Non-constancy, estimation bias, biocreep, and an alternative to current methods used in non-inferiority trials

This dissertation concerns the evaluation of the efficacy of an experimental treatment (EXP) in the setting of a non-inferiority (NI) clinical trial where some loss of effect of the standard treatment (STD) is acceptable. In NI trials without a placebo (PLA) arm, estimation of EXP effect relies on a historical STD to PLA effect estimate that is (i) substantially positive, (ii) precise, and (iii) unbiased in the setting of the NI trial[1]. Aspects related to these assumptions have been subject to recent debate [1, 2, 3, 4].;First, validity of assumption (iii) relies on "constancy" of STD to PLA effect across historical and NI trials. The "95-95" margin currently advocated by the FDA makes an adjustment which depends on variability of efficacy estimates. We introduce an alternative margin which makes a bias adjustment, and we compare this method to the 95-95 margin and to Synthesis margins.;Next, selection of the data for estimating the effect of STD is at the discretion of trial investigators. This flexibility is one of several factors which may lead to a biased estimate of STD effect. Furthermore, when very few, small trials are used for estimation of effect, the precision of the estimate is poor violating (ii).;Finally, repeated use of NI trials may lead to "biocreep", whereby there is an erosion in the level of improvement provided by newly approved therapies after multiple generations of NI trials. Simulations suggested that biocreep may be great, especially when applying large fixed margins traditionally employed in Anti-Infective Drug trials.;Simulations suggest that bias in formulation of the NI margin may lead to substantial risk for false conclusion of NI. Very often, each of the many sources of bias will tend to overestimate the effect of standard, and the accumulation of biases will lead to even greater false positive rates. Whenever possible, superiority tests should be used when comparing new therapies to active controls. When testing for non-inferiority is deemed appropriate, use of an approved treatment which has been directly compared to placebo is ideal, and the proposed bias-adjusted margin makes an accommodation for the many likely though unknown sources of bias.