Why Separate Size From Maturation?
In pediatric pharmacokinetics, body size and developmental age often change together. A young child is usually smaller than an older child, and clearance may increase both because the child is larger and because drug-metabolizing or eliminating pathways are more mature. A model can fit observed concentrations well while still having difficulty deciding how much of a clearance difference belongs to each process. The key question is therefore not only whether clearance varies with age and weight, but whether the available data contain enough independent variation to distinguish their effects.
The earlier material on allometric scaling and maturation functions provides standard ways to represent size and age effects. Here the focus is on their joint interpretation: how to diagnose confounding, how parameterization affects predictions, and what cannot be learned when age and size nearly determine one another.
A Joint Size-and-Maturation Model
For illustration, consider clearance \(CL_i\) for individual \(i\), body weight \(WT_i\) in kilograms, and postmenstrual age \(PMA_i\) in weeks. A common multiplicative model is
Here \(CL_{\mathrm{ref}}\) is clearance at reference weight \(WT_{\mathrm{ref}}\) when the maturation factor is one, \(F_{\mathrm{mat}}\) is a dimensionless maturation function, and \(\eta_{CL,i}\) represents an individual deviation on the log scale. For the examples, use \(WT_{\mathrm{ref}}=3.2\) kg and a Hill maturation function with fixed Hill coefficient 3 and \(TM_{50}=50\) weeks:
The allometric exponent and maturation function are choices that should be supported by the model’s scientific purpose and data. The equation does not establish that the exponent is exactly 0.75 or that a particular maturation function is correct. It makes explicit the two contributions whose separation is under assessment.
Interpretation of the product requires assumptions. The size and maturation terms are assumed to act multiplicatively on clearance; the chosen functional forms are assumed adequate over the studied range; and the relevant covariates must be measured with sufficient accuracy. A population model also needs an appropriate description of between-individual variability and residual error. Most importantly for this tutorial, there must be enough variation in weight at similar ages, and in age at similar weights, to distinguish the terms from one another.
Worked Example: A Covariate Plot Reveals Near-Perfect Confounding
Suppose four hypothetical infants have PMA values of 32, 40, 48, and 56 weeks. Their weights were generated by \(WT=3.2\exp[0.05(PMA-40)]\). Calculate the weights and assess whether these observations can separate age and size effects.
Substitution gives the following values:
A plot of weight against PMA would show a strictly increasing relationship, with every age paired with only one weight. There is no observation at, for example, 40 weeks with a different weight, or at 48 weeks with a weight near 3.2 kg. These data provide no direct covariate overlap with which to compare age while holding size approximately fixed, or size while holding age approximately fixed.
The four points can describe a combined trend in clearance along this particular growth trajectory. Without adequate overlap, separating the effects is often weakly identified and depends heavily on the assumed functional forms; whether the parameters are identifiable under a particular parametric model must be assessed for that model. More observations on the same deterministic relationship add information along that trajectory but do not create independent covariate variation, so their ability to distinguish effects depends on the model’s functional forms.
Why a Single-Covariate Model Can Mislead
A size-only model may attribute age-related maturation to weight, inflating or otherwise distorting the apparent size exponent. The reverse can happen when a maturation-only model attributes size-related clearance differences to age. The following derivation shows the issue for the size-only case.
Ignoring individual variability for the moment, take logs of the joint model:
If PMA and weight lie on a single trajectory, a size-only analysis can absorb the maturation term into its apparent weight slope. Between two points on that trajectory, the slope is
This is an apparent slope along the observed trajectory, not necessarily a biological size exponent. In a broader data set with imperfect correlation, the estimated slope will also depend on the observations’ distribution, model error, and other covariates.
Worked Example: Maturation Inflates an Apparent Size Exponent
Use the 32- and 56-week observations from the previous example. Suppose the true clearance follows the joint model with a size exponent of 0.75 and the stated maturation function. What size exponent would a size-only model imply from the two endpoints?
At 32 weeks, \(F_{\mathrm{mat}}=32^3/(32^3+50^3)=0.2077\). At 56 weeks, \(F_{\mathrm{mat}}=56^3/(56^3+50^3)=0.5842\). The log weight ratio is \(0.05(56-32)=1.200\), because the generating weight equation gives \(WT_{56}/WT_{32}=\exp(1.200)\). Thus,
A size-only model would therefore describe the endpoint trend with an apparent exponent of about 1.61, even though the generating size exponent was 0.75. The difference arises because maturation also increased along the same trajectory. This two-point calculation is a demonstration of confounding, not a recommended method for estimating an exponent from sparse data.
Use Contrasting Predictions to Check the Parameterization
A useful model diagnostic is to compare predictions where one covariate changes and the other is held fixed. This does not substitute for data support: a contrast can expose what a model assumes, but if the data contain no comparable individuals, the contrast may be extrapolation. Still, it clarifies whether the selected parameterization has the intended implications.
Worked Example: Separating the Model’s Age and Weight Predictions
Set \(CL_{\mathrm{ref}}=4.00\) L/h and \(\eta_{CL}=0\). Compare predictions at the same weight of 3.2 kg for PMA 32 and 56 weeks, then at the same PMA of 40 weeks for weights 3.2 and 6.4 kg.
