When EWOC Has No Feasible Dose
EWOC controls dose assignment by limiting the posterior probability that the dose under consideration is above the maximum tolerated dose (MTD). Usually, this leaves at least one dose available for assignment. But after accumulating toxicity data, the posterior distribution can shift downward: even the lowest dose on the trial’s grid may exceed the MTD with probability greater than the prespecified feasibility bound. In that situation, choosing the lowest dose simply because it is the lowest is not a valid application of EWOC. The rule has identified that no dose currently meets its safety constraint.
The appropriate response is a prespecified stop on new dose assignments, followed by a careful check of the data, model, and calculation. This tutorial formalizes that decision and explains why it is important to distinguish an operational pause from a final decision to end a trial.
Definition of the No-Feasible-Dose Condition
Suppose the trial has an ordered grid of available doses \(d_1<d_2<\cdots<d_J\). Let \(x^*\) denote the MTD under the trial’s dose-toxicity model: the dose at which the modeled probability of dose-limiting toxicity reaches the target toxicity probability. Let \(D\) denote the observed trial data, and let \(\alpha\) be the prespecified EWOC feasibility bound. As in the earlier material on EWOC feasibility bounds, a dose is feasible when its posterior probability of exceeding the MTD is no greater than \(\alpha\).
The MTD and overdose probability are defined by the model and its posterior, not by a simple count of toxicities. The usual setup assumes a prespecified target toxicity probability, an appropriate dose-toxicity model, a prior distribution, and a likelihood for the observed outcomes. The model should represent the dose ordering appropriately: higher doses are not assumed to have lower toxicity probabilities than lower doses. Outcomes used in the update must be correctly recorded and sufficiently observed for the likelihood being used. If follow-up is incomplete, the analysis must use a model that accounts for that fact rather than treating pending outcomes as observed non-toxicities.
The inequality is strict for the event \(x^*<d_j\): a dose exactly equal to the MTD is not counted as exceeding it. Under a continuous posterior for \(x^*\), exact equality has probability zero. With a discrete posterior representation, the trial’s convention for ties should be stated and applied consistently. A dose with overdose probability exactly \(\alpha\) meets the feasibility condition.
Why Failure at the Lowest Dose Is Decisive
For an ordered grid, the overdose events are nested. If \(d_j\geq d_1\), then every MTD below \(d_1\) is also below \(d_j\). Therefore, posterior overdose probability cannot decrease as the candidate dose rises. This gives a simple check on the no-feasible-dose decision.
Consequently, if the lowest dose fails, every higher dose fails as well. This argument depends on comparing the same posterior MTD distribution across the dose grid. If software reports a higher feasibility probability at a lower dose than at a higher one, check the dose ordering, calculation, and implementation rather than treating that pattern as a reason to assign a dose.
Worked Examples
Worked Example: Every Dose Fails the Bound
Consider a hypothetical trial with available doses \(d_1=10\), \(d_2=20\), and \(d_3=30\) mg, feasibility bound \(\alpha=0.25\), and posterior probabilities for the MTD shown below. For simplicity, suppose the posterior puts mass only at 5, 15, 25, and 35 mg.
| Posterior MTD value \(x^*\) (mg) | Posterior probability |
|---|---|
| 5 | 0.32 |
| 15 | 0.28 |
| 25 | 0.25 |
| 35 | 0.15 |
At 10 mg, only the posterior mass at 5 mg is below the candidate dose, so \(P(x^*<10\mid D)=0.32\). Since \(0.32>0.25\), the lowest dose is infeasible.
At 20 mg, the MTD values 5 and 15 mg are below the dose: \(P(x^*<20\mid D)=0.32+0.28=0.60\). At 30 mg, the values 5, 15, and 25 mg are below the dose: \(P(x^*<30\mid D)=0.32+0.28+0.25=0.85\). Thus the feasibility probabilities are 0.32, 0.60, and 0.85, respectively, all greater than 0.25. No dose is feasible, so EWOC calls for stopping new assignments rather than enrolling at 10 mg.
Worked Example: Checking a Posterior-Drawing Calculation
Suppose a hypothetical analysis uses 10,000 independent draws from the posterior distribution of \(x^*\). In 3,020 draws, \(x^*<d_1\), where \(d_1\) is the lowest available dose. With \(\alpha=0.25\), estimate the lowest-dose overdose probability and assess whether Monte Carlo error plausibly explains the result.
The estimated posterior probability is the proportion of draws below \(d_1\):
For independent posterior draws, the Monte Carlo standard error of this estimated probability is approximately \(\sqrt{\hat{q}(1-\hat{q})/M}\), where \(\hat{q}=0.302\) and \(M=10000\). Substitution gives \(\sqrt{0.302(0.698)/10000}=\sqrt{0.0000210796}\approx0.00459\). The estimate is 0.052 above the bound, about 11.3 Monte Carlo standard errors (\(0.052/0.00459\)). Ordinary Monte Carlo fluctuation is therefore not a plausible explanation for crossing the bound in this example.
The estimated probability exceeds \(\alpha\), so the calculation indicates no feasible dose, provided the posterior draws and model are valid. This numerical check does not alter the rule: Monte Carlo uncertainty is a computational diagnostic, not permission to relax the prespecified bound. For correlated MCMC draws, the independent-draw formula is not appropriate; use diagnostics and an effective sample size that account for autocorrelation.
