Evaluating Goodness-of-Fit Measures for Synthetic Microdata

Abstract

Goodness-of-fit tests are widely used by geographers, but choice remains difficult. Our overview starts with a conceptual examination of the nature of fit. Three important approaches to assessing the goodness-of-fit of categorical data are reviewed and appraised: statistics tested against the h ² distribution, the normal Z score and its variants, and measures derived from information theory. The forms of the phi and psi statistics highlighted in earlier geographical work are shown to be closely approximated by a simple measure of absolute error. Empirical examples help to illustrate the relative utility of these tests for a range of purposes.