A Predictive Model Selection Criterion

SUMMARY A procedure for model selection is presented which chooses the model that gives the best prediction of the future observation. The predictive distributions associated with each model are compared by means of the logarithmic utility function. For two nested normal linear models, the choice criterion is the product of the posterior odds ratio and a factor depending on the design point of the future observation. Taking appropriate mean values, a criterion is obtained which is independent of the particular design point. This average criterion differs from the ones proposed by Akaike, Schwarz and others in that it adjusts the likelihood ratio statistic by taking into account not only the difference in dimensionality, but also the estimated distance of the two models.