What Is Being Transported?
A covariate model estimated in one population may be used to predict pharmacokinetic (PK) parameters in another. The central question here is narrower than whether an entire population PK model transports: does the estimated relationship between a covariate and clearance or volume remain applicable over the target population’s covariate values? A coefficient can be estimated precisely yet be used to predict far outside the data that informed it. Conversely, a target population may lie within the source data’s range while the covariate effect remains imprecise.
The assessment therefore needs more than a point estimate of a covariate coefficient. It considers uncertainty in the relevant parameter contrast, the source and target covariate distributions, and how well predictions agree with appropriate external PK observations. As in the earlier material on transporting population pharmacokinetic models, the result is conditional on the assumptions connecting the source and target populations.
State the Model and the Transport Assumptions
For a continuous covariate \(W_i\), one possible model for individual clearance is
Here \(CL_i\) is individual clearance, \(CL_{\mathrm{ref}}\) is typical clearance at reference covariate value \(W_{\mathrm{ref}}\), \(\beta\) is the covariate coefficient, and \(\eta_i\) represents between-individual variation under the specified model. The same structure can be used for volume \(V_i\), with its own reference value and coefficient. For this model, a contrast in typical clearance between \(W_a\) and \(W_b\) is
Transporting this effect requires assuming, among other things, that the covariate is defined and measured comparably, the functional form is suitable over the target values, and the conditional relationship represented by \(\beta\) is sufficiently similar in source and target. It also requires that important effect modifiers or population differences have not been omitted in a way that changes the relationship. The structural PK model and data collection must support estimation of the covariate effect; for example, a covariate should not be so confounded with another model feature that its coefficient is weakly identified.
These are substantive assumptions, not consequences of a small standard error. The target covariate distribution must also be considered. A coefficient supported by source observations near the target values is being interpolated; applying it well beyond the observed source support is extrapolation. With several covariates, support concerns their joint combinations, not just each covariate’s separate minimum and maximum.
Quantifying Uncertainty in a Covariate Contrast
For the log-linear model above, let \(\hat{\beta}\) be the source estimate and \(SE(\hat{\beta})\) its standard error. If the reference parameter is common to both covariate values, it cancels from the ratio. The estimated log ratio is \(\hat{\beta}\log(W_a/W_b)\), with standard error \(|\log(W_a/W_b)|SE(\hat{\beta})\). An approximate 95% interval uses the estimate plus or minus 1.96 standard errors, provided the estimator is approximately normal and the model-based uncertainty is adequate.
This contrast formulation matters: it reports the difference relevant to the transport question, not a coefficient in isolation. In a model with multiple covariates, the uncertainty of a linear contrast depends on the covariance among parameter estimates. If \(\boldsymbol{\theta}\) is the parameter vector and \(\mathbf{c}\) defines the contrast, then the contrast variance is \(\mathbf{c}^{T}\mathbf{\Sigma}\mathbf{c}\), where \(\mathbf{\Sigma}\) is the estimated covariance matrix. Ignoring covariance can overstate or understate uncertainty.
Worked Example: Uncertainty in a Clearance Ratio
Suppose a hypothetical source analysis uses \(W_{\mathrm{ref}}=40\) kg and estimates \(\hat{\beta}=0.60\) for body weight in the log-clearance model. Assume \(SE(\hat{\beta})=0.20\), the estimate is approximately normal, and the model is considered suitable for this contrast. Calculate the typical clearance ratio at 20 kg versus 40 kg and an approximate 95% interval.
The point estimate on the log scale is
The standard error of the log contrast is
Thus the approximate 95% log-scale interval is
Exponentiating the estimate gives \(\exp(-0.4158883)=0.6598\). Exponentiating the unrounded interval endpoints gives a clearance-ratio interval of approximately \([0.503,0.866]\). Under the model, typical clearance at 20 kg is estimated to be 0.660 times typical clearance at 40 kg; the approximate interval corresponds to ratios from 0.503 to 0.866. This quantifies uncertainty in the model-based contrast. It does not establish that the relationship is transportable to a target population containing 20-kg individuals.
