Bias in Conditional and Unconditional Fixed Effects Logit Estimation

Fixed-effects logit models can be useful in panel data analysis,when N units have been observed for T time periods. There aretwo main estimators for such models: unconditional maximum likelihoodand conditional maximum likelihood. Judged on asymptotic properties,the conditional estimator is superior. However, the unconditionalestimator holds several practical advantages, and thereforeI sought to determine whether its use could be justified onthe basis of finite-sample properties. In a series of MonteCarlo experiments for T < 20, I found a negligible amountof bias in both estimators when T 16, suggesting that a researchercan safely use either estimator under such conditions. WhenT < 16, the conditional estimator continued to have a verysmall amount of bias, but the unconditional estimator developedmore bias as T decreased.