Crossover Bioequivalence Analysis

Why Crossover Bioequivalence Uses a Log-Scale Contrast

A crossover study gives each participant both formulations, in different periods and in randomized order. Because the same person contributes measurements under test and reference, the comparison can account for stable differences between people in exposure. The analysis must also account for the period in which each formulation was given and for the randomized sequence.

Exposure measures such as AUC and Cmax are positive and are often interpreted proportionally: a change from 10 to 12 has the same ratio as a change from 20 to 24. Taking natural logarithms turns ratios into differences, which fit an additive statistical model. Exponentiating the estimated log-scale contrast then returns the comparison to the ratio scale used to interpret bioequivalence.

For example, \(\theta=0\) corresponds to a ratio of 1, or 100%; \(\theta>0\) corresponds to a test geometric mean greater than the reference geometric mean. A log-scale difference is not an absolute exposure difference in units such as mg·h/L. It represents a multiplicative comparison after exponentiation.

Design and Model

Consider a randomized two-sequence, two-period design. One sequence receives test in period 1 and reference in period 2; the other receives reference in period 1 and test in period 2. Let \(Y_{ijk}\) be the natural logarithm of the measured exposure for subject \(i\), in sequence \(j\), during period \(k\). A standard additive model is:

Here, \(\mu\) is an overall mean; \(S_j\) is the sequence effect; \(B_{i(j)}\) is the subject effect for subject \(i\) nested within sequence \(j\); \(P_k\) is the period effect; \(F_{ijk}\) is the formulation effect for the formulation given in that period; and \(\varepsilon_{ijk}\) is residual error. With reference-coded formulation, the test-versus-reference effect is \(\theta\). The model treats sequence, period, and formulation as systematic effects, and subject within sequence as a subject-level effect.

The usual analysis relies on independent subjects, appropriate randomization, adequate washout so that a prior dose does not materially affect the next period, and a correctly specified additive model. For the conventional small-sample t-based analysis, log-scale errors are assumed approximately normal with a common variance. The treatment effect is assumed not to change across periods or sequences. Missing observations must not create a biased comparison; the simple paired-difference calculation below requires a complete pair for each included subject.

The sequence effect is not the treatment effect: sequence labels describe treatment order. The subject effect accounts for stable between-person differences. Period accounts for a systematic shift from the first to the second study period. The formulation contrast is the target of interest.

How the Period-Adjusted Contrast Is Obtained

The key calculation can be seen by forming each subject’s test-minus-reference difference on the log scale. Write \(d_i=Y_{i,T}-Y_{i,R}\), regardless of which period contained test. Let the period-2 mean shift relative to period 1 be \(\pi\), and let the true formulation contrast be \(\theta\). In the test-then-reference sequence, test occurs first, so the expected paired difference is \(\theta-\pi\). In the reference-then-test sequence, test occurs second, so the expected difference is \(\theta+\pi\). Thus, the equally weighted average of the two sequence-specific mean differences removes the period shift:

For a balanced study with complete pairs, the treatment estimate is therefore the average of the two sequence-specific mean paired differences. If both sequences have the same variance of paired differences, a pooled estimate \(s_d^2\) of that variance gives:

Here \(n_1\) and \(n_2\) are the numbers of complete subjects in the two sequences, and \(\bar d_1,\bar d_2\) are their respective mean test-minus-reference log differences. For equal sequence sizes \(n_1=n_2=m\), this simplifies to \(SE(\hat{\theta})=s_d/\sqrt{2m}\). The pooled variance is calculated from the within-sequence variation in the paired differences, with \(df=n_1+n_2-2\) under the equal-variance t analysis. This calculation makes explicit how the crossover removes stable subject differences and balances the period effect.

Confidence Intervals and the Bioequivalence Decision

Bioequivalence is assessed against prespecified ratio limits, not by asking whether a conventional test finds a difference from a ratio of 1. A commonly used illustrative interval is 0.80 to 1.25, although the applicable limits and any endpoint-specific requirements depend on the study context and governing guidance. On the log scale, these limits are \(\log(0.80)\) and \(\log(1.25)\).

A two-sided 90% confidence interval for \(\theta\) corresponds to two one-sided tests, each at significance level 0.05, for the usual symmetric t-based procedure. Bioequivalence is demonstrated for an endpoint only when both confidence limits, after exponentiation, lie within the prespecified ratio limits. A confidence interval that overlaps a limit does not demonstrate bioequivalence, even if its point estimate is close to 1. This is a failure to establish equivalence under the chosen criterion, not proof that the formulations are different.

Diagnostics, Alternatives, and Limitations

A valid interval depends on the design and model, not just the arithmetic. Review the sequence allocation, period-specific records, formulation administration, washout, adherence, and completeness of exposure measurements. Plot log-scale residuals or paired differences to look for outliers, skewness, unequal variation between sequences, and values that may indicate data or assay problems. A strong period pattern or sequence imbalance deserves investigation; it should not be hidden by reporting only the ratio interval.

Carryover is especially important to handle carefully. If exposure in period 2 is materially affected by the period-1 formulation, the standard crossover contrast may be biased. In a conventional two-sequence, two-period design, carryover effects are difficult to distinguish reliably from sequence-related effects. A non-significant carryover test does not prove that carryover is absent. The primary safeguards are suitable washout, sound study conduct, and a design appropriate to the drug’s disposition.

The paired-difference formula is transparent and efficient for complete, balanced data under its assumptions. A mixed-effects model can be more flexible for unbalanced data or some missing observations, but it does not automatically remove bias from informative missingness or a misspecified model. A parallel-group comparison avoids within-subject carryover between periods but usually loses the precision advantage of comparing each participant with themself. Neither alternative rescues a poor design or unreliable measurements.

The t interval can be sensitive to severe departures from the assumed error structure, particularly with very small samples. Inspect results for influential subjects and confirm that the analysis population and exclusions follow the prespecified plan. Analyze each required endpoint on its log scale, preserve the original units when describing the exposure measure, and report the ratio and interval with clear units or percentage interpretation. A ratio is dimensionless; the underlying endpoint, such as AUC in concentration·time units or Cmax in concentration units, should still be named.

A common mistake is to conclude bioequivalence because the estimated ratio is near 1, or because a test for difference is not statistically significant. Neither is the decision rule. The point estimate can be central while its interval is too wide, and failure to reject a difference does not establish that the true ratio lies within acceptable bounds. The correction is to calculate the prespecified confidence interval and check that both limits are inside the permitted range.