Why Evaluate Dose Performance by Subgroup?
A dose can perform well on average while leaving an important subgroup with a materially lower chance of meeting the clinical target. This can occur when groups differ in covariates that affect exposure or response, or when subgroup-specific target distributions differ. A population average alone can obscure that imbalance. Robust dose selection therefore asks two related questions: how often is the target met across the intended population, and how often is it met in the least-served clinically important subgroup?
The groups, performance endpoint, and acceptable performance threshold must be specified before comparing candidate doses. The endpoint may be a probability of attaining a defined efficacy target, a joint efficacy-safety criterion, or another prespecified performance measure. The analysis below treats that measure as a probability; it does not prescribe which clinical target is appropriate.
The weights represent the target population composition, not the relative clinical importance of the groups. A small subgroup may have a small weight in \(A(d)\) but still be decisive through \(W(d)\). If a subgroup is clinically important, it should not be omitted merely because it is uncommon.
Selection Rules and Their Trade-Offs
A transparent primary rule is to require every subgroup to meet a prespecified threshold \(q\), then select the least costly or otherwise preferred dose among those that pass. In the simplest dose-only comparison, this means selecting the lowest candidate dose for which \(W(d)\geq q\). The rule expresses an explicit minimum standard for each included subgroup. It does not guarantee that the dose is optimal for every other objective, such as minimizing toxicity or treatment burden; those objectives need their own prespecified role in the decision.
The minimum in this expression assumes that lower dose is preferred among doses that pass. If another preference applies, such as lower treatment burden or a particular regimen constraint, replace the minimum rule with that explicit preference. The essential robustness condition is that every subgroup satisfy the criterion.
A different rule maximizes weighted overall attainment, \(\arg\max_d A(d)\), while a maximin rule maximizes \(W(d)\). These can choose different doses. The average-oriented rule permits a poor result in a low-weight group if gains elsewhere compensate. The maximin rule gives the weakest subgroup direct priority, even if improving it slightly requires a substantial loss in the others. A threshold-constrained rule lies between them: it protects every subgroup up to the required level, then uses a stated preference to choose among passing doses.
Worked Examples
Worked Example: Finding the Lowest Dose That Passes in Every Subgroup
Suppose three hypothetical candidate doses have the following simulated target-attainment probabilities. These are illustrative model-based estimates, not results from a real study. The intended population weights are 0.50, 0.30, and 0.20 for subgroups A, B, and C, respectively. The prespecified minimum is \(q=0.80\), and lower dose is preferred among passing options.
| Dose (mg) | Subgroup A | Subgroup B | Subgroup C |
|---|---|---|---|
| 100 | 0.92 | 0.72 | 0.60 |
| 150 | 0.96 | 0.88 | 0.78 |
| 200 | 0.98 | 0.94 | 0.90 |
At 100 mg, the weighted overall attainment is \(A(100)=0.50(0.92)+0.30(0.72)+0.20(0.60)=0.460+0.216+0.120=0.796\). The worst-subgroup attainment is \(W(100)=0.60\), so this dose fails the all-subgroup rule.
At 150 mg, \(A(150)=0.50(0.96)+0.30(0.88)+0.20(0.78)=0.480+0.264+0.156=0.900\). But \(W(150)=0.78\), because subgroup C is below 0.80. Thus a 90.0% overall attainment does not make 150 mg acceptable under the prespecified subgroup rule.
At 200 mg, \(A(200)=0.50(0.98)+0.30(0.94)+0.20(0.90)=0.490+0.282+0.180=0.952\), and \(W(200)=0.90\). All three subgroup estimates meet the threshold, so 200 mg is the lowest candidate dose that passes. This conclusion depends on the candidate doses shown and on the stated preference for the lowest passing dose.
The difference between average and weakest-group attainment can be written directly as \(A(d)-W(d)=\sum_g w_g[p_g(d)-W(d)]\). Since \(p_g(d)\geq W(d)\) for every subgroup and the weights are nonnegative, this difference cannot be negative. It is zero only if every positively weighted subgroup has the same attainment as the minimum. This identity shows why a high overall average does not establish that every group is adequately served.
Worked Example: Average, Threshold-Constrained, and Maximin Choices
Consider two hypothetical subgroups, common and uncommon, with intended-population weights 0.90 and 0.10. The minimum acceptable attainment is \(q=0.80\).
| Dose (mg) | Common subgroup | Uncommon subgroup |
|---|---|---|
| 120 | 0.94 | 0.62 |
| 160 | 0.91 | 0.82 |
| 200 | 0.88 | 0.90 |
At 120 mg, \(A(120)=0.90(0.94)+0.10(0.62)=0.846+0.062=0.908\), and \(W(120)=0.62\). At 160 mg, \(A(160)=0.90(0.91)+0.10(0.82)=0.819+0.082=0.901\), and \(W(160)=0.82\). At 200 mg, \(A(200)=0.90(0.88)+0.10(0.90)=0.792+0.090=0.882\), and \(W(200)=0.88\).
Maximizing the weighted average selects 120 mg, but that dose fails the subgroup threshold. The threshold-constrained rule selects 160 mg, the lowest dose for which both subgroup estimates are at least 0.80. The maximin rule selects 200 mg, since its worst-subgroup attainment of 0.88 is higher than 0.82 or 0.62. Thus the three approaches answer different questions; none should be described as universally best without stating the decision objective.