At 3.2 kg, the size factor is one. Therefore the predictions at 32 and 56 weeks are:
At 40 weeks, \(F_{\mathrm{mat}}=40^3/(40^3+50^3)=0.3386\). Doubling weight from 3.2 to 6.4 kg multiplies the size term by \(2^{0.75}=1.682\). Hence:
These contrasts show the separate effects encoded by the model: at fixed weight the maturation function changes clearance, and at fixed age the allometric term changes clearance. They do not show that the hypothetical data can estimate those effects. For the perfectly linked covariates in the prior example, predictions for age-weight combinations away from the observed trajectory depend heavily on the chosen structure.
Diagnosing Confounding in Practice
Begin with the covariates, not just the fitted coefficients. Plot weight against the developmental-age measure used in the model and inspect the density of observations across the two-dimensional space. Then examine clearance-related predictions or residual patterns against each covariate while accounting for the other. Useful displays include weight versus PMA with points colored by study or age band, and stratified plots of clearance or model residuals against weight within relatively narrow age bands.
A weakly identified model may show large standard errors, strong correlations between estimated covariate parameters, sensitivity to starting values, or materially different estimates when a few observations are removed. Profile likelihoods or bootstrap distributions may reveal broad, skewed, or boundary-seeking estimates. These patterns are warnings, not unique proof of confounding: misspecified residual error, influential observations, or incorrect structural assumptions can cause similar symptoms.
Individual empirical Bayes clearance estimates can be useful for exploratory displays, but they are shrunk toward the population distribution and have unequal precision. A smooth-looking plot of those estimates does not establish a covariate effect. Prefer diagnostics based on the fitted model and its residuals, and interpret parameter uncertainty alongside the covariate support.
Plot age against weight and identify whether observations overlap across ages and sizes, rather than relying only on a correlation coefficient.
Compare size-only, maturation-only, and joint models, keeping the observation and variability structures comparable.
Check parameter uncertainty and dependence, sensitivity to influential records, and systematic residual patterns across age and weight.
Compare model predictions at fixed age or fixed weight, while distinguishing interpolation within observed support from extrapolation beyond it.
Parameterization Choices and Their Limits
| Approach | Useful when | Main risk |
|---|---|---|
| Size-only | Age variation is narrow, or maturation is not relevant over the studied range. | Age-related changes can be absorbed into the apparent size effect. |
| Maturation-only | Weight variation is narrow or already handled by a justified design or model component. | Size-related differences can be misattributed to developmental age. |
| Separate multiplicative size and maturation terms | There is sufficient age-weight overlap and a defensible basis for both functions. | Highly correlated covariates can make parameters unstable even when the model fits well. |
| Age-dependent size exponent or interaction | There is a specific scientific rationale and data support for the size effect changing with development. | Extra flexibility increases identification demands and can overfit sparse pediatric data. |
Fixing a well-supported allometric exponent, using an informative prior, or simplifying one covariate function may stabilize estimation. These choices reduce flexibility; they do not create information absent from the data. A fixed exponent should be treated as an assumption and examined in sensitivity analyses when conclusions depend on it. Likewise, adding an interaction term is not a general cure for confounding: it requires even more independent information.
When the age-weight relationship is nearly deterministic, report the limitation directly. The data may support predictions for children along the observed growth trajectory but not reliable predictions for unusual combinations of age and size. New data that broaden covariate overlap—for example, observations at similar ages but different weights—can be more informative for separating effects than simply adding more individuals along the same trajectory.
Common Pitfall
Wrong reasoning: “Both weight and age have significant-looking trends with clearance, so each must have an independently estimated effect.” When weight and age move together, two separate marginal trends can reflect the same underlying covariate pattern. A good overall fit does not prove that the model has distinguished the contributions.
Correction: Examine the joint covariate distribution and the conditional comparisons the model needs to make. If the data do not contain age variation at similar weights and weight variation at similar ages, separate effects may be weakly identified or unidentified. State which conclusions are supported by observations and which depend on structural assumptions or extrapolation.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- Why can adding many observations along an exact age-weight relationship fail to identify separate size and maturation effects?
- In a log-clearance model, what does the apparent size slope represent when maturation is omitted and age increases with weight?
- For the joint model in this tutorial, what happens to predicted clearance when weight doubles at fixed age?
- Name two diagnostics that can indicate weak separation of covariate effects, and explain why neither alone proves confounding.
Solutions. (1) The added observations improve information about the combined trend along that relationship but add no independent variation in either covariate. (2) It includes the size effect plus the maturation change associated with the age-weight trajectory, so it need not equal the biological size exponent. (3) With the exponent fixed at 0.75 and maturation unchanged, clearance is multiplied by \(2^{0.75}=1.682\). (4) Examples include strong parameter correlation, broad profile likelihoods, and sensitivity to influential observations. These can also arise from model misspecification, error structure, or data quality problems, so they must be interpreted with covariate plots and model checks.