MODEL SELECTION FOR MULTIVARIATE REGRESSION IN SMALL SAMPLES

We develop a small-sample criterion (AIC(C)) for selecting multivariate regression models. This criterion adjusts the Akaike information criterion to be an exact unbiased estimator for the expected Kullback-Leibler information. A small-sample comparison shows that AIC(C) provides better model order choices than other available model selection methods. Data from an agricultural experiment are analyzed.