Why Use Model-Estimated AUCs?
A formulation comparison usually concerns exposure, not the particular curve-fitting method used to summarize it. In a crossover study, an individual AUC can be estimated for each subject under each formulation using a population pharmacokinetic model. Those estimates can then be compared within subject. The resulting analysis is useful when a prespecified model provides a suitable representation of concentration-time data and produces meaningful individual exposure estimates.
The method does not make bioequivalence a property of a model output alone. The estimand, analysis population, model, treatment contrast, and acceptance criterion must be appropriate for the study question. A model can reduce noise or use information across subjects, but it can also impose incorrect assumptions or understate uncertainty. In particular, a confidence interval based only on variation in estimated AUCs may omit uncertainty from estimating the population PK model itself.
The AUCs must refer to the same exposure definition, such as AUC over a specified dosing interval, and to compatible doses and conditions. In the examples below, the study criterion is the conventional 80.00% to 125.00% range for the test-to-reference geometric mean ratio. A different prespecified criterion must replace it when the study requires one.
Model and Assumptions for the Paired Comparison
For a two-period crossover, a straightforward analysis starts with one pair of model-estimated AUCs per subject and calculates \(d_i\). The sample mean \(\bar d\) is the estimated average treatment contrast on the log scale. A paired \(t\) interval uses the sample standard deviation \(s_d\) of the \(d_i\) values and standard error \(SE=s_d/\sqrt{n}\). This simple analysis assumes the pairs are independent across subjects and that the distribution of the paired log contrasts is sufficiently close to normal for the \(t\)-based interval to be appropriate.
These assumptions concern both the comparison and the upstream PK analysis. The model should adequately describe the concentration-time profiles for the intended AUC calculation; treatment, period, and relevant design features should be handled appropriately; and the individual estimates should be sufficiently informed by each subject’s data. In a crossover, the interpretation also relies on suitable randomization, adequate washout, and no material carryover. If the study has important period or sequence effects, incomplete pairs, or a design other than a simple two-period crossover, a model that represents those features is generally more appropriate than an unadjusted paired \(t\) analysis.
Population PK estimation often uses empirical Bayes estimates of subject-specific parameters. Such estimates are partly informed by the population distribution: when an individual’s data are weakly informative, estimates can be pulled toward the population mean, a phenomenon called shrinkage. Shrinkage may make individual estimates look less variable than the underlying individual values. If the amount of shrinkage differs across treatments or subjects, a simple comparison of the estimated AUCs can distort the treatment contrast. The fit should therefore be evaluated, and uncertainty from estimating the model should be considered in the inferential procedure.
From a Log Contrast to the BE Criterion
For the paired analysis, calculate the mean and standard deviation of the subject-level log contrasts. A two-sided 90% confidence interval for the mean log contrast is
Exponentiating the two endpoints gives a 90% confidence interval for the test-to-reference geometric mean ratio. Under the 80.00% to 125.00% criterion, the comparison satisfies the average BE interval rule only when the entire interval is inside \((0.80,1.25)\). Equivalently, the log-scale interval must lie within \((\log(0.80),\log(1.25))\). Merely having the point estimate inside the range is not sufficient.
The connection with two one-sided tests explains the use of a 90% interval. At an overall one-sided significance level of \(\alpha=0.05\), equivalence requires rejecting both hypotheses that the mean log contrast is at or beyond a boundary. Each one-sided test uses a 5% tail. The corresponding pair of tests rejects both boundary hypotheses exactly when the two-sided 90% confidence interval lies wholly between the boundaries. Thus the decision is about the interval’s location, not about whether a test of no difference is statistically significant.
Worked Analysis of Paired Model Estimates
Worked Example: A Paired AUC Comparison That Meets the Criterion
Consider a hypothetical 12-subject, two-period crossover. A prespecified population PK model yields one AUC estimate for each subject under each formulation, both in ng·h/mL and for the same exposure interval. The twelve paired log contrasts \(d_i=\log(AUC_{iT}/AUC_{iR})\) comprise six values of \(-0.01489\) and six values of \(0.21489\). These values are hypothetical model outputs, not data from a real study.
