EWOC When the Current Dose Becomes Infeasible

When New Data Change the Dose Decision

An EWOC recommendation is based on the posterior distribution of the maximum tolerated dose (MTD), not on a fixed label attached to a dose. After a new cohort has completed toxicity follow-up, the evidence can shift that distribution toward lower MTD values. A dose that was feasible before the update can then exceed the permitted posterior overdose probability. The appropriate response is to update the posterior and reassess the full dose grid, rather than to keep using the previous recommendation.

This tutorial considers complete binary toxicity outcomes and a dose grid ordered from lower to higher doses. It focuses on the decision after the current dose becomes infeasible, assuming that at least one lower dose remains feasible. If no dose is feasible, the trial needs a separate prespecified safety action; that situation is not resolved by pretending that the current dose still passes the bound.

Definition: Let \(M\) denote the MTD, defined as the dose at which the modelled toxicity probability reaches the prespecified target \(\theta\). For a candidate dose \(d\), its posterior overdose probability after data \(D\) is \(P(M<d\mid D)\). Under EWOC with feasibility bound \(\gamma\), dose \(d\) is feasible when \(P(M<d\mid D)\leq\gamma\). The bound \(\gamma\) is fixed by the trial design; it is not a significance level.

The event \(M<d\) means that the candidate dose lies above the MTD. Equality is not overdose under this definition. In a continuous-dose model, a single equality point generally has posterior probability zero; in a discrete MTD representation, the strict inequality matters and should be applied consistently. The EWOC rule considered here recommends the highest feasible dose on the grid.

Updating the Feasibility Calculation

Suppose the posterior distribution after previously observed data \(D\) assigns probabilities to possible values of \(M\). A new cohort is treated at dose \(x\), and its outcome is \(Y\). Bayes' rule updates each posterior probability by multiplying it by the likelihood of the new outcome under that MTD value, then normalizing. Thus the updated overdose probability at candidate dose \(d\) is

P(M<d\mid D,Y) = \frac{\sum_{m<d}P(Y\mid M=m,D)\,P(M=m\mid D)} {\sum_m P(Y\mid M=m,D)\,P(M=m\mid D)}.

This expression makes an important point: the observed toxicity count does not itself determine whether a dose is feasible. Its effect depends on how likely that count is under each possible MTD value, as specified by the model. In the examples below, patient outcomes are conditionally independent given the dose-toxicity model, and a cohort's likelihood is binomial. These are modelling assumptions for the examples, not guarantees about every trial.

Assumptions for these updates: The dose-toxicity model is correctly specified for the calculation; toxicity probability increases with dose; the MTD and target \(\theta\) have the stated definitions; and the included outcomes are observed over the protocol-defined toxicity window. Conditional on the model parameters and assigned doses, patient outcomes are treated as independent. In practice, the analysis must also follow the protocol's rules for missing or incomplete toxicity assessments.

Worked Example: A Previously Feasible Dose Crosses the Bound

Consider four ordered dose levels and let \(M\) take values 1, 2, 3, or 4. Before the latest cohort, the posterior probabilities are \(0.05, 0.15, 0.50,\) and \(0.30\), respectively. The current dose is level 3, and the prespecified feasibility bound is \(\gamma=0.25\). Before the new data, its overdose probability is \(P(M<3\mid D)=0.05+0.15=0.20\), so it is feasible.

Three patients are treated at level 3, and two experience a dose-limiting toxicity (DLT). For this hypothetical model, the toxicity probabilities at level 3 for \(M=1,2,3,4\) are \(0.90,0.70,0.30,0.10\). They decrease with increasing \(M\), as expected: if the MTD is low, level 3 is more likely to be overly toxic. The binomial likelihood for two DLTs among three patients is \(\binom{3}{2}p^2(1-p)\).

The likelihoods for \(M=1,2,3,4\) are therefore \(0.243, 0.441, 0.189,\) and \(0.027\). Multiplying by the prior posterior weights gives unnormalized weights \(0.01215, 0.06615, 0.09450,\) and \(0.00810\), with total \(0.18090\). The updated overdose probability at level 3 is calculated from these unrounded weights:

P(M<3\mid D,Y) = \frac{0.01215+0.06615}{0.18090} = \frac{0.07830}{0.18090} = 0.4328358\ldots \approx 0.4328.

