Why Identifiability Matters
A population PK model may include both a body-size effect and a maturation effect on clearance. These mechanisms are biologically distinct, but the data must contain enough information to distinguish them. In pediatric populations, older children are often heavier than younger children. If age and weight move together in the observed sample, a model may explain the same clearance pattern with a stronger size effect, a stronger maturation effect, or a combination of the two.
This is an identifiability problem: it concerns whether the available data can determine the model parameters separately. A model can converge, produce plausible predictions, and still estimate individual covariate effects poorly. The key evidence is not convergence alone, but whether materially different parameter combinations fit the data almost equally well.
A Covariate Model for Separating Effects
Consider a simplified covariate submodel for typical clearance. Let \(W_i\) be an individual's weight and \(W_{\mathrm{ref}}\) a reference weight. Define \(x_i=\log(W_i/W_{\mathrm{ref}})\). Let \(z_i\) be a prespecified, dimensionless maturation score calculated from age; larger values represent greater maturation. A simple model for log typical clearance is
Here, \(\beta_0\) is log typical clearance at the reference weight and at \(z_i=0\); \(\gamma\) is the size coefficient; and \(\delta\) is the maturation coefficient per unit of the chosen maturation score. Both predictors are dimensionless, so the coefficients are interpretable on the log-clearance scale. The model is a useful local or simplified representation: in an actual population PK model, the maturation function can be nonlinear and its parameters may also be estimated.
To see how the two effects compete, suppose the model is fitted to a response \(y_i\) that carries information about log typical clearance, with independent errors of mean zero and common variance \(\sigma^2\). The usual regression calculation assumes the model form is correct, the predictor values are treated as known, and the design matrix has full column rank. These assumptions make the calculations below transparent. Concentration data in a nonlinear PK likelihood do not generally reduce to this exact regression, but the same principle applies: the data must distinguish changes due to size from changes due to maturation.
For centered predictors, define \(S_{xx}=\sum_i(x_i-\bar{x})^2\), \(S_{zz}=\sum_i(z_i-\bar{z})^2\), and \(S_{xz}=\sum_i(x_i-\bar{x})(z_i-\bar{z})\). Their sample correlation is \(r_{xz}=S_{xz}/\sqrt{S_{xx}S_{zz}}\). With an estimated intercept, the information about the two slopes is represented by the matrix
Its determinant is \(S_{xx}S_{zz}-S_{xz}^2=S_{xx}S_{zz}(1-r_{xz}^2)\). If this determinant is zero, the centered predictors are exactly linearly dependent and the two slope coefficients cannot be uniquely estimated. If it is positive but small, the coefficients are mathematically estimable, but their uncertainty can be large.
Why Correlation Inflates Uncertainty
Under the regression assumptions above, the estimated slope coefficients have a covariance matrix equal to \(\sigma^2\) times the inverse of the predictor cross-product matrix. Inverting a two-by-two matrix gives
Thus, even if there is substantial variation in each predictor considered alone, high correlation leaves little variation that is unique to either predictor. For example, the size coefficient is informed by the part of weight variation that cannot be predicted from maturation. As \(|r_{xz}|\) approaches 1, \(1-r_{xz}^2\) approaches zero, so the standard errors grow. This describes uncertainty, not bias: high correlation does not automatically force the estimates in a particular direction, but it makes them sensitive to small changes in the data or model.
Worked Example: Broad Age-Weight Overlap
Suppose a hypothetical dataset has broad age-weight overlap: within age ranges, participants include a range of weights, and similar weights occur across different ages. For the centered predictors, let \(S_{xx}=0.500\), \(S_{zz}=0.500\), and \(S_{xz}=0.100\). Assume the regression approximation has residual standard deviation \(\sigma=0.100\) log-clearance units. These are illustrative design summaries, not results from a real study.
First calculate the predictor correlation:
For the size coefficient, the variance and standard error are
The correlation is modest, and the standard error is not greatly inflated by predictor overlap. The maturation coefficient has the same standard error here because \(S_{xx}=S_{zz}\). This does not prove that the full PK model is well identified: sampling times, concentration information, random-effects structure, and residual error also matter. It does show that this covariate design provides useful independent variation for separating the two slopes.
Worked Example: Narrow Age-Weight Overlap
Now consider a hypothetical dataset in which age and weight are tightly linked. Suppose \(S_{xx}=0.0254\), \(S_{zz}=0.1000\), \(S_{xz}=0.0500\), and \(\sigma=0.100\). The size predictor has limited spread, and much of its variation tracks maturation.
The correlation, calculated from the cross-products, is
The determinant of the slope information matrix is
Using that determinant directly, the size-coefficient variance and standard error are
Equivalently, \(SE(\hat{\gamma})=0.100/\sqrt{0.0254(1-0.9921^2)}\), which is approximately \(5.00\). This very large uncertainty means the dataset gives little precision about the size effect separately, despite having a full-rank design in principle. A fitted combination of size and maturation effects may still predict clearance over the observed age-weight range reasonably well, while the separate coefficients vary widely.
Exact Dependence and Finite Data
The next example separates exact non-identifiability from severe but finite uncertainty. Four observations are enough to illustrate the algebra, although they are far too few to represent a typical population PK analysis.
Worked Example: Near Dependence With Four Observations
Consider hypothetical centered predictor values \(x=(-1,-1,1,1)\) and \(z=(-1.1,-0.9,0.9,1.1)\). Suppose the noiseless responses follow \(y_i=2+0.4x_i-0.3z_i\). The resulting responses are \(1.93, 1.87, 2.13,\) and \(2.07\), respectively. The \(z\) values are close to \(x\), but they are not exactly a constant multiple of \(x\).
