in fact time-varying. Using data from the British Offending Crime and Justice panel, we estimate a dynamic factor model of the occurrence of a range of illicit activities as outcomes of young people's development processes.  This structure is then used to demonstrate that relying on the assumption of time-invariant individual effects to deal with confounding factors in a conventional dynamic panel data model is likely to lead to spurious "gateway" effects linking cannabis use to subsequent hard drug use.