The Bioequivalence Crossover Estimand

What Does a Crossover Study Aim to Compare?

In a bioequivalence crossover, the same participant receives both the test formulation and the reference formulation. The study therefore provides a response under each formulation for each participant, rather than responses from two wholly separate groups. The central target is the population treatment contrast those paired responses represent. Defining that target precisely clarifies what a study estimate means and what assumptions are needed to interpret it as a formulation effect.

For a positive pharmacokinetic endpoint such as AUC or Cmax, bioequivalence comparisons are commonly defined on the natural-log scale. The resulting treatment contrast is transformed back to the original scale as a ratio of geometric means. This is the log-scale counterpart of the ratio-scale criteria introduced in the earlier material on regulatory acceptance limits; here the focus is the quantity a crossover is designed to estimate.

Definition: Let \(Y_i(T)\) and \(Y_i(R)\) denote participant \(i\)'s positive endpoint under the test and reference formulations, respectively, under the administration conditions and target population specified for the study. Define \(Z_i(T)=\ln Y_i(T)\) and \(Z_i(R)=\ln Y_i(R)\). The participant-level log contrast is \(D_i=Z_i(T)-Z_i(R)\). The population crossover estimand is \(\Delta=E[D_i]\), where the expectation is over the target population. Its ratio-scale representation is \(GMR=\exp(\Delta)\), the ratio of the population geometric mean endpoint under test to that under reference.

The estimand is a property of the target population and the specified treatment conditions, not a number that depends on which participants happened to enroll. The observed sample supplies an estimate of it. In a complete crossover, each participant contributes both formulation responses, so the observed log contrast is formed within participant. The sample average of these paired contrasts estimates the population average when the sample represents the target population and the design and analysis support the stated treatment comparison.

This definition is endpoint-specific: an estimand for AUC is not automatically an estimand for Cmax. It is also condition-specific. Formulation, dose, administration conditions, endpoint definition, and target population all need to be specified sufficiently to say what responses \(Y_i(T)\) and \(Y_i(R)\) mean. The estimand does not by itself determine a regulatory decision; regulatory acceptance criteria are applied to an estimate and its uncertainty for the relevant endpoint.

Why the Log-Scale Contrast Becomes a Ratio

Taking logs turns a ratio of positive outcomes into a difference. The population average difference on the log scale can therefore be expressed as a ratio of geometric means on the original scale. This connection is why the log-scale contrast is useful for endpoints whose bioequivalence comparison is naturally expressed as a ratio.

\begin{aligned} \Delta &=E[\ln Y_i(T)-\ln Y_i(R)]\\ &=E[\ln Y_i(T)]-E[\ln Y_i(R)],\\ \exp(\Delta) &=\frac{\exp(E[\ln Y_i(T)])}{\exp(E[\ln Y_i(R)])} =\frac{GM_T}{GM_R}. \end{aligned}

Here \(GM_T=\exp(E[\ln Y_i(T)])\) and \(GM_R=\exp(E[\ln Y_i(R)])\) are the population geometric means. The equalities use the log rule \(\ln a-\ln b=\ln(a/b)\) for positive \(a\) and \(b\), and the fact that exponentiation converts a difference into a ratio. Thus, if \(\Delta=0\), the ratio of geometric means is 1; if \(\Delta>0\), it exceeds 1; and if \(\Delta<0\), it is below 1. These statements describe the direction of the population contrast, not whether a particular regulatory criterion has been met.

Worked Example: From Paired AUC Ratios to a Geometric Mean Ratio

Suppose four hypothetical participants complete a crossover, and their individual test-to-reference AUC ratios are 1.10, 0.90, 1.20, and 0.80. Assume the endpoint is positive and the paired responses are measured under comparable conditions, with no residual influence from an earlier administration. Calculate the sample mean log contrast and its ratio-scale representation.

The participant-level log contrasts are the natural logs of the listed ratios. Thus the sample mean log contrast is

\bar{D} =\frac{\ln(1.10)+\ln(0.90)+\ln(1.20)+\ln(0.80)}{4} =\frac{0.09531-0.10536+0.18232-0.22314}{4} =-0.01272.

Exponentiating gives the sample geometric mean ratio:

\exp(\bar{D})=\exp(-0.01272)=0.9874.