Since \(0.4328>0.25\), level 3 is now infeasible. Check the lower candidate, level 2:

P(M<2\mid D,Y) = \frac{0.01215}{0.18090} = 0.0671641\ldots \approx 0.0672.

Level 2 is feasible, so the highest feasible level is 2. The recommendation for the next cohort is to de-escalate from level 3 to level 2, subject to any additional protocol rules and clinical review. The patients already treated at level 3 are not “undone”; the calculation changes the next-dose recommendation.

Why Feasibility Changes in an Ordered Way

For a fixed posterior distribution of \(M\), the overdose events are nested across increasing doses. If \(d_1<d_2\), then every MTD value below \(d_1\) is also below \(d_2\). Therefore,

\{M<d_1\}\subseteq\{M<d_2\} \quad\Longrightarrow\quad P(M<d_1\mid D)\leq P(M<d_2\mid D).

This establishes that, on an ordered grid, the feasible doses form a lower segment of the grid: once a dose is infeasible, every higher dose is infeasible under the same posterior and bound. It also explains why the standard EWOC selection rule uses the highest dose that passes, rather than choosing a higher dose based on its proximity to a target toxicity probability.

Worked Example: Check the Boundary and Choose the Grid Maximum

After a posterior update, suppose the probabilities for \(M=1,2,3,4\) are \(0.10, 0.15, 0.50,\) and \(0.25\). With \(\gamma=0.25\), calculate the overdose probabilities for candidate doses 2, 3, and 4.

At dose 2, only \(M=1\) is below the dose, so \(P(M<2\mid D)=0.10\). At dose 3, the relevant MTD values are 1 and 2, giving \(P(M<3\mid D)=0.10+0.15=0.25\). Because the feasibility definition uses “less than or equal to,” dose 3 passes exactly at the bound. At dose 4, the probability is \(0.10+0.15+0.50=0.75\), so dose 4 fails.

The highest feasible dose is level 3. Rounding should not turn a value just above the bound into a pass, or a value just below it into a failure. Compare using unrounded probabilities whenever possible, and state the numerical precision used when displaying results.

What the Trial Team Should Do

When an updated calculation makes the current dose infeasible, the operational response is to stop recommending that dose for new assignments and identify the highest dose that still satisfies the prespecified bound. The posterior should be recomputed using all eligible observed outcomes, not just the latest cohort, because the decision is conditional on the accumulated data. Any escalation restriction, cohort timing rule, or clinical safety review in the protocol also remains in force; EWOC's feasibility calculation does not replace those requirements.

A fixed de-escalation rule based only on the number of DLTs is an alternative kind of decision rule. It may be operationally simple, but it does not directly calculate posterior overdose probability and may respond differently to the same observed count when prior information or the dose-toxicity model differs. EWOC instead uses the model, the prior, and all included outcomes to quantify uncertainty about \(M\). That model-based advantage also creates a limitation: a misspecified dose-toxicity model or a poorly calibrated prior can make the reported posterior risk misleading.

Worked Example: No DLTs Can Restore Feasibility

Use the same prior posterior probabilities and dose-3 toxicity probabilities as in the first example: \((0.05,0.15,0.50,0.30)\) for \(M=1,2,3,4\), and \((0.90,0.70,0.30,0.10)\) for toxicity at dose 3. Suppose instead that none of three patients has a DLT. The likelihood is \((1-p)^3\), giving likelihoods \(0.001,0.027,0.343,\) and \(0.729\).

The unnormalized posterior weights are \(0.00005,0.00405,0.17150,\) and \(0.21870\). Their sum is \(0.39430\). Hence the updated overdose probability at dose 3 is

P(M<3\mid D,Y) = \frac{0.00005+0.00405}{0.39430} = \frac{0.00410}{0.39430} = 0.0103987\ldots \approx 0.0104.

