Worked Comparison of Competing Regimens

From Two Regimens to a Decision

Comparing regimens is not just a matter of asking which one produces a higher exposure or a larger efficacy probability. A regimen can improve the chance of meeting an efficacy target while also increasing the chance of an unacceptable safety outcome. The comparison therefore needs clinical criteria, a target population, and a decision rule specified before looking at the simulated results.

This tutorial works through a hypothetical comparison of two oral regimens. The goal is not to identify a universally best regimen, but to show how a prespecified rule can combine efficacy, safety, and subgroup performance. All simulated counts and thresholds below are illustrative, not results from a real study.

A joint PTA answers a patient-level question: under the model and the stated conditions, how often are both criteria met together? It is not generally equal to the product of the separate efficacy and safety probabilities. That product would require independence of the two outcomes, which should not be assumed merely because two probabilities are reported.

Specify the Comparison Before Simulating

Suppose regimen A is 400 mg every 24 hours and regimen B is 300 mg every 12 hours. A hypothetical population pharmacokinetic/pharmacodynamic model is used to simulate each regimen at steady state. The efficacy target is \(AUC_{24}/MIC \geq 100\), and the safety criterion is \(C_{\max} \leq 12\) mg/L. A simulated patient is a joint success only if both conditions hold.

The target population is divided into two clinically relevant subgroups, with prespecified population weights of 60% for subgroup 1 and 40% for subgroup 2. For each regimen and subgroup, 5,000 virtual patients are simulated. The decision rule is:

• Joint PTA must be at least 70% in each subgroup.
• The probability of exceeding the safety limit must be no more than 10% in each subgroup.
• The population-weighted joint PTA must be at least 78%.

These thresholds are decision criteria for this hypothetical exercise, not universal standards. Specifying them in advance makes the final choice auditable and avoids changing the success definition to favor a regimen after seeing its results.

Worked Comparison: Subgroup PTA and the Decision Rule

Why the Joint Probability and Weights Matter

The two separate probabilities are useful diagnostics, but the joint PTA is the direct estimate of the proportion meeting both requirements. For any two events \(E\) (efficacy target met) and \(S\) (safe), the joint probability must satisfy:

The upper bound follows because an event that meets both criteria must be contained in each event separately. For the lower bound, the probability of failing at least one criterion is at most \(P(E^c)+P(S^c)\). Therefore \(P(E\cap S)=1-P(E^c\cup S^c)\geq 1-P(E^c)-P(S^c)=P(E)+P(S)-1\); probabilities cannot be negative, giving the stated maximum with zero. These bounds are a useful arithmetic check on reported counts.

For example, regimen A in subgroup 1 has efficacy probability 0.880 and safety probability 0.940. Its joint PTA must lie between \(0.880+0.940-1=0.820\) and \(\min(0.880,0.940)=0.880\). The observed joint PTA, 0.840, is within those bounds.

Subgroup weights matter because a simple average gives the two subgroups equal influence, which would describe a different target population when their proportions differ. If a decision concerns a population with different subgroup proportions, the weights should be changed to represent that population and the weighted PTA recalculated. The subgroup estimates should still be shown: an acceptable weighted average can conceal poor performance in a smaller subgroup.

Monte Carlo Precision and Diagnostics

A simulated PTA is an estimate based on a finite number of virtual patients. Conditional on the simulation model and independent simulated patients, the success count has a binomial distribution with success probability \(p\). If \(\hat{p}\) is the simulated proportion from \(N\) patients, its Monte Carlo standard error is estimated by:

This follows because the variance of a binomial count is \(Np(1-p)\). Dividing the count by \(N\) divides its variance by \(N^2\), giving \(\operatorname{Var}(\hat{p})=p(1-p)/N\); substituting \(\hat{p}\) for the unknown \(p\) gives the estimated standard error.

Useful checks include confirming that each denominator matches the intended subgroup simulation size, that joint successes do not exceed either component success count, and that the joint PTA satisfies the probability bounds. Also inspect subgroup results rather than only a weighted total, and verify that target definitions, units, dosing schedules, and population weights are consistent across candidates. Implausible subgroup differences or an unexpected regimen ranking can indicate a coding error, an incorrect target definition, or model assumptions that deserve review.

Monte Carlo error is only one source of uncertainty. A small Monte Carlo standard error does not account for uncertainty in model parameters, model structure, adherence, or the future population. Those issues require additional evaluation; they are not resolved simply by increasing the number of virtual patients.

Decision Rules, Alternatives, and Limitations

This example uses a constrained rule: a regimen must pass subgroup safety and joint-PTA requirements, as well as a population-level requirement. That approach makes the clinical constraints visible. A weighted average alone would have ranked A above B here, but would not, by itself, explain B’s unacceptable safety exceedance probabilities. Conversely, ranking solely by efficacy PTA would favor B in both subgroups and ignore the safety constraint.

Another possible approach is to summarize benefit and risk with a utility function. That can be appropriate when trade-offs can be specified credibly, but it answers a different decision question: a sufficiently large benefit may compensate for a specified harm under the chosen utility. The threshold rule in this example does not permit that compensation. The method should follow the clinical decision being represented, not be chosen because it produces a preferred ranking.

A decision based on simulated PTA is conditional on the endpoint definitions, model, simulated population, and weights. A regimen that qualifies is not thereby proven safe or effective, and a regimen that fails this particular rule may warrant reconsideration if the clinical targets or intended population change. Report the criteria, subgroup estimates, population-weighted result, and rule outcome together so readers can see what the recommendation does and does not establish.