Separating Disease Progression From Developmental Change

Why Separating Trajectories Matters

A pediatric outcome can change over time even if treatment has no effect. A developmental score may increase as a child matures, while a disease process simultaneously slows that increase or causes decline. The observed trajectory is therefore not, by itself, evidence of either disease progression or treatment benefit. A useful analysis distinguishes normal developmental change, disease-related change, and the effect of treatment.

This distinction is especially important when the outcome scale is age-sensitive. For example, a raw score may rise with age in healthy children, while the same score in affected children rises more slowly. A treated child’s score might still rise, yet less than it would have under untreated disease. Conversely, a stable score could represent benefit if untreated children of the same age and disease duration typically decline.

The earlier material on developmental PK-PD models of pediatric response considers how response relates to drug exposure and developmental characteristics. Here the focus is the longitudinal outcome trajectory and its counterfactual: what would be expected at the same ages and follow-up times without treatment?

Representing Development, Disease, and Treatment

Let \(Y_{ij}\) denote the outcome for child \(i\) at visit \(j\), \(a_{ij}\) the child’s age, \(t_{ij}\) time since the study baseline, and \(d_{ij}\) disease duration. Let \(A_i=1\) indicate assignment to treatment and \(A_i=0\) indicate assignment to placebo or no treatment. A conceptual model is:

Here \(b_i\) is the child’s expected baseline level, \(M(a)\) describes normal development on the chosen outcome scale, \(G(d)\) describes disease-related change in the untreated state, and \(\tau\) is the treatment effect relative to that untreated state. The residual \(\varepsilon_{ij}\) captures remaining measurement and model error; repeated measurements within a child generally require an appropriate within-child covariance model. The functions may be nonlinear, and the treatment effect may depend on follow-up time or age at treatment initiation.

This decomposition is a modeling framework, not a guarantee that each component can be uniquely estimated. Its interpretation depends on the data and assumptions. In particular, \(M\) and \(G\) may be difficult to distinguish if age and disease duration move together, or if there is no suitable reference information about normal development. A model can fit observed data well and still fail to identify the components of scientific interest.

What a Placebo-Controlled Study Identifies

In a randomized placebo-controlled study, the placebo group provides an estimate of the untreated disease trajectory under the study’s background care and measurement schedule. It does not necessarily represent healthy development: placebo participants have the disease, so their observed change can include both normal maturation and disease progression. Randomization makes the treatment and placebo groups comparable on average at baseline, including on measured and unmeasured prognostic factors, subject to chance imbalance and trial conduct.

For a simple illustration, suppose the expected change in each arm is the sum of a shared developmental component and a shared untreated disease component, plus a treatment effect in the active arm. The difference in mean changes then removes the shared components. This is the basis of a treatment-versus-placebo contrast; the components need not be estimated separately to estimate that contrast.

A longitudinal model can use all available visits rather than only baseline and endpoint. It may include a flexible visit-time function, treatment-by-time terms, baseline outcome, and prespecified age terms. The treatment-by-time component represents how the treatment-placebo contrast evolves. If treatment is expected to alter the rate of progression, the contrast can grow with time; if onset is delayed, a model with an immediate constant effect may be unsuitable.

Using Natural-History Information

When a concurrent placebo group is unavailable or too small, an untreated natural-history model may be used to predict the outcome trajectory expected without treatment. Such information can come from an untreated cohort with the disease, a prior study, or a suitable registry. A healthy reference trajectory can help characterize normal development, but it is not a substitute for affected untreated natural history: healthy children do not supply the disease component.

An external natural-history comparison requires assumptions beyond those needed for a randomized concurrent control. The external and treated populations must be sufficiently comparable in age, disease duration, baseline severity, outcome definition, measurement schedule, supportive care, and factors influencing prognosis. Differences in enrollment, assessment, or care can create apparent treatment effects. Modeling measured covariates can reduce some differences but cannot ensure comparability on unmeasured factors.

If healthy reference data support an age-specific expected outcome \(M(a)\), one may express an affected child’s observed outcome relative to that reference. For example, an age-standardized score can remove the average age trend represented by the reference. But it does not remove disease progression, and it does not automatically make an external disease cohort comparable to a trial population. The reference’s population, scale, and measurement procedures must be appropriate to the children being analyzed.

Age, Follow-Up, and Disease Duration Can Be Confounded

For each child, age at a visit equals baseline age plus time since baseline: \(a_{ij}=a_{i0}+t_{ij}\). Thus, within a child, age and follow-up time are not independent variables. If every child has the same disease duration at enrollment and disease duration increases one-for-one with follow-up, duration and follow-up are also closely linked. A model that assigns separate flexible effects to these quantities may not be able to distinguish them from the observed data alone.

Identification can improve when there is meaningful variation in baseline age or disease duration, repeated observations across a suitable age range, and reliable information on untreated development. Even then, age and disease effects may remain correlated. Scientific knowledge can motivate a parsimonious shape for \(M(a)\) or \(G(d)\), but imposing a shape is an assumption and should be checked against the available observations.

A practical way to expose the issue is to ask what comparisons the design actually supplies. Does it observe affected children of similar ages but different disease durations? Does it include untreated observations over the ages represented in the treated cohort? Are there observations before and after a developmental transition? If the answer is no, the separate developmental and disease components may be weakly identified even if the overall predicted trajectory is precise.

Model Checks, Alternatives, and Limitations

Diagnostics should examine both the overall fit and whether the model captures the distinctions it is intended to make. Plot observed outcomes over time by treatment and age group, and compare them with model-predicted trajectories. For natural-history analyses, also display the external untreated observations and the region over which predictions are being used. Look for systematic residual patterns by age, disease duration, baseline severity, visit, or treatment arm; such patterns may signal an inadequate trajectory shape, missing effect modification, or a measurement issue.

Check whether predictions outside the observed age or disease-duration range are driving the treatment contrast. Examine sensitivity to reasonable alternative developmental and progression functions, baseline adjustment, and assumptions about external comparability. For longitudinal data, assess whether dropout or missed visits are related to health status: a model based on observed measurements can be biased if worsening children are more likely to leave and the analysis does not adequately address that process. Diagnostics cannot prove assumptions such as exchangeability with an external cohort, but they can reveal incompatibilities and dependence on modeling choices.

The strongest option is usually the one that supplies the most credible untreated counterfactual for the intended population. A concurrent randomized control often supports the clearest causal comparison. Natural-history models can be valuable when a concurrent control is infeasible, but their conclusions are more dependent on transportability and model assumptions. A pre-post analysis alone is generally insufficient to attribute change to treatment.

Common Pitfall

A related error is to assume that age adjustment automatically separates development from disease progression. Age standardization can account for a specified reference trend, but disease-related change remains, and the result depends on how well the reference applies. The treatment contrast must still be defined against an appropriate untreated outcome.