We propose jackknife estimators for nonlinear dynamic panel data models with fixed effects that reduce the asymptotic bias of the maximum likelihood estimator (MLE) from O(T −1) to O(T −2) or smaller. The estimators are linear combinations of the MLE computed from the full panel and the MLE’s computed from two or more shorter subpanels.
The relative lengths of the subpanels determine the order of bias reduction that can be achieved. The jackknife can in a similar manner be applied to correct the score or the likelihood. Preliminary simulation results for the probit and logit binary AR(1) models are very encouraging. Even
in small, short panels such as N = 25 and T = 8, the jackknife is very effective in reducing the bias of the MLE and has smaller mean squared error.