EWOC Feasibility Boundaries for Dose Escalation

From MTD Uncertainty to a Feasibility Boundary

A dose-escalation decision must account for uncertainty in the dose-toxicity relationship, not just the estimated toxicity probability at a candidate dose. EWOC (escalation with overdose control) expresses this uncertainty through the posterior distribution of the maximum tolerated dose (MTD). A candidate dose is feasible when the posterior probability that it lies above the MTD does not exceed a chosen tolerance.

The boundary is therefore a feature of a posterior distribution, not a fixed dose that is intrinsically safe. It changes as toxicity outcomes update the model, and it depends on both the target toxicity probability and the chosen feasibility tolerance.

Definition: Let \(p(d;\theta)\) be the probability of dose-limiting toxicity (DLT) at dose \(d\), given model parameters \(\theta\), and let \(\tau\) be the target DLT probability. The MTD, denoted \(\nu(\theta)\), is the dose at which the model toxicity probability reaches \(\tau\), when such a dose exists. If the curve is continuous and strictly increasing on the dose range and attains \(\tau\) there, the MTD is the unique solution to \(p(\nu;\theta)=\tau\). For a nondecreasing curve that may not attain the target, a useful generalized definition is \(\nu(\theta)=\inf\{d:p(d;\theta)\geq\tau\}\), with \(\nu=+\infty\) if this set is empty. This generalized definition need not satisfy \(p(\nu;\theta)=\tau\).

Let \(D\) denote the data observed so far. For a candidate dose \(d\), the EWOC feasibility probability is

q(d)=P\{\nu(\theta)<d\mid D\}.

The strict inequality encodes the event that the candidate dose is above the MTD. Given a chosen tolerance \(\alpha_{\mathrm{EWOC}}\), the dose is feasible if \(q(d)\leq\alpha_{\mathrm{EWOC}}\). This tolerance is a design choice: it is not automatically the significance level \(\alpha\) used in a hypothesis test.

Deriving the Boundary

As the candidate dose increases, the event \(\{\nu<d\}\) can only grow: if an MTD draw is below one dose, it is also below any higher dose. Thus \(q(d)\) is nondecreasing in \(d\). Feasible doses form a lower portion of the dose range, and the EWOC boundary is the largest dose that meets the feasibility condition.

d_{\mathrm{EWOC}} = \sup\{d:q(d)\leq\alpha_{\mathrm{EWOC}}\}.

If the posterior distribution of \(\nu\) is continuous, then \(q(d)=F_{\nu\mid D}(d)\), where \(F_{\nu\mid D}\) is its posterior cumulative distribution function. When the level \(\alpha_{\mathrm{EWOC}}\) is attained within the available dose range, the boundary is the rightmost dose at which \(F_{\nu\mid D}(d)=\alpha_{\mathrm{EWOC}}\), equivalently \(\sup\{d:F_{\nu\mid D}(d)\leq\alpha_{\mathrm{EWOC}}\}\); this can differ from the conventional lower \(\alpha_{\mathrm{EWOC}}\) quantile if the CDF is flat at that level. This gives a direct derivation: the boundary is chosen so that the posterior probability the MTD lies below it equals the tolerated probability of overdose.

With a discrete posterior distribution, or an empirical distribution formed from a finite number of posterior draws, equality at a quantile may not be possible. In that case, evaluate \(P(\nu<d\mid D)\) at each allowable candidate dose and apply the feasibility condition directly. This also handles atoms at a candidate dose correctly: an MTD equal to \(d\) is not counted in the strict event \(\nu<d\).

Important distinction: EWOC directly controls \(P(\nu<d\mid D)\), under the stated MTD definition. It equals \(P\{p(d;\theta)>\tau\mid D\}\) when, for each relevant parameter value, the dose-toxicity curve is continuous and strictly increasing and reaches the target on the dose range. Without those conditions, the two probabilities need not be equal. For any particular curve, calculate the direct event using its actual \(p(d;\theta)\); do not infer it from the MTD event without checking the curve.

Calculating MTD Draws in a Logistic Model

For illustration, consider the two-parameter logistic model used in BLRM work, with a positive dose \(d\), reference dose \(d_{\mathrm{ref}}\), intercept \(a\), and slope \(b>0\):

p(d;a,b)=\frac{\exp\{a+b\log(d/d_{\mathrm{ref}})\}} {1+\exp\{a+b\log(d/d_{\mathrm{ref}})\}}.

For a target \(0<\tau<1\), setting \(p(\nu;a,b)=\tau\) and taking logits gives \(\log\{\tau/(1-\tau)\}=a+b\log(\nu/d_{\mathrm{ref}})\). Since \(b>0\), this curve is continuous and strictly increasing for positive doses. Its probability ranges from 0 toward 1, so the target is attained on the positive dose axis, though the root may fall outside a trial's allowed dose range. Solving yields

\nu(a,b)=d_{\mathrm{ref}}\exp\left[ \frac{\log\{\tau/(1-\tau)\}-a}{b} \right].

