The Empirical Performance of Measures of Association in the 2×2 Table

Abstract

A comparison of various measures of association for 2 x 2 data was carried out by computer analysis. Bivariate normal population distributions of known p were generated and divided into quadrants, yielding a total of 1,539 2 x 2 tables. For each of these tables, 14 measures of association were calculated and evaluated against the underlying p. A recently developed measure of association based on the odds-ratio (Chambers, 1982) appeared to be the most accurate in capturing the p underlying a majority of 2×2 tables. The phi coefficient, despite its continuing popularity, was found to be relatively inaccurate under most conditions studied.