AIC for the non-concave penalized likelihood method (Q2414941)

A generalized linear model is considered based on a natural exponential family. A regularization method is presented which provides an estimator with nice asymptotic properties. The asymptotic bias of the estimator is found which yields an information criterion based on the original definition of the Akaike information criterion by considering minimization of the prediction error rather than model selection consistency. In fact a function of the score statistic is derived that is asymptotically equivalent to the non-concave penalized maximum likelihood estimator, and then an estimator is provided of the Kullback-Leibler divergence between the true distribution and the estimated distribution. A simulation study shows the validity of the proposed information criterion for several models.