Automatic Model Selection

Automatic Model Selection is used to algorithmically choose the terms to keep in the model.

The Criterion is the statistic used to make the decision for how to choose the best model. There are four choices:

AICc stands for Akaike’s Information Criterion with a correction for a small design.

BIC is an alternate to AICc that performs better for larger designs. BIC stands for Bayesian Information Criterion.

For AICc, BIC, the Selection direction can be either Forward or Backward with Forward being the default. See the topic on Likelihood AICc and BIC.

p-value is the standard method looking for significant terms to keep and/or insignificant terms to remove from the model. P-value Selection directions of Forward, Backward or Stepwise are available.

The fourth criterion is Adjusted R-Squared. Only All Hierarchical Subsets selection is used with Adjusted R-Squared. The adjusted R-squared selection is followed by a backward step to determine if there is an insignificant term in the model. If so, the most insignificant term is removed from the model.

For most designs AICc using forward and/or p-value using backward work best. It is recommended to use multiple combinations of criterion and selection to help pick the best model for the system. Automatic Model Selection is not intended to replace the analyst’s decisions. Please take the time to review the results on the ANOVA and Diagnostics before using the model to make decisions.

Stat-Ease will remember the last criterion and selection method used and reuse it on the next use of automatic model selection.