Sylvia Frühwirth-Schnatter
Recent direction toward improving statistical estimation using MCMC methods
***************************************************************************

Statistical estimation using MCMC  methods is known to be sometimes extremely
slow. In this talk some of the reasons for this behavior will be analyzed. It
is a common strategy in statistical modeling to start with a rather general
model structure and to let the data tell us which special structure to choose.
I will demonstrate in this talk that a mayor cause for poor convergence of MCMC
methods stems from the attempt to estimate nearly unidentified models, whereas
MCMC convergence  is fine, whenever the data are informative about all unknown
parameters in the model. This indicates two potential direction toward improving
MCMC methods:

  1. Re-parameterizing the model: this is a standard technique for improving
     MCMC, which is well-known for random-effects models. It will be shown
     that this strategy could be extended  to much more general model
     structures.

  2. Combining MCMC estimation with simultaneous model selection: a second
     strategy which is currently under investigation is to find a
     parsimonious model structure simultaneously with MCMC estimation.