For these four hypothetical participants, the geometric mean of the paired test-to-reference AUC ratios is about 0.9874. It is slightly below 1, corresponding to a geometric mean test AUC about 1.26% lower than the reference AUC. This is a descriptive sample result; inference about a target population requires accounting for sampling uncertainty and the study design.

How Each Participant Contributes

The paired construction is central to the estimand. For participant \(i\), subtracting the reference log response from the test log response gives \(D_i\). The population target averages these participant-level contrasts, giving each participant equal weight in the expectation unless a different target population or weighting scheme is explicitly defined. This is not the same operation as first combining all test responses and reference responses on the original scale and then taking their arithmetic means.

In a realized study, both responses are observed at different occasions, not simultaneously. The estimand uses the potential-response idea to define the comparison: it asks what the same target participant’s endpoint would be under each formulation in the specified study conditions. The crossover obtains observations under both treatments from that participant, but the interpretation as a formulation contrast depends on the conditions being sufficiently comparable. A response affected by the prior treatment, a systematic difference between occasions that is not addressed by the design or analysis, or treatment effects that vary in ways the target does not represent can undermine that interpretation.

The following distinction is useful:

Estimand: the defined population quantity, such as \(\Delta\) or \(\exp(\Delta)\).
Estimate: a value calculated from the observed participants and their data.
Inference: a statement about the uncertainty in that estimate and the population quantity, using a method appropriate to the design.

A crossover is valuable in part because each participant serves as a source of information under both formulations. Stable differences among participants, such as a tendency for one person to have higher exposure than another under either formulation, can cancel in the within-participant contrast. This does not guarantee an unbiased or precise estimate: it depends on the design assumptions, the variability of the paired differences, and the completeness and quality of the observed responses.

Worked Example: Why the Target Is Not an Arithmetic Mean Ratio

Consider two hypothetical participants with reference AUC values of 10 and 100 mg·h/L, and test AUC values of 20 and 100 mg·h/L. Their paired test-to-reference ratios are 2 and 1.

The mean of the individual ratios is \((2+1)/2=1.50\). But the mean log contrast is

\bar{D}=\frac{\ln(2)+\ln(1)}{2} =\frac{0.6931+0}{2}=0.3466, \qquad \exp(\bar{D})=\exp(0.3466)=1.414.

The ratio of the arithmetic mean AUCs is a third quantity:

\frac{(20+100)/2}{(10+100)/2} =\frac{60}{55}=1.091.

Thus the arithmetic mean of individual ratios is 1.50, the geometric mean of those ratios is 1.414, and the ratio of arithmetic mean AUCs is 1.091. The crossover estimand defined here is the mean log contrast, represented on the ratio scale by 1.414. The values differ because these summaries weight and combine the participant responses differently; they are not interchangeable.

Assumptions and Conditions for Interpretation

The estimand can be defined without claiming that every assumption is satisfied. To interpret a crossover estimate as an estimate of the defined formulation contrast, however, the study and analysis need to support that interpretation. In particular:

Well-defined treatments and endpoint. Test and reference must refer to specified formulations and administration conditions, and the endpoint must be defined consistently. The log contrast requires positive endpoint values.
Comparable response conditions. The observed response under each formulation must represent the corresponding potential response under the target conditions. Residual effects of an earlier administration or unaddressed systematic occasion differences can violate this condition.
Appropriate population and sampling. The participants and analysis must support inference to the stated target population. A study of a restricted population does not automatically identify an effect in a broader one.
Usable paired information. The analysis must account for incomplete pairs. If missingness is related to treatment response, simply averaging the available pairs can misrepresent the target.
Suitable treatment-effect summary. The mean log contrast is a population average. If participant-level contrasts vary, that average does not imply that every participant has the same response or that all individual ratios are close to the population ratio.

The exact design and statistical model determine how observed period-specific outcomes are used to estimate the target. A raw average of paired differences is directly interpretable under conditions where the relevant occasion contributions do not distort the treatment contrast. If the design or analysis must account for systematic period differences, the estimator should reflect that. These considerations are developed in detail in the subsequent material on crossover comparisons and period effects.

