This paper shows that popular linear fixed-effects panel-data estimators (first-differences, within-transformation) are biased and inconsistent when applied in a discrete-time hazard setting, that is, one with the outcome variable being a binary dummy indicating an absorbing state, even if the data generating process is fully consistent with the linear discrete-time hazard model. Besides conventional survival bias, these estimators suffer from another source of - potentially severe - bias that originates from the data transformation itself and is present even in the absence of any unobserved heterogeneity. We suggest an alternative, computationally very simple, adjusted first-differences estimator that cures the data-transformation driven bias of the classical estimators. The theoretical line of argument is supported by evidence from Monte Carlo simulations and is illustrated by an empirical application.
linear probability model individual fixed effects short panel discrete-time hazard duration analysis survival analysis non-repeated event absorbing state survival bias misscaling bias