For these data, \(\sum x_i=0\), \(\sum z_i=0\), \(S_{xx}=4.00\), \(S_{zz}=4.04\), and \(S_{xz}=4.00\). Therefore,
The design matrix containing an intercept and the two predictors has rank 3: the predictors are highly correlated but not exactly dependent. With four observations and three estimated parameters, the residual degrees of freedom are \(4-3=1\). Fitting the stipulated noiseless data gives residual sum of squares zero; this is a property of these exact responses, not an absence of residual degrees of freedom. If the intercept were fixed at its stated value of 2, two slopes would be estimated and the residual degrees of freedom would be \(4-2=2\); the fitted residual sum of squares would still be zero.
If instead \(z_i=x_i\) for every observation, the prediction would reduce to \(2+(\gamma+\delta)x_i\). The data could identify the sum \(\gamma+\delta\), but not the two coefficients individually: increasing \(\gamma\) by a quantity and decreasing \(\delta\) by that same quantity would leave every prediction unchanged. That is exact non-identifiability. In the near-dependent example, unique estimates exist, but even small measurement or model perturbations could substantially change the separate estimates.
Diagnosing the Problem in a Population PK Analysis
A regression cross-product calculation is a diagnostic illustration, not a substitute for evaluating the PK likelihood. In a nonlinear population PK model, the relevant information depends on how predicted concentrations change when each parameter changes, as well as on the covariate distribution. The local sensitivity of predictions to parameters can be summarized by a Jacobian; if the sensitivity columns for size and maturation are nearly redundant, the likelihood contains little information to distinguish them. Approximate covariance matrices, parameter correlations, standard errors, and profile likelihoods can reveal this difficulty.
Useful warning signs include:
- Large estimated parameter correlation: a correlation near \(+1\) or \(-1\) between size and maturation estimates indicates that one can compensate for the other over the observed data.
- Large standard errors or wide profile likelihoods: estimates may be unstable even when the optimizer reports convergence.
- Sensitivity to starting values or small data changes: materially different parameter estimates with similar objective function values suggest a weakly identified direction.
- Little covariate overlap: plots of age against weight may show that the sample occupies a narrow band, with few observations that separate the effects.
- Unrealistic parameter estimates or boundary solutions: these may reflect competing effects or a model that the observed covariate range cannot support.
No single diagnostic proves that the model is unusable. Parameter correlations depend on parameterization and scale, and a large correlation can occur alongside adequate precision for a particular prediction. Conversely, a moderate correlation does not guarantee good estimation if the dataset has little total information or the PK sampling is uninformative. Assess the joint uncertainty and the intended use of the estimates, not just one summary statistic.
What Can Improve Separation?
The most direct remedy is information in the study design. Recruit or combine participants so that ages and weights overlap more broadly, where ethically and clinically appropriate. A design with younger participants across a range of weights and older participants at overlapping weights gives the model more opportunity to distinguish the effects. Increasing the number of observations only helps separation if it adds relevant independent information; many additional participants along the same narrow age-weight relationship may improve precision less than expected.
Other options include using external knowledge to fix or constrain one effect, or simplifying the model when the data cannot support separate parameters. These choices change what is being estimated and should be scientifically justified. Fixing a maturation parameter can stabilize the size estimate, for example, but its uncertainty is then not reflected as if that parameter had been estimated freely. A model that estimates both effects may be retained for prediction within the observed covariate range while acknowledging that the individual effects are weakly determined.
The main alternatives have different purposes:
| Approach | Strength | Limitation |
|---|---|---|
| Collect data with broader age-weight overlap | Directly adds information that distinguishes the effects | May be difficult or infeasible; does not solve all PK information problems |
| Fix or constrain one effect using external evidence | Can stabilize estimation when the sample is weakly informative | Results depend on the external assumption and may understate uncertainty if treated as known |
| Estimate both effects and report joint uncertainty | Displays what the current dataset can and cannot determine | Separate estimates may be too imprecise for mechanistic interpretation |
| Use a simpler covariate model | Can avoid unsupported parameter complexity | May omit a real effect and should not be chosen solely to obtain convenient estimates |
Common Mistake
Wrong reasoning: “The model converged and both coefficients have numerical estimates, so size and maturation have been identified separately.” Convergence only indicates that the estimation algorithm met its numerical stopping conditions. It does not establish that the data distinguish the parameters.
Correction: Inspect the covariate overlap, parameter correlation, uncertainty, and stability of the estimates. When predictors are highly correlated, the combined prediction over the observed data can be useful even when separate size and maturation coefficients are not precise. Avoid interpreting unstable coefficients as definitive evidence about the relative biological contributions.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- If the centered size and maturation predictors are exactly proportional, what part of the model can be estimated, and what cannot?
- With \(r_{xz}=0.90\), what is the variance inflation factor \(1/(1-r_{xz}^2)\) relative to uncorrelated predictors, holding \(S_{xx}\) and \(\sigma^2\) fixed?
- Why might adding many participants fail to separate size and maturation effects?
- A model converges but has a size-maturation parameter correlation of \(-0.98\). What is a reasonable interpretation and what should be checked next?
Solutions. (1) The data can estimate a combination of the two coefficients, but not their separate values, because different coefficient pairs yield identical predictions. (2) \(1/(1-0.90^2)=1/0.19\approx5.263\); the coefficient variance is about 5.263 times its uncorrelated value under the stated comparison. (3) If the added participants follow the same narrow age-weight relationship, they may add little independent covariate variation. (4) The estimates likely compensate for each other, so separate effects may be unstable. Check covariate overlap, standard errors or joint uncertainty, profile likelihoods, and sensitivity to model choices or data perturbations.