Testing Robustness To Model Misspecification

Why Model Robustness Matters

A model-based bioequivalence analysis connects measured concentrations to a formulation contrast. Its conclusion depends not only on the observed data but also on how the model represents absorption, disposition, within-subject variability, and residual error. If that representation is wrong, the analysis can estimate the wrong contrast, report uncertainty that is too small or too large, or both. A model can fit the observed data reasonably well and still give a fragile decision, especially when sampling is sparse or the fitted model imposes restrictive variability assumptions.

Robustness assessment asks a focused question: would the bioequivalence conclusion remain credible under plausible data-generating structures that differ from the fitted model? It is not a search for a model that produces a preferred decision. The estimand, decision limits, and plausible departures should be specified before interpreting the results.

For the conventional limits used here, let \(\theta\) be the population mean contrast on the log scale, defined as test formulation minus reference formulation for the chosen exposure measure. The estimated geometric mean ratio is \(\exp(\hat{\theta})\). At significance level \(\alpha=0.05\), the usual two one-sided tests (TOST) decision is equivalent to requiring the 90% confidence interval for \(\theta\) to lie wholly inside \((\log(0.80),\log(1.25))\). This tutorial holds those limits and the estimand fixed; robustness concerns the model and inference used to assess them.

What Can Go Wrong

It is useful to distinguish two broad consequences. Estimation misspecification can shift \(\hat{\theta}\), for example if a fitted concentration-time structure cannot represent a feature of the data that differs by formulation. Uncertainty misspecification can leave the estimated contrast approximately centered on \(\theta\) but produce an incorrect \(SE\). The first threatens the location of the interval; the second threatens its width and coverage. Either can change the decision.

The distinction is conceptual rather than absolute. A misspecified mean or covariance structure can affect both the estimate and its reported uncertainty. For instance, with sparse sampling, an incorrect disposition shape may change the fitted area under the curve. With repeated observations, a covariance model that ignores within-subject dependence may misstate how much information the data contain. A variance model that averages across groups with different variability may also obscure uncertainty in the formulation contrast.

A basic justification comes from a sample mean. Suppose independent subject-level contrasts have true variance \(\sigma^2\), but the fitted model uses variance \(\tilde{\sigma}^2\). The estimated mean can still be centered correctly, while the fitted standard error is wrong:

If \(\tilde{\sigma}<\sigma\), the fitted interval is too narrow. Its nominal coverage then need not hold, and a test calibrated to that interval can make too many false declarations of bioequivalence. If \(\tilde{\sigma}>\sigma\), the interval may be too wide and reduce the chance of demonstrating bioequivalence. In more complex models, the same principle applies: uncertainty must reflect the actual variability and dependence of the estimator, not just the model's assumed structure.

Assessing Robustness in Practice

A useful assessment separates the target of inference from the model used to reach it. State the exposure measure and formulation contrast of interest first. Then identify which fitted assumptions could plausibly fail given the design, sampling schedule, and observed data. Candidate departures may include alternative concentration-time structures, different residual-error forms, nonconstant variability across formulations or subjects, and different within-subject dependence. This is a structured stress test, not an instruction to explore every conceivable model.

Model checks can reveal warning signs, but no single diagnostic proves that a model is correctly specified. Examine residuals against time, fitted concentration, formulation, period, and subject; inspect whether variability changes systematically; and check whether fitted curves miss consistent features of the observed concentration profiles. In crossover or replicate data, examine whether the assumed covariance structure is consistent with the pattern of repeated observations. Convergence problems, parameters at boundaries, or large changes under small fitting choices are additional warnings. Good residual plots do not rule out misspecification that the design cannot distinguish.

A direct way to assess decision stability is simulation-estimation. Specify a set of plausible data-generating scenarios, simulate studies with the planned design, fit the intended analysis model to each, and record the estimate, reported uncertainty, and bioequivalence decision. Include scenarios in which the true contrast is within the limits and scenarios at or outside a limit. This evaluates how the complete analysis procedure behaves under stated departures; it does not prove that the scenarios include the actual data-generating process.

Alternatives, Limitations, and Common Pitfalls

Several approaches answer related but different questions. Fitting plausible alternative models shows whether conclusions depend on a particular structural choice, but it does not identify which model is true. A sandwich variance estimator can provide uncertainty that is less dependent on a fully specified likelihood under appropriate regularity conditions, but it does not repair bias in the estimated contrast or make an incorrect exposure model correct. Bootstrap resampling can propagate sampling uncertainty under the resampling design, but ordinary resampling from the observed data does not automatically represent unobserved misspecification. Simulation-estimation directly tests a specified analysis procedure under specified generating scenarios, but its conclusions are conditional on those scenarios.

The main limitations are practical and scientific. Sparse or uninformative sampling may make alternative structures difficult to distinguish. Simulation results can be imprecise, and no finite set of scenarios covers every possible departure. A stable decision across the scenarios examined is evidence of stability over that set, not a guarantee of validity in all settings. Robustness also cannot rescue a poorly defined estimand or an analysis that violates basic design assumptions.

• Mistake: treating a good fit as proof of robustness. A model may reproduce observed concentrations while still misrepresenting uncertainty or extrapolated exposure. Check decision stability under plausible departures as well as fit diagnostics.
• Mistake: comparing only point estimates. Similar estimates can produce different decisions if standard errors differ. Compare intervals, calibration, and decisions.
• Mistake: choosing scenarios after seeing which one passes. That makes the assessment outcome-driven. Define scientifically credible scenarios and report them, including scenarios that change the conclusion.
• Mistake: interpreting an unchanged decision as proof the model is correct. Robustness concerns stability of the decision, not proof of the generating mechanism.