The current regulatory criterion for bioequivalence of narrow therapeutic index (NTI) drugs in the European Union requires that the 90% confidence interval for the ratio of the population geometric means of the test product compared with the reference for area under the plasma concentration-time curve (AUC), and in certain cases maximum plasma drug concentration (C max ), to be included within the tighter acceptance range of 90.00-111.11%. As a consequence, sponsors need to recruit a higher number of subjects to demonstrate bioequivalence and this may be seen as increasing the burden for the development of generics. This "one-size-fits-all" criterion is particularly questionable when the within-subject variability of the reference product is moderate to high. As an alternative, we propose a further refined statistical approach where the acceptance range is narrowed based on the withinsubject variability of the reference product of the NTI drug, similar to the one used for widening the standard 80.00-125.00% acceptance range for highly variable drugs. The 80.00-125.00% acceptance range is narrowed, only if the within-subject variability is lower than 30%, down to the current NTI acceptance range of 90.00-111.11% when the within-subject variability is 13.93% or lower. Examples within the current European Medicines Agency list of NTI drugs show a considerable reduction in required sample size for drugs like tacrolimus and colchicine, where the predicted within-subject variability is 20-30%. In these cases, this approach is less sample size demanding without any expected increase in the therapeutic risks, since patients treated with reference products with moderate to high within-subject variability are frequently exposed to bioavailability differences larger than 10%.

Bioequivalence (BE) of products containing narrow therapeutic index (NTI) drugs in the European Union is currently established by demonstrating that the 90% confidence interval for the ratio of the population geometric means of the test compared to the reference product’s AUC, and in certain cases Cmax, is included within the tighter acceptance range of 90.00–111.11%. An alternative criterion, consisting of narrowed limits based on the within-subject variability of the reference product, was recently proposed. Its performance for a three-period partial replicate design was tested by simulation in terms of power to show BE, type I error (T1E) and sample size requirements. A new condition, a constraint on the test-to-reference geometric mean ratio (cGMR) to be contained within the range of 90.00–111.11%, was also tested. The probability of showing BE when the products differ more than 10% was increased, but only if the reference product’s within-subject variability was moderate-to-high. The inclusion of the additional cGMR limited this. An increase in the T1E (<7%) was observed. The inclusion of the additional cGMR did not change the highest inflation of the T1E. Finally, a significant sample size reduction was observed and the inclusion of the cGMR usually did not increase the required sample size.