A posterior draw of \((a,b)\) therefore gives one MTD draw. Applying this calculation to all posterior draws approximates the posterior MTD distribution; its lower tail determines the EWOC boundary. The requirement \(b>0\) matters: if the model permits nonpositive slopes, the curve may not be strictly increasing, and the usual root and overdose interpretation require extra care.

Worked Example: Feasibility From Posterior MTD Draws

Suppose eight equally weighted posterior draws of the MTD are 20, 30, 35, 40, 50, 60, 70, and 80 mg. For this hypothetical decision, let \(\alpha_{\mathrm{EWOC}}=0.25\). Apply the strict feasibility event \(P(\nu<d\mid D)\leq0.25\) to candidate doses of 30, 35, and 40 mg.

At 30 mg, only the 20 mg MTD draw is strictly below the candidate dose, so \(q(30)=1/8=0.125\). At 35 mg, the 20 mg and 30 mg draws are below it, so \(q(35)=2/8=0.250\). At 40 mg, the 20, 30, and 35 mg draws are below it, so \(q(40)=3/8=0.375\).

Thus 30 mg and 35 mg satisfy the feasibility condition, while 40 mg does not. Among these candidates, 35 mg is the highest feasible dose. The MTD draw equal to 35 mg is not counted as below the 35 mg candidate; replacing the strict inequality with \(\nu\leq d\) would change this calculation.

Worked Example: Computing the Boundary From Logistic Posterior Draws

Suppose three equally weighted posterior draws have \(d_{\mathrm{ref}}=10\) mg, \(\tau=0.30\), and parameter values \((a,b)=(-2.0,1.2)\), \((-1.5,0.9)\), and \((-2.5,1.5)\). The target logit is \(\log(0.30/0.70)\approx-0.8473\). Calculate the MTD for each draw, then use \(\alpha_{\mathrm{EWOC}}=0.50\) to assess candidate doses 20, 25, and 30 mg.

For the first draw, \(\nu=10\exp\{(-0.8473+2.0)/1.2\}=10\exp(0.9606)\approx26.13\) mg. For the second, \(\nu=10\exp\{(-0.8473+1.5)/0.9\}=10\exp(0.7252)\approx20.65\) mg. For the third, \(\nu=10\exp\{(-0.8473+2.5)/1.5\}=10\exp(1.1018)\approx30.10\) mg.

At 20 mg, no MTD draw is below the candidate dose, giving \(q(20)=0/3=0\). At 25 mg, only 20.65 mg is below it, giving \(q(25)=1/3\approx0.333\). At 30 mg, the 20.65 mg and 26.13 mg draws are below it, giving \(q(30)=2/3\approx0.667\). Therefore 20 mg and 25 mg are feasible, and 25 mg is the highest feasible of these candidates under the 0.50 tolerance.

The continuous-posterior median would be the 0.50 quantile of the MTD distribution. These three equally weighted draws give only a coarse empirical approximation, so candidate-dose feasibility is more transparent than treating a single interpolated empirical quantile as an exact boundary.

When MTD Feasibility and Toxicity Risk Differ

The MTD-based event and direct toxicity event coincide under a continuous, strictly increasing curve that reaches the target. They can differ if the curve has a plateau, a jump, or a target that is not attained on the dose range. For example, with \(\tau=0.30\), consider one curve with \(p(d)=0.20\) for \(d<1\) and \(p(d)=0.30\) for \(d\geq1\), and a second curve with \(p(d)=0.20\) for \(d<1\) and \(p(d)=0.40\) for \(d\geq1\). Both are nondecreasing and have generalized MTD \(\nu=1\).

Worked Example: A Plateau Changes the Overdose Probability

Suppose the two curves in the preceding example each have posterior probability 0.50, and consider candidate dose \(d=2\). For both curves, \(\nu=1<2\), so the posterior MTD-based probability is \(P(\nu<2\mid D)=1\). But the first curve has \(p(2)=0.30\), which does not strictly exceed the target, while the second has \(p(2)=0.40\), which does. Therefore \(P\{p(2)>\tau\mid D\}=0.50\), not 1.

With a hypothetical tolerance of 0.50, the direct strict-toxicity probability meets that threshold, while the MTD-based probability does not. This is not a contradiction: the rules assess different events. A decision protocol must specify which event it controls and how equality with the target is treated.

