Abstract:
|
Like some brand of mathematical mentalist, the skilled statistician instills awe in their collaborators with eerily accurate predictions obtained from carefully curated statistical models. However, without informed selection, the statistician’s predictive performance almost surely falls flat. For this reason, an integral part of any statistician’s box of tricks are information criteria used for model selection. Of these criteria, it is well known that the Akaike information criterion, AIC, and its corrected counterpart, AICc, facilitate the selection of models with good predictive ability. However, the development of these and other AIC variants presume a panel of validation data that follows an identical distribution to that of the fitting sample, which is conceptually misaligned with the objective of predicting future or new observations, latent data, or missing data. This presentation introduces a new class of selection criteria that addresses this potential shortcoming and allows for the selection of good predictive models, regardless of the nature of data one would like to predict.
|