Why a Dose Must Meet Both Criteria
A dose can produce strong efficacy for many individuals and still be unacceptable if it also produces toxicity too often. Conversely, a dose with little toxicity may fail to provide adequate efficacy. An efficacy–safety dose window makes these requirements explicit: at a candidate dose, an individual is counted as a success only when the efficacy criterion and the safety criterion are both met.
The key is to evaluate both outcomes for the same simulated individual. The population probability of meeting both criteria depends on how efficacy and toxicity co-occur, not just on their separate rates. The earlier material on joint efficacy targets and dependence establishes why this pairing matters; here the joint event is used to define an acceptable dose and a dose window.
The endpoints must be defined before calculating this probability. For example, efficacy might be an exposure-response index evaluated over a dosing interval, while toxicity might be a peak concentration or a binary adverse-event outcome over a stated period. The efficacy target, toxicity limit, dose regimen, evaluation time, and target population are part of the estimand. Changing any of them can change the probability and therefore the resulting window.
The probability describes population variability under the model: individuals differ through the random effects and any within-individual variability included in the endpoint simulation. Unless parameter uncertainty is deliberately included, it is conditional on the fitted model parameters. Do not interpret ordinary simulated individuals as parameter draws; adding parameter uncertainty is a separate modeling choice.
Estimating the Joint Probability by Simulation
Suppose \(M\) simulated individuals represent draws from the target population. For each individual, simulate both endpoints at dose \(d\), retaining their pairing. Define an indicator \(I_i(d)\) equal to 1 when both criteria are met and 0 otherwise. The Monte Carlo estimate is the proportion of simulated individuals with joint success:
If the \(M\) simulated individuals are independent draws from the intended population distribution, the indicators are Bernoulli draws with success probability \(p_{\mathrm{joint}}(d)\). Thus the estimate is unbiased conditional on the model and simulation setup, and its Monte Carlo standard error is estimated by
This standard error measures finite-simulation error, not uncertainty about whether the model is correct or whether its parameters are known. A small Monte Carlo standard error does not repair a misspecified model, an unrepresentative target population, or an incorrectly defined endpoint.
From Joint Probability to a Dose Window
A dose window requires a decision rule in addition to the joint probability. Let \(q\) be the prespecified minimum probability required for a dose to be acceptable. For a set of candidate doses \(\mathcal{D}\), the estimated acceptable set is
The value of \(q\) is a decision threshold, not a confidence level. It should be set according to the dose-selection context rather than chosen after inspecting which doses pass. If candidate doses are evaluated only on a grid, the result identifies acceptable grid points; it does not prove that every dose between two passing points is acceptable. Nor is a dose window necessarily a single continuous interval: the joint probability need not change monotonically with dose.
Worked Example: Counting Joint Successes
Consider 10 hypothetical simulated individuals at a candidate dose. Suppose the efficacy target is \(E^*=5\) units and the toxicity limit is \(T^*=8\) units. The simulated paired outcomes are:
| Individual | Efficacy \(E_i(d)\) | Toxicity \(T_i(d)\) | Both criteria met? |
|---|---|---|---|
| 1 | 6 | 4 | Yes |
| 2 | 7 | 9 | No |
| 3 | 5 | 3 | No |
| 4 | 8 | 7 | Yes |
| 5 | 4 | 2 | No |
| 6 | 9 | 8 | No |
| 7 | 6 | 6 | Yes |
| 8 | 7 | 5 | Yes |
| 9 | 3 | 9 | No |
| 10 | 10 | 7 | Yes |
Because the criteria are strict, an efficacy value of exactly 5 does not exceed the target, and toxicity of exactly 8 is not below the limit. Individuals 1, 4, 7, 8, and 10 meet both criteria. Therefore,
The estimated joint success probability is 50.0% in this small hypothetical simulation. This calculation is an estimate for the specified population and model, not a statement that the dose will succeed for exactly half of any future group of 10.
Worked Example: Why the Joint Probability Is Not the Product of Marginal Rates
In a larger hypothetical simulation of 1,000 individuals at one dose, 900 meet the efficacy target, 850 meet the safety limit, and 810 meet both. The marginal estimates are \(900/1000=0.900\) for efficacy and \(850/1000=0.850\) for safety. The observed joint estimate is
Multiplying the two marginal probabilities would give
which is not the simulated joint probability. The product would represent the joint probability only under independence of the two success indicators. In this hypothetical set, 810 of 900 efficacy successes also meet the safety criterion, so the conditional safety success rate among efficacy successes is \(810/900=0.900\). The paired outcomes, rather than the product of the marginals, give the required estimate.
If the decision threshold is \(q=0.800\), this dose passes because \(0.810\ge0.800\). That conclusion is based on the estimated joint probability; it does not establish that the dose is acceptable under every plausible parameter value or every target population.