Worked Example: Updating Evidence at the Lowest Dose
Consider a hypothetical two-state model used only to illustrate a posterior update at the lowest dose. State A says the MTD is below \(d_1\), with prior probability 0.20 and toxicity probability 0.60 at \(d_1\). State B says the MTD is at or above \(d_1\), with prior probability 0.80 and toxicity probability 0.10 at \(d_1\). Three participants treated at \(d_1\) have two dose-limiting toxicities. Assume their outcomes are conditionally independent given the state and its toxicity probability. Find the posterior probability that the MTD is below \(d_1\), and apply \(\alpha=0.25\).
Under State A, the likelihood of exactly two toxicities among three participants is \(\binom{3}{2}(0.60)^2(0.40)=3(0.36)(0.40)=0.432\). Under State B, it is \(\binom{3}{2}(0.10)^2(0.90)=3(0.01)(0.90)=0.027\).
Multiplying by the prior probabilities gives unnormalized posterior weights of \(0.20(0.432)=0.0864\) for State A and \(0.80(0.027)=0.0216\) for State B. The total is \(0.0864+0.0216=0.1080\). Therefore,
In this illustrative model, State A is precisely the event that \(x^*<d_1\). The posterior overdose probability at the lowest dose is 0.800, greater than 0.25. The lowest dose therefore fails, and the nesting result means every higher dose fails too. A real trial would use its prespecified dose-toxicity model and the full observed data; this small two-state calculation illustrates the update, not a substitute model for clinical use.
Operational Checks and Failure Modes
A no-feasible-dose result is a safety decision under a specified model, not proof that every dose is toxic or that the model is certainly correct. Before acting on an unexpected result, verify the items that can change the calculated posterior:
- Data and follow-up: confirm dose assignments, toxicity definitions, event dates, and which outcomes are complete. An outcome still pending must not silently be coded as no toxicity.
- Model inputs: check the target toxicity probability, prior, skeleton or dose-toxicity parameterization, and dose units against the protocol.
- Posterior calculation: verify that the event \(x^*<d_1\) was evaluated using the correct lowest dose and the stated tie convention.
- Numerical reliability: examine convergence and Monte Carlo error if posterior simulation was used. A noisy estimate near \(\alpha\) calls for better computation, not an improvised change to \(\alpha\).
- Protocol response: follow the prespecified hold and review process. Any resumption or change in dose availability must be authorized and scientifically justified, not made solely to obtain a feasible assignment.
A temporary enrollment hold and permanent trial termination are different decisions. The mathematical result immediately rules out assigning a new participant to any dose under the current posterior and bound. Whether the study then resumes after a valid review, remains paused, or ends depends on prespecified governance and safety procedures. The stopping rule itself should not be silently modified after seeing the data.
Common Mistakes and Alternatives
The most consequential mistake is to reason, “The lowest dose is safer than the others, so assign it anyway.” It may be lower risk than the higher doses, but EWOC’s condition is not a ranking rule: it requires the posterior probability of exceeding the MTD to be at most \(\alpha\). If the lowest dose exceeds that bound, assigning it violates the rule just as surely as assigning a higher dose.
Another mistake is to increase \(\alpha\) after observing that no dose is feasible. That changes the prespecified safety constraint in response to the data. A different bound can be considered only through an appropriate prospective design or formal protocol amendment, not as an informal way to continue enrollment.
| Response | What it means | Relation to the no-feasible-dose rule |
|---|---|---|
| Stop new assignments and review | Pause enrollment or dosing while checking the analysis and protocol response. | Preserves the current EWOC constraint while allowing a governed review. |
| Assign the lowest dose anyway | Continue despite the lowest dose violating the bound. | Does not satisfy EWOC and should not be treated as an acceptable workaround. |
| Terminate the dose-finding stage | End dose assignments under the study’s stopping provisions. | May be the required final action after review, depending on the protocol. |
This stopping decision is specific to EWOC’s feasibility criterion. Other dose-selection rules may use different quantities and thresholds, but switching rules after seeing that EWOC has no feasible dose does not preserve the original design. Comparisons among dose-selection methods require their own prespecified operating characteristics and safety criteria.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(P(x^*<d_1\mid D)=0.18\) and \(\alpha=0.20\), does the no-feasible-dose stopping condition hold?
- For an ordered grid, explain why \(P(x^*<d_1\mid D)>\alpha\) implies that \(d_3\) is also infeasible.
- A simulation estimates the lowest-dose overdose probability as 0.27 with Monte Carlo standard error 0.03, and \(\alpha=0.25\). What should be checked before treating this as a settled stopping decision?
- Why is “assign the lowest dose because it is safer than the other doses” not a valid EWOC justification when the lowest dose fails the bound?
- Does a no-feasible-dose result, by itself, prove that the trial must be permanently terminated? Explain.
Solutions. (1) No. Since \(0.18\leq0.20\), the lowest dose is feasible, so the feasible set is not empty. (2) The event \(x^*<d_1\) is contained in \(x^*<d_3\), so the latter probability is at least as large; if the first exceeds \(\alpha\), so does the second. (3) Check the posterior computation, model inputs, dose ordering, data and follow-up, and simulation convergence. Here the estimate is only \(0.02/0.03\approx0.67\) standard errors above the bound, so Monte Carlo uncertainty is substantial; improve the computation and follow the protocol’s review procedure rather than changing \(\alpha\). (4) EWOC requires the overdose probability at the assigned dose to be at most \(\alpha\); being the least risky available dose does not make a dose feasible when it violates that condition. (5) No. It requires stopping new assignments under the current calculation, but a prespecified review may lead to a valid resumption, continued pause, or termination as the protocol directs.