Age-by-Size Interactions in PK Models

Why Model an Age-by-Size Interaction?

In a population PK model, age and body size can both be associated with a parameter such as clearance. A model with independent age and size effects assumes that the size relationship is the same at every age, and that the age relationship is the same at every weight. An age-by-size interaction relaxes that restriction: the size association can change with age, or equivalently, the age association can change with size.

This is a question about the form of a covariate relationship, not simply whether age and weight are correlated. For example, the model may describe a stronger size association at older ages even if age and weight are measured independently. Conversely, a data set can have strongly correlated age and weight without supporting an interaction. Evidence for an interaction depends on the observations and the model, not on the correlation alone.

A Reference-Based Interaction Model

For a positive parameter such as clearance, one convenient specification models the logarithm of its typical value. Let \(CL_{\mathrm{typ}}\) denote typical clearance for individuals with specified covariates, \(W\) weight in kg, and \(A\) age in years. Choose reference values \(W_{\mathrm{ref}}\) and \(A_{\mathrm{ref}}\), and define the dimensionless covariates

The age covariate is expressed per decade from the reference age; the size covariate is the natural logarithm of the weight ratio. A model that includes both main effects and their interaction is

Here \(CL_{\mathrm{ref}}\) is typical clearance at \(A=A_{\mathrm{ref}}\) and \(W=W_{\mathrm{ref}}\). The coefficients \(\theta_W\) and \(\theta_A\) describe the size and age effects at their respective reference covariate values. The coefficient \(\theta_{AW}\) describes how those effects change across the other covariate. Because both covariates in this specification are dimensionless, the coefficients in the exponential are dimensionless as well.

At the reference age and weight, both covariates are zero, so the model gives \(CL_{\mathrm{typ}}=CL_{\mathrm{ref}}\). This centering makes the reference parameter interpretable and ensures that the interaction term does not alter the prediction at that point. Changing the reference values changes the interpretation of the main-effect coefficients, but does not by itself imply a different biological relationship.

What the Interaction Changes

The effect of size at a fixed age can be seen by comparing two weights while holding age constant. If weight changes from \(W_1\) to \(W_2\), the model gives

Thus, the coefficient multiplying the log weight contrast is \(\theta_W+\theta_{AW}z_A\). It varies with age when \(\theta_{AW}\ne 0\). In particular, the predicted ratio for a doubling of weight at age \(A\) is \(2^{\theta_W+\theta_{AW}z_A}\).

The same idea can be expressed in the other direction. Comparing two ages at fixed weight gives

The age effect therefore varies with weight. If \(\theta_{AW}=0\), the model reduces to independent multiplicative age and size effects: the weight contrast does not depend on age, and the age contrast does not depend on weight. A nonzero interaction means those contrasts vary according to the specified model.

Derivation and Interpretation of the Contrasts

The contrast formulas follow by taking the ratio of two predictions. At fixed age, subtract the log predictions at \(W_1\) and \(W_2\). Terms containing only age cancel, leaving

The difference of logarithms is \(\ln(W_2/W_1)\). Exponentiating both sides gives the size-contrast formula above. The age-contrast formula follows in the same way by subtracting log predictions at \(A_1\) and \(A_2\), holding weight fixed. This derivation also shows why the interaction term cannot be interpreted in isolation: its contribution depends on both covariate values.

A positive \(\theta_{AW}\) means that the coefficient for the size contrast increases as \(z_A\) increases. A negative \(\theta_{AW}\) means it decreases as \(z_A\) increases. Neither sign alone says that clearance itself rises or falls with age or weight: the main-effect coefficients and the covariate values also contribute. For example, the size contrast coefficient at a particular age is \(\theta_W+\theta_{AW}z_A\), not \(\theta_{AW}\) alone.

Assumptions, Diagnostics, and Alternatives

The interaction is useful only if its functional form and the data support the intended interpretation. The displayed model assumes a log-linear relationship in the chosen covariates, positive typical clearance, and a stable set of coefficients over the range where predictions are made. It also assumes that the selected age measure and weight measure are appropriate for the modeling question. Other covariates, structural PK components, and between-subject variability are not represented in the displayed covariate function.

In a full population model, individual clearance may be represented as \(CL_i=CL_{\mathrm{typ}}(A_i,W_i)\exp(\eta_i)\), where \(\eta_i\) is an individual random effect. The examples calculate \(CL_{\mathrm{typ}}\), not \(CL_i\). They also treat the coefficients as fixed supplied values; parameter uncertainty is not included in the arithmetic.

A useful comparison is the nested model with the interaction coefficient fixed at zero. The two models share the same reference values and main effects; the restricted model asserts that the age and size effects are independent on the specified log scale. In practice, model assessment should consider uncertainty in the interaction estimate, the fit to the data, and whether the interaction improves predictions in a way that is scientifically and clinically relevant. A numerical nonzero estimate alone is not sufficient.

Potential warning signs include systematic residual patterns by age and weight, unstable interaction estimates, large changes when a few observations are removed, and wide uncertainty in regions with few subjects. A fitted interaction may appear compelling if the data contain little overlap in age and weight. The earlier material on Identifiability of Size and Maturation Effects in Population PK is relevant to that concern: covariate patterns can make separate effects difficult to distinguish. Examine the joint age-weight coverage and uncertainty, rather than relying on a single coefficient or fit statistic.

Other model forms may be preferable when the age relationship is curved, maturation is better represented by a maturation function, or the size relationship is not adequately represented by the chosen term. An interaction is not a substitute for checking the main-effect functions. It also does not establish a mechanism or a causal effect. Predictions far beyond the observed covariate range are particularly uncertain, since the interaction can amplify or reverse a modeled contrast as age or weight changes.

Common Mistakes and Practical Checks

• Leaving out a main effect. A product term alone does not generally represent the intended model. Include the age and size main effects when specifying their interaction.
• Calling any age-weight correlation an interaction. Correlation describes the covariate data; an interaction describes how the modeled parameter relationship varies.
• Interpreting \(\theta_{AW}\) as the size effect. The size-contrast coefficient at age \(A\) is \(\theta_W+\theta_{AW}z_A\). Calculate it at the age of interest.
• Ignoring the reference values or units. State \(A_{\mathrm{ref}}\), \(W_{\mathrm{ref}}\), and the scaling of age. The main effects depend on those choices.
• Extrapolating without qualification. An algebraically valid prediction can still be scientifically unsupported if that age-weight combination is poorly represented in the data.

Before interpreting an interaction, verify the model equation, calculate the relevant contrast at the covariate values of interest, and check whether those combinations are covered by the data. Report the predicted ratio or parameter value with units and context, rather than describing the interaction coefficient alone.