Short Justification of the Simulation Estimate
The indicator \(I_i(d)\) equals 1 with probability \(p_{\mathrm{joint}}(d)\) and 0 otherwise. Its expectation is therefore \(E[I_i(d)]=p_{\mathrm{joint}}(d)\). Taking expectations of the sample average gives
For independent individuals, the variance of each indicator is \(p_{\mathrm{joint}}(d)[1-p_{\mathrm{joint}}(d)]\). The variance of their average is consequently \(p_{\mathrm{joint}}(d)[1-p_{\mathrm{joint}}(d)]/M\), which yields the plug-in standard-error formula above. This justification depends on independent, identically distributed draws from the defined population simulation. Correlated simulation replicates or unequal sampling weights require an appropriate variance calculation rather than this simple formula.
Worked Example: Identifying an Acceptable Dose Window
A hypothetical simulation evaluates four candidate doses. At each dose, 10,000 paired individuals are simulated using the same efficacy and safety definitions. Suppose the prespecified decision threshold is \(q=0.800\).
| Dose (mg) | Joint successes | Estimated joint probability | Passes \(q=0.800\)? |
|---|---|---|---|
| 50 | 6,200 | 0.620 | No |
| 100 | 8,400 | 0.840 | Yes |
| 150 | 8,800 | 0.880 | Yes |
| 200 | 7,300 | 0.730 | No |
For example, at 100 mg the estimate is \(8,400/10,000=0.840\), which exceeds the threshold. At 200 mg it is \(7,300/10,000=0.730\), which does not. The estimated acceptable set on this candidate grid is therefore \(\{100,150\}\) mg. It is reasonable to describe these as the passing candidate doses, but the grid alone does not establish a continuous acceptable interval from 100 to 150 mg.
At 100 mg, the estimated Monte Carlo standard error is
The simulation estimate is thus subject to about 0.00367 of Monte Carlo standard error, or 0.367 percentage points. This quantifies simulation precision under the assumed model; it is not a confidence assessment of the model assumptions themselves.
Diagnostics, Alternatives, and Limitations
Before interpreting a dose window, check that the simulation implements the intended question. The same simulated individual must contribute both outcomes at each dose, with any dependence induced by shared random effects or residual structure retained. Verify that the endpoint definitions, units, strict or inclusive boundary rules, regimen, and evaluation period match the decision specification. Check marginal efficacy and safety rates as well as the joint count: unexpected marginal rates can reveal a coding or model-input error even when the joint estimate appears plausible.
- Too few simulations near the decision threshold. If an estimate is close to \(q\), Monte Carlo error may affect whether it passes. Increase the number of independent simulated individuals and reassess the estimate and its simulation error.
- Unpaired outcomes. Simulating efficacy and toxicity in separate populations and combining them discards their dependence. Simulate both outcomes for each individual, then count joint successes.
- Boundary mistakes. Applying \(\ge\) where the specified rule is \(>\), or \(>\) where it is \(\ge\), can change the count, especially for discrete endpoints or rounded values. Implement the stated decision rule exactly.
- Over-interpreting a dose grid. Passing doses do not guarantee that untested doses between them pass. Evaluate additional doses if the shape or location of the window matters.
- Confusing variability with parameter uncertainty. A large number of simulated individuals can estimate the population probability precisely conditional on fixed parameters, while uncertainty about those parameters remains. More individuals do not resolve that distinct source of uncertainty.
The paired simulation estimate is generally preferable to multiplying separate efficacy and safety probabilities because it directly estimates the event of interest without assuming independence. If a valid joint probability model is available, analytic calculation may be more efficient, but it still must represent the same dependence and endpoint definitions. In nonlinear mixed-effects applications with complex exposure-response relationships, simulation is often a practical way to preserve those features. The estimate remains only as credible as the model, target population, and endpoint assumptions; simulation does not validate them.
Check Your Understanding
Answer each question, then compare your reasoning with the concise solutions below.
- In 2,000 paired simulations, 1,500 individuals meet the efficacy target, 1,600 meet the safety limit, and 1,300 meet both. What is the estimated joint probability?
- Why is the product of the marginal success estimates not automatically the joint success estimate?
- A dose has estimated joint probability 0.79 when the prespecified threshold is 0.80. Does it pass the stated rule?
- What does the Monte Carlo standard error describe, and what important source of uncertainty does it not automatically describe?
- Only doses of 80 mg and 120 mg were simulated, and both pass. What can be concluded about doses between them?
Solutions. (1) The estimate is \(1,300/2,000=0.650\), or 65.0%. (2) The product equals the joint probability only under independence of the two success indicators; paired simulation directly estimates their joint event. (3) No. Since \(0.79<0.80\), the dose does not meet the specified threshold. (4) It describes finite-simulation variability conditional on the simulation setup and model; it does not by itself account for model misspecification or parameter uncertainty. (5) The two tested doses pass on the simulated grid. No conclusion follows for untested doses between them without additional evaluation or justified assumptions about the dose-response pattern.