The mean is
Each value is \(0.11489\) from the mean, so the sample standard deviation is approximately
With \(df=11\), \(t_{0.95,11}\approx1.796\). The 90% confidence interval on the log scale is
The estimated geometric mean ratio is \(\exp(0.1000)=1.105\), or 110.5%. Exponentiating the interval endpoints gives \((\exp(0.03779),\exp(0.1622))\approx(1.039,1.176)\), or approximately 103.9% to 117.6%. Both limits lie inside 80.00% to 125.00%, so this hypothetical paired analysis meets the stated interval criterion.
This conclusion depends on the suitability of the PK model and the paired analysis assumptions. The calculation alone does not verify that the model is adequate or that its AUC uncertainty has been incorporated.
Worked Example: A Point Estimate Inside the Limits Is Not Enough
Suppose another hypothetical 12-subject comparison has \(\bar d=0.1600\) and \(s_d=0.1600\). These are summaries of paired log AUC contrasts from the model; assume for this calculation that the pairs are independent and the log contrasts are reasonably compatible with the paired \(t\) method.
The standard error is \(0.1600/\sqrt{12}=0.04619\). With \(df=11\), \(t_{0.95,11}\approx1.796\), so the margin is \(1.796(0.04619)\approx0.08295\). Therefore the 90% log-scale interval is
The geometric mean ratio is \(\exp(0.1600)\approx1.174\), which is within the ratio limits. The confidence interval, however, is approximately \((\exp(0.07705),\exp(0.24295))=(1.080,1.275)\). Its upper limit exceeds 1.25. The result therefore does not meet the 80.00% to 125.00% interval criterion, despite the point estimate being inside the limits.
This is not proof that the formulations differ, nor is it a declaration that the formulations are inequivalent. It means this analysis has not demonstrated BE under the specified criterion: the interval is too wide at its upper end to rule out a ratio above the allowed boundary.
Using a Design Model for the Treatment Contrast
A paired \(t\) analysis is most directly justified for a complete, simple crossover with independent subjects and no need for additional design adjustment. A model-based analysis can instead represent the treatment contrast while accounting for design terms such as period and sequence. For example, a crossover model may include treatment, period, and sequence as fixed effects and subject nested within sequence as a random effect. The treatment coefficient then estimates a contrast on the log AUC scale, provided the model is correctly specified and its covariance assumptions are appropriate.
The treatment coefficient’s standard error and degrees of freedom depend on the fitted model and design. They should not be borrowed from the simple paired calculation if the analysis model differs. The same decision principle applies: form the prespecified 90% interval for the log treatment contrast, exponentiate its endpoints, and compare the whole ratio interval with the study limits.
Worked Example: Interpreting an Adjusted Model Contrast
In a hypothetical crossover, a prespecified model for log individual AUC estimates includes treatment and period effects and accounts for within-subject pairing. Its fitted test-minus-reference treatment contrast is \(0.0400\) log units, with \(SE=0.0500\) and \(df=18\). Assume the model fit is adequate and that these reported uncertainty estimates are appropriate for the design.
For a 90% interval, \(t_{0.95,18}\approx1.734\). The margin is \(1.734(0.0500)=0.08670\), giving the log-scale interval
The point estimate corresponds to a geometric mean ratio of \(\exp(0.0400)=1.041\), or 104.1%. Exponentiating the interval gives approximately \((\exp(-0.04670),\exp(0.1267))=(0.9544,1.135)\), or 95.44% to 113.5%. The entire interval is within 80.00% to 125.00%, so the adjusted model contrast meets the stated interval criterion, subject to the model and study assumptions.
This example illustrates how a design-adjusted contrast is interpreted; it does not show that any model can be selected after seeing which one gives a favorable result. The model and analysis should be specified and justified independently of the desired conclusion.
