Deriving EWOC From a Posterior MTD Constraint

From an MTD to an Overdose Event

EWOC can be derived from a direct safety question: given the data observed so far, how probable is it that the proposed dose is above the maximum tolerated dose? The central step is to express that question as a posterior probability about the MTD. This is a probability about uncertainty in the dose-toxicity model, not the probability that a particular patient will experience a dose-limiting toxicity.

Let $d$ denote a proposed dose, $p_T$ the prespecified target toxicity probability, and $p(d;\theta)$ the modelled probability of toxicity at dose $d$ under parameter value $\theta$. Let $Y$ denote the accrued trial data. The posterior distribution of $\theta$, and hence of the MTD, is conditional on $Y$.

Definition: For a specified target $p_T$, define the parameter-dependent maximum tolerated dose $M(\theta)$ as the dose at which the modelled toxicity probability reaches the target. In the derivation below, assume the dose-toxicity curve is strictly increasing and crosses $p_T$ once, so that $p(M(\theta);\theta)=p_T$. The posterior MTD is the random quantity $M=M(\theta)$ induced by the posterior distribution of $\theta$ given $Y$.

The single-crossing assumption matters. Under it, doses below $M(\theta)$ have modelled toxicity probability below $p_T$, while doses above $M(\theta)$ have modelled toxicity probability above $p_T$. Thus, for a fixed parameter value, proposing $d$ is an overdose relative to the target precisely when $M(\theta)<d$. The strict inequality means that a dose exactly equal to the MTD is not counted as exceeding it.

Deriving the Posterior Constraint

Because $\theta$ is uncertain after observing $Y$, the event $M(\theta)<d$ has a posterior probability. Applying the event equivalence across the posterior distribution gives

P\{p(d;\theta)>p_T\mid Y\} = P\{M(\theta)<d\mid Y\}.

The equality is a statement about the model’s parameter uncertainty: each posterior draw of $\theta$ either places the MTD below $d$ or does not. EWOC requires that the resulting posterior probability not exceed a prespecified feasibility bound $\gamma$.

EWOC posterior MTD