Why Separate Size and Maturation Effects?
Pediatric clearance can change with both body growth and developmental changes in enzyme or transporter activity. A PBPK model may represent these contributions separately, but the data do not necessarily identify them separately. If the children with greater body weight are also older, a fitted increase in clearance could be attributed to size scaling, maturation, or a combination. The model may predict the observed clearances well while giving very different values for those individual effects.
This is a problem of confounding in parameter estimation: two or more model parameters can compensate for one another because the observed data provide similar information about their combined effect. Parameter correlations, profile plots, and residual patterns provide complementary ways to diagnose the problem. None, on its own, proves that the model is correct or that a specific biological mechanism is responsible.
Model and Conditions for Diagnosis
Consider a clearance model in which size scaling and maturation enter multiplicatively:
Here \(CL_i\) is clearance for individual \(i\), \(W_i\) is body weight, \(A_i\) is age, \(b\) is the size-scaling exponent, and \(M(A_i;\theta)\) is a positive maturation multiplier with parameters \(\theta\). \(W_{\mathrm{ref}}\) is a reference weight and \(CL_{\mathrm{ref}}\) is clearance at that reference weight when the maturation multiplier equals one. For clarity, suppose the maturation function is normalized so that \(M(A_{\mathrm{ref}};\theta)=1\). In an actual analysis, the reference definitions and the form of \(M\) must match the PBPK model being fitted.
A common observation model assumes that measured clearances are positive and that their logarithms have independent, mean-zero errors with constant variance:
This is a working model, not a universal assumption. If clearance estimates have different precision, the fitting method should account for that; if errors are correlated or the residual variance changes systematically, the simple error model may be inappropriate. Age and weight are observed covariates in this expression. The diagnostic question is whether the available observations distinguish their modeled contributions, conditional on the chosen maturation function and error model.
Why Parameter Estimates Can Compensate
Write \(x_i=\log(W_i/W_{\mathrm{ref}})\) and \(\eta_i=\log(CL_i)\). The model prediction is \(\eta_i=\log(CL_{\mathrm{ref}})+b x_i+\log M(A_i;\theta)\). For small changes in the parameters, a first-order approximation gives:
The transpose indicates that the maturation sensitivity and parameter change are vectors when \(\theta\) has more than one component. This expression includes the change in the scale parameter \(CL_{\mathrm{ref}}\): changing it affects every predicted log-clearance by the same amount. If \(CL_{\mathrm{ref}}\) is held fixed, its contribution is zero, but it must not be omitted when it is estimated jointly.
The approximation follows by taking the differential of the model’s log prediction. The derivative with respect to \(b\) is \(x_i\), and the derivative with respect to \(\theta\) is the gradient of \(\log M\). Therefore, changes in \(b\) and \(\theta\) can cancel if their sensitivity patterns across individuals are similar. If age and weight also move together in the data, the observed rows may contain little information for distinguishing those patterns. In matrix terms, columns of the prediction-sensitivity matrix can become nearly linearly dependent; estimates may then be highly correlated and imprecise.
This is a local justification: it describes small parameter changes near a fitted value. A nonlinear maturation function can behave differently farther away, which is one reason to inspect the full profile likelihood rather than relying only on a local covariance matrix.
Worked Example: Reading a Parameter-Correlation Warning
Suppose a hypothetical fit estimates \(b=0.82\) and a maturation parameter \(\theta=1.40\). The estimated standard errors are \(SE(b)=0.28\) and \(SE(\theta)=0.31\), and their estimated correlation is \(-0.96\). These quantities come from a local covariance approximation around the fitted values.
The corresponding covariance is:
The strong negative correlation means that, near this fitted solution, a larger estimate of \(b\) tends to occur with a smaller estimate of \(\theta\). That pattern is consistent with compensation between the size and maturation contributions. It is a warning to investigate identifiability, not evidence that either estimate is biologically wrong.