Applying the Boundary in a Dose Decision

In practice, posterior simulation supplies parameter draws conditional on the observed DLT data. For each draw, calculate the MTD according to the model and the prespecified MTD definition. Then estimate \(q(d)\) at each candidate dose as the proportion of draws for which \(\nu<d\). A dose is feasible only if this estimated probability is no larger than \(\alpha_{\mathrm{EWOC}}\). If doses are selected from a discrete grid, recommend the highest feasible candidate only after applying any additional protocol constraints, such as restrictions on dose skipping.

Specify the target and MTD definition.
State the target DLT probability, dose range, curve assumptions, and what happens if the target is not attained within that range.

Transform posterior draws.
For each draw of the model parameters, solve for the MTD or apply the prespecified generalized definition.

Evaluate candidate doses.
Estimate \(P(\nu<d\mid D)\) separately at each allowable dose and compare it with \(\alpha_{\mathrm{EWOC}}\).

Apply trial constraints and review.
Select from feasible doses according to the protocol, and check posterior computation, model behavior, and the direct toxicity probability when relevant.

The posterior sample should be large and well-mixed enough to estimate a tail probability reliably. If \(M\) approximately independent posterior draws are used and \(k\) have MTD below \(d\), the estimate is \(\hat q(d)=k/M\); its approximate Monte Carlo standard error is \(\sqrt{\hat q(d)\{1-\hat q(d)\}/M}\). Correlated simulation draws reduce the effective information, so this formula should not be interpreted as an exact precision measure for a poorly mixing chain.

Diagnostics, Alternatives, and Limitations

Check that the model's posterior draws yield finite, meaningful MTDs and that the slope and dose-response assumptions support the calculation. For a logistic model, inspect whether draws satisfy \(b>0\), whether the implied MTDs lie far outside the allowed dose range, and whether the posterior tail near the feasibility boundary is stable across simulation runs. Compare estimated feasibility probabilities at neighboring candidates: they should not decrease as dose increases. A decrease can signal a coding error, inconsistent MTD calculations, or inadequate Monte Carlo precision.

EWOC's MTD-based probability rule differs from applying a posterior toxicity-risk rule directly at each dose. Under the strictly increasing, target-attaining conditions stated above, the events are equivalent and so are their probabilities. Outside those conditions, direct evaluation of \(P\{p(d;\theta)>\tau\mid D\}\) answers a distinct question. The earlier material on posterior overdose-control rules provides the general setting for risk-based decisions; here, the key issue is that EWOC obtains its boundary through the posterior distribution of the MTD.

Do not equate a quantile with certainty. The boundary is based on a chosen posterior tail probability; it does not guarantee that a dose below it is safe for every possible parameter value.
Do not ignore the dose grid. A continuous boundary may fall between available doses. Assess the actual candidates rather than rounding upward without checking feasibility.
Do not assume every model draw has a usable root. If the target is not reached on the allowed range, use the prespecified convention or report that the MTD is outside the range; do not silently discard the draw.
Do not treat the tolerance as a universal constant. The choice of \(\alpha_{\mathrm{EWOC}}\) reflects the trial's safety and escalation objectives and should be justified in the design.

EWOC's boundary is only as useful as the model and posterior supporting it. Model misspecification, weakly identified slopes, limited data, or poorly calibrated priors can make the posterior MTD distribution misleading. A feasibility threshold is not a substitute for reviewing the model, the posterior risk at the proposed dose, the clinical context, and the protocol's stopping and escalation rules.

Check Your Understanding

Answer each question, then compare with the brief solutions.

If posterior MTD draws are 10, 20, 30, and 40 mg, what is \(P(\nu<30\mid D)\)?
For a continuous posterior MTD distribution with CDF \(F_{\nu\mid D}\), how is the EWOC boundary related to \(\alpha_{\mathrm{EWOC}}\)?
Why can a dose-toxicity curve's nondecreasing property alone fail to make \(P(\nu<d\mid D)\) equal \(P\{p(d;\theta)>\tau\mid D\}\)?
In the logistic MTD formula, what assumption on \(b\) ensures the curve is strictly increasing in dose?
If estimated feasibility probabilities are 0.10, 0.20, and 0.18 at increasing doses, name a possible concern.

Solutions. (1) Two of the four draws are strictly below 30 mg, so the probability estimate is \(2/4=0.50\); the draw equal to 30 mg is excluded. (2) For a continuous posterior, \(d_{\mathrm{EWOC}}=\sup\{d:F_{\nu\mid D}(d)\leq\alpha_{\mathrm{EWOC}}\}\); if the level is attained in range, this is the rightmost dose with \(F_{\nu\mid D}(d)=\alpha_{\mathrm{EWOC}}\), not necessarily the conventional lower quantile when the CDF is flat. (3) A plateau at the target can make \(\nu<d\) true while \(p(d;\theta)>\tau\) is false. (4) Require \(b>0\). (5) Feasibility should be nondecreasing with dose; the decrease may indicate Monte Carlo variability, a calculation error, or a problem in the MTD implementation.