A Note on Bayes Factors for Log-Linear Contingency Table Models with Vague Prior Information

Abstract

The approximate Bayes factor, B  01, for log-linear contingency table models proposed by Spiegelhalter and Smith (1982) is indeterminate if any of the cell frequencies is zero. It is noted that use of a standard Jeffreys prior overcomes this difficulty. It is pointed out that – 2 log B  01 is approximately equivalent to Schwarz's (1978) model selection criterion in large samples.