Why a Covariate Model Can Stop Transporting
A population PK model may describe its development population well and still make poor predictions in a target population. The problem is not necessarily that the PK structure is wrong. A covariate may have been omitted, its effect may have the wrong functional form, two covariates may be too strongly associated to estimate their separate effects, or the target population may occupy covariate combinations poorly represented in the source data. These failures can look similar in a summary of average exposure, but they have different causes and call for different checks.
The aim here is to diagnose those causes in the covariate component of a population PK model. We use a simple log-clearance model to make the reasoning transparent; in a full nonlinear mixed-effects analysis, the same questions apply to the structural PK parameters, random effects, residual model, and resulting concentration-time predictions. A failure of transport is evidence that the current model and data do not support the intended predictions. By itself, it does not identify the cause or establish a biological explanation.
The assumptions in this working model are deliberately strong: the conditional mean is linear on the log-clearance scale, coefficients are shared across populations, and the residual has conditional mean zero. A real analysis must also assess whether the population PK structure, parameterization, sampling design, and variability assumptions are adequate. The simple model isolates covariate-model failure; it is not a substitute for checking concentration-time predictions.
Omitted Covariates and Conditional Shifts
Suppose the true mean depends on both \(X\) and \(Z\), but a fitted model includes \(X\) alone. In a given population, if the conditional mean of \(Z\) is approximately linear in \(X\), write \(E(Z\mid X)=a+bX\). Substitution into the true model gives the omitted-variable conditional mean:
Thus the slope in the omitted model is not generally \(\beta_X\). It combines the effect of \(X\) with the association between \(X\) and the omitted covariate. For a linear projection with an intercept, the same point follows directly from the covariance definition of the slope.
where the last equality uses \(\operatorname{Cov}(X,\varepsilon)=0\). If the association between \(X\) and \(Z\) differs across populations, the omitted-model slope can differ even when \(\beta_X\) and \(\gamma\) are unchanged. This is why an omitted-covariate problem can appear to be a change in the effect of \(X\).
For a comparison at a fixed value of \(X=x\), the relevant omitted-covariate shift is conditional, not automatically an unconditional difference in average \(Z\). Under the working model, the target-minus-source difference in conditional mean outcome attributable to \(Z\) is:
An unconditional difference in mean \(Z\) would represent this shift only under extra conditions, such as a constant conditional shift across \(X\) and a comparison standardized appropriately over \(X\). Equal overall means of \(Z\) do not rule out different conditional distributions.
Worked Example: An Omitted Covariate Changes the Transported Prediction
Hypothetical simulated data. Suppose source-population observations follow \(Y=1+0.4X+0.3Z\), with no residual error for this illustration, and the observed source covariate values satisfy \(Z=X\). The fitted source model omits \(Z\). In the target population, suppose \(Z=0\) at each value of \(X\).
Source fit. Since \(Z=X\) in the source, \(Y=1+(0.4+0.3)X=1+0.7X\). The omitted model therefore has intercept 1 and slope 0.7 in these data.
Prediction at \(X=1\). Transporting that fitted model predicts \(Y=1+0.7(1)=1.7\). The target conditional mean under the stated data-generating model is \(Y=1+0.4(1)+0.3(0)=1.4\). The prediction is high by \(1.7-1.4=0.3\) log units. On the clearance scale, the ratio of predicted to target clearance is \(\exp(0.3)=1.350\), so the transported prediction is about 35.0% higher.
At \(X=x\), the conditional \(Z\) difference is \(0-x=-x\); multiplying by \(\gamma=0.3\) gives a target-minus-source conditional mean shift of \(-0.3x\). This example shows why the source association between \(X\) and \(Z\), not just their separate marginal distributions, matters when a covariate is omitted.
Wrong Functional Form and Covariate-Range Shifts
A model can include the right covariate and still fail if it represents the relationship incorrectly. For example, a linear effect may be inadequate when the true relationship is curved, thresholded, or saturable. A flexible curve can fit the source data better, but a more flexible form is not automatically more transportable: its predictions outside the observed range may be especially uncertain.
Worked Example: A Linear Form Misses Curvature Beyond the Source Range
Hypothetical simulated data. Suppose the true mean is \(Y=0.2+0.5X+0.4X^2\). In the source, three equally weighted observations have \(X=-1,0,1\), giving \(Y=0.1,0.2,1.1\). A linear model is fitted to these three source observations.
The source means are \(\bar{x}=0\) and \(\bar{y}=(0.1+0.2+1.1)/3=0.4667\). The fitted slope is:
The intercept is \(\bar{y}-\hat{\beta}_X\bar{x}=0.4667\). At a target value \(X=2\), which is outside the source range, the linear model predicts \(0.4667+0.5(2)=1.4667\). The true mean under the stated quadratic model is \(0.2+0.5(2)+0.4(2^2)=2.8\). The linear prediction is therefore \(1.3333\) log units below the true mean. The example combines a wrong functional form with extrapolation beyond the data; source fit alone cannot distinguish these risks.
A range shift is not itself proof of model failure. If the relationship is truly linear and the coefficients are stable, the linear model can remain valid beyond the source range. The difficulty is evidential: source data cannot check the form where they contain no observations. Assess both marginal ranges and joint support, meaning whether target combinations of covariates are represented in the source. A target value of \(X\) may be common in the source overall yet unusual among people with the target’s values of \(Z\), or vice versa.
Collinearity: Uncertain Attribution and Fragile Extrapolation
Collinearity occurs when covariates in the fitted model are strongly associated. It can make the separate coefficients imprecise, even when predictions for the source covariate combinations are stable. In severe cases, the data do not identify separate effects at all. If a target population has different combinations of those covariates, predictions can become sensitive to coefficient estimates that fit the source equally well.
