Master of Science / Department of Statistics / Christopher Vahl / Non-inferiority trials have been widely used in many medical areas. The goal of a non-inferiority trial is to show that a new test therapy is either better or not too much worse than the active control rather than showing the test therapy is superior to a negative control (i.e. placebo). The appeal of a non-inferiority trial is that it is often unethical to give some patients a treatment with no therapeutic benefit. When designing a non-inferiority trial, the issues of assay sensitivity, sample size, constancy condition, and a suitable non-inferiority margin need to be considered. A poor choice of the non-inferiority margin is a major reason that many non-inferiority trials fail. A numerical example is presented to show how to estimate the non-inferiority margin without historical data.

To assess non-inferiority of an experimental product to an active control in a clinical trial, an ideal design is to include a placebo arm to ensure both the experimental product and the active control is superior to placebo. We aim to identify methodology to control Type I error rate and maintain adequate power in a superiority/non-inferiority seamless clinical trial defined as: 1. selecting the best experimental treatment dose vs. placebo out of multiple treatment doses in Stage I; and 2. assessing non-inferiority of the chosen experimental dose to an active control, after adding subjects to yield adequate power for non-inferiority, in Stage II. The trial design here is an antihypertensive trial with change in systolic blood pressure as the outcome. The trial has three experimental treatment doses arms of experimental, a placebo control arm, and an active control arm. A simulation study of 20,000 such trials was conducted. We apply multiple comparison methodologies in Stage I to detect the most beneficial experimental treatment dose versus placebo, and test non-inferiority of the selected experimental dose to the active control in Stage II. Simulated Type I error rate and power for claiming non-inferiority are calculated for various dose-response trends. The need to adjust alpha to control Type I error either stage is assessed, seeking the optimal approach for doing so. Next, type I error and power for various fixed and variable non-inferiority margins are evaluated, exploring a range of margins informed by the first stage results of the study. A variable non-inferiority margin informed completely by the first stage of the trial approach results in inflated error rate which cannot be controlled by suggested multiplicity adjustments. We assess a synthesis approach between the fixed and variable margins, to both control the family-wise error rates and reach adequate power, depending on a tuning parameter defined in our work. We conclude that well-designed and adequately controlled seamless superiority/non-inferiority trials are possible with appropriate multiple comparisons adjustments and could result in less development time and fewer subjects needed to assess efficacy than separate trials.