From a PTA Estimate to a Decision
Probability of target attainment (PTA) summarizes how often a specified target is attained in the population and under the dosing conditions represented by the model. A dose-selection decision adds a requirement: how large must that probability be before a candidate dose is considered adequate? That requirement is the PTA decision threshold. It is not estimated from the simulation; it is specified as part of the decision rule.
The threshold is distinct from the target used to define attainment. For example, a target might define a pharmacokinetic/pharmacodynamic criterion that an individual must meet. The PTA threshold then states what proportion of the modeled population must meet that criterion for a dose to qualify. A required PTA of 80% does not mean the target itself is set at 80%; it means the decision requires at least 80% population attainment of that target.
The inequality is inclusive: a dose with PTA exactly equal to \(\tau\) qualifies. In simulation, \(p(d)\) is usually not known exactly; it is estimated by \(\widehat p(d)\). A direct plug-in rule classifies candidates using \(\widehat p(d)\geq\tau\), while recognizing that finite simulation creates Monte Carlo error. Whether a decision should instead require a conservative bound to exceed \(\tau\) is a separate policy choice, not part of the definition of the threshold rule.
Why the Threshold Is a Decision Input
A model can estimate PTA across doses, but it cannot determine the acceptable minimum probability by itself. The threshold reflects the decision context, including the consequences of inadequate target attainment, the purpose of the analysis, and any relevant clinical or regulatory guidance. It should be specified and justified before selecting a dose from the simulated curve. There is no single threshold that is automatically correct for every target or decision.
The lowest-qualifying-dose rule also makes an objective explicit: among candidates that meet the requirement, prefer the smallest dose. This may be a sensible dose-minimization objective, but it is not the only possible objective. For example, a decision could prioritize another criterion or require additional constraints. A PTA threshold alone does not establish that a dose is acceptable on every relevant dimension.
Worked Example: A Threshold Selects Among Crossing Curves
Consider two hypothetical candidate regimens, A and B, evaluated at the same three dose levels. For each regimen and dose, 1,000 simulated individuals are assessed for a fixed target. The counts below are the numbers attaining that target.
| Dose | Regimen A: count / 1,000 | Regimen A: estimated PTA | Regimen B: count / 1,000 | Regimen B: estimated PTA |
|---|---|---|---|---|
| 100 mg | 520 | 0.520 | 650 | 0.650 |
| 200 mg | 780 | 0.780 | 820 | 0.820 |
| 300 mg | 900 | 0.900 | 860 | 0.860 |
For example, regimen A at 200 mg has estimated PTA $$ \widehat p_A(200)=\frac{780}{1000}=0.780. $$ The estimated curves cross between 200 mg and 300 mg: regimen B has higher PTA at the first two doses, but regimen A has higher PTA at 300 mg.
First set the required threshold to \(\tau=0.80\). Regimen A at 200 mg does not qualify because \(0.780<0.80\), while regimen B at 200 mg does because \(0.820\geq0.80\). Regimen B at 100 mg does not qualify, so its lowest qualifying dose is 200 mg. Regimen A first qualifies at 300 mg, where \(0.900\geq0.80\). Under the lowest-qualifying-dose objective across both regimens, regimen B at 200 mg is selected.
Now set \(\tau=0.88\). Regimen B does not qualify at any candidate dose because its largest estimated PTA is \(0.860<0.88\). Regimen A qualifies at 300 mg, since \(0.900\geq0.88\), but not at 200 mg. The selected candidate is therefore regimen A at 300 mg. Raising the threshold changed the choice, and the crossing curves explain why the higher-dose candidate with the best PTA is not the same regimen that first met the lower threshold.
Plateaus and the Lowest Qualifying Dose
A PTA curve may rise rapidly at lower doses and then change little with further increases. Such a plateau matters because a higher dose may add little estimated attainment once the threshold has already been met. Under the lowest-qualifying-dose rule, a dose on the plateau is not selected merely because it has the largest PTA: the rule selects the first candidate that reaches the requirement.
