Why Multiple Endpoints Need a Joint View
A dose can meet a target for one clinical endpoint while missing a target for another. For example, a simulated population might have high attainment of a symptom-response target but lower attainment of a functional-improvement target. Reporting only one endpoint can therefore conceal an important part of the dose profile. Reporting the marginal attainment of each endpoint is informative, but it does not by itself show how often the same individuals attain all the required targets.
The central task is to define the clinical success event for each endpoint and the rule that combines those events. When the rule requires several targets to be met by the same individual, the probability of joint attainment must reflect the dependence among that individual’s simulated outcomes. The earlier material on joint efficacy targets and dependence establishes why this matters; here the focus is on translating distinct clinical endpoints into a dose-selection calculation.
Endpoints may have different scales, directions, and clinically justified thresholds. One endpoint might require a response indicator to equal 1; another might require a continuous score change to be at least a specified amount. If a larger value is unfavorable for an endpoint, its event must instead use the appropriate upper bound. Each event should be defined in clinical terms before comparing doses. A model output is not a target merely because it is available.
The combination rule is also part of the estimand. “All endpoints must meet their targets” means an intersection. “At least two of three targets must be met” is a different event. A composite rule should be specified from the clinical decision question, not chosen after inspecting which rule makes a candidate dose look best.
Estimating Attainment From Joint Simulations
For a candidate dose, simulate a representative set of individuals from the intended population. Each individual should have a complete vector of endpoint outcomes generated under that dose. The simulation should retain the shared sources of variation that connect endpoints, such as common individual characteristics or common pharmacokinetic and pharmacodynamic quantities. Apply each endpoint’s target to each simulated outcome, then apply the pre-specified combination rule to that individual’s resulting success indicators.
If \(N\) individuals are simulated and \(I_i(d)\) equals 1 when individual \(i\) satisfies the joint rule and 0 otherwise, the estimated joint probability is the proportion satisfying that rule. The same calculation, using the endpoint-specific indicator, estimates each marginal attainment probability.
This calculation is a direct Monte Carlo estimate: each simulated individual contributes one indicator for the event of interest. Conditional on the model and dose, if the simulated individuals are independent draws from the target population, the indicators are independent Bernoulli variables with success probability \(p\), where \(p\) is the joint attainment probability. Therefore the estimate has expectation \(p\) and variance \(p(1-p)/N\). Its Monte Carlo standard error is estimated by \(\sqrt{\hat p(1-\hat p)/N}\). This standard error describes finite simulation error conditional on the model; it does not measure uncertainty in the model or its parameter estimates.
Worked Example: Two Distinct Targets in Ten Simulated Individuals
Suppose a hypothetical dose is assessed in 10 simulated individuals. Endpoint 1 meets its target when the response score is at least 4 points. Endpoint 2 meets its target when the functional score improves by at least 2 points. The two endpoints are simulated together for each individual, producing these success patterns:
| Endpoint 1 met? | Endpoint 2 met? | Number of individuals |
|---|---|---|
| Yes | Yes | 6 |
| Yes | No | 2 |
| No | Yes | 1 |
| No | No | 1 |
Endpoint 1 is attained by \(6+2=8\) individuals, so its estimated marginal attainment is \(8/10=0.80\). Endpoint 2 is attained by \(6+1=7\) individuals, giving \(7/10=0.70\). If the clinical rule requires both targets in the same individual, the joint estimate is \(6/10=0.60\).
The product of the marginal estimates is \(0.80(0.70)=0.56\), not the observed joint estimate of \(0.60\). Mutual independence would guarantee the product formula for the joint probability. Without an independence assumption, the marginals generally do not determine the joint probability, although their product can coincide with the joint probability by chance. Here the paired simulation directly estimates the quantity needed for the all-targets rule.
Deriving the Joint Estimate and Its Precision
The indicator formulation explains both how to calculate the probability and what the simulation error means. For each individual, \(I_i(d)\) is a Bernoulli random variable: it is 1 with probability \(p=P(\text{joint rule met at }d)\), and 0 otherwise. Thus \(E[I_i(d)]=p\) and \(\operatorname{Var}(I_i(d))=p(1-p)\). For independent simulated individuals, the variance of their average is
The first equality uses independence across simulated individuals, not independence among endpoints within an individual. Endpoint dependence is preserved inside each joint indicator. Replacing the joint indicator with a product of separate marginal probabilities would discard that within-individual dependence unless the relevant probability factorization is justified.
Worked Example: Applying Marginal and Joint Criteria to Two Doses
Consider two hypothetical candidate doses, each evaluated using 1,000 simulated individuals. Endpoint 1 and endpoint 2 have distinct clinical success criteria, and the pre-specified decision rule requires estimated attainment of at least 0.80 for each endpoint separately and at least 0.75 for both endpoints in the same individual. The simulated counts are:
| Dose | Endpoint 1 met | Endpoint 2 met | Both met |
|---|---|---|---|
| 40 mg | 840 | 810 | 760 |
| 60 mg | 910 | 890 | 820 |
For 40 mg, the marginal estimates are \(840/1000=0.840\) and \(810/1000=0.810\); the joint estimate is \(760/1000=0.760\). All three estimates meet their respective criteria. For 60 mg, the marginal estimates are \(0.910\) and \(0.890\), and the joint estimate is \(0.820\); these also meet the criteria. Under a rule that selects the lowest candidate dose meeting all specified criteria, 40 mg is selected from these two candidates.
