Nonlinear Age Effects Beyond Childhood

Why Model Age Beyond Childhood?

Age can remain associated with pharmacokinetic parameters during adolescence and adulthood, even when the process of rapid developmental maturation is not the modeling question. For example, a parameter may change gradually with age, flatten over part of the observed range, or show a curved association. A straight line can miss such patterns, while applying a pediatric maturation curve to adults can impose a shape and biological interpretation that the data do not justify.

The objective here is to represent the observed age relationship flexibly and interpretably, not to assume that every age effect is caused by a specific biological mechanism. Age is often a proxy for several changing characteristics. An estimated age association is conditional on the other covariates in the model and does not, by itself, establish causation.

Let \(A_i\) denote the age in years of individual \(i\), and let \(P_i\) be a positive PK parameter such as clearance (CL). One convenient specification models the logarithm of the parameter, so predictions remain positive. Define a reference age \(A_{\mathrm{ref}}\), and write

Here \(P_{\mathrm{ref}}\) is the typical parameter at the reference age when the age function is zero, \(f(A_i)\) is the fixed age effect, and \(\eta_i\) represents individual variation in the parameter. Other covariate terms can be included in the model; age predictions discussed below are interpreted with those terms held fixed. Exponentiating the fixed-effect prediction gives \(P_{\mathrm{ref}}\exp\{f(A_i)\}\). Thus, an age effect on the log scale is multiplicative on the parameter scale.

Quadratic Age Effects

A quadratic term is a compact way to represent a curved relationship across a limited, well-observed age range. Centering and scaling age makes coefficients easier to interpret and often improves numerical stability. For example, define \(x=(A-A_{\mathrm{ref}})/10\), so \(x\) is the age difference from the reference in decades. A quadratic model is

At the reference age, \(x=0\), so the age contribution is zero. The coefficient \(\beta_1\) is the local slope of log parameter with respect to \(x\) at the reference age; \(\beta_2\) determines curvature. Neither coefficient alone is the percent change per year throughout the age range. The age-specific slope is \(\beta_1+2\beta_2x\) per decade, so it varies with age.

What the Quadratic Turning Point Means

A quadratic model has a stationary point where its slope is zero. Because exponentiation is monotone, the turning point of the predicted parameter occurs at the same age as the turning point of its log prediction. Differentiating the age contribution with respect to \(x\) gives

Setting this derivative to zero and solving yields \(x^*=-\beta_1/(2\beta_2)\), provided \(\beta_2\ne 0\). If \(\beta_2>0\), the point is a minimum; if \(\beta_2<0\), it is a maximum. For the example above, \(x^*=-(-0.060)/(2(0.015))=2\), which corresponds to age \(40+10(2)=60\) years. This matches the calculated predictions.

The turning point is meaningful only if it falls within a reasonably supported part of the observed age range. A fitted quadratic may mathematically turn far outside that range, but that does not establish a real age pattern there. Also, a small change in coefficients can move the turning point substantially when curvature is weak. Report predictions and their uncertainty at relevant ages rather than presenting the turning-point estimate as a precise biological threshold.

Alternatives for Curved Relationships

A quadratic is not the only option. A transformed age term can produce a nonlinear relationship on the age scale while remaining linear in its coefficient. For instance, \(\log(A/A_{\mathrm{ref}})\) gives a curve that changes more rapidly at younger ages and more gradually at older ages. This form is monotonic when its coefficient is nonzero; unlike a quadratic, it cannot represent a turning point.

A continuous piecewise-linear model allows the slope to change at a chosen knot while keeping predictions continuous. For a knot \(K\), one hinge term is \(\max(0,A-K)\). Multiple knots or a spline basis can provide more flexibility. A spline represents the age effect as a weighted sum of basis functions; the resulting curve is generally smooth, but its shape depends on the basis and knot choices. It is not necessary to interpret each basis coefficient in isolation. Instead, examine predicted parameters across age.

Choosing a Representation and Checking Its Behavior

The model should be flexible enough to capture a plausible pattern but not so flexible that it follows sampling noise. A linear age term is a useful baseline. A quadratic adds one curvature term; a prespecified transformed age or a spline can represent other shapes. Candidate models should be compared on the same data and with the same non-age model structure. For nested models, a likelihood-ratio test can assess whether added terms improve fit, under the usual regularity assumptions for that test. Information criteria such as AIC can compare models with different numbers of parameters, but neither a small p-value nor a lower AIC establishes biological truth.

Before interpreting the fitted relationship, inspect the age distribution and plot model predictions against age with uncertainty, alongside appropriately summarized data. Check whether the curve is supported throughout the range or is driven by a few observations near an endpoint. Examine residuals or individual parameter estimates for remaining age trends, and assess whether the model’s random-effects and residual assumptions remain adequate. A smooth-looking curve alone is not evidence of adequate fit.

Common warning signs include unstable coefficients when the polynomial order changes, extreme predictions near the edges, a turning point outside the data range, and large differences between plausible functional forms. Sparse coverage in a section of the age range can make even a low-dimensional curve poorly identified there. When these occur, consider a simpler form, report the uncertainty and supported range clearly, or avoid making predictions in the unsupported region.

A pediatric maturation function is appropriate when the scientific question and data concern developmental maturation and its assumptions are defensible. For adolescent or adult age patterns, polynomial, transformed-age, or spline terms may be more suitable descriptions of the observed association. A more flexible spline can reduce dependence on a particular global shape, but it also requires enough age information and careful control of complexity. No functional form removes the need to consider what ages were observed and how uncertainty changes across them.