Simulation Error and a Conservative Subgroup Check
In practice, \(p_g(d)\) is usually estimated from simulated individuals. If \(N_g\) independent simulated individuals are generated for subgroup \(g\), and \(K_g(d)\) meet the criterion, then \(\hat p_g(d)=K_g(d)/N_g\). Conditional on the model and simulation setup, the attainment indicators are Bernoulli outcomes with probability \(p_g(d)\), so the estimated Monte Carlo standard error is \(\sqrt{\hat p_g(d)[1-\hat p_g(d)]/N_g}\). This quantifies simulation noise; it does not include uncertainty in the model, its parameters, or the clinical criterion.
A point estimate just above \(q\) can therefore be a fragile basis for selection. One conservative simulation rule is to require a lower confidence bound for each subgroup to meet \(q\). Such a rule should specify the confidence level and interval method in advance. If many subgroups are checked, separate 95% bounds do not jointly provide 95% coverage for all subgroup probabilities; the joint decision requires a multiplicity-aware procedure or an explicitly justified alternative.
Worked Example: A Subgroup Estimate Near the Threshold
Suppose a hypothetical dose is evaluated in \(N_g=400\) independently simulated individuals from a clinically important subgroup, and 324 meet the target. The point estimate is \(\hat p_g=324/400=0.810\). The prespecified threshold is \(q=0.80\).
The estimated Monte Carlo standard error is \(SE=\sqrt{0.810(1-0.810)/400}=\sqrt{0.1539/400}=\sqrt{0.00038475}=0.01961\). Using a simple normal approximation, the approximate 95% interval is \(0.810\pm1.96(0.01961)=0.810\pm0.03844\), or approximately \((0.772,0.848)\).
The point estimate exceeds 0.80, but the approximate lower limit does not. Therefore, the dose passes a point-estimate rule for this subgroup but does not pass a rule requiring this approximate lower bound to be at least 0.80. The approximation is reasonable as a simple illustration because the estimated expected counts of successes and failures are \(400(0.810)=324\) and \(400(0.190)=76\), both well above 10. A Wilson or other prespecified binomial interval may be preferable in actual analysis. Neither interval addresses uncertainty in the model parameters or in the appropriateness of the simulated population.
Diagnostics, Alternatives, and Limitations
Before interpreting a robust-dose result, verify that each subgroup is defined consistently in the simulation and that its target and outcome calculation match the prespecified criterion. Report subgroup-specific estimates alongside \(A(d)\) and \(W(d)\); a single summary can conceal which group drives the decision. Inspect dose-response patterns across groups. If attainment falls or rises unexpectedly with dose, investigate the exposure-response model, safety constraints, simulation implementation, and Monte Carlo precision rather than assuming the pattern must be monotone. Monotonicity should not be imposed when the performance criterion includes competing efficacy and safety conditions that can move in different directions.
The main alternatives are useful when stated as distinct decision rules. Weighted-average optimization represents a population-average objective but may tolerate low attainment in a small subgroup. Maximin optimization directly protects the weakest subgroup but can sacrifice overall attainment to improve a group that is already near the acceptable range. A threshold-constrained rule makes the minimum standard explicit and then applies a separate preference among doses that meet it. Showing all three summaries can clarify the trade-off, but the primary selection rule should be chosen before inspecting results to avoid selecting whichever rule favors a preferred dose.
Robustness is always relative to the subgroups, weights, model, and threshold included in the analysis. A subgroup average can hide heterogeneity within that subgroup; a finite list of subgroups cannot guarantee performance in every individual or in an unrepresented population. Estimates also depend on the assumed covariate distributions and model structure. Consequently, a dose that passes the simulated criterion is evidence of performance under the specified assumptions, not a guarantee of clinical success. Parameter and structural uncertainty should be considered in addition to Monte Carlo error when they are relevant to the decision.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Two subgroups have weights 0.80 and 0.20 and attainment probabilities 0.95 and 0.70. Calculate weighted overall and worst-subgroup attainment. If the required threshold is 0.75, does the dose pass an all-subgroup rule?
- A dose has subgroup estimates 0.90 and 0.82. Another has estimates 0.86 and 0.88. Which dose has the larger worst-subgroup attainment?
- Why can a dose with high weighted overall attainment still fail a robust subgroup criterion?
- For \(\hat p_g=0.75\) from \(N_g=100\) simulated individuals, calculate the approximate Monte Carlo standard error using \(\sqrt{\hat p_g(1-\hat p_g)/N_g}\).
Solutions. (1) \(A=0.80(0.95)+0.20(0.70)=0.76+0.14=0.90\); \(W=\min(0.95,0.70)=0.70\). It fails because 0.70 is below 0.75. (2) The first dose has \(W=0.82\); the second has \(W=0.86\), so the second has the larger worst-subgroup attainment. (3) A high-weight group can dominate the weighted average and compensate numerically for low attainment in a less common group, while the robust rule requires every subgroup to meet the threshold. (4) \(SE=\sqrt{0.75(0.25)/100}=\sqrt{0.001875}=0.04330\), or about 4.33 percentage points.