Dissertation, Otto-Friedrich-Universität Bamberg, 2017

Longitudinal data are collected over several time periods for the same units and therefore allow for modeling both latent heterogeneity and dynamics. Unfortunately, the analyst is often confronted with serious estimation problems that arise due to missing data. As the dependent variable within dynamic setups additionally serves as explaining variable in later periods, missing values cause an even higher loss of information and potential to inefficient estimation than in static panel models. For linear dynamic panel models with fixed or random effects, the thesis suggests a Bayesian approach to deal with missing values. Therefore, a Gibbs sampling scheme providing a sample from the model parameter posterior distribution is augmented by draws from the full conditional distribution of the missing values. While the full conditional distribution for missing values in the dependent variable is implied by the model setup, a flexible non-parametric approximation of the full conditional posterior distribution of missing values in the explaining variables is incorporated.
The thesis also provides non-nested model comparison in terms of the marginal likelihood from the resulting Gibbs sampling output. The properties and efficiency gains of the suggested estimation approach are evaluated by means of a simulation study. We compare the biases, coverage rates, root mean squared errors and model selections resulting from the data analyses before deletion and with missing values. Two empirical applications, the first modeling human growth, and the second investigating current account balances, illustrate the wide field of application of the developed approach.