This is below \(\gamma=0.25\), so dose 3 is feasible under this updated posterior. The contrast with the two-DLT outcome is not a rule that “zero DLTs always permits escalation”: it is the consequence of the stated likelihood and posterior. The decision still requires checking all other protocol constraints.

Diagnostics and Common Pitfalls

A dose becoming infeasible can be a real consequence of new evidence, but errors in data handling or computation can produce the same appearance. Before acting on a changed recommendation, verify the cohort assignment, the number of evaluable patients, and the DLT coding against the protocol. Confirm that the outcomes have completed the required assessment window and that no patient has been counted twice or omitted. Then check that the likelihood corresponds to the actual cohort size and observed count.

Using the old posterior. A dose that passed before the cohort does not automatically pass after it. Recompute the posterior and the risk for each candidate dose after incorporating the new data.
Rounding before comparing. Rounded posterior weights may not sum to exactly one, and summing rounded terms can change the last reported digit. Compute the risk from unrounded weights, then round for display.
Changing the bound after seeing the result. Raising \(\gamma\) to keep the current dose feasible defeats the prespecified safety constraint. A change to the design requires a formal, prospectively governed decision, not an ad hoc response to an inconvenient posterior.
Confusing a DLT count with overdose probability. Two DLTs do not by themselves imply a particular EWOC risk. The risk is \(P(M<d\mid D)\), obtained by the model-based posterior calculation.
Ignoring ordering or normalization. Posterior probabilities must be nonnegative and sum to one before rounding. With an ordered dose grid and the same posterior, overdose probability should not decrease as candidate dose increases; a decrease signals a likely indexing or calculation error.

If a computed risk is unexpectedly high or the feasible set is empty, first verify the inputs, likelihood, posterior weights, and dose ordering. Do not silently substitute a lower risk calculation or select an infeasible dose. When at least one lower feasible dose remains, the highest such dose is the EWOC recommendation under the rule taught here; clinical monitoring and the protocol's other protections remain necessary.

Key takeaway: New toxicity data can make a previously feasible current dose infeasible by shifting posterior probability toward MTD values below that dose. Recompute from the full updated posterior, apply the unchanged feasibility bound, and—when a lower feasible dose exists—recommend the highest dose that still passes.

Check Your Understanding

Answer each question using the EWOC definitions in this tutorial. Solutions follow the questions.

For posterior probabilities \(P(M=1,2,3,4\mid D)=(0.08,0.17,0.45,0.30)\), bound \(\gamma=0.25\), and current dose 3, is the current dose feasible? What is the highest feasible dose?
For posterior probabilities \((0.06,0.24,0.50,0.20)\) and \(\gamma=0.25\), calculate the overdose probability at doses 2 and 3. What should be recommended if the current dose is 3?
A posterior before a new cohort assigns probabilities \((0.20,0.30,0.50)\) to \(M=1,2,3\). The new outcome has likelihoods \((0.50,0.20,0.10)\) under those MTD values. Calculate the updated probability \(P(M<3\mid D,Y)\).
A colleague says, “The current dose had two DLTs, so it is automatically unsafe; we do not need the posterior calculation.” What is wrong with this reasoning?

Solutions. (1) At dose 3, \(P(M<3\mid D)=0.08+0.17=0.25\), so it passes at equality. Dose 4 has overdose probability \(0.70\) and fails; the highest feasible dose is 3. (2) At dose 2 the risk is \(0.06\). At dose 3 it is \(0.06+0.24=0.30\), which exceeds \(0.25\); dose 3 is infeasible and dose 2 is the highest feasible choice on this grid. (3) The unnormalized weights are \(0.20(0.50)=0.10\), \(0.30(0.20)=0.06\), and \(0.50(0.10)=0.05\), with total \(0.21\). Thus \(P(M<3\mid D,Y)=(0.10+0.06)/0.21=0.7619\), approximately. (4) A DLT count alone does not equal posterior overdose probability. The EWOC calculation depends on the likelihood, prior posterior, dose, and all included data; the count is evidence to update the posterior, not a substitute for calculating the risk and comparing it with the prespecified bound.