Why Diagnose Formulation-by-Subject Effects?
A population-average comparison can summarize whether the test formulation tends to produce higher or lower exposure than the reference. It does not, by itself, show whether that average difference is similar across subjects. In an NTI setting, this distinction matters: a small average difference may coexist with substantial subject-to-subject variation in the test-versus-reference response. Replicate observations can help determine whether the apparent variation reflects genuine formulation-by-subject heterogeneity or ordinary within-subject measurement and occasion variability.
The target here is a subject-specific contrast in a pharmacokinetic endpoint, such as log AUC or log Cmax. The analysis is not a replacement for the population-average bioequivalence assessment. It is a diagnostic of whether subjects differ in their underlying formulation responses, and of how much confidence to place in that interpretation given the data and model.
A Replicate Crossover Model
Let \(Y_{ip}\) be the log-transformed endpoint for subject \(i\) in period \(p\), and let \(x_{ip}=+1/2\) when the subject receives test and \(x_{ip}=-1/2\) when the subject receives reference. This coding makes the difference between test and reference equal to one unit of \(x\). A model for a complete replicate crossover design is
Here, \(\mu\) is an overall mean, \(\delta\) is the population-average test-minus-reference difference on the log scale, \(\pi_p\) is a period effect, and \(q_{r(i)}\) is the effect of the randomized sequence assigned to subject \(i\). The subject random intercept \(u_i\) represents a subject's general exposure level. The random treatment slope \(b_i\) represents that subject's deviation from the population-average formulation difference. Thus subject \(i\)'s underlying log test-minus-reference contrast is \(\delta+b_i\). The residual \(\epsilon_{ip}\) is the remaining within-subject variation for that observation.
A standard model assumes that subjects are independent, that \((u_i,b_i)\) follows a mean-zero bivariate normal distribution with an estimable covariance matrix, and that residuals are conditionally independent of those subject effects and of one another. Residual variances may differ by formulation: write them as \(\sigma_T^2\) and \(\sigma_R^2\). The model also requires an adequate design and data: formulation and period effects must be estimable, the randomization and period structure must be represented correctly, and carryover must be negligible under the study's washout and design assumptions. The log transformation and normal random-effects assumptions should be plausible enough for the intended variance interpretation.
Replicate administrations of each formulation are important because they provide information about the residual variation within each formulation, separately from variation across subjects' formulation contrasts. A conventional two-period crossover, with only one observation per formulation per subject, generally cannot make that separation reliably without stronger assumptions or external information.
How Replication Separates the Variance Components
Consider a subject whose period-adjusted log observations are averaged separately across replicate test and reference administrations. If there are \(m_T\) test observations and \(m_R\) reference observations, the difference of those averages can be expressed as
If residuals are independent across occasions and have formulation-specific variances, then averaging \(m_T\) test observations gives residual variance \(\sigma_T^2/m_T\), and averaging \(m_R\) reference observations gives residual variance \(\sigma_R^2/m_R\). The two averages are independent under the model. Therefore, across subjects,
This identity is the central diagnostic idea. The observed spread of subject contrasts combines true subject-to-subject formulation heterogeneity with residual noise in each contrast. Replicate observations estimate the residual contributions; the random-slope model uses that information to estimate \(\tau^2\). The subtraction can be unstable in a small study, so the variance component is usually estimated within the full model rather than treated as a precise standalone calculation.
A useful simple estimator illustrates the logic. With complete balanced replication and equal residual variance \(\sigma^2\), if each formulation is observed twice, the residual contribution to the contrast variance is \(\sigma^2/2+\sigma^2/2=\sigma^2\). A rough method-of-moments estimate is then the observed sample variance of the adjusted subject contrasts minus \(\sigma^2\), truncated at zero if that difference is negative. It is an illustration, not a substitute for checking model assumptions or uncertainty.
Worked Example: Separating Contrast Spread From Residual Noise
Suppose a hypothetical complete replicate study gives four period-adjusted subject contrasts \(d_i\), in log units: \(-0.04,\ 0.00,\ 0.04,\ 0.16\). Each contrast is the mean log test exposure minus the mean log reference exposure. Assume two observations per formulation and an estimated common within-formulation residual variance of \(0.0020\) log-units squared.
