Narrow-Therapeutic-Index BE With Replicate Data

Why Replicate Data Matter for NTI Bioequivalence

For a drug with a narrow therapeutic index (NTI), comparatively small differences in exposure may be clinically important. The acceptance limits used for an NTI assessment can therefore be tighter than the conventional bioequivalence limits. A replicate crossover design adds an important feature: participants receive the reference formulation more than once, allowing within-subject reference variability to be estimated as well as the test-to-reference treatment effect.

This tutorial uses a specified pair of fixed limits, 90.00% and 111.11%, where the upper limit is defined as the exact reciprocal of 0.90 before rounding. These limits are an assumption for the worked analyses, not a claim that every regulator or product uses the same NTI procedure. The applicable product guidance and protocol determine the endpoint, limits, analysis population, and model.

Replicate Design and Analysis Model

Consider a four-period, two-sequence replicate design. One sequence receives test, reference, test, reference (TRTR); the other receives reference, test, reference, test (RTRT). Each participant receives each formulation twice. The endpoint is analyzed on the natural-log scale, as is standard for positive exposure measures such as AUC or Cmax. The analysis is performed separately for each endpoint subject to the bioequivalence decision.

A suitable mixed-effects model for an observation from participant \(i\), period \(j\), and formulation \(k\) is

where \(Y_{ijk}\) is the log-transformed exposure, sequence, period, and treatment are fixed effects, and \(b_{i(\text{sequence})}\) is a participant-specific random effect nested within sequence. The residual \(\epsilon_{ijk}\) represents within-participant occasion variation. In a basic model, participant effects are normally distributed with mean zero and variance \(\sigma_b^2\); residuals are normally distributed with mean zero and variance \(\sigma_e^2\), independent across occasions and of participant effects. The treatment contrast is the estimated log mean for test minus the estimated log mean for reference.

The model and its assumptions matter. The planned washout must be adequate to prevent relevant carryover; the formulation and assay conditions should be consistent across periods; and the analysis must account for the design’s sequence and period structure. The residual model should be appropriate for the data. Strong period patterns, influential observations, unexplained sequence differences, or unequal residual variability may signal that the basic model needs investigation. Replication does not itself correct carryover, data errors, or a misspecified model.

The two reference observations help estimate within-subject reference variability. In a simple setting with two independent reference occasion residuals of variance \(\sigma_{wR}^2\), their difference has variance \(2\sigma_{wR}^2\). Thus a sample standard deviation \(s_d\) of appropriately adjusted paired reference differences estimates the within-subject log-scale standard deviation as \(s_{wR}=s_d/\sqrt{2}\). In a full analysis, estimation should account for period effects and the specified mixed model rather than blindly applying this formula to unadjusted observations.

From the Treatment Contrast to the Decision

Let \(\hat{\theta}\) be the estimated treatment contrast on the log scale, with standard error \(SE\) and degrees of freedom \(df\). The estimated geometric mean ratio is \(\exp(\hat{\theta})\). A two-sided 90% confidence interval on the log scale is

Exponentiating both endpoints gives the 90% confidence interval for the geometric mean ratio. The limits on the log scale are \(\ln(0.90)\) and \(\ln(1/0.90)\). The decision is based on the entire interval: both endpoints must be at least 0.90 and at most \(1/0.90\). This is equivalent to two one-sided tests at significance level \(\alpha=0.05\): the null represents a ratio at or beyond either unacceptable boundary, and both one-sided null hypotheses must be rejected.

Worked Analyses

Deriving the Interval Decision

The fixed-limit rule follows directly from the two one-sided tests. Write the log-scale lower and upper boundaries as \(L=\ln(0.90)\) and \(U=\ln(1/0.90)\). The lower-bound test evaluates \(H_0:\theta\leq L\) against \(H_a:\theta>L\); the upper-bound test evaluates \(H_0:\theta\geq U\) against \(H_a:\theta<U\). At \(\alpha=0.05\), the lower test rejects when \(\hat{\theta}-t_{0.95,df}SE>L\), and the upper test rejects when \(\hat{\theta}+t_{0.95,df}SE<U\). Both inequalities must hold. Taken together, they say exactly that the 90% log-scale confidence interval lies wholly between \(L\) and \(U\). Exponentiation preserves order, so the ratio-scale interval must lie wholly between 0.90 and \(1/0.90\).

Diagnostics, Alternatives, and Limitations

Before interpreting the interval, inspect whether the analysis is credible. Check treatment and period coding, sequence allocation, duplicate reference records, missing observations, and the log transformation. Examine period and sequence patterns and residuals for unusual skew, outliers, or changing variability. A strong period pattern may reflect operational or biological changes and can undermine a simple model; it should not be dismissed solely because a treatment comparison is available. Missing occasions or imbalance may require a model that handles the observed replicate structure correctly. Any sensitivity analysis should be prespecified or clearly identified as exploratory.

The main alternative is a conventional nonreplicate crossover analysis, which estimates the treatment contrast but cannot estimate within-subject reference variability from repeated reference dosing. Another alternative, reference-scaled bioequivalence, uses reference variability in a different decision rule and may involve scaled limits and additional constraints. It is not interchangeable with the fixed tighter NTI procedure illustrated here. The applicable regulatory framework determines which approach is appropriate.

The confidence interval also has limitations. Its coverage relies on the model, design, and variance estimates being adequate; it does not prove identical exposure or guarantee identical clinical outcomes. The estimated within-subject CV is itself uncertain, especially in a small study. Finally, passing one endpoint does not imply that all required pharmacokinetic endpoints pass, and a result from one product, strength, or study does not automatically establish bioequivalence in another setting.

Common Mistake

A frequent error is to argue, “The ratio is close to 100%, so the products are bioequivalent.” Bioequivalence is not established by the point estimate alone: the full confidence interval must fit within both limits. A second error is to use the estimated reference variability to widen the fixed NTI limits. In the procedure used here, variability affects estimation and precision, but the limits remain 90.00% and 111.11%. Use the exact upper limit \(1/0.90=1.111111\ldots\) in calculations and round only for reporting; if the limit is instead explicitly defined as 1.1111, its logarithm is approximately 0.10535.