Desempenho da medida L na seleção de modelos normais

Abstract

Statistical models attempt to explain phenomena, natural or experimental. It is common to formulate more than one model to the same phenomenon and thus it is necessary to choose that one the best describes it. There are many criteria in the literature for comparison of models such as the Akaike information criterion (AIC), corrected Akaike criterion (AIC), Bayesian information criterion (BIC), among others, that try to minimize the loss of information in the modeling process. These criteria have asymptotic results. The L-measure is a measure for comparison of models concerned with the prediction values arising from the same or similar experiments using concepts such as predictive density in its definition, and thus, by comparing what is predicted to what is observed to make choice between models. In this work were calculated the rate of true positives (TP), false positives (FP), false negatives (FN) and true negatives (TN) for L-measure, as well as sensitivity to different sample sizes, smaller than 60. When considered predictive distributions quite close to the true predictive distribution, the results of the rates of TP and TN were low as well as the results for sensitivity. In other configurations considered for the study, with different predictive distributions from true predictive distribution, the results of the rates of TP and TN were high as well as the results for sensitivity. In general, the L-measure presented best performance than the AIC criteria, AIC and BIC for samples smaller than 60.