A comparison of measures used for Configural Frequency Analysis of large, sparse cross-classifications
Alexander von Eye · 2005
In this article, the measures relative risk, RR, transformed versions of Log P and Pearson’s 2 X , and an approximation of Log P are compared. The measures had been discussed by DuMouchel (1999) and von Eye and Gutierrez-Pena (in this issue) for use in CFA of large, sparse contingency tables. Simulation results suggest that the four measures can be expected to yield the same rankings of cells as extreme if the discrepancies between observed and expected cell frequencies are extreme. With the exception of RR, this applies to the search for both types and antitypes. The approximation of Log P is good in particular for large discrepancies between observed and expected cell frequencies. Numerical issues and implications for users of CFA are discussed.