Algorithms for evaluating reference scaled average bioequivalence: power, bias, and consumer risk

Abstract

The determination of the bioequivalence between highly variable drug products involves the evaluation of reference scaled average bioequivalence. The European and US regulatory authorities suggest different algorithms for the implementation of this approach. Both algorithms are based on approximations reflected in lower than the achievable power or higher than the nominal consumer risk of 5%. To overcome these deficiencies, a new class of algorithms, the so-called Exact methods, was earlier introduced. However, their applicability was limited. We propose 2 modifications which make their computation simpler and also applicable with any study design. Four algorithms were evaluated in simulated 3-period and 4-period bioequivalence studies: Hyslop's approach recommended by the US FDA, the method of average bioequivalence with expanding limits requested by the European EMA, and 2 versions of the new Exact methods. At small sample sizes, the Exact methods had substantially higher statistical power than Hyslop's algorithm and had lower consumer risk than the method of average bioequivalence with expanding limits. Similarly to the Hyslop's algorithm, higher than 5% consumer risk was observed only with either unbalanced study design or with additional regulatory requirements. The improved Exact algorithms compare favorably with the alternative procedures. They are based on the bias correction method of Hedges. The recognition that the scaled difference statistics is measured with bias has important practical implications when results of pilot bioequivalence studies are evaluated and, at the same time, calls for the revision of the statistical theory of RSABE and its related methods.

Critical Remarks on Reference-Scaled Average Bioequivalence

“…Many products were approved in Europe according to such more liberal rules. However, the observation that the consumer risk in reference-scaled average bioequivalence is compromised is not new (22,23,(49)(50)(51)(52)(53)(54)(55)(56)(57)(58)(59). It is not surprising that the Type I Error can be inflated, since in the current approaches  is undefined and the conclusion whether a product passes or fails is based on the within-subject variability of the reference treatment realised in the same study.…”

“…The current situation can be regarded as selecting the final statistical model based on intermediate datadependent decisionswhere up to three decisions are possible in the respective frameworks also utilizing unblinded treatment information. In such situations multiplicity concerns could arise where the influence on the overall Type I Error is difficult to assess (62, section 5.3) but were shown to lead to an inflation of the overall Type I Error (22,23,(49)(50)(51)(52)(53)(54)(55)(56)(57)(58)(59). To ensure compliance with the International Conference on Harmonisation (ICH) guideline E9 (63) regarding adjustment for multiplicity (e.g., section 5.6, which states that 'adjustment should always be considered and the details of any adjustment procedure or an explanation of why adjustment is not thought to be necessary should be set out in the analysis plan') the level of the test should be adjusted accordingly.…”

“…To ensure compliance with the International Conference on Harmonisation (ICH) guideline E9 (63) regarding adjustment for multiplicity (e.g., section 5.6, which states that 'adjustment should always be considered and the details of any adjustment procedure or an explanation of why adjustment is not thought to be necessary should be set out in the analysis plan') the level of the test should be adjusted accordingly. Various methods have been proposed to address this topic (52)(53)(54)(55)(56)58,59). Among those approaches, two do not entirely control the Type I Error (52,56), one requires the assumption that the observed CVwR is the true one (53), two require abandoning the PE constraint (54,55), one is computationally intensive (58), some require a modification of the regulatory approach of ABEL (52,54,55,58), and others lead to a substantial loss in power (52,59).…”

Association of Variability and Pharmacogenomics With Bioequivalence of Gefitinib in Healthy Male Subjects

“…In any case, the CI is directly proportional to the intra-CV and inversely proportional to the number of participants ( N ). Therefore, in study 5, we enlarged the Diff and sample size ( N ) without changing the drug batch or production process (Haidar et al, 2008 ; Ramirez et al, 2008 ; Tothfalusi et al, 2009 ; Baek et al, 2010 ; Tothfalusi and Endrenyi, 2016 , 2017 ). Then, we were able to get the bioequivalence result between the T and R formulations.…”

A study of confidence intervals for Cohen's dp in within-subject designs with new proposals

“…It is motivated by the numerous personal communications that we received following the publication of a review paper on the subject (Goulet-Pelletier & Cousineau, 2018) as well as by a follow-up article (Fitts, 2020). From these, it became apparent that there was to this day no satisfactory confidence interval for Cohen's d p in within-subject designs (see Tothfalusi & Endrenyi, 2017;and Viechbauer, 2007, for explorations). We consequently wish to document the strengths and limitations of these past approaches.…”