Jackknifing then model averaging: investigating the improvements to fitness to data and prediction accuracy of two-input under-fitted and just-fitted response models

Abstract

The possibility of improving the fitness to data and prediction accuracy of models in a multi-response surface methodology environment of under and just-fitted ordinary least squares response models by jackknifing then combining the resultant partial estimates and the pseudo-values using arithmetic averaging or criterion-based frequentist model averaging was investigated. Jackknifing is known to reduce parametric and model bias. Model averaging is known to reduce model bidirectional bias and variance. A typical multi-response surface methodology dataset and resultant validation dataset were used as example. Results suggest that it is possible to obtain better fitness to data and prediction accuracy by jackknifing a just-fitted response model of interest and combining the resultant partial estimates using arithmetic averaging. The combining of pseudo-values using arithmetic averaging or criterion-based frequentist model averaging gave mixed results. The actual jackknife model estimators gave good performance with under-fitted models.