Parsimony and Model Evaluation

Abstract Marsh and Hau (1996) argued that certain models should not be penalized for having low parsimony because an appropriate model for the data may require estimating more parameters. Mulaik argues that Marsh and Hau misunderstand the concept of parsimony, particularly its role in testing a hypothesis about an incompletely specified model to establish its objective validity. More parsimonious models represent more complete hypotheses having more ways of being tested and possibly being disconfirmed. Mulaik also shows that even within the context of the models used in Marsh and Hau's examples, there are much more parsimonious versions of those models that could have been hypothesized and tested, with good fit.