Covariate Support: Interpolation or Extrapolation?
A useful first diagnostic is to compare source and target covariate distributions directly. Plot or tabulate their ranges, quantiles, and density, and identify the portion of the target population that lies outside source support. The target’s covariate values should be evaluated against the values that actually informed the model, not only a nominal range. A few source observations at an extreme may provide little information about the slope there.
For multiple covariates, marginal overlap can be misleading. For example, the source data may include older individuals and may include individuals with impaired renal function, while containing almost no older individuals with impaired renal function together. Predictions for that combination are extrapolative in the joint covariate space. A covariate-effect estimate also should not be treated as evidence for target values if the source analysis had poor precision or inadequate design support in that region.
Worked Example: Assessing a Target Weight Range
In a hypothetical PK analysis, the source data contain body weights from 35 to 75 kg, with most observations between 45 and 65 kg. A target population has weights from 50 to 70 kg. The estimated weight coefficient is \(\hat{\beta}=0.45\) with \(SE(\hat{\beta})=0.12\). Assume the log-linear form is appropriate in this region and that the source and target measurements are comparable. What can be said about predicting typical clearance at 60 kg relative to 40 kg?
Both 60 kg and 40 kg fall within the source range, so this particular contrast is interpolation by the range criterion, although 40 kg may be less densely represented than the central source values. The estimated log contrast and its standard error are
The typical clearance ratio is \(\exp(0.1824593)=1.200\). The approximate 95% log interval is \(0.1824593\pm1.96(0.0486558)=[0.0871,0.2778]\), which exponentiates to approximately \([1.091,1.320]\). Thus the model estimates typical clearance at 60 kg to be about 20.0% higher than at 40 kg, with an approximate interval of ratios from 1.091 to 1.320. This interval reflects coefficient uncertainty under the stated model; it does not account for every possible structural or population difference. The target range is largely represented by the source range, but the density and joint covariate support still need review before extending that conclusion to the full target population.
External Predictive Checks
An external predictive check asks whether the transported model predicts relevant PK data in the target population adequately. The check should be defined before comparing predictions and observations: specify the target subgroup, dose and regimen, sampling design, PK summary or concentration-time feature, and the uncertainty expected in that summary. Useful quantities may include clearance or volume summaries, concentration quantiles at prespecified times, or exposure summaries derived from the PK model.
Predictions must include the uncertainty relevant to the check. Depending on the aim, that can include parameter uncertainty, between-individual variability, target covariate sampling, and the observation or estimation process. Comparing an observed target summary only with a single typical prediction can make an uncertain prediction appear more precise than it is. Conversely, a predictive interval for individual observations should not be compared as though it were an interval for a population mean.
Worked Example: Checking a Target Clearance Summary
Suppose a hypothetical external target cohort has an observed geometric mean clearance of 6.2 L/h in a prespecified subgroup. Simulations from the transported model produce a predictive distribution for the cohort’s log geometric mean with mean \(\log(5.0\text{ L/h})\) and standard deviation 0.18. Assume this predictive standard deviation includes the parameter and cohort-summary uncertainty relevant to this comparison, and that the log summary is adequately approximated by a normal distribution.
The observed difference from the predicted center on the log scale is
The standardized discrepancy is \(0.2151114/0.18=1.195\). A central approximate 95% predictive interval on the log scale is
On the clearance scale this is \(5.0[\exp(-0.3528),\exp(0.3528)]\), or approximately \([3.51,7.12]\) L/h. The observed 6.2 L/h lies within that interval. This is evidence that this particular target summary is not unusually discrepant under the stated predictive distribution. It is not proof that the covariate effect is correct or transportable: other target subgroups, exposure features, or data sources could reveal a mismatch, and a broad predictive interval may accommodate substantial uncertainty.
