Why Identifiability and Fit Must Come Before Interpretation
A model can return a numerical estimate of a formulation effect even when the data do not determine the model parameters reliably, or when the model systematically misses the observed concentration-time pattern. These are different problems. Identifiability asks whether the available model and data can determine the parameters; model fit asks whether the model describes the observed data adequately. A converged optimizer and a plausible-looking estimate do not, by themselves, answer either question.
This distinction matters in model-based bioequivalence because exposure metrics and formulation contrasts are calculated from fitted model parameters. If those parameters are weakly determined, or the concentration model is systematically wrong, a precise-looking geometric mean ratio (GMR) can give false reassurance. Diagnostics should therefore precede interpretation of the bioequivalence estimate.
If two or more parameter combinations yield indistinguishable predictions even with ideal observations, the model has a structural identifiability problem. By contrast, a structurally identifiable parameter can still be practically weakly identified: the available sampling times may contain little information about it, or its likelihood may be nearly flat over a broad range. A profile can reveal practical weakness, but cannot alone prove structural identifiability for every possible observation design.
Likelihood Profiles: Looking Beyond a Standard Error
Suppose a fitted concentration model has parameter vector \(\boldsymbol{\theta}\), and its likelihood for the observed data is \(L(\boldsymbol{\theta})\). For a parameter of interest \(\psi\), the profile likelihood at a fixed value \(\psi_0\) is the largest likelihood achievable while holding \(\psi=\psi_0\) and re-estimating all remaining parameters. Writing the log likelihood as \(\ell=\log L\), define the profile likelihood-ratio statistic as follows.
The profile allows nuisance parameters to adjust as \(\psi\) changes. This matters because a flat profile can show that other parameters compensate for changes in \(\psi\), even when a local standard error looks moderate. Under a correctly specified likelihood, regularity conditions, and an interior parameter value, \(D(\psi_0)\) is approximately distributed as a chi-square random variable with one degree of freedom when \(\psi_0\) is the true value. The approximate 95% profile-likelihood interval contains values satisfying \(D(\psi_0)\leq 3.84\). This is an asymptotic approximation, not a universal cutoff: boundaries, sparse data, or nonregular models can undermine it.
Worked Example: Profiling an Absorption-Rate Parameter
Consider hypothetical concentration data fitted with a one-compartment model. The parameter of interest is the absorption-rate constant \(k_a\), measured in \(h^{-1}\). Assume the concentration likelihood uses the previously selected residual-error model, the optimizer reaches the same maximum from multiple starting values, and the other parameters are re-estimated at each fixed \(k_a\). The fitted value is \(1.10\ h^{-1}\). The following table gives profile likelihood-ratio values.
| \(k_a\) (\(h^{-1}\)) | 0.80 | 0.90 | 1.10 | 1.30 | 1.40 |
|---|---|---|---|---|---|
| \(D(k_a)\) | 4.80 | 3.30 | 0.00 | 2.20 | 4.00 |
For a one-parameter 95% profile interval, use the approximate cutoff \(3.84\). The lower crossing lies between 0.80 and 0.90. Linear interpolation gives
The upper crossing lies between 1.30 and 1.40:
Thus the interpolated approximate interval is \(0.864\) to \(1.391\ h^{-1}\). It is broad relative to the fitted value \(1.10\ h^{-1}\), so \(k_a\) is not precisely determined by these hypothetical data. The stated crossings are approximations: a finer profile grid or numerical root-finding would be preferable for a final analysis. The optimizer checks support the calculation but do not establish that the model is correct.
Parameter Correlations as a Warning Signal
When parameters can compensate for one another, their estimates may be strongly correlated. Near an interior maximum, the inverse of the observed information matrix is often used to approximate the covariance matrix of the parameter estimates. For estimates \(\hat{\theta}_i\) and \(\hat{\theta}_j\), their estimated correlation is
A correlation near \(1\) or \(-1\) can indicate that the likelihood is elongated along a direction in which one parameter increases as another changes. This is a useful warning, but not a diagnosis on its own. Correlation is local to the fitted solution and depends on parameterization; nonlinear likelihood shapes, multiple modes, or a nearly flat direction may not be summarized well by one correlation matrix.
Worked Example: Interpreting a Strong Parameter Correlation
In a hypothetical fit, the estimated standard errors are \(0.18\ h^{-1}\) for \(k_a\) and \(0.040\ h^{-1}\) for an elimination-rate parameter \(k_e\). Their estimated covariance is \(-0.00698\ (h^{-1})^2\). The corresponding estimated correlation is
The strong negative correlation warns that the data may allow increases in one rate parameter to be offset by decreases in the other. It does not prove that either parameter is structurally non-identifiable, nor does it show that the concentration model fits poorly. The next diagnostic is to inspect profiles for the rate parameters and relevant derived quantities, while also checking whether predicted concentrations reproduce the observed pattern.
Residual Patterns and Prediction Checks
A residual is the difference between an observed concentration and the corresponding model prediction. Its scale and distribution should be assessed using the residual-error model used in fitting: for example, raw residuals are not directly comparable across observations if their expected variances differ. Standardized or appropriately scaled residuals help make that comparison. Under an adequate model and error specification, residuals should not show a systematic pattern against time, predicted concentration, formulation, period, or other relevant design features. They need not be exactly normal or pattern-free in a finite sample.
