|
Abstract:
|
Consider a clinical trial designed to establish the non-inferiority of two treatments based on a binary endpoint. If randomization is pre-stratified by a prognostic factor (e.g., gender), then the Z-statistic used to test for non-inferiority is a function of (i) the non-inferiority bound, (ii) the estimated overall treatment effect, d, obtained as a weighted average of the stratum-specific treatment effects, and (iii) the estimated variance of d, V_d. The power of the test for non-inferiority is influenced by the choice of weights, and by whether the null or observed stratum-specific variances are used to determine V_d. We use simulations to show the power gains achieved by using the minimum risk weights of Mehrotra and Railkar (2000) in tandem with observed variances, relative to common approaches such as the Cochran-Mantel-Haenszel weights in combination with null variances.
|
