Why Joint Attainment Requires Subject-Level Results
A dose may need to produce adequate exposure on more than one measure. For example, a regimen might be required to attain a minimum AUC while also keeping a peak concentration below a toxicity-related limit. The relevant question is not how many simulated subjects meet each criterion separately. It is how many meet all of them in the same simulated profile.
That distinction matters because exposure measures for one subject are generally related. A subject’s clearance, volume, covariates, and other model inputs can affect multiple exposure metrics together. AUC and a concentration measure computed from the same simulated profile therefore should not be treated as if they came from unrelated subjects. The earlier material on simulating correlated random effects and computing PTA provides the basic simulation setting; here the focus is the simultaneous event.
The criteria need not have the same form. One might be a lower bound, another an upper bound, and another a therapeutic window. For instance, if the requirements are \(AUC_i \geq 30\) mg·h/L and \(C_{\min,i} \geq 1.0\) mg/L, subject \(i\) contributes 1 only if both inequalities hold. If either one fails, that subject does not attain the joint target.
Calculation and Its Assumptions
Joint PTA is calculated by applying each criterion to each subject’s simulated endpoints, combining those subject-level pass/fail results with a logical “and,” and then taking the proportion that pass. The exposure values used together must belong to the same virtual subject under the same regimen and evaluation scenario. This paired construction is what preserves their simulated dependence.
Interpretation relies on several conditions. The virtual cohort should represent the intended population, the population PK model and simulation inputs should be suitable for the dose scenario, and the endpoint definitions and thresholds should match the decision question. The simulation must also treat variability consistently across endpoints: a shared subject-level random effect should remain shared when it influences both endpoints. If stochastic residual error is part of the defined exposure outcome, it must be incorporated according to that definition, not added inconsistently to only some endpoints.
For the estimate to behave as a Monte Carlo proportion across independent draws, simulated subjects should be independently sampled from the intended population distribution, even though endpoints within each subject may be dependent. The dependence among AUC, \(C_{\min}\), or other endpoints is not a violation of this condition. It is part of the joint event being estimated.
Worked Example: Counting Simultaneous AUC and Cmin Attainment
Consider a hypothetical simulation of 10 virtual subjects for one regimen. The criteria are \(AUC \geq 30\) mg·h/L and \(C_{\min} \geq 1.0\) mg/L. The table shows whether each subject passes each criterion; it does not represent a real study.
| Subjects | AUC criterion | Cmin criterion | Pass both |
|---|---|---|---|
| 1–6 | Pass | Pass | Yes |
| 7–8 | Pass | Fail | No |
| 9 | Fail | Pass | No |
| 10 | Fail | Fail | No |
Eight subjects meet the AUC criterion, so its marginal PTA is \(8/10=0.800\). Seven meet the \(C_{\min}\) criterion, giving marginal PTA \(7/10=0.700\). Six subjects meet both. Therefore, $$ \widehat{PTA}_{\mathrm{joint}}=\frac{6}{10}=0.600. $$
The result is a simulated joint attainment of 60.0% for this regimen and these two criteria. Multiplying the marginal PTAs would give \(0.800\times0.700=0.560\), which is not the observed joint proportion in this simulated cohort. That product would be an appropriate probability calculation only if the two attainment events were independent; the paired results here do not establish that assumption.
Why Marginal PTAs Do Not Determine Joint PTA
Let \(A\) be the event that one endpoint meets its criterion and \(B\) the event that another endpoint meets its criterion. The joint event is \(A \cap B\). Its probability depends on how often the events occur together, not just on how often each occurs separately. If \(A\) and \(B\) are independent, then \(P(A \cap B)=P(A)P(B)\). Without independence, that product need not equal the joint probability.
A short derivation gives useful bounds even when the dependence is unknown. Since \(A \cap B\) is contained in both \(A\) and \(B\), its probability cannot exceed either marginal probability. Also, the probability of the union satisfies \(P(A \cup B)=P(A)+P(B)-P(A \cap B)\leq 1\). Rearranging that inequality gives the lower bound below.
These bounds are a useful check, not a substitute for paired simulation results. For more than two criteria, the joint event is the intersection of all the events, and it must be counted by checking all criteria for each subject. Multiplying several marginal PTAs is justified only if the attainment events are mutually independent. Even pairwise independence among some events does not, in general, establish mutual independence of an entire set.
Worked Example: Comparing Two Regimens on the Same Joint Target
Suppose two hypothetical regimens are each evaluated using 12 simulated subjects. The joint target is \(AUC\) within 25–60 mg·h/L and \(C_{\min} \geq 0.8\) mg/L. The table groups subjects by whether they pass each criterion.
| Regimen | Pass both | Pass AUC only | Pass Cmin only | Pass neither |
|---|---|---|---|---|
| A | 7 | 2 | 1 | 2 |
| B | 9 | 1 | 1 | 1 |
For regimen A, 9 of 12 pass the AUC window and 8 of 12 pass the \(C_{\min}\) criterion. The joint estimate is $$ \widehat{PTA}_{\mathrm{joint,A}}=\frac{7}{12}=0.583. $$ For regimen B, 10 of 12 pass each marginal criterion, while 9 pass both: $$ \widehat{PTA}_{\mathrm{joint,B}}=\frac{9}{12}=0.750. $$
In these hypothetical simulations, regimen B has the larger joint PTA by 0.167, or 16.7 percentage points. This comparison concerns the defined simultaneous target; it does not establish that regimen B is preferable under other clinical, safety, or practical considerations. The marginal product for A would be \((9/12)(8/12)=0.500\), while for B it would be \((10/12)(10/12)=0.694\). Neither product is the corresponding observed joint proportion because the paired event counts, rather than an independence assumption, determine these estimates.