A correlation alone does not show whether the likelihood is flat over a practically important range, whether the result depends on the maturation function, or whether one unusual observation drives the estimate. A profile plot and residual checks are needed before drawing a substantive conclusion.
Profile Plots: Testing How Much the Data Pin Down Size Scaling
A profile likelihood for \(b\) fixes \(b\) at a sequence of values and, at each value, refits the other free parameters, including \(\theta\), \(CL_{\mathrm{ref}}\), and any error parameters. The resulting curve shows how much worse the best fit becomes when \(b\) is constrained. Profiling permits the other parameters to compensate, so it directly probes whether the data distinguish a particular value of \(b\) from alternatives.
Under regular large-sample conditions, a likelihood-ratio interval for one profiled parameter retains values satisfying \(2[\ell_{\max}-\ell_{\mathrm{prof}}(b)]\leq 3.84\), the approximate 95th percentile of a chi-square distribution with one degree of freedom. This approximation can be unreliable with small samples, parameters on boundaries, or severe non-identifiability. The curve should be considered alongside the model assumptions and the data coverage.
Worked Example: Interpreting a Profile for the Size Exponent
For a hypothetical pediatric dataset, an analyst fixes \(b\) at several values and refits the remaining parameters. The table gives the increase in deviance, \(2[\ell_{\max}-\ell_{\mathrm{prof}}(b)]\), relative to the best fit.
| Fixed \(b\) | Profile deviance increase | Interpretation using the approximate 3.84 cutoff |
|---|---|---|
| 0.60 | 0.40 | Retained |
| 0.80 | 0.00 | Best fit |
| 1.00 | 0.50 | Retained |
| 1.20 | 2.10 | Retained |
| 1.40 | 5.00 | Not retained by this approximate cutoff |
The approximate likelihood-ratio interval includes \(b=0.60\), \(0.80\), \(1.00\), and \(1.20\), but not \(1.40\). Thus, this fit does not tightly determine \(b\) within the displayed range. The profile’s broad, shallow section is consistent with other parameters—particularly maturation parameters—adjusting to preserve predictions. The best-fitting value \(0.80\) should not be reported as though the data precisely established it.
The cutoff is an asymptotic guide, not a universal rule. A profile that is asymmetric, discontinuous, or sensitive to starting values may also signal numerical or model-fitting problems; those should be addressed before interpreting an interval.
Age-Stratified Residuals: Looking for Systematic Age Patterns
For observed and predicted clearances, define the log residual \(r_i=\log(CL_i^{\mathrm{obs}})-\log(CL_i^{\mathrm{pred}})\). A positive residual means the observed clearance exceeds its prediction; a negative residual means the prediction exceeds the observation. Grouping these residuals into clinically meaningful age strata can reveal systematic patterns hidden by an overall average. Age strata should be chosen to represent relevant developmental periods and should not be selected after inspecting residuals solely to manufacture a pattern.
A sequence of positive mean residuals at younger ages and negative means at older ages suggests that the model overpredicts relative clearance at older ages relative to younger ages, or underpredicts it at younger ages. Such a pattern is compatible with an inadequate maturation function, a size–age allocation problem, or another age-related model misspecification. Residuals diagnose discrepancy; they do not identify its biological cause on their own.
Worked Example: Calculating Age-Stratified Mean Residuals
Consider six hypothetical observations. The residuals are on the log-clearance scale and are grouped into three age strata, with two observations in each stratum.
| Age stratum | Log residuals | Mean residual calculation |
|---|---|---|
| Infants | 0.18, 0.12 | \((0.18+0.12)/2=0.15\) |
| Younger children | 0.06, \(-0.02\) | \((0.06-0.02)/2=0.02\) |
| Older children | \(-0.10\), \(-0.14\) | \((-0.10-0.14)/2=-0.12\) |
The infant mean residual of \(0.15\) means that the geometric mean of observed clearance is about \(e^{0.15}=1.162\), or 16.2%, above the geometric mean prediction in that stratum. The older-child mean of \(-0.12\) corresponds to \(e^{-0.12}=0.8869\), about 11.3% below prediction. The change from positive to negative means is a systematic age pattern in these hypothetical residuals, not merely an overall offset.
