A bioequivalence study that misses the Cmax interval with 42% within-subject variability
A two-sequence, two-period, two-treatment crossover bioequivalence study of a 400 mg modified-release tablet was run in 24 healthy fasted volunteers with a seven-day washout and full-profile sampling to 72 h. Twenty-three subjects completed; one withdrew after period 1 because of vomiting. In one subject, the period 2 pre-dose sample contained 6% of that subject's own Cmax. Results: AUC0-t geometric mean ratio 1.06 with 90% CI 0.97-1.15; Cmax geometric mean ratio 1.11 with 90% CI 0.92-1.34. Within-subject coefficients of variation were 18% for AUC and 42% for Cmax. The protocol pre-specified the conventional 80.00-125.00% acceptance interval for both parameters and stated that no subject would be excluded after the blind review. The sponsor asks whether the study can be reported as showing bioequivalence, and whether the two problem subjects can be removed from the Cmax analysis.
1. Interpret the results against the pre-specified acceptance criteria.
Read the worked answer
The two parameters must be judged separately. AUC passes comfortably: the geometric mean ratio of 1.06 with a 90% CI of 0.97-1.15 sits entirely within 80.00-125.00%. Cmax fails: the point estimate of 1.11 is within the interval, but the upper confidence limit of 1.34 exceeds 1.25. Bioequivalence for the rate of absorption is therefore not demonstrated, and because both AUC and Cmax must meet the criteria, the product cannot be declared bioequivalent.
The reason is visible in the variability: 42% within-subject CV on Cmax means the 90% CI will be wide regardless of the true difference, so this is as much a design failure (underpowered for a highly variable parameter) as a formulation failure. Under the conventional unscaled analysis the study is simply inconclusive-negative.
Crucially, the conclusion is not a matter of interpretation: the acceptance interval was fixed before unblinding, so re-analysis under a different rule after seeing the data would be a protocol violation and would undermine the study's validity in any regulatory review. Whether reference-scaling is even available depends on the applicable jurisdiction's guideline and must have been specified in the protocol in advance.
2. Evaluate the sponsor's proposal to remove the withdrawn subject and the subject with the 6% pre-dose concentration from the Cmax analysis.
Read the worked answer
Both requests should be refused as post hoc manoeuvres. The withdrawal after vomiting is a dropout unrelated to formulation performance in a way that permits exclusion from the dataset: the subject contributes valid period 1 data, and the standard analysis handles the imbalance either by including the period 1 data in an appropriate mixed model or by excluding the subject entirely from both parameters symmetrically. Excluding them only from Cmax while retaining them for AUC is data-driven and indefensible.
The pre-dose concentration is a genuine scientific problem: a value of 6% of Cmax exceeds the usual 5% limit and indicates carryover from period 1, given a seven-day washout against a modified-release product whose terminal phase may be long. The correct handling is pre-specified and symmetric: either exclude that subject's period 2 data (which leaves only period 1 data and requires an unpaired or mixed-model analysis) or exclude the subject entirely, and do the same for every parameter. Applying the rule only to the parameter that fails is the definition of bias.
ICH E9 requires that all exclusions be justified, pre-specified and documented before unblinding. The honest report includes a per-protocol analysis alongside the intention-to-treat analysis and states that neither changes the Cmax conclusion.
3. The sponsor wants to repeat the study. Advise on design, sample size and the statistical and ethical requirements before the repeat is run.
Read the worked answer
First, decide whether the true Cmax ratio is genuinely 1.11 or whether the estimate is noise. If the formulation truly differs by 11%, the repeat will need far more subjects: with 42% CV and a target of 80% power at alpha 0.05, the approximate total sample size is about 40 subjects for a true ratio of 1.00 but roughly 140 subjects for a true ratio of 1.11, since the margin to the 1.25 limit shrinks. The 24-subject study was therefore adequate only for a perfectly matching product.
Options, in order of preference: (1) reformulate to reduce the Cmax difference, then run a properly powered study; (2) if a reference-scaled or widened acceptance criterion for highly variable drugs is permitted in the target jurisdiction, design the study prospectively around that pathway, with a within-subject CV estimate from a pilot and an explicit pre-specified statistical analysis plan, noting that scaling widens limits symmetrically and the point estimate must still lie within 80.00-125.00%; (3) use a two-stage or group-sequential adaptive design with a pre-specified blinded sample-size re-estimation and a single interim analysis to control the type I error.
Ethically, a repeat exposure of healthy volunteers carries no benefit to them and must be justified by the value of the question, reviewed by an ethics committee to Declaration of Helsinki and GCP standards, with a data safety monitoring plan, a maximum blood-loss limit, and clear stopping rules for vomiting or postural hypotension. Repeating an underpowered study without changing the design is both unethical and likely to fail again.