For a rough measure of finite simulation error at 40 mg, the estimated Monte Carlo standard error for joint attainment is \(\sqrt{0.760(1-0.760)/1000}=\sqrt{0.0001824}\approx0.01351\). This quantifies simulation variability conditional on the model and parameters; it does not establish that the underlying clinical targets, model, or parameter values are correct. The selection here follows the stated estimated-probability rule, not a claim that the true probability is known exactly.
Endpoint Rules Can Define Different Events
An all-endpoints rule is not the only possible clinical rule, but the chosen rule must be explicit. If the decision requires at least two of three endpoint targets, the relevant event is the union of the success patterns containing two or three successes. It is not the same as the intersection requiring all three. A pattern table makes the distinction auditable and prevents a marginal summary from being mistaken for a composite probability.
Worked Example: An “At Least Two of Three” Rule
For a hypothetical dose, 200 simulated individuals have the following joint success-pattern counts. “1” means that the endpoint-specific target is met; “0” means it is not.
| Endpoint pattern | Count |
|---|---|
| 111 | 84 |
| 110 | 26 |
| 101 | 18 |
| 011 | 22 |
| 100 | 16 |
| 010 | 12 |
| 001 | 10 |
| 000 | 12 |
The counts sum to \(84+26+18+22+16+12+10+12=200\). Endpoint 1 is met by \(84+26+18+16=144\) individuals, so its marginal estimate is \(144/200=0.720\). Endpoint 2 is met by \(84+26+22+12=144\), also \(0.720\). Endpoint 3 is met by \(84+18+22+10=134\), or \(0.670\).
For an all-three rule, only pattern 111 qualifies, giving \(84/200=0.420\). For the specified at-least-two rule, patterns 111, 110, 101, and 011 qualify. The estimated attainment is \((84+26+18+22)/200=150/200=0.750\). The marginal probabilities alone would not reveal either composite estimate; the pattern counts preserve the joint outcomes.
Diagnostics, Alternatives, and Limitations
A useful review begins with the target definitions and the simulated records, not just the final percentages. Check that each endpoint’s threshold, direction, assessment time, and handling of an unavailable or undefined outcome match the clinical question. Confirm that the joint rule is applied to outcomes from the same simulated individuals. If each endpoint is simulated in a separate unpaired run, the marginal estimates may still be available, but their joint attainment is not identified by simply combining those runs.
Inspect the pattern counts or a cross-tabulation of the endpoint success indicators. They should sum to the number of simulated individuals, and the marginal counts reconstructed from the patterns should agree with the reported endpoint-specific probabilities. Large discrepancies often indicate mismatched denominators, an incorrect event direction, or an implementation error. If results seem clinically implausible, inspect the endpoint outcomes and the shared simulation quantities that generate their dependence. A smooth or plausible marginal probability for each endpoint does not guarantee a plausible joint probability.
The main approaches answer different questions:
- Marginal attainment: estimate \(P(A_j(d))\) for each endpoint. This shows how often each target is met separately, but not how often they are met by the same individual.
- Joint simulation: apply the specified rule to each individual’s full outcome vector. This is the direct approach for an all-targets or other composite rule when the simulation can retain the relevant dependence.
- Product of marginals: mutual independence guarantees that the product of marginal probabilities equals the intersection probability. Without that assumption, the product generally does not determine the joint probability, even though equality can occur in a particular dependent configuration.
- Analytical joint calculation: if a suitable joint outcome distribution is available and justified, it can provide a joint probability without Monte Carlo sampling. It still requires a correct dependence model and a clearly defined clinical event.
Simulation-based estimates are conditional on the specified model, parameters, target population, dosing regimen, and endpoint definitions. A finite run adds Monte Carlo error, and rare joint success patterns can be estimated imprecisely unless enough individuals are simulated. Moreover, a high joint attainment probability does not by itself establish that the targets are clinically appropriate or that the model is adequate. Dose comparisons should therefore report the endpoint-specific and joint criteria, the simulated population and sample size, and the results needed to apply the pre-specified rule.
Check Your Understanding
Answer each question using the stated event definitions and counts.
- In 500 paired simulations, endpoint A is met by 410 individuals, endpoint B by 360, and both by 300. Calculate the two marginal estimates and the joint estimate for an all-targets rule.
- What does mutual independence guarantee about the probability of an intersection, and why is multiplying marginals not generally a substitute for estimating joint attainment?
- In 200 simulations, 70 individuals meet all three targets, 40 meet exactly two, and 30 meet exactly one. What is the estimated probability of meeting at least two targets?
- State one reason why a Monte Carlo standard error does not represent uncertainty in the dose recommendation arising from uncertainty in model parameters.
Solutions. 1. The marginal estimates are \(410/500=0.820\) for A and \(360/500=0.720\) for B. The joint estimate is \(300/500=0.600\). 2. Mutual independence guarantees that the intersection probability is the product of the marginal probabilities. Without independence, the marginals generally do not determine the intersection probability; the joint outcomes or a justified joint model are needed. 3. At least two targets are met by \(70+40=110\) individuals, so the estimate is \(110/200=0.550\). 4. The Monte Carlo standard error quantifies finite-simulation variation conditional on the model and parameter values used. It does not account for uncertainty about those parameter values or other model assumptions.