Abstract:
|
Longitudinal data are collected over several time periods for the same units and allow for modelling both latent heterogeneity and dynamics. Unfortunately, the analyst is often faced with serious estimation problems that arise due to missing data. In a dynamic setup the dependent variables also appear as explaining variables in later periods, and item nonresponse induces an even higher loss of information and potential to inefficient estimation, if the missing-data mechanism is not taken into account. The suggested linear Bayesian fixed and random effects estimation provides a Gibbs sampler routine with a data augmentation step to treat the item nonresponse by drawing the missing values iteratively from their full conditional posterior distribution. To handle possible nonresponse within additional covariates, we incorporate a Classification and Regression Trees (CART)-imputation into the sampler. Also, we provide non-nested model comparisons in terms of the marginal likelihood from the Gibbs output. The properties and efficiency gains of the suggested approaches are illustrated by means of a simulation study and an empirical application.
|