This paper considers estimation of panel data models with fixed effects. First, we will show that a consistent "unrestricted fixed effects" estimator does not exist for autoregressive panel data models with initial conditions. We will derive necessary and sufficient conditions for the consistency of estimators for these models. In particular, we will show that various widely used GMM estimators for the conditional AR(1) panel model are inconsistent under trending fixed effects sequences. Next, we derive, justify, and compare restricted Fixed Effects GMM and (Q)ML estimators for this model. We find that the FEML estimator is asymptotically efficient, whereas the Modified ML estimator is not. We will also compare the fixed effects approach for estimating the conditional AR(1) panel model and covariance parameters in static panel data models with the correlated random effects approach.