From Simulated Exposures to a Regimen Decision
Candidate regimens can have different strengths: one may attain an efficacy target in more simulated subjects, while another may more often remain within a safety limit. A high median exposure does not by itself resolve that trade-off. The decision therefore needs to use the relevant exposure criteria together, state which compromises are acceptable, and distinguish a clear pass from a close or uncertain result.
The example below compares three hypothetical regimens using simulated steady-state exposures. It uses efficacy targets of AUC over a dosing interval at least 50 mg·h/L and trough concentration at least 1 mg/L, together with a safety limit of peak concentration no greater than 12 mg/L. These numerical targets are illustrative, not clinical recommendations. The same simulated subjects contribute all exposure measures for a given regimen, so the joint result preserves the relationships among them.
For clarity, let \(E_i=1\) when simulated subject \(i\) meets both efficacy criteria, and \(S_i=1\) when that subject meets the safety limit. The joint pass indicator is \(J_i=E_iS_i\). For \(N\) simulated subjects, the estimated joint PTA is the number with \(J_i=1\), divided by \(N\). It is not generally valid to multiply the separate efficacy and safety proportions: that would require an independence assumption that may not hold.
Set the Rule Before Comparing Candidates
A transparent example of a decision rule is to declare a regimen eligible only if the lower Monte Carlo confidence bound for joint PTA is at least 0.90 and the lower bound for safety attainment is at least 0.95. If more than one regimen qualifies, choose the eligible regimen with the greatest lower bound for joint PTA; use a prespecified practical consideration, such as fewer administrations per day, to break a tie. If none qualifies, report that none meets the rule rather than selecting the “best” candidate by default.
These cutoffs are decision thresholds for this hypothetical exercise, not universal standards. The lower bounds address simulation error: they make the rule less likely to label a regimen as qualifying merely because of random fluctuation in a finite simulation. They do not, by themselves, account for uncertainty in the estimated model parameters, misspecification, or uncertainty about whether the target values are clinically appropriate.
Worked Example: Turning Simulated Counts Into PTA
Suppose 10,000 subjects are simulated under each of three hypothetical regimens. For each subject, the simulation records whether both efficacy targets are met and whether the peak concentration is no greater than 12 mg/L. The counts are:
| Regimen | Meets both efficacy targets | Meets safety limit | Meets efficacy and safety together |
|---|---|---|---|
| R1: 600 mg every 24 h | 9,300 | 9,700 | 8,600 |
| R2: 400 mg every 12 h | 9,500 | 9,600 | 9,100 |
| R3: 800 mg every 24 h | 9,700 | 9,000 | 8,500 |
For R2, the estimated joint PTA is
Thus, 91.0% of the simulated R2 subjects meet both efficacy targets and the safety limit. Its efficacy attainment is \(9{,}500/10{,}000=0.950\), and its safety attainment is \(9{,}600/10{,}000=0.960\). These marginal proportions answer separate questions; the joint count directly answers whether all required criteria are met by the same simulated subject. The joint count cannot exceed either marginal count, as it does not here.
Quantify Monte Carlo Error and Apply the Rule
With a fixed model and independent simulated subjects, each joint pass indicator is a Bernoulli random variable: it equals 1 with probability \(p\), the regimen’s simulated joint attainment probability, and 0 otherwise. The sample proportion estimates \(p\). Since a Bernoulli variable has variance \(p(1-p)\), the estimated standard error of the proportion is \(\sqrt{\hat{p}(1-\hat{p})/N}\). This calculation describes Monte Carlo error under the specified simulation model; it is not a measure of all uncertainty in the regimen’s true clinical performance.
The displayed interval is a normal-approximation Monte Carlo interval. It is appropriate here as a simple illustration because the simulated pass and fail counts are both large. When counts are sparse or estimates are near 0 or 1, a more suitable binomial interval can be used. Any such interval describes simulation variability conditional on the model and simulation setup; it is not automatically a confidence interval for real-world attainment.
Worked Example: Checking Whether R2 Clears the Thresholds
For R2, the joint estimate is \(\hat{p}=0.910\) from \(N=10{,}000\) simulated subjects. Its Monte Carlo standard error is
The approximate 95% Monte Carlo interval is
The lower bound, 0.9044, exceeds the joint PTA requirement of 0.90. For safety attainment, \(\hat{p}=0.960\), giving
The safety lower bound also exceeds 0.95. R2 therefore meets both parts of the example rule. The interval endpoints are not probabilities that the model or regimen is correct; they quantify Monte Carlo error under the stated simulation assumptions.
Worked Example: Selecting Among All Three Regimens
Apply the same rule to the counts above. For R1, joint PTA is \(8{,}600/10{,}000=0.860\), below the required 0.90 even before allowing for Monte Carlo error. Its safety attainment is \(0.970\), so R1 passes the safety criterion but is not eligible because its joint attainment is inadequate.
For R2, the joint point estimate is 0.910 and the safety point estimate is 0.960. As calculated above, their approximate lower Monte Carlo bounds are 0.9044 and 0.9562, respectively. Both clear the thresholds, so R2 is eligible.
For R3, joint PTA is \(8{,}500/10{,}000=0.850\), and safety attainment is \(9{,}000/10{,}000=0.900\). Both point estimates are below their respective requirements, so R3 is ineligible. In this comparison, R2 is the only regimen that qualifies and is selected. The conclusion is conditional on the specified target definitions, decision rule, model, and simulated population.
