Model evaluation based on the sampling distribution of estimated absolute prediction error

摘要: The construction of a reliable, practically useful prediction rule for future responses is heavily dependent on the 'adequacy' fitted regression model. In this article, we consider absolute error, expected value difference between and predicted responses, as model evaluation criterion. This error easier to interpret than average squared equivalent misclassification binary outcome. We show that can be consistently estimated via resubstitution crossvalidation methods even when not correctly specified. Furthermore, resulting estimators are asymptotically normal. When 'nonsmooth', variance above normal distribution well with perturbation-resampling method. With two real examples an extensive simulation study, demonstrate interval estimates obtained from approximation errors provide much more information about adequacy their point-estimate counterparts.