The mean contrast is \(( -0.04+0.00+0.04+0.16 )/4=0.040\). The sample variance is
With two observations per formulation, the residual contribution to each contrast variance is \(0.0020/2+0.0020/2=0.0020\). The rough variance-component estimate is therefore
This calculation suggests that the contrast spread is larger than the estimated residual contribution. It does not establish a precise or clinically important interaction: four subjects provide little information about a variance, and the result depends on the period adjustment and residual variance estimate. A fitted mixed-effects model and uncertainty assessment are needed before drawing a substantive conclusion.
Diagnostics and Interpretation
A variance estimate is not self-validating. Begin with the study structure and the endpoint used in the model. Check that each subject's formulation and period records agree with the randomized design, that replicate measurements are assigned to the correct subject and formulation, and that the endpoint was transformed and defined consistently. Period patterns or an implausible sequence pattern can signal design or data problems; a random slope cannot repair a mis-specified period structure or unaddressed carryover.
Then examine model-based diagnostics. Plot conditional residuals against fitted values, formulation, and period to look for changing spread, systematic patterns, or unusual observations. Compare within-formulation replicate differences across subjects and inspect subject-level adjusted contrasts. A few extreme contrasts may dominate the estimated random-slope variance, so investigate influential subjects and verify their data rather than deleting them solely because they are influential. A normal quantile plot of estimated random effects can flag severe departures from a normal random-slope distribution, but empirical random effects are shrunken estimates; the plot is a diagnostic, not a definitive distribution test.
Check whether the model estimates are stable. A random-slope variance estimated at or very near zero may indicate little information for detecting heterogeneity, not proof that every subject responds identically. Convergence warnings, singular covariance estimates, or large changes after small model adjustments are reasons to qualify the estimate. Likelihood-ratio testing of a zero variance component also needs care: zero is on the boundary of the allowable parameter space, so a routine chi-square reference distribution is not generally justified. Profile-likelihood or bootstrap-based uncertainty methods can be more appropriate, depending on the design and available software.
Worked Example: Translating a Fitted Heterogeneity Estimate
Consider a separate hypothetical model fit for log AUC. Suppose it estimates a population-average log difference \(\hat{\delta}=0.020\), a subject-specific formulation variance \(\hat{\tau^2}=0.0064\), and residual variances \(\hat{\sigma_T^2}=\hat{\sigma_R^2}=0.0016\). Assume the model diagnostics do not reveal a major period pattern, strong residual trend, or influential data error.
The population-average test-to-reference geometric mean ratio is \(\exp(0.020)=1.020\), or about 1.020. The estimated standard deviation of the subject-specific log differences is \(\hat{\tau}=\sqrt{0.0064}=0.080\). A one-standard-deviation shift in the underlying individual log contrast corresponds to a multiplicative factor of \(\exp(0.080)=1.083\) relative to the population-average ratio.
Under the fitted normal random-slope model, an approximate central 95% range for the underlying subject-specific log contrast is \(\hat{\delta}\pm1.96\hat{\tau}\), or \(0.020\pm0.1568\), giving \([-0.1368,\ 0.1768]\). Exponentiating gives a range of about \(0.872\) to \(1.193\) for the model-implied subject-specific ratios. This is a model-based description of the distribution of underlying subject effects, not a confidence interval for the population mean and not a regulatory acceptance interval. Its breadth suggests that the modest population-average difference alone does not characterize every subject's response.
The calculation also depends on the random-effects distribution and the fitted variance estimate. In a small study, the range may be highly uncertain. It should be reported with appropriate uncertainty and diagnostics, not treated as an exact prediction for any named subject.
Different Variability Questions
Formulation-by-subject heterogeneity is not the same as a difference between test and reference within-subject variances. A test-versus-reference variability comparison asks whether the residual spread among replicate observations differs by formulation, as reflected in \(\sigma_T^2\) and \(\sigma_R^2\). The random-slope variance \(\tau^2\) asks whether subjects differ in their underlying test-minus-reference response. These components can behave differently: residual variances can be similar while subject-specific contrasts vary, or residual variances can differ without clear evidence of a random treatment slope.
A population-average analysis answers another question: it estimates the average formulation difference and its uncertainty under the chosen design and analysis. A small average difference does not imply small \(\tau^2\). Conversely, evidence of a random slope does not by itself show that the population-average comparison fails its prespecified criterion. The population-average estimand and the distribution of individual effects should be described separately.
