Modeling Individual Bioequivalence And Subject-by-Formulation Effects

Why Individual Bioequivalence Needs More Than a Mean Contrast

A formulation comparison can have a small average exposure difference while some individuals consistently show a larger difference than others. It can also have similar average exposure but greater within-person fluctuation under the test formulation. A model intended to describe individual bioequivalence therefore needs to distinguish these patterns, rather than summarizing the comparison only by a population mean contrast.

The key distinction is between subject-by-formulation interaction—variation between individuals in their underlying formulation response—and within-subject variability—variation between repeated administrations of the same formulation to the same individual. These components have different interpretations and require repeated observations to separate reliably.

A Model for Subject-Specific Formulation Responses

Let \(Y_{ijk}\) be a log-transformed exposure outcome, such as log AUC, for subject \(i\), formulation \(j\), and replicate administration \(k\). Use \(x_j=0\) for reference and \(x_j=1\) for test. A compact model is

Here, \(\mu\) is the population mean log exposure under reference, and \(\tau\) is the population mean test-minus-reference contrast on the log scale. The subject effect \(b_i\) accounts for persistent differences in exposure between people. The interaction \(u_i\) is subject \(i\)’s deviation from the population formulation contrast. The residual \(\varepsilon_{ijk}\) represents within-subject variation not explained by those terms.

A common working assumption is that \(b_i\) and \(u_i\) are independent, mean-zero normal random effects, with variances \(\sigma_S^2\) and \(\sigma_D^2\), respectively. Conditional residuals are assumed independent across replicate administrations, independent of the random effects, and normally distributed with formulation-specific variances:

The residual variances \(\sigma_{WR}^2\) and \(\sigma_{WT}^2\) describe within-subject variability under reference and test, respectively. In this model, the underlying subject-specific log contrast is \(\tau+u_i\). Its population mean is \(\tau\), and its between-subject variance is \(\sigma_D^2\). The population geometric mean ratio associated with the log contrast is \(\exp(\tau)\).

These assumptions are a working model, not facts guaranteed by randomization. The design must include repeated administrations of the formulations for residual variation to be separated from a persistent subject-by-formulation response. Systematic design effects, such as period or sequence effects, must also be represented appropriately when relevant; omitting them can distort the estimated contrast or variance components. The model further assumes that the log-scale outcome and the chosen random-effects structure adequately describe the data.

What the Interaction and Residual Variances Contribute

The interaction is not simply another name for residual error. If a subject repeatedly shows a larger test-versus-reference contrast than other subjects, that pattern contributes to \(\sigma_D^2\). If repeated administrations of a formulation fluctuate within the same subject, that contributes to the relevant within-subject variance. Distinguishing them matters because the first describes heterogeneity of individual responses, while the second describes repeatability of exposure under a formulation.

In a replicate design, observations can be grouped by subject and formulation. Repeated values within a subject-formulation group provide information about within-subject variation. Differences between the subject-level test and reference responses provide information about subject-by-formulation heterogeneity, after within-subject noise is accounted for. With only one observation per subject-formulation combination, a large observed spread in subject contrasts can reflect either interaction or residual variability. Without additional assumptions or information, those sources cannot generally be disentangled.

From Components to an Individual-Bioequivalence Criterion

One model-based individual-bioequivalence criterion combines the squared population log contrast, interaction variance, and the difference in within-subject variances. Define

The terms are on the squared log-exposure scale. The squared contrast \(\tau^2\) penalizes a population shift in either direction. The interaction variance penalizes variation between individuals in their underlying formulation response. The difference in residual variances compares within-subject variability under test with that under reference: greater test variability increases the criterion, while lower test variability decreases it.

A decision rule requires a prespecified limit, say \(\Theta_{\mathrm{lim}}\), defined on the same scale and under the applicable procedure. A point estimate below that limit is not, by itself, an inferential demonstration of individual bioequivalence: uncertainty in the components must be addressed using the specified testing or confidence-bound procedure. Some formulations of the criterion scale or standardize terms differently, so the exact criterion and its limit must be stated together. Do not compare an unscaled estimate to a limit defined for a scaled criterion.

A Short Derivation of the Individual Contrast Component

The subject-specific underlying log contrast is \(\tau+u_i\). Since \(u_i\) has mean zero and variance \(\sigma_D^2\), its mean squared value across subjects is

This shows why both a mean shift and between-subject heterogeneity contribute to an individual-focused summary. The criterion adds the test-minus-reference difference in within-subject variances, \(\sigma_{WT}^2-\sigma_{WR}^2\), to compare test repeatability with the reference benchmark. That subtraction is part of the defined criterion; it is not the variance of a test-minus-reference observation. Consequently, \(\Theta_I\) need not behave like a variance and can be negative when the test’s within-subject variance is sufficiently lower than the reference’s.

How This Differs From Other Bioequivalence Summaries

Average bioequivalence focuses on the population mean formulation contrast. On the log scale, that contrast is \(\tau\), often expressed as the geometric mean ratio \(\exp(\tau)\). It does not, by itself, describe whether individuals differ in their underlying response or whether repeated test administrations are more variable.

Population bioequivalence, covered in the earlier material on predicting population bioequivalence from exposure distributions, asks about population exposure distributions rather than only their mean. The individual criterion considered here specifically includes a subject-by-formulation component and compares within-subject variability against reference. These approaches answer related but distinct questions; one criterion cannot be substituted for another without matching the estimand, scale, and decision rule.

The individual-focused model is useful when the scientific or regulatory question concerns whether formulation substitution preserves exposure patterns within individuals, not merely whether average exposures are close. Its usefulness depends on having a design and data that support the variance-component distinctions.

Diagnostics, Assumptions, and Limitations

A plausible fit should be checked against the structure of the study and the behavior of the data. Inspect residuals by formulation, subject, and relevant design factors for systematic patterns, extreme observations, changing spread, or departures from the assumed log-scale distribution. Compare observed repeated values within subjects with model-implied within-subject variation. Check whether the fitted interaction variance is supported by variation across subject-specific formulation responses rather than being driven by a few subjects.

Warning signs include a variance estimate at or near zero when the design suggests substantial variation, unstable estimates when a small number of subjects are removed, markedly different residual spreads from those assumed, and strong residual correlation that the model omits. Sparse replication can make the components weakly identified and estimates imprecise. A non-convergent fit or a covariance estimate at a boundary should prompt investigation of model specification, design support, and data quality; it should not be treated as evidence that the component is exactly zero.

• Do not conflate interaction with repeatability. Between-subject differences in formulation response and repeat-to-repeat noise are different sources of variation.
• Do not infer individual equivalence from the mean ratio alone. A ratio near one says little by itself about \(\sigma_D^2\) or the within-subject variance contrast.
• Do not treat a favorable point estimate as a decision. A prespecified inferential procedure must account for uncertainty in estimated components.
• Do not overlook design effects. Period, sequence, carryover, or other systematic features can contaminate formulation comparisons if not handled by an appropriate design and model.

The model also has limits beyond sampling uncertainty. Its conclusions depend on the selected outcome, the log transformation, the random-effects distribution, and the representativeness of the study population. A single interaction variance summarizes heterogeneity but does not identify which clinical characteristics explain it. Further, a variance-component model does not by itself determine the appropriate regulatory limit or establish that an exposure difference has a particular clinical consequence.