Worked Example: Interpreting a Sample Mean Log Contrast

In a hypothetical study, three participants have test and reference AUC values, in mg·h/L, as follows: participant 1 has 12 and 10; participant 2 has 18 and 20; participant 3 has 30 and 25. Assume all pairs are complete, the values are positive, and the administration conditions allow the paired differences to represent the stated treatment contrast. Calculate the sample log contrast and interpret it.

The paired ratios are \(12/10=1.20\), \(18/20=0.90\), and \(30/25=1.20\). Therefore,

\bar{D} =\frac{\ln(1.20)+\ln(0.90)+\ln(1.20)}{3} =\frac{0.18232-0.10536+0.18232}{3} =0.08643.

Transforming back to the ratio scale gives

\exp(\bar{D})=\exp(0.08643)=1.090.

The sample geometric mean test-to-reference AUC ratio is approximately 1.090. In these hypothetical data, that is about 9.0% higher on the geometric mean scale. The calculation does not establish that the population ratio is 1.090, nor does it establish bioequivalence; it is a sample estimate whose uncertainty and regulatory interpretation require further analysis.

Diagnostics, Alternatives, and Limitations

The estimand itself is not a diagnostic test, but several data and design checks can reveal when a planned estimate may not support it. Confirm that each participant’s test and reference records are correctly matched and that the endpoint, units, and log transformation are consistent. Review missing responses and the reasons for missingness. Examine whether observed results show substantial differences across occasions or patterns suggesting that the response under one formulation depends on what preceded it. Such patterns require design-aware assessment; a simple paired calculation cannot repair a mismatch between the data and the target.

A parallel-group comparison is a main alternative: different participants receive test and reference. It can estimate a formulation contrast, but it does not obtain both responses from each participant. A crossover can use within-participant information and reduce the impact of stable between-participant differences, while requiring that the conditions for a meaningful within-participant comparison hold. Neither design is universally preferable; the choice depends on the treatments, endpoint, population, and feasibility of controlling occasion-related influences.

A common mistake is to say, “Because every participant received both formulations, the observed paired difference is automatically the causal effect for that participant, and averaging ratios is the bioequivalence estimand.” The correction has two parts. First, the population estimand is defined as the mean of individual log contrasts, with \(\exp(\Delta)\) giving the geometric mean ratio; it is not generally the arithmetic mean of ratios. Second, observing both formulations at different occasions does not by itself ensure comparable conditions. The interpretation depends on the study design, data, and analysis supporting the paired treatment contrast.

Finally, the population mean contrast can conceal heterogeneity. A geometric mean ratio near 1 does not mean that every participant’s test and reference responses are nearly equal. Nor does a correctly defined estimand guarantee precision: small samples, highly variable paired differences, or incomplete data can leave substantial uncertainty. The estimand specifies what is being targeted; it does not replace evaluation of uncertainty or the applicable acceptance criteria.

Key takeaway: In a two-formulation crossover, the basic bioequivalence target is the population mean of the participant-level test-minus-reference log contrasts. Exponentiating that target gives the ratio of population geometric means. Each participant informs the comparison through both formulation responses, provided the design and analysis make those responses comparable for the intended interpretation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A population mean log contrast is \(\Delta=0.04\). What is the corresponding test-to-reference geometric mean ratio, approximately?
Two participants have paired test-to-reference ratios of 0.80 and 1.25. Calculate the mean log contrast and its ratio-scale representation.
Why is the mean of individual ratios not generally the crossover estimand defined in this tutorial?
Name two conditions that can undermine interpreting observed paired responses as the intended formulation contrast.

Solutions. (1) \(\exp(0.04)=1.041\), so the corresponding geometric mean ratio is about 1.041. (2) The mean log contrast is \([\ln(0.80)+\ln(1.25)]/2=[-0.22314+0.22314]/2=0\); exponentiating gives \(\exp(0)=1.00\). The ratios are reciprocals, so their geometric mean is 1. (3) The estimand averages log contrasts and then exponentiates; this yields a geometric mean ratio, whereas an arithmetic mean of ratios uses a different operation and can give a different value. (4) Examples include residual effects from an earlier administration, unaddressed systematic occasion differences, incomplete pairs handled without accounting for relevant missingness, or participants who do not represent the target population.