Dose Selection Under Covariate Distribution Shift

Why the Covariate Distribution Matters

A dose can perform differently across covariate profiles because covariates such as body size, organ function, or disease severity influence exposure or response. Even if the relationship between covariates and target attainment is unchanged, the population probability of target attainment (PTA) can change when the intended treatment population has a different mix of profiles from the development sample. A dose recommendation based on the development sample may therefore not meet the same population target after transport to the intended population.

The issue is not resolved by comparing a few covariate subgroups alone. The goal is to average the model's covariate-conditional prediction over the covariate distribution of the population in which the dose will be used. This tutorial considers that change in covariate distribution while treating the fitted model and its conditional predictions as applicable to both populations.

The function \(g_d(x)\) can represent a probability that already integrates over individual random effects and relevant residual variability in the fitted model. Those distributions and the definition of target attainment must be appropriate for the intended population. Here the transport step changes \(F_T\), not the conditional model \(g_d(x)\). In particular, this approach does not by itself address a change in the covariate-response relationship, model parameters, adherence, or other features of care.

Two Ways to Average Over the Target Population

If covariate data are available from the intended population, the most direct approach is to simulate or predict for those covariate profiles and average their conditional PTAs. If only development-sample covariates are available, reweighting can estimate the target average, provided the target distribution is represented in the development sample.

The identity follows by changing the measure in the expectation:

For a development sample \(x_1,\ldots,x_n\), a common finite-sample estimate is the normalized weighted average

Normalization is useful when weights are estimated or do not sum exactly to \(n\). If covariates fall into strata \(k\), the same calculation can be written \(\sum_k \pi_{T,k}\hat g_{d,k}\), where \(\pi_{T,k}\) is the target-population proportion in stratum \(k\) and \(\hat g_{d,k}\) is the estimated conditional PTA there. This is direct standardization over strata; it does not require individual-level density estimates when stratum proportions and conditional PTAs are available.

Worked Dose-Selection Examples

With individual development-sample records, estimated density ratios can perform the same transport. For discrete strata, the ratio for stratum \(k\) is \(\pi_{T,k}/\pi_{S,k}\). These weights make profiles that are underrepresented in the development sample count more heavily in the target average.

A second route is to generate covariate profiles directly from the intended population distribution, then evaluate \(g_d(x)\) at each profile. This avoids estimating density ratios, but it requires a credible target-population covariate distribution. If the conditional model uses simulated random effects or other stochastic components, those should be generated according to the model assumptions for each profile.

Assumptions, Diagnostics, and Alternatives

Reweighting or direct target-population simulation is meaningful only when several conditions are credible. First, the covariates that determine differences in conditional target attainment must be represented: omitted effect modifiers can make \(g_d(x)\) wrong for transport. Second, conditional on the modeled covariates, the response model and relevant random-effect assumptions must apply to the target population. Third, reweighting requires overlap: target profiles cannot occur in regions wholly absent from the development sample. Finally, the target distribution itself must be estimated or specified well enough for the averaging step.

Useful diagnostics focus on whether the target average is supported by the available profiles and whether the numerical estimate is stable:

• Compare covariate distributions in the development and target populations, including joint patterns and clinically important tails, not only separate marginal means.
• Inspect the range and concentration of weights. A few extremely large weights mean the estimate depends heavily on a small number of profiles.
• Calculate the effective sample size \(n_{\mathrm{eff}}=(\sum_i w_i)^2/\sum_i w_i^2\). It is at most \(n\) for nonnegative weights and can be far smaller when weights are concentrated.
• Repeat the Monte Carlo calculation with more profiles or simulation draws and check whether estimated PTA and the dose decision are stable. Also inspect conditional predictions for implausible values or abrupt patterns across covariates.

When overlap is poor, trimming or capping extreme weights may stabilize an estimate, but it changes the effective target distribution and can introduce bias. It is not a substitute for support. Consider collecting target-population covariate data, restricting the target estimand to a defensible region of common support, or explicitly acknowledging that the model-based estimate extrapolates beyond the data.

Direct simulation from target covariates is often preferable when representative target profiles are available: it makes the target averaging explicit and avoids density-ratio estimation. Reweighting is useful when development profiles and a target distribution description are available but new individual profiles are not. Stratified standardization can be more transparent when a few meaningful strata capture the relevant differences. All three methods rely on appropriate conditional predictions; they do not repair a misspecified model. Monte Carlo error, uncertainty in the target covariate distribution, and uncertainty in fitted parameters are also distinct sources of uncertainty and should not be conflated.

Common Mistake and Practical Interpretation

A common mistake is to reweight a development-sample PTA by changing only the overall proportion of a covariate while ignoring how that covariate is associated with other modeled covariates. For example, changing the marginal proportion with impaired renal function while leaving body size fixed at its development-sample distribution may create a joint covariate mix that does not represent the target population. The correction is to define the target distribution of the full covariate vector, or use a justified stratification that preserves the relevant joint structure.

Another mistake is to interpret a shifted estimate as evidence that the conditional drug response changed. A difference between \(\operatorname{PTA}_S(d)\) and \(\operatorname{PTA}_T(d)\) can arise solely because the populations weight the same profile-specific predictions differently. Conversely, if the conditional relationship itself differs, covariate reweighting alone is insufficient. A dose recommendation should therefore name the target population and its covariate distribution, state the PTA threshold and other decision constraints, and describe the support and uncertainty behind the estimate.