Dose-Weighted Adjusted Mantel-Haenszel Tests for Numeric Scaled Strata in a Randomized Trial

Abstract

A recent three-arm parallel groups randomized clinical prevention trial had a protocol deviation causing participants to have fewer active doses of an in-office treatment than planned. The original statistical analysis plan stipulated a minimal assumption randomization-based extended Mantel-Haenszel (EMH) trend test of the high frequency, low frequency, and zero frequency treatment groups and a binary outcome. Thus a dose-weighted adjusted EMH (DWAEMH) test was developed with an extra set of weights corresponding to the number of active doses actually available, in the spirit of a pattern mixture model. The method can easily be implemented using standard statistical software. A set of Monte Carlo simulations using a logistic model was undertaken with (and without) actual dose-response effects through 1000 replicates for empirical power estimates (and 2100 for empirical size). Results showed size was maintained and power was improved for DWAEMH versus EMH and logistic regression Wald tests in the presence of a dose effect and treatment by dose interaction.

Figures

Figure 1

Schematic of one stratum (h=1 Clinic 1 with eligible dosing pattern AAAI)

Figure 2

Empirical Power and Size for Extended Mantel-Haenszel (EMH), Dose Adjusted EMH, Wald, and Conditional Logistic Regression (CLR) Wald Methods with No Treatment by Eligible Dosing Interaction (Φ=0) No Dose Effect is for the scenario with γ = Φ = 0.

Figure 3

Empirical Power and Size for Extended Mantel-Haenszel (EMH), Dose Adjusted EMH, Wald, and Conditional Logistic Regression (CLR) Wald Methods with Treatment by Eligible Dosing Interaction (Φ=0.25)

Figure 4

Empirical Power and Size for Extended Mantel-Haenszel (EMH), Dose Adjusted EMH, Wald, and Conditional Logistic Regression (CLR) Wald Methods with Simple Treatment Effect Equal to Treatment by Eligible Dosing Interaction (γ =Φ)