A plateau also does not guarantee that every threshold can be met. If the plateau lies below the required threshold, none of the evaluated doses qualifies. A decision-maker should report that result rather than treating the largest available dose as adequate by default. Likewise, a threshold set exactly at the plateau can be sensitive to simulation error or to small changes in model assumptions.
Worked Example: A Plateau Limits the Benefit of Increasing Dose
Suppose a hypothetical simulation evaluates four doses, with 1,000 simulated individuals at each dose. The estimated PTAs are 0.680, 0.840, 0.910, and 0.920 at 50, 100, 150, and 200 mg, respectively. The PTA rises by 0.070 from 100 to 150 mg, but by only 0.010 from 150 to 200 mg.
For a required threshold of \(\tau=0.90\), the 100 mg candidate fails because \(0.840<0.90\). The 150 mg candidate qualifies because \(0.910\geq0.90\). Since it is the first qualifying candidate, the selected dose is 150 mg. The increase from 150 to 200 mg changes estimated PTA by $$ 0.920-0.910=0.010, $$ or 1 percentage point, and is not needed to meet this particular threshold.
If the required threshold is instead \(\tau=0.93\), even the highest estimated PTA is too low: \(0.920<0.93\). There is no qualifying dose in the evaluated set. Selecting 200 mg and describing it as meeting the requirement would contradict the rule. The appropriate conclusion is that no evaluated candidate meets the 0.93 threshold; a different candidate set or a reconsidered decision requirement would need separate justification.
Justifying the Selection Rule
The lowest-qualifying-dose rule can be justified directly from its objective. It chooses the least dose among the candidates that satisfy the prespecified PTA requirement. For a finite ordered set of doses, this requires evaluating which candidates qualify and then selecting the smallest qualifying one; it does not require the PTA curve to be monotone.
If the set on the right is empty, \(d^*\) is undefined within the candidate set: no dose qualifies. If it is nonempty, its minimum qualifies by construction, and every smaller evaluated dose fails the requirement. This proves that \(d^*\) is the lowest evaluated dose meeting the plug-in rule. The conclusion applies only to the candidates evaluated; it does not establish that untested doses between them also pass or fail.
Monotonicity is sometimes assumed to simplify dose selection: if PTA never decreases as dose increases, one can search for the first crossing of the threshold. But monotonicity must be justified for the particular candidates and estimand. Regimens with different schedules, or curves that differ in how exposure changes with dose, can have different rankings across the candidate range. When curves cross, a method that assumes one candidate is uniformly superior can give the wrong selection. With a finite candidate set, evaluating every candidate and applying the stated rule avoids that assumption.
Monte Carlo Error Near the Cutoff
A simulated PTA is an estimate based on a finite number of draws. If \(M\) independent simulated individuals are used at a dose and \(K\) attain the target, then \(\widehat p=K/M\). Even when the underlying model-based PTA is fixed, a different simulation run can produce a different \(\widehat p\). Consequently, an estimate just above or below \(\tau\) may not clearly resolve whether the model-based PTA is above or below the requirement.
Worked Example: A Candidate Close to the Threshold
At a hypothetical dose, 828 of 1,000 simulated individuals attain the target. The estimated PTA is $$ \widehat p=\frac{828}{1000}=0.828. $$ If the prespecified threshold is \(\tau=0.83\), the direct plug-in rule says this candidate does not qualify, because \(0.828<0.83\). That is the mechanical result of the stated rule, not proof that the model-based PTA is below 0.83.
For independent Bernoulli attainment indicators, the estimated Monte Carlo standard error is $$ SE_{\mathrm{MC}}=\sqrt{\frac{\widehat p(1-\widehat p)}{M}} =\sqrt{\frac{0.828(1-0.828)}{1000}} =\sqrt{0.0001424} \approx 0.01193. $$ A rough normal-approximation interval for Monte Carlo uncertainty is $$ 0.828\pm1.96(0.01193)=(0.8046,\ 0.8514). $$ This interval straddles 0.83, indicating that this simulation run does not estimate the cutoff precisely. It describes simulation error under the assumed model and simulation procedure; it is not an interval for every source of uncertainty in the model or target population.
