Alternative Tests for the Selection of Model Variables

Abstract

This paper contrasts two approaches to testing the importance of model variables: single-equation statistical tests, such as are used in regression analysis, and model-behavior tests. The paper attempts to show that tests which analyze the impact of individual variables on model behavior are better suited, both theoretically and operationally, to the task of selecting model variables. Conversely, the analysis shows that statistical tests should not be viewed as tests of model specification per se, but as tests of a particular type of data usefulness: they warn the modeler when available data do not permit accurate estimation of a model parameter. However, as a detailed example illustrates, a model relationship may be difficult to estimate yet extremely important for overall behavior. The paper concludes by summarizing two recent applications of model-behavior testing to analyze alternative business-cycle theories and alternative models for capital investment.