Linear fixed-effects estimation with nonrepeated outcomes

Abstract We demonstrate that popular linear fixed-effects panel-data estimators are biased and inconsistent when applied in a discrete-time hazard setting, even if the data-generating process is consistent with the linear model. The bias is not just survival bias, but originates from the impossibility to transform the model such that the remaining disturbance term becomes conditional mean independent of the explanatory variables. The bias is hence present even in the absence of unobserved heterogeneity. We discuss instrumental variables estimation, using first-differences of the explanatory variables as instruments, as alternative estimation strategy. Monte Carlo simulations and an empirical application substantiate our theoretical results.