Correcting the corrected AIC
Timothy DelSole, Michael K. Tippett
Abstract
The standard correction to Akaike’s Information Criterion, AICc, assumes the same predictors for training and verification and therefore underestimates prediction error for random predictors. A corrected AIC for regression models containing a mix of random and fixed predictors is derived.
Topics & Concepts
Akaike information criterionMathematicsStatisticsRegressionRandom errorStepwise regressionMean squared prediction errorStandard errorRegression analysisEconometricsStatistical Methods and InferenceAdvanced Statistical Methods and ModelsStatistical Distribution Estimation and Applications