Residual plots and prediction checks address different views of model adequacy. Residuals emphasize discrepancies at observed records and can reveal trends hidden by a summary statistic. Prediction checks compare observed concentrations with the model's predictions on the concentration scale and in the sampling-time and formulation context. In-sample predictions are useful diagnostics, but agreement with the same data used for fitting is not independent validation. Neither a good-looking plot nor a single fit statistic establishes that every model assumption is correct.
Worked Example: Detecting Late-Time Underprediction
Suppose a hypothetical model uses a constant residual standard deviation of \(0.50\ \mathrm{mg/L}\). At four sampling times, the following are observed concentrations and fitted predictions. Define the standardized residual as \((\text{observed}-\text{predicted})/(0.50\ \mathrm{mg/L})\).
| Time (h) | Observed (mg/L) | Predicted (mg/L) | Standardized residual |
|---|---|---|---|
| 1 | 4.5 | 5.0 | -1.0 |
| 2 | 7.5 | 8.0 | -1.0 |
| 4 | 8.0 | 7.0 | 2.0 |
| 8 | 5.5 | 4.0 | 3.0 |
For example, at 8 h the standardized residual is \((5.5-4.0)/0.50=3.0\). The residuals are negative early and positive later, with larger positive values at the later times. That ordered pattern suggests the fitted curve may be underpredicting late concentrations rather than merely missing observations at random. Four points are not enough to establish a general lack-of-fit conclusion, but the pattern warrants checking the full residual-versus-time plot and observed-versus-predicted concentrations by formulation. If the pattern is repeated across subjects, a systematic model discrepancy is more concerning than one isolated residual.
A Practical Diagnostic Sequence
These diagnostics are most useful when treated as complementary checks, not interchangeable pass-or-fail tests. A practical sequence is:
Confirm convergence, inspect whether different starting values lead to comparable maxima, and verify that estimates are not at unintended parameter bounds.
Inspect standard errors and correlations, then profile parameters that are scientifically important or visibly uncertain. Look for broad, asymmetric, flat, or disconnected profile regions.
Plot appropriately scaled residuals against time, prediction, formulation, and relevant design factors. Look for trends, changing spread, clusters, and influential observations.
Compare observed and predicted profiles across the sampling schedule and formulations. Investigate whether discrepancies occur where they could materially affect the derived exposure metric or formulation contrast.
The likelihood profile is often more informative than a Wald interval when the likelihood is asymmetric or a parameter is near a boundary. A Wald interval uses a local standard error and a roughly symmetric approximation around the estimate; a profile interval allows nuisance parameters to adjust and reflects the likelihood shape more directly. Neither replaces residual and prediction checks: a well-determined parameter can belong to a misspecified model, and a model can fit the observed data while some parameters remain weakly identified.
Likewise, a single goodness-of-fit summary or information criterion cannot establish structural identifiability or show where a model misses the data. Correlation matrices are quick local summaries, whereas profiles explore how likelihood changes over a parameter range. Residual plots localize discrepancies; prediction comparisons show their scale and context. The appropriate alternative depends on the question, and conclusions are stronger when these checks tell a consistent story.
Consequences for Bioequivalence and Common Pitfalls
If parameters are weakly identified, the model-derived exposure or GMR can also be uncertain, especially when the derived quantity changes substantially along a flat likelihood direction. Conversely, individual parameters can be uncertain while a particular derived contrast is relatively stable. Therefore, do not infer the precision of a bioequivalence estimate solely from the standard errors of individual parameters; examine the uncertainty and sensitivity of the quantity that will actually be interpreted.
If residual patterns indicate systematic lack of fit, the issue is not repaired by reporting a narrower interval for the formulation contrast. A wrong concentration-time shape may distort exposure estimates differently across formulations or sampling schedules. Investigate the source of the pattern and reassess the model before interpreting its bioequivalence result. Diagnostics can flag concern, but they do not automatically identify the correct replacement model.
- Mistake: treating optimizer convergence as evidence of identifiability. Convergence means the numerical procedure met its stopping criteria; a broad profile can still show that many parameter values fit almost as well.
- Mistake: treating a high parameter correlation as proof of non-identifiability. It is a warning about local compensation, not proof. Inspect profiles and the model's predictions.
- Mistake: treating random-looking residuals as proof of a correct model. Plots have limited power, especially with sparse data. Consider sampling times, formulation-specific patterns, and whether discrepancies affect the derived exposure.
- Mistake: interpreting a precise GMR before checking the concentration model. Numerical precision conditional on the fitted model does not protect against weak identification or systematic misspecification.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In one sentence, distinguish structural identifiability from practical identifiability.
- A one-parameter profile has \(D=2.7\) at a tested value, and the regularity conditions for the usual approximation are considered reasonable. Is that value inside the approximate 95% profile interval?
- Two estimated parameters have correlation \(-0.98\). What does this suggest, and what does it not establish?
- Standardized residuals are mostly negative early after dosing and positive late after dosing. Give one concern and one diagnostic follow-up.
Solutions. (1) Structural identifiability concerns whether ideal observations uniquely determine the parameter under the model and design; practical identifiability concerns how precisely the actual finite, noisy data determine it. (2) Yes. Since \(2.7<3.84\), the value satisfies the approximate profile-likelihood interval criterion. (3) It suggests strong local parameter compensation and possible weak practical determination; it does not prove structural non-identifiability or poor concentration fit. (4) The fitted time-course shape may be systematically wrong, such as underpredicting later concentrations. Inspect residuals against time and compare observed with predicted profiles, separated by formulation where appropriate.