On robustness of noninferiority clinical trial designs against bias, variability, and nonconstancy

The regulatory guidelines on noninferiority (NI) trials emphasize constancy not only in the treatment effect over time but also in the trial design, clinical practice, and quality of the trial conduct and execution. In practice, the constancy assumption is generally impossible to justify; often, there are clear reasons to expect a loss of efficacy over time. There are also concerns about the inherent and publication bias in the historical data, and various sources of selection bias in the NI trial design. Thus, a conservative NI margin is often considered. However, different NI margin approaches are largely evaluated under the assumption of constancy and absence of bias, and therefore, controversies arise and are unresolved on the necessary degree of conservativeness. We develop a framework to quantify the robustness of any NI margin approach against inherent and publication bias in historical data, selection bias in trial design, and nonconstancy in reference effects. We introduce a consistency principle to address variability in the historical data. We control across-trial conditional error rates given a final NI trial design over a design specific robust range for reference effects. Following a conditionality principle, we provide a theoretical justification of the framework and the conditions for controlling across-trial unconditional type 1 error rates. We raise the issue of inherent bias in historical data with an illustrative example.