Why Joint Targets Require Joint Probabilities
A regimen may be considered efficacious only when more than one exposure criterion is met—for example, when both AUC is high enough and trough concentration remains above a minimum. Each criterion can be assessed separately, but the dose-selection question is about the same simulated individual meeting all required criteria. The association between that individual's exposure metrics therefore matters.
The key operational rule is to calculate every metric for each simulated individual and evaluate all criteria on that individual's paired values. Separate marginal probabilities do not, by themselves, determine the probability of joint attainment. In particular, multiplying them assumes independence, which may be implausible when metrics arise from the same concentration-time profile.
The simulated values must represent the target population and dosing conditions in the estimand. The two metrics must be computed from the same simulated individual, using the same relevant parameter, covariate, random-effect, and residual-error draw wherever the estimand includes those sources of variation. If the target is conditional on fixed model parameters, the resulting probability is conditional on those parameters; if parameter uncertainty is part of the estimand, the simulation must include it in the prescribed way.
Estimating Joint Attainment From Paired Simulations
Suppose \(N\) simulated individuals each have a pair \((A_i,C_{\min,i})\). Define an indicator \(I_i=1\) when individual \(i\) meets both criteria and \(I_i=0\) otherwise. The direct Monte Carlo estimate is the fraction of simulated individuals meeting both targets.
This calculation treats each pair as one outcome: pass both or fail at least one. It does not require AUC and trough to be independent. For independent, identically distributed complete simulation replicates, the indicators are Bernoulli with success probability \(P(J)\), so the Monte Carlo standard error is estimated by \(\sqrt{\widehat{P}(J)[1-\widehat{P}(J)]/N}\). This measures simulation error in estimating the probability under the specified simulation setup; it does not automatically include uncertainty in the fitted model or in the target definition.
Worked Example: Counting Paired AUC and Trough Attainment
In a hypothetical simulation at one dose, the minimum AUC target is \(50\ \mathrm{mg\cdot h/L}\) and the minimum trough target is \(1.0\ \mathrm{mg/L}\). Ten simulated individuals have the following results. “Pass” means the value is at or above its target.
| Individual | AUC result | Trough result | Meets both? |
|---|---|---|---|
| 1 | Pass | Pass | Yes |
| 2 | Pass | Pass | Yes |
| 3 | Pass | Pass | Yes |
| 4 | Pass | Pass | Yes |
| 5 | Pass | Pass | Yes |
| 6 | Pass | Pass | Yes |
| 7 | Pass | Fail | No |
| 8 | Pass | Fail | No |
| 9 | Fail | Pass | No |
| 10 | Fail | Fail | No |
There are six joint successes among ten pairs, so
In this illustrative sample, the estimated probability that a simulated individual meets both targets is 60.0%. The marginal AUC and trough attainment estimates are \(8/10=0.800\) and \(7/10=0.700\), respectively. Their product is \(0.800(0.700)=0.560\), not the observed joint fraction of 0.600. The difference arises because the paired pass/fail outcomes are associated in these simulated data. With only ten replicates, these proportions are pedagogical calculations, not precise estimates.
If the ten complete replicates were independent, the estimated Monte Carlo standard error would be
That large value reflects the very small simulation size. A practical analysis would generally use enough complete replicates to make Monte Carlo error acceptably small for the intended decision.
How Dependence Changes the Joint Probability
Let \(E_A=\{A\geq a_0\}\) and \(E_C=\{C_{\min}\geq c_0\}\). The multiplication rule gives
The joint probability equals \(P(E_A)P(E_C)\) only when the two attainment events are independent—that is, when \(P(E_C\mid E_A)=P(E_C)\). Correlation between the continuous exposure metrics can be informative, but the relevant dependence for the joint target is dependence between the threshold-crossing events. A correlation coefficient for AUC and trough alone does not determine the joint probability unless a sufficiently specified joint distribution is also assumed.
A useful identity makes the role of association explicit. For the two pass indicators \(I_A\) and \(I_C\), each taking values zero or one, their product is one exactly when both criteria pass. Thus
Positive covariance between the pass indicators raises joint attainment above the product of the marginal probabilities; negative covariance lowers it. The identity concerns the indicators at the specified cutoffs, so dependence can differ across target thresholds even when the underlying exposure metrics have a stable correlation.
Worked Example: Same Marginals, Different Joint Attainment
Consider two hypothetical paired simulations, each summarized over ten individuals. In both, eight individuals pass the AUC target and six pass the trough target. In the first pairing, six pass both; in the second, only four pass both. These combinations are possible: the first has six both, two AUC-only, zero trough-only, and two neither; the second has four both, four AUC-only, two trough-only, and zero neither.
Both pairings have marginal estimates \(\widehat{P}(E_A)=0.800\) and \(\widehat{P}(E_C)=0.600\), so the independence product would be \(0.800(0.600)=0.480\). But direct paired counting gives joint estimates of \(6/10=0.600\) and \(4/10=0.400\). The marginals alone cannot distinguish these outcomes.