Why the Joint Result and the Rule Matter
A regimen may have strong efficacy attainment but still expose too many subjects to a safety-limit violation. R3 illustrates this: 97.0% meet both efficacy targets, yet only 90.0% meet the safety limit and 85.0% meet all criteria together. Reporting only the efficacy result would conceal the reason it fails the decision rule. Conversely, a strong safety proportion does not compensate for inadequate efficacy attainment, as R1 illustrates.
The intersection is important because exposure measures can be related. Subjects with higher AUC may also have higher peak concentrations, for example, so the probability of meeting both efficacy and safety requirements cannot be inferred from their separate probabilities without additional assumptions. Simulating all required measures for each subject and counting simultaneous attainment retains their simulated dependence.
Alternative selection approaches answer different questions. Choosing the maximum joint PTA ranks candidates by one summary but can select a regimen that still fails a minimum safety requirement. Comparing median exposure with a target can hide a clinically important tail of subjects who do not attain efficacy or exceed a safety limit. A utility-based approach can balance benefits and harms, but it requires explicit and defensible values for those outcomes. A threshold-based rule is easier to audit when minimum efficacy and safety requirements are central, but its answer can depend on the chosen thresholds. No selection rule removes the need to justify those choices.
Diagnostics, Sensitivity, and Limitations
Before accepting a comparison, check that each regimen was simulated using the intended dose, interval, population characteristics, and relevant dosing conditions. Confirm that the concentration-time profiles and exposure summaries are compatible with the model and that the simulation reflects the intended state, such as steady state when steady-state targets are used. Verify that all criteria use consistent units and definitions, and that joint attainment was counted at the subject level rather than assembled from unrelated marginal summaries.
- Inspect the exposure distributions. Implausible ranges, unexpected modes, or extreme tails may indicate a coding, parameter, or simulation problem. Summaries should be checked against concentration-time behavior, not treated as self-validating outputs.
- Check simulation precision. Increase the number of simulated subjects if Monte Carlo intervals are too wide to resolve eligibility. A result just above a cutoff deserves more scrutiny than one well above it.
- Separate sources of uncertainty. If simulations draw individual variability but hold population parameters fixed, the PTA interval above reflects only finite simulation error. Parameter uncertainty must be propagated separately if the decision is intended to reflect it.
- Revisit the targets. A mathematically precise PTA does not establish that a threshold is clinically justified. The regimen choice should be tested against reasonable alternative target definitions and assumptions.
A useful sensitivity check is to tighten the safety limit and recompute both safety attainment and joint PTA from the same simulated exposure records. For example, changing the hypothetical peak limit from 12 mg/L to 11 mg/L can only leave the safety pass count unchanged or reduce it. The joint pass count cannot increase, because every joint pass must also be a safety pass. This monotonicity is a basic check on a recalculation.
Worked Example: A Tighter Safety Limit Changes Eligibility
Suppose the safety criterion is revised to peak concentration no greater than 11 mg/L. Recounting the same hypothetical simulated records gives R1 9,000 safety passes and 8,100 joint passes; R2 9,200 safety passes and 8,700 joint passes; and R3 7,600 safety passes and 7,300 joint passes, out of 10,000 for each regimen. Each revised count is no larger than its corresponding count under the 12 mg/L limit.
For R2, revised safety attainment is \(9{,}200/10{,}000=0.920\), and revised joint PTA is \(8{,}700/10{,}000=0.870\). Both point estimates are below the original rule’s respective requirements of 0.95 and 0.90. Under this tighter target, no candidate is eligible. The appropriate conclusion is that the specified candidates do not meet the revised rule, not that R2 remains acceptable because it was selected under the earlier safety definition.
The final recommendation is therefore a model-based decision conditional on the simulated population, the exposure model, target definitions, and the prespecified rule. It does not prove a causal clinical benefit or guarantee individual outcomes. If no candidate qualifies, reasonable next steps include evaluating alternative regimens, checking whether the model or targets need review, and quantifying relevant model uncertainty before making a stronger recommendation.
Check Your Understanding
Answer each question using the definitions and decision rule in this tutorial. Solutions follow each question.
- Why can’t the joint PTA generally be found by multiplying the separate efficacy and safety attainment proportions?
Solution: The exposure criteria may be dependent. Joint PTA must be calculated from subjects who meet all criteria simultaneously; multiplying marginal proportions assumes independence. - A regimen has 8,400 joint passes among 10,000 simulated subjects. What is its estimated joint PTA, and does its point estimate meet a 0.90 minimum?
Solution: \(\hat{p}=8{,}400/10{,}000=0.840\), or 84.0%. It does not meet the 90% minimum. - What does the Monte Carlo interval for a simulated PTA quantify, and what does it not automatically quantify?
Solution: Under the simulation assumptions, it quantifies error from using a finite number of simulated subjects. It does not automatically quantify parameter uncertainty, model misspecification, or uncertainty in the clinical target. - If a safety limit is tightened, can the joint pass count increase when the same simulated exposure records are used? Explain.
Solution: No. Tightening the safety criterion can only preserve or remove safety passes. Since joint passes must also satisfy safety, the joint count can only stay the same or decrease. - If no candidate satisfies the prespecified eligibility rule, what is the appropriate conclusion?
Solution: State that none of the evaluated candidates qualifies under that rule. Do not automatically choose the candidate with the largest PTA; consider alternative regimens or review the assumptions and targets.