Derivation and a Practical Assessment
The uncertainty calculation follows directly from the fitted contrast. Write \(z_a=\log(W_a/W_{\mathrm{ref}})\) and \(z_b=\log(W_b/W_{\mathrm{ref}})\). The contrast is \(\delta=\beta(z_a-z_b)\). If \(\hat{\beta}\) is approximately normal with variance \(SE(\hat{\beta})^2\), multiplying it by the fixed quantity \(z_a-z_b\) gives
This explains why a coefficient’s uncertainty cannot be interpreted without specifying the contrast: as the covariate values move farther apart on the model scale, uncertainty in the estimated difference increases. For several parameters, the same reasoning gives \(\operatorname{Var}(\mathbf{c}^{T}\hat{\boldsymbol{\theta}})=\mathbf{c}^{T}\mathbf{\Sigma}\mathbf{c}\).
A defensible assessment can be organized as a sequence:
Specify the target population, covariate contrasts or distribution, PK parameter, and intended prediction or decision.
Calculate intervals for relevant parameter contrasts using the estimated covariance, and report the scale and units of the resulting ratio or difference.
Compare source and target covariate distributions, including important combinations; label predictions beyond meaningful source support as extrapolations.
Compare prespecified target PK summaries with predictive distributions that reflect the uncertainty appropriate to those summaries.
Alternatives, Diagnostics, and Limitations
Transporting the source coefficient is one strategy, not the only one. Re-estimating a covariate relationship using target data may be preferable when sufficiently informative target PK data are available. A partial update can retain the source model structure while estimating selected target parameters; this changes the evidence supporting the relationship and should be stated clearly. Comparing source-only predictions with target observations is an external check, whereas fitting the target observations and then reporting their agreement with the fitted model is not an independent predictive check.
Warning signs include large coefficient uncertainty, sparse source observations near target values, covariate combinations absent from the source, unstable estimates under reasonable model or data perturbations, and systematic external predictive discrepancies. A model can pass a check for one summary but miss time-dependent concentration patterns or another target subgroup. These diagnostics indicate where the transport claim is weak; none alone identifies the cause.
A common mistake is to say, “The source coefficient is statistically precise, so it applies to the target.” Precision concerns uncertainty under the source model and data. It does not establish adequate target support, correct functional form, comparable covariate measurement, or an invariant conditional relationship. The correction is to qualify the claim by the target covariate region, report uncertainty in the relevant contrast, and evaluate target predictions with external evidence where possible.
| Evidence | What it supports | What it cannot establish alone |
|---|---|---|
| Narrow interval for a covariate contrast | Precision of that contrast under the fitted model | Correct model form or target applicability |
| Target values within source support | Interpolation rather than range-based extrapolation | Support for every joint covariate combination or invariant effects |
| External target summary within a predictive interval | Compatibility of that summary with the predictive distribution | Overall model validity or transportability to all target outcomes |
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A log-linear model has \(\hat{\beta}=0.30\) and \(SE(\hat{\beta})=0.10\). What is the estimated typical clearance ratio at \(W=80\) versus \(W=40\), and its approximate 95% interval?
- The source range for a renal covariate is 10 to 80 units. The target range is 20 to 70 units, but nearly all source observations with high age have renal values below 30. Why may a high-age, renal-value-65 prediction still be extrapolative?
- An external observed target summary lies inside a broad predictive interval. What conclusion is justified, and what conclusion is not?
- Why should parameter covariance be used when calculating uncertainty for a contrast involving two estimated covariate coefficients?
Solutions. (1) The log ratio is \(0.30\log(80/40)=0.2079\), so the ratio is \(\exp(0.2079)=1.231\). Its log-scale standard error is \(0.10\log 2=0.06931\); the interval is \(0.2079\pm1.96(0.06931)=[0.0721,0.3437]\), or approximately \([1.075,1.410]\) after exponentiation. (2) The individual renal value is within the marginal range, but the high-age/high-renal combination is poorly represented. Support must be considered jointly. (3) The observed summary is compatible with that predictive distribution for that check; agreement does not prove the model or the covariate effect is transportable more broadly. (4) The contrast variance includes covariance terms, so omitting them can produce an incorrect standard error and interval.