For these marginals, the possible joint probability is bounded by
which here is \(0.400\leq P(E_A\cap E_C)\leq0.600\). The lower bound follows because the probability of the union cannot exceed one: \(P(E_A\cup E_C)=P(E_A)+P(E_C)-P(E_A\cap E_C)\leq1\). The upper bound follows because the intersection is a subset of each event. The two hypothetical joint estimates reach the bounds, illustrating how much the association can matter.
Dependence-Preserving Workflow and Checks
A reliable computation keeps the pairing intact from simulation through classification. For every complete replicate, calculate both metrics from the same simulated individual and then apply the thresholds. A practical sequence is:
Write the target criteria, including units and whether equality counts as attainment. For minimum targets, use “greater than or equal to” consistently.
Retain each individual's AUC and trough together, with the same simulation identity and the variability required by the estimand.
Mark a joint success only if every criterion is met, then divide the number of joint successes by the number of valid complete replicates.
Compare joint attainment with both marginal estimates and the probability bounds; assess Monte Carlo precision and inspect the paired values near each threshold.
The bounds provide a simple implementation check: a computed joint proportion cannot exceed either marginal proportion or fall below the lower bound formed from them. A scatterplot of paired AUC and trough values, with both target thresholds drawn, shows which individuals occupy the joint-attainment region. Tabulating the four possible pass/fail combinations also helps detect a lost pairing, reversed comparison, or incorrect denominator.
A common mistake is to multiply the marginal attainment probabilities because both are estimated from the same simulation. Using the first example, that would report \(0.800\times0.700=0.560\) instead of the paired estimate \(0.600\). The multiplication is valid only under independence of the two attainment events; sharing the same simulated individual does not establish independence. Conversely, assuming positive dependence without checking can also mislead. Use the paired joint count unless a justified joint probability model is being used deliberately.
The ordinary binomial Monte Carlo standard error also has conditions. It is appropriate when the complete simulation replicates used for the count are independent draws from the target distribution. If replicates are dependent—for example, because a simulation design reuses shared draws in a way that couples outcomes—the binomial formula can understate or otherwise mischaracterize Monte Carlo error; uncertainty should then reflect the simulation design. Likewise, a probability based on weighted simulations is not generally the unweighted success count divided by \(N\).
Worked Example: Monte Carlo Precision for a Joint Estimate
Suppose 1,000 independent hypothetical simulation replicates at a specified dose produce 720 individuals meeting both efficacy targets. The estimate is \(\widehat{P}(J)=720/1000=0.720\). Under the independent-replicate assumption, its estimated Monte Carlo standard error is
A rough normal-approximation interval for Monte Carlo error is \(0.720\pm1.96(0.01420)\), or approximately \(0.6922\) to \(0.7478\). This interval describes uncertainty from finite simulation sampling under the specified simulation setup. It is not a 95% probability statement about the fixed joint probability, and it does not incorporate model uncertainty unless that uncertainty has been represented in the simulation and the interval calculation matches that design.
The estimate's precision should be judged against the intended decision. If the decision threshold is close to 0.720, this amount of Monte Carlo variation may make the estimate insufficiently decisive, suggesting more simulation or a more explicit decision rule. A narrow Monte Carlo interval cannot correct a misspecified model, an inappropriate target population, or a poorly chosen efficacy criterion.
Alternatives and Limitations
Direct paired simulation is often attractive because it preserves the dependence produced by the model without requiring a separate parametric description of the joint distribution. Other approaches can be useful in defined circumstances, but they require additional justification:
- Product of marginal probabilities: simple, but valid for joint attainment only when the two pass/fail events are independent.
- Joint-distribution calculation: a specified bivariate distribution or copula can calculate joint probabilities, but its dependence and tail behavior must be appropriate and supported. A correlation estimate by itself is not a complete joint model.
- Probability bounds: the bounds above require only the marginals, but may be too wide for a dose decision. They are useful as checks or when dependence information is unavailable, not as a precise substitute for it.
Direct simulation is not automatically correct merely because it uses many replicates. It still depends on the model, the population distribution, the regimen, the definition of each exposure metric, and the chosen target values. Rare joint successes or failures may need especially large simulation sizes to estimate reliably. If targets are changed, the paired values can often be reclassified, but the joint probability must be recomputed at the new thresholds.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- If the estimated marginal attainment probabilities are 0.75 for AUC and 0.60 for trough, what is the independence-based product, and when would it equal joint attainment?
- With those same marginals, give the lower and upper probability bounds for joint attainment.
- In 500 independent replicates, 340 meet both targets. Calculate the joint estimate and its estimated Monte Carlo standard error.
- Why should AUC and trough be calculated from the same simulated individual when estimating joint attainment?
Solutions. (1) The product is \(0.75(0.60)=0.450\); it equals joint attainment only if the two attainment events are independent. (2) The bounds are \(\max(0,0.75+0.60-1)=0.350\) and \(\min(0.75,0.60)=0.600\). (3) The estimate is \(340/500=0.680\), and \(SE=\sqrt{0.680(0.320)/500}=0.02086\). (4) Joint attainment asks whether both criteria hold for the same individual; separating or re-pairing the metrics can destroy their dependence and change the joint count.