ABSTRACT
Model selection is the process of finding the 'best' model supported by data. This chapter presents the idea for balanced experiments which are applicable in general to unbalanced experiments and other linear models. Since there may be a large number of models, most selection procedures involve comparing two models that differ only by one term, either an interaction or a main effect. Different model selection procedures return two disparate models which support the data equally well but are mutually incompatible. Balanced designs with one observation per cell raise interesting dilemmas about model selection. Plots can provide insight into the model selection process. It is possible use one or a few degrees of freedom to investigate special forms of interaction. There are few automated tools for model selection in statistical analysis of variance packages. Factors with only two levels can be recast as regressors and used in a stepwise regression procedure.
