Evaluation of non-inferiority is based on ruling out a threshold for what would constitute unacceptable loss of efficacy of an experimental treatment relative to an active comparator "Standard". This threshold, the "non-inferiority margin", is often based on preservation of a percentage of Standard's effect. To obtain an estimate of this effect to be used in the development of the "non-inferiority margin", data are needed from earlier trials comparing Standard to placebo if the non-inferiority trial does not have a placebo arm. This approach often provides a biased over-estimate of Standard's true effect in the setting of the current non-inferiority study. We describe two commonly used non-inferiority margin methods that adjust for this bias, the two-confidence interval (95-95) and the Synthesis margins. However, the added 'variance inflation' adjustment made by 95-95 margin diminishes with increasing information from historical trial(s), and the Synthesis margin is based on a strong assumption that the relative bias is known. We introduce an alternative "Bias-adjusted" margin addressing vulnerabilities of each by attenuating the estimate and by accounting for uncertainty in the true level of bias. Examples and asymptotic estimates of non-inferiority hypothesis rejection rates in the proportional hazards setting are used to compare methods.
Keywords: Non-inferiority; active control; bias; constancy; margin.