Bioequivalence: design before arithmetic

Conventional average BE

Log-transform each prespecified positive PK metric and estimate the test/reference contrast under the actual design. Back-transform the 90% confidence interval. Conventional 80–125% limits are not universal; select metric, design, region and product-specific guidance before analyzing data.

The helper supports complete two-period RT/TR crossover and independent parallel observations. Crossover requires exactly one T and one R per subject, correct period/sequence labels and both sequences. It accounts for sequence in the within-subject contrasts. Parallel analysis requires one value per subject. Neither path handles missing-period replicate data, carryover, multiple groups, adaptive designs or repeated independent-looking values.

Reference scaling

Replicate BE needs a design-specific statistical model. Period/sequence-adjusted reference variance and treatment-contrast uncertainty cannot generally be obtained by pooling each subject's raw replicate variance or averaging repeats. The CLI therefore rejects replicate analysis and reference scaling; it must not output a regulatory pass from those shortcuts.

FDA's May 2026 guidance replaces the 2001 statistical guidance. Appendix G supplies HVD models: sWR >= 0.294, a reference-scaled upper-bound criterion and an 80–125% point-estimate constraint. In its scalar bound, x=estimate²-SE² and the negative variance term's bound uses the 0.95 chi-square quantile. The retained rsabe_bound helper only combines externally justified scalar inputs; it cannot fit the required design or determine applicability. FDA NTI analysis is distinct; --nti merely changes this CLI's conventional limits. FDA statistical guidance.

EMA widening for eligible highly variable Cmax uses exp(±0.760*sWR), capped at CVwR 50%, with 80–125% point-estimate constraints and prespecified clinical justification. It is not generic permission to widen AUC or NTI limits. ABEL and FDA RSABE are different procedures. The scalar abel_limits is arithmetic only; use a qualified design-specific implementation. EMA BE guideline.

Planning

Power depends on GMR, within-subject CV, allocation, design, dropout and analysis assumptions. The bundled calculation integrates over estimated variance using a finite quantile grid. It is a numerical approximation, not an exact Owen-Q implementation. With balanced 2x2, CV 30%, GMR 0.95 and target power 80%, it returns N=40 evaluable participants in synthetic checks. Probe sensitivity, then independently check the intended design using PowerTOST.

Current harmonization boundary

ICH M13A concerns immediate-release solid oral products; EU implementation began 25 January 2025. Regional provisions for replicate and NTI studies still require attention. M13B is now EMA Step 5, adopted September 2026 with effect March 2027; it concerns additional-strength biowaivers, not HVD statistics. M13C remains under development. See regulatory guidance.