Why Age and Weight Effects Can Be Confounded

Why Correlated Covariates Create a Separation Problem

In pediatric data, older children are often heavier than younger children. This pattern is physiologically expected, but it creates an inferential challenge when a pharmacokinetic model includes both body size and maturation. An observed increase in clearance across the sample could reflect greater body weight, increasing organ function with age, or both. If age and weight carry much the same information in the observed participants, the data may not clearly distinguish their separate contributions.

This is confounding of model effects: changes attributed to one covariate can be partly exchanged for changes attributed to another while producing similar predictions over the observed data. It does not mean that age and weight are the same variable, nor does it establish that either effect is absent. It means that the available pattern of ages and weights may provide weak evidence about how much each contributes.

The issue is especially important when interpreting clearance models. A plausible model structure is:

Here \(CL_i\) is individual clearance, \(WT_i\) is body weight, \(A_i\) is age, \(b\) is the allometric exponent, and \(M(A_i;\theta)\) is a maturation function with parameters \(\theta\). \(CL_{\mathrm{ref}}\) is clearance at the reference weight when the maturation multiplier is one, and \(\eta_{CL,i}\) is an individual random effect, conventionally modeled with mean zero. The equation represents a multiplicative combination of size, maturation, and individual variability; the particular reference values and maturation function must be specified for an analysis.

Taking logarithms separates the two covariate contributions additively:

If weight rises systematically with age, then the weight term and the maturation term can rise together. The concern is not simply that two variables are correlated in a table: it is whether the data contain enough contrasting combinations—for example, children of different weights at similar ages and children of similar weights at different ages—to distinguish their model contributions.

How the Effects Become Difficult to Separate

A short linear approximation makes the source of the problem clear. Write \(x_i=\log(WT_i/WT_{\mathrm{ref}})\), and let \(z_i\) represent a local age-related maturation score. Near a particular parameter value, the age contribution can be approximated by a coefficient times this score. The relevant part of the log-clearance model then has the form \(b x_i+\gamma z_i\), where \(\gamma\) summarizes a local maturation effect. This approximation is for understanding the confounding mechanism; it does not replace the specified nonlinear maturation function.

If the observed data follow approximately \(z_i=c+d x_i\), then:

The term \(\gamma c\) can be absorbed into the intercept. Along this observed pattern, the data mainly inform the combined slope \(b+d\gamma\), rather than \(b\) and \(\gamma\) separately. A different pair of parameter values can preserve that combined slope and produce nearly the same predictions. In a nonlinear mixed-effects model, the corresponding issue appears through similar changes in the model predictions and likelihood as parameters vary together.

This reasoning does not require age and weight to be perfectly correlated. Strong but imperfect correlation can still make the separate effects uncertain. Conversely, a noticeable age–weight correlation does not automatically make the model unusable if the data include enough overlap and informative outcome measurements.

Using Age-Stratified Parameter Plots

An age-stratified parameter plot is a diagnostic display, not a substitute for fitting the joint model. Divide the data into clinically interpretable age bands, then examine the relationship between estimated individual clearance and body weight within each band. A log–log display is often useful because an allometric relationship becomes approximately linear: if maturation is nearly constant within a narrow age band, the slope of log clearance against log weight is informative about the size effect in that band.

A related plot can show estimated clearance after dividing by the assumed weight multiplier, such as \(CL_i/(WT_i/WT_{\mathrm{ref}})^b\), against age. Under the proposed model, the remaining systematic age pattern should be compatible with the chosen maturation function, subject to random effects and estimation error. These plots should use individual parameter estimates cautiously: shrinkage, sparse observations, and uncertainty in individual clearance estimates can make apparent patterns misleading. Where possible, show uncertainty and the underlying age and weight distributions rather than only a fitted line.

The most useful diagnostic is whether each age band contains enough weight spread to estimate or inspect a weight relationship, and whether different age bands overlap in weight. If each band contains only a narrow range of weights and the ranges shift upward with age, the data offer little within-age leverage on the size effect. If substantial weight overlap exists across age bands, comparisons at similar weights can help reveal whether an age-related contribution remains.

Practical Diagnostics and Alternatives

A useful assessment combines plots with model-based uncertainty. First inspect age against weight to see whether there are overlapping values or distinct clusters. Then inspect age-stratified clearance–weight relationships and the age trend after applying the proposed size adjustment. Finally, examine whether estimates of \(b\) and maturation parameters are stable when the model is refit under reasonable alternative specifications. Large changes in one parameter accompanied by compensating changes in another, with little change in fit, are consistent with weak separation.

These checks rely on assumptions that should be made explicit. The clearance estimates or observations must be informative about clearance; the selected maturation function must be a reasonable representation of the age trend; the allometric form must be appropriate over the observed weight range; and the random-effects and residual-error models must not create systematic patterns that are mistakenly attributed to age or weight. Age bands should be chosen for interpretability and adequate data, not repeatedly adjusted until a desired slope appears. Within-band slopes are especially uncertain when a band has few participants or little weight spread.

Several approaches answer different questions and should not be mistaken for interchangeable solutions:

• Fit the joint size-and-maturation model. This directly represents both mechanisms, but its separate parameter estimates depend on covariate overlap and model assumptions.
• Fix a scientifically justified allometric exponent. This can reduce competition between the exponent and maturation parameters, but it does not create information absent from the data. Results should be interpreted conditional on the fixed value.
• Fit simpler comparison models, such as size-only or maturation-only forms. Comparing their fit can show whether each contribution is supported, but a fit comparison alone does not prove that the retained effect is uniquely responsible for the pattern.
• Use age-stratified plots. These make overlap and within-age weight information visible, but small strata can be noisy and the plots do not replace joint estimation.

A common mistake is to see that a combined model fits well and conclude that it has separately identified both the allometric exponent and maturation parameters. A good fit establishes that the model can describe the observed data reasonably; it does not show that the individual effects are precisely estimated. The correction is to inspect covariate overlap, parameter uncertainty and compensation, and predictions at combinations of age and weight supported by the data. Avoid strong extrapolation to combinations scarcely observed in the study.