A MODEL SELECTION APPROACH IN STATISTICAL MODELING

It is argued that quantitative results from statistical surveys and experiments should be communicated as inferences of the model maximising the log Bayes factor against a reference model penalised by a subjectively chosen constant times the difference in model complexity. Model complexity is measured by the degrees of freedom. In this study, an efficient algorithm is proposed to select a model from among a large set of models with unit penalties in some interval. The algorithm utilizes the penalised log Bayes factor with only the likelihood ratio statistic, model dimensions and a constant. This approach seems to be a more realistic screening device than related criteria similar to the Bayesian information criterion.