Abstract
The Mantel–Haenszel (M–H) procedure is commonly used to compare two treatments in a stratified binomial trial. However, this procedure is asymptotically optimal only if the odds ratio is constant across strata. We propose an alternative analytic strategy based on the simultaneous use of two statistics, Z S and Z I, each involving a weighted averaging of within-stratum differences between proportions. The two treatments are declared significantly different at overall level α if either min (|ZS |, |ZI |) > Z α/2 or max (|ZS |, |ZI |) > Z α*/2, where α* is data dependent. Our strategy is shown to be more powerful than the M–H and other related procedures. Numerical examples are provided for illustration.