Because each stratum contains only two observations, these means are unstable and are not sufficient for a formal conclusion. The pattern should prompt inspection of individual residuals, uncertainty, data quality, and the model’s age and weight coverage. If the pattern persists with adequate information, the assumed maturation or size relationship may not describe the data well.
Combining Diagnostics and Comparing Approaches
The three diagnostics answer different questions. Parameter correlation summarizes local co-movement around one solution. A profile plot shows how fit changes across values of a selected parameter after other parameters are re-optimized. Age-stratified residuals show whether predictions have systematic errors across age groups. Agreement among them strengthens the case that the data do not cleanly separate size and maturation effects, but none establishes the cause without further evaluation.
| Diagnostic | Useful signal | Important limitation |
|---|---|---|
| Parameter-correlation matrix | Large absolute correlation between \(b\) and maturation parameters | Local summary; can depend on parameterization and fit location |
| Profile likelihood | Broad or irregular range with little loss of fit | Requires reliable refitting and appropriate likelihood assumptions |
| Age-stratified residuals | Persistent age-related pattern in observed-minus-predicted values | Indicates lack of fit, but does not isolate its biological source |
A practical alternative is to constrain one contribution using external information—for example, a separately justified size exponent or maturation profile—and then estimate the remaining parameters. This can improve numerical stability, but it does not make the pediatric data themselves informative about the constrained quantity. The constraint and its source must be reported, and sensitivity to reasonable alternatives should be examined. Simply comparing a freely estimated model with a constrained one is not a cure for confounding if the data cannot discriminate between them.
Other explanations must also be checked. A residual trend could reflect measurement error, an omitted covariate, an unsuitable maturation curve, or an incorrect structural assumption. Sparse observations in an age range, a narrow range of body weights within age strata, or lack of overlap in weight across ages can make separation especially difficult. Conversely, observations that cover overlapping weights at different ages can help distinguish the effects, provided their precision and model structure are adequate.
Common Pitfall and Practical Interpretation
Incorrect reasoning: “The estimated correlation between \(b\) and \(\theta\) is \(-0.96\), so the model proves that maturation and size have opposite biological effects.” The correlation describes statistical co-movement of estimates near the fitted solution; it does not establish the direction or cause of biological effects.
Correction: Treat strong parameter correlation as a warning. Inspect the profile for \(b\), examine age-stratified residuals, check parameter uncertainty and numerical stability, and evaluate whether the data cover enough independent variation in age and weight. Report the uncertainty in the separate contributions. If distinct effects remain weakly identified, say so rather than presenting one fitted decomposition as uniquely established.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the log-clearance model, what term represents the local change in predicted log-clearance when \(CL_{\mathrm{ref}}\) changes?
- A profile deviance increase is 2.7 at a fixed value of \(b\). Under the stated large-sample, one-parameter approximation, is that value retained by the 3.84 cutoff?
- Mean log residuals are positive in infants and negative in older children. What does this indicate, and what does it not establish?
- Why does a strong negative correlation between \(b\) and a maturation parameter not prove that their biological effects oppose each other?
Solutions. (1) The contribution is \(\Delta\log(CL_{\mathrm{ref}})\); it shifts every predicted log-clearance by the same amount. (2) Yes. Since \(2.7<3.84\), the value is retained by that approximate likelihood-ratio criterion, subject to its regularity assumptions. (3) It indicates an age-related prediction pattern: the model tends to predict too little relative clearance in infants and too much in older children. It does not identify whether the cause is maturation, size scaling, another covariate, or model misspecification. (4) The correlation describes how fitted parameter estimates compensate near the solution. It is a local statistical relationship, not a direct statement about biological mechanisms or causal effects.