Assessing prediction error at interpolation and extrapolation points (Q2286369)

In the context of predictive statistical modeling and inference, the authors introduce two new prediction error estimators which are particularly suitable for situations where the prediction error has to be assessed at so-called transduction points (points that do not appear in the training sample). The first estimator is based on normality assumptions and penalized likelihood, while the second estimator is based on the squared error loss and expected optimism correction. Based on the proposed estimators, also two new model selection criteria are suggested. Relevant use cases include spatial (kriging) models and repeated measures studies. Properties of the suggested methodology are illustrated in computer simulations as well as by analyzing two real data sets.