Monte Carlo Precision and Practical Checks
The joint indicator for each simulated subject is a binary result: 1 for passing all criteria, 0 otherwise. When the \(N\) virtual subjects are independent draws and the joint probability is \(p_{\mathrm{joint}}\), the joint count follows a binomial model with success probability \(p_{\mathrm{joint}}\). Replacing the unknown probability by its simulated estimate gives the estimated Monte Carlo standard error:
This describes simulation error in estimating the probability under the chosen model and inputs. It does not quantify uncertainty about whether the model, population assumptions, or target thresholds are correct.
Worked Example: Reporting Monte Carlo Error for Joint PTA
A hypothetical simulation of \(N=400\) independent virtual subjects finds 312 that meet all specified exposure criteria. The estimated joint PTA is \(312/400=0.780\). Its estimated Monte Carlo standard error is $$ SE\approx\sqrt{\frac{0.780(1-0.780)}{400}} =\sqrt{\frac{0.1716}{400}} =\sqrt{0.000429} =0.0207. $$
Using a normal approximation, an approximate 95% Monte Carlo interval is \(0.780\pm1.96(0.0207)\), or approximately 0.739 to 0.821. This interval describes the precision of the Monte Carlo estimate under the simulation assumptions; it is not a 95% probability statement about the fixed model-based joint PTA. With very small counts or estimates near 0 or 1, a normal approximation may be poor, so a suitable binomial interval or more simulation replicates may be preferable.
Several checks help detect mistakes. Confirm that each endpoint is evaluated using the intended units, time point or summary measure, and threshold, and that the logical rule really is “all criteria.” Recompute the joint count directly from subject-level records; it cannot exceed any constituent marginal count. For two criteria, it must also lie within the bounds derived above. If a joint estimate exceeds a marginal PTA, the event coding, denominator, or subject matching is inconsistent.
A further warning sign is an implausibly large change in joint PTA after endpoints have been simulated separately or reordered. Results must remain attached to their subject identifiers. Sorting each endpoint independently and then pairing rows creates artificial pairings and can alter the apparent dependence. Likewise, using the product of marginal PTAs silently replaces the simulated dependence with an independence assumption.
Approaches, Limitations, and a Common Pitfall
| Approach | What it answers | Main limitation |
|---|---|---|
| Paired, subject-level joint count | What proportion meet all criteria together? | Depends on the validity of the simulation and the specified targets. |
| Separate marginal PTAs | What proportion meet each criterion considered alone? | Does not determine the proportion meeting all criteria. |
| Product of marginal PTAs | What joint probability follows under independence? | Can be wrong when endpoint attainment events are dependent. |
| Minimum marginal PTA | An upper bound on joint PTA | Is not generally the joint PTA itself. |
The minimum marginal PTA is an upper bound because passing all criteria necessarily means passing each one. It may be much larger than the true joint PTA. Marginal PTAs and their shifts across dose levels can still be informative, but they answer distinct questions from simultaneous attainment.
A common mistake is to argue: “80% meet the AUC target and 70% meet the \(C_{\min}\) target, so \(80\%\times70\%=56\%\) meet both.” The correction is that 56% follows only under independence of the two attainment events. If both endpoints were derived from the same simulated subject profiles, the appropriate estimate is the fraction of those profiles passing both criteria. The dependence could make that fraction higher or lower than the product, subject to the probability bounds.
Joint PTA is specific to the set of endpoints, thresholds, regimen, population distribution, and simulation definitions chosen. Adding another criterion can only maintain or reduce the joint attainment proportion, since fewer subjects can pass all conditions. A high joint PTA does not by itself validate the model, establish a clinical benefit, or resolve uncertainty in the target definitions. It is a compact summary of simultaneous attainment under the stated assumptions, and should be reported alongside enough information to identify those assumptions and the marginal results.
Check Your Understanding
Answer each question using the subject-level definition of joint PTA.
- In 100 virtual subjects, 72 meet criterion A, 65 meet criterion B, and 50 meet both. What are the two marginal PTAs and the joint PTA?
- If two marginal PTAs are 0.80 and 0.70, what range of joint probabilities is compatible with those marginals?
- A simulation reports joint PTA 0.62 and marginal PTA 0.55 for one of the required endpoints. What should be checked?
- In a simulation of 250 independent virtual subjects, 200 pass all criteria. Calculate the joint PTA and its approximate Monte Carlo standard error.
- Why can multiplying two marginal PTAs give a different result from counting subjects who meet both criteria?
Solutions. (1) The marginal PTAs are \(72/100=0.720\) and \(65/100=0.650\); the joint PTA is \(50/100=0.500\). (2) The lower bound is \(\max(0,0.80+0.70-1)=0.500\), and the upper bound is \(\min(0.80,0.70)=0.700\). Thus the joint probability must be between 0.500 and 0.700, inclusive. (3) This is impossible for correctly defined nested events: passing all criteria requires passing that endpoint’s criterion, so joint PTA cannot exceed its marginal PTA. Check event coding, denominators, and whether results were matched by subject. (4) The joint PTA is \(200/250=0.800\). The estimated standard error is \(\sqrt{0.800(0.200)/250}=\sqrt{0.000640}=0.0253\). (5) The product is the joint probability only when the two attainment events are independent. Otherwise their within-subject dependence changes how often the criteria are met together.