Mean squared errors of forecast for selecting nonnested linear models and comparison with other criteria (Q1174643)

This paper considers the Bayesian and sampling distribution of the mean squared error of forecast (MSEF) of a linear regression model. A new derivation is given for the distribution of MSEF using a degenerate hypergeometric function. The paper discusses the use of the smallest posterior mean of MSEF as a criterion for model selection. Performing sampling experiments with various nonnested linear models, the Bayesian MSEF criterion has been evaluated with Akaike's information criterion (AIC), Efron's confidence interval, the \(N\), \(J\) and encompassing tests. These experiments show that the Bayesian criterion performs as well as or sometimes better than the AIC except in case of small sample size when the dimension of the true model exceeds that of the competing model. Efron's \(C I\) tends to be more conservative than the other criteria when the difference between the models is not large; tables of empirical probabilities of correct selection by the criteria obtained from sampling experiments with various nonnested linear models have been presented.