Dynamic models for discrete response data extend the usual GLM framework by introducing time-dependent parameters. They can also be viewed as modified state space models with normal transition but non-normal observation models.
These models are a useful tool for the analysis of categorical time series or panel-data. It might not be obvious that other types of data also fit into this framework:
-Discrete time survival data with one or multiple terminating events and time-varying effects;
-Ordered paired comparison data, observed over time.
Throughout my talk I will focus on two applications analyzing unemployment data from the German socioeconomic panel (SOEZP) and results from the German soccer league. It will be illustrated that an analysis via MCMC in a full Bayesian setting has some distinct advantages compared to other approaches:
-It allows for calculating pointwise or simultaneous credible regions of time-dependent parameters or functionals of those, e.g. hazard rates or predictive survival functions.
-Missing values and restrictions on the parameter space are easily incorporated.
Finally, some generalizations of these models will be discussed.