Worked Example: Collinearity Makes a Target Prediction Unstable
Hypothetical simulated data. Suppose the true mean is \(Y=1+0.2X+0.3Z\), but in the source every observed pair satisfies \(Z=X\). The observed source relationship is therefore \(Y=1+0.5X\). A model with separate coefficients for \(X\) and \(Z\) cannot identify their contributions from these source observations alone.
For example, each of the coefficient pairs \((\hat{\beta}_X,\hat{\gamma})=(0.5,0)\), \((0,0.5)\), and \((0.2,0.3)\) predicts the same source mean \(1+0.5X\). Now consider a target individual with \(X=1,Z=0\). The three fitted pairs predict, respectively, \(1.5\), \(1.0\), and \(1.2\). The true mean under the stated model is \(1+0.2(1)+0.3(0)=1.2\). Source fit cannot select among these target predictions because the source did not contain independent variation in \(X\) and \(Z\).
This exact-collinearity example is intentionally stark. With near-collinearity, coefficients are estimable but may have large uncertainty and strong dependence, so target predictions can still vary substantially. A small standard error for source fitted values does not establish precision for a target combination outside source support.
A Practical Diagnostic Sequence
When target predictions disagree with observed target data, first establish that the comparison is meaningful: the PK endpoint, time scale, dosing information, assay, and population definitions must be sufficiently aligned. Then investigate covariate-model failure without assuming that any one diagnostic proves its cause.
Compare observed and predicted PK summaries or concentration-time profiles in the target, including relevant covariate strata. Look for systematic residual patterns rather than relying only on an overall mean error.
Compare source and target ranges, joint combinations, and conditional distributions such as \(E(Z\mid X)\). Identify target strata with little or no source representation.
Plot residuals or prediction errors against candidate covariates and relevant transformations. A trend, curvature, or subgroup pattern may motivate a prespecified covariate extension, but should be assessed against validation data where possible.
Review covariate correlations, coefficient uncertainty, and sensitivity of target predictions to plausible coefficient values. If effects cannot be separated in the source, do not interpret a particular coefficient allocation as established.
Consider a supported alternative form, a reduced parameterization, or additional target information. Recheck predictions and uncertainty; document any target covariate regions where the evidence remains weak.
Residual trends and target prediction errors are diagnostic signals, not automatic tests of the mechanism. A model comparison or improved fit in the source may help assess functional form, but it cannot demonstrate transportability on its own. When target observations are sparse, report that limitation rather than treating the absence of a detected discrepancy as evidence of equivalence.
Choosing a Response and Recognizing Its Limits
| Failure pattern | Useful response | Important limitation |
|---|---|---|
| Omitted covariate with a plausible, supported relationship | Add the covariate or standardize predictions using its conditional target distribution. | The covariate’s effect and its relationship to included covariates must be estimable and relevant in the target. |
| Residual pattern suggesting curvature or a threshold | Evaluate a scientifically plausible nonlinear form, such as a transformation or spline, and validate it. | Extra flexibility can increase uncertainty and does not justify extrapolation beyond observed support. |
| Strong collinearity | Use a parameterization supported by the data, report joint prediction uncertainty, or seek observations that separate covariate effects. | Dropping one variable may simplify prediction but can leave an omitted-covariate bias if the target association differs. |
| Target covariates outside source support | Restrict claims to supported regions, obtain relevant data, or explicitly qualify model-based extrapolation. | No diagnostic based only on source fit can verify behavior in unobserved regions. |
The choice among a mechanistic covariate model, a simpler supported model, and a more flexible empirical form depends on scientific rationale and data support. A mechanistic explanation can guide parameterization but does not eliminate the need for validation. A reduced model may be more stable but can conceal a relevant omitted predictor. Recalibrating an intercept or slope to target observations can improve local prediction, yet it does not establish that the full covariate relationship transports to other target strata.
Common Pitfall
A second frequent error is to label every target prediction discrepancy a covariate effect. Discrepancies can also reflect misspecified structural PK or variability components, differences in data collection, or random sampling. Covariate diagnostics narrow the investigation; they do not replace assessment of the full model and the data-generating context. The practical objective is to identify where predictions are supported, where they are fragile, and what additional evidence would distinguish plausible explanations.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the omitted-covariate model, what source-population quantity contributes to the slope fitted using \(X\) alone?
- At fixed \(X=x\), what conditional comparison determines the target-minus-source mean shift due to an omitted \(Z\)?
- Why can a model have a good source fit but an unreliable prediction for a target combination of two covariates?
- Does a shift in the target covariate range prove that a linear model is wrong? Explain.
- What does strong collinearity imply about fitted coefficients and source predictions?
Solutions. (1) Under the stated linear model and zero covariance between \(X\) and the residual, the omitted-model slope is \(\beta_X+\gamma\operatorname{Cov}(X,Z)/\operatorname{Var}(X)\); the source association between \(X\) and \(Z\) therefore contributes. (2) It is \(\gamma\{E(Z\mid X=x,\text{target})-E(Z\mid X=x,\text{source})\}\), not generally a difference in unconditional means. (3) The source may contain little or no support for that joint combination, and collinearity can leave separate coefficients weakly identified even when fitted source values are stable. (4) No. A range shift creates an extrapolation concern, but the linear form could remain correct; the source data cannot validate it in regions they do not cover. (5) Separate coefficients may be imprecise or unidentified, while predictions for covariate combinations represented in the source can still be stable. Predictions for new combinations may be sensitive to the uncertain coefficient allocation.