The analyst could increase the number of independent simulation draws and reassess the Monte Carlo uncertainty, or follow a prespecified decision policy for borderline estimates. If a policy requires a lower uncertainty bound to exceed 0.83, this candidate would not pass that more conservative rule. That policy differs from simply applying \(\widehat p\geq\tau\), and should not be silently substituted after seeing the result.
Diagnostics, Alternatives, and Common Pitfalls
Before reporting a selected candidate, check that the threshold, target, population, and candidate set are clearly stated and held consistent across the comparisons. Inspect the PTA estimates across doses rather than relying only on the selected value. Sudden reversals or irregular changes may reflect genuine differences among candidates, Monte Carlo variation, or a problem requiring investigation. In particular, verify that candidates were simulated and summarized comparably and that borderline classifications are not being driven by imprecise estimates.
The threshold rule is one way to summarize a decision, not a replacement for examining the curve. Common alternatives include choosing the dose with the largest PTA, choosing a dose at a prespecified exposure level, or applying a rule that accounts for other decision criteria. Maximizing PTA alone may favor the highest dose even when several lower doses already satisfy the requirement. Conversely, a dose may have the highest PTA and still fail a demanding threshold. Any alternative answers a different decision question and should be stated explicitly.
| Approach | Decision question | Important limitation |
|---|---|---|
| Threshold with lowest qualifying dose | What is the smallest evaluated dose with PTA at least \(\tau\)? | Depends on the chosen threshold and candidate set. |
| Maximum estimated PTA | Which candidate has the largest estimated attainment? | Does not prefer a lower dose that already meets a requirement. |
| Threshold plus uncertainty policy | Does the estimate, or a specified uncertainty bound, meet the requirement? | The bound-based rule is more conservative and must be defined in advance. |
Common mistake: “The 200 mg dose has the highest PTA for this regimen, so it must be selected.” That conclusion does not follow from a lowest-qualifying-dose rule. If 150 mg already meets the stated threshold, the rule selects 150 mg, not 200 mg. The correction is to identify the decision objective first, then apply it consistently to all evaluated candidates.
A related mistake is to interpret a threshold as a universal scientific constant or to adjust it until a preferred dose qualifies. The threshold is a decision input that requires contextual justification. Report it alongside the estimated PTAs and the selection rule, and state plainly when no evaluated candidate meets it. Also distinguish the point-estimate classification from uncertainty about the estimated PTA.
Check Your Understanding
Answer each question using the stated threshold rule. Solutions follow the questions.
- At doses 40, 80, and 120 mg, estimated PTAs are 0.62, 0.81, and 0.89. If \(\tau=0.80\) and the objective is the lowest qualifying dose, which dose is selected?
- Two regimens have estimated PTAs of 0.75 and 0.83 at 100 mg, then 0.91 and 0.87 at 200 mg. What does this pattern imply about their ranking, and why should one not assume a single regimen is always superior?
- A candidate has estimated PTA 0.895 and the threshold is 0.90. Does it qualify under the plug-in rule? What additional issue should be considered if the estimate is imprecise?
- All evaluated doses have estimated PTA below a threshold of 0.95. Is the highest-PTA dose a qualifying dose? State the conclusion accurately.
Solutions. (1) The selected dose is 80 mg: \(0.81\geq0.80\), while the lower-dose estimate \(0.62\) is below the threshold. (2) The first regimen is lower at 100 mg but higher at 200 mg, so the rankings reverse across the candidates; selection must apply the threshold rule to each candidate rather than assume uniform superiority. (3) It does not qualify because \(0.895<0.90\). If Monte Carlo uncertainty is material, assess its precision and apply any prespecified borderline-decision policy; do not treat the point estimate as exact. (4) No evaluated dose qualifies. The highest-PTA candidate is only the best among those evaluated, not a dose that meets the 0.95 requirement.