Lindley paradox

• Probabilities on models tend to emerge as zero or one, and to be sensitive to the particular data set or to apparently minor aspects of specification.
• Sparse sets of models.
• The importance of choice of priors on parameters within models.
• Training sample priors.

References:

• sims.princeton.edu/yftp/OttawaRemarks on Bayesian Model Comparison
• http://www3.interscience.wiley.com/cgi-bin/fulltext/77001972/PDFSTART Schorfheide (2000)