The main alternatives are therefore complementary rather than interchangeable. A population-average model without a random treatment slope is simpler and may be adequate for its estimand, but it cannot quantify subject-specific treatment-effect variance. A paired analysis of individual contrasts is more transparent, but without replication it mixes true heterogeneity with residual noise. A replicate mixed-effects model can separate these sources under its assumptions, though it is more dependent on design adequacy, model specification, and sufficient information.
Worked Example: Recognizing Heterogeneity in Replicate Patterns
Suppose a hypothetical study has two subjects, each with two log AUC observations on each formulation. Subject A has reference values \(-0.02,\ 0.00\) and test values \(0.08,\ 0.10\). Subject B has reference values \(0.00,\ 0.02\) and test values \(-0.08,\ -0.10\).
For subject A, the reference mean is \(-0.01\), the test mean is \(0.09\), and the contrast is \(d_A=0.09-(-0.01)=0.10\). For subject B, the means are \(0.01\) and \(-0.09\), so \(d_B=-0.09-0.01=-0.10\). The contrasts point in opposite directions even though the within-formulation replicate values are close together.
Within each of the four subject-by-formulation cells, the two observations differ from their cell mean by \(0.01\) in each direction. The pooled residual variance is \(0.0008/4=0.0002\): there are eight squared deviations of \(0.01^2\), and four residual degrees of freedom. The sample variance of the two contrasts is \(\{(0.10-0)^2+(-0.10-0)^2\}/(2-1)=0.020\). For two replicates per formulation, the residual contribution to contrast variance is \(0.0002/2+0.0002/2=0.0002\), giving a rough \(\hat{\tau^2}\) of \(0.0198\).
This pattern is consistent with marked subject-to-subject contrast variation relative to within-formulation noise. However, two subjects cannot support a dependable population variance estimate. The example shows what replication can reveal, not a sufficient sample size or a basis for a clinical conclusion.
Common Pitfalls and Limitations
- Calling all contrast variability an interaction. The observed contrast variance includes residual contributions from both formulations. Use replication and the fitted variance structure to separate them.
- Equating a small average difference with uniform subject response. The population mean and random-slope variance describe different features of the response distribution.
- Interpreting a near-zero estimate as proof of no heterogeneity. A small study or an uninformative design may yield a boundary estimate with substantial uncertainty.
- Using test-versus-reference residual variability as a surrogate for subject interaction. Different variance components answer different questions and must be interpreted separately.
- Treating a model-implied individual range as a regulatory decision rule. It describes effects under model assumptions; it does not replace the prespecified population-average assessment.
The method is limited by the number of subjects, the replicate structure, data quality, and the plausibility of the assumed random-effects and residual distributions. A genuine subject-specific response pattern may also reflect unmodeled design or measurement features. For NTI medicines, substantial heterogeneity can qualify what a reassuring population-average result says about consistency across subjects, but it does not on its own determine clinical outcomes or establish that a formulation change will harm a particular patient.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a replicate crossover model, what does the variance \(\tau^2\) describe, and how does it differ from \(\sigma_T^2\)?
- With two replicates per formulation and equal residual variance \(\sigma^2=0.0030\), what residual variance contributes to a subject's mean test-minus-reference contrast?
- A study has a small estimated population-average log difference but a substantial estimated random-slope variance. What can and cannot be concluded from these two results?
- Why is a likelihood-ratio test of a zero random-slope variance not automatically interpreted using an ordinary chi-square reference distribution?
- List two diagnostic findings that should prompt qualification of a formulation-by-subject variance estimate.
Solutions. (1) \(\tau^2\) describes between-subject variation in underlying test-minus-reference effects; \(\sigma_T^2\) describes residual variation among test observations after the model terms are accounted for. (2) The contribution is \(\sigma^2/2+\sigma^2/2=0.0015+0.0015=0.0030\) log-units squared. (3) The average difference may be small while subjects' underlying effects vary; this does not alone establish failure of the population-average criterion, nor does it identify an individual's clinical outcome. (4) Zero lies on the boundary of the allowable nonnegative variance values, so the usual regularity conditions for the ordinary chi-square reference distribution do not generally hold. (5) Examples include influential subjects dominating the estimate, residual patterns by period or fitted value, convergence or singular-fit warnings, and severe instability under reasonable model checks.