Uncertainty, Diagnostics, and Failure Modes
The two-stage procedure—first estimate individual AUCs, then analyze their log contrasts—can be convenient, but it treats the first-stage estimates as if they were known unless their uncertainty is explicitly carried forward. That can produce confidence intervals that are too narrow. A joint approach or a suitable bootstrap can propagate uncertainty from population PK model estimation into the treatment contrast. For a bootstrap, the resampling scheme must respect the study design, typically resampling subjects and refitting the PK model and treatment comparison in each replicate. The choice of method should be justified for the model and data, rather than treated as a routine correction.
Useful checks include examining observed and model-predicted concentration-time profiles by formulation, residual patterns, influential subjects, convergence, parameter plausibility, and the precision of individual AUC estimates. Check whether AUC estimates are defined over the intended interval, whether both formulations have comparable data support, and whether any subjects have extreme or poorly determined estimates. For the contrast analysis, inspect the distribution of \(d_i\), influential paired differences, missing periods, and any relevant period or sequence patterns. The paired \(t\) interval is especially questionable when a few extreme contrasts dominate the mean or the log contrasts show marked departures from a roughly symmetric distribution.
A failed diagnostic does not automatically imply that the formulations are not bioequivalent. It signals that the planned model or interval may not reliably answer the question. Reconsidering the model or using a robust, design-appropriate inferential method may be warranted, with the reason for the change documented. Conversely, a model that converges is not necessarily adequate: convergence does not establish that its structural or statistical assumptions are credible.
| Approach | What it compares | Main consideration |
|---|---|---|
| Paired analysis of individual model AUCs | Within-subject log AUC ratios from model-derived estimates | Simple and transparent for complete crossover pairs, but first-stage uncertainty may be omitted. |
| Design-adjusted model contrast | A treatment effect on log AUC with period, sequence, or other design terms represented | Can handle design structure, but depends on the correctness of its fixed- and random-effects assumptions. |
| Noncompartmental AUC analysis | Exposure summarized from observed concentrations using a prespecified noncompartmental method | Provides a different route to AUC estimation; agreement with a model-based result is not guaranteed. |
The model-based approach is limited by the information and assumptions that produce its AUC estimates. It should not be used to conceal sparse or weakly informative data, and this analysis does not by itself establish individual-level interchangeability, safety, or efficacy. In particular, the fact that all subjects’ model-estimated ratios are not identical is not a failure of average BE; the criterion is applied to the prespecified population-level contrast and its uncertainty.
Common Mistake
Check Your Understanding
For each problem, decide what the calculation supports and identify the relevant limitation or assumption.
- A paired analysis has \(\bar d=0.0200\), \(SE=0.0400\), and \(df=9\). Using \(t_{0.95,9}\approx1.833\), calculate the 90% log-scale interval and the ratio-scale interval. Does it meet 80.00% to 125.00%?
- A model-based ratio estimate is 0.98, but its 90% interval is 0.77 to 1.18. State the BE conclusion under the conventional criterion.
- Why might a paired analysis of empirical Bayes AUC estimates understate uncertainty?
- Name two checks that should be made before relying on individual AUC estimates from a population PK model.
Solutions. (1) The margin is \(1.833(0.0400)=0.07332\), so the log interval is \((0.0200-0.07332,0.0200+0.07332)=(-0.05332,0.09332)\). Exponentiating gives approximately \((0.9481,1.0978)\). Both endpoints are inside the limits, so this interval meets the criterion, assuming the analysis is valid. (2) It does not meet the criterion: the lower endpoint, 0.77, is below 0.80. The point estimate cannot replace the interval decision. (3) The calculation may treat estimated AUCs as fixed and omit uncertainty from estimating the population PK parameters and individual effects. (4) Appropriate checks include comparing observed and predicted concentration-time profiles, evaluating residual patterns and influential subjects, confirming convergence and plausible parameters, and checking whether individual AUC estimates are sufficiently precise.