| Many empirical studies involving dynamic panels employ the first-order linear dynamics without providing obvious justification for doing so. Hence, these models are most likely misspecified. This dissertation seeks to tackle this problem by providing more general approaches to econometric specification, estimation and bias correction in dynamic panel models with fixed effects.; The first chapter considers fixed-effects autoregressive models with unknown lag orders and examines asymptotic bias formulae in them. When the lag orders are unknown, first-order models are most likely misspecified and attempts to correct the bias using formulae that correct for first-order dynamics may even exacerbate the bias. To address these concerns, we extend Nickell (1981) bias formula to the case where the dynamics follow a general autoregressive form. We also examine higher order approximations for the bias, and develop a limit distribution for estimators that allows for lag order misspecification. The results reveal additional biases under misspecification, and therefore model specification should precede any correction for bias in dynamic panel modelling. We suggest a general form for bias correction, in the context of dynamic model specification, which incorporates a modified lag order selection method. An empirical study on habit formation in consumption preferences is presented to illustrate the use of the general bias correction.; The second chapter investigates nonlinear dynamic structures in panel models and develops nonparametric estimation of dynamic panel models using series approximations. We extend the standard linear dynamic panel model to a nonparametric form that maintains additive fixed effects. Nonlinear homogeneous Markov process is properly conditioned to be stationary beta-mixing. Convergence rates and the asymptotic distribution of the series estimator are derived, in which an asymptotic bias is present and it reduces the mean square convergence rate compared with the cross section case. To tackle this problem, bias correction is developed using a heteroskedasticity and autocorrelation consistent type estimator. Some extensions of this framework are also considered under exogenous variables and partial linear models. Finally, an empirical study on nonlinearity in the cross-country growth regression is presented to illustrate the use of the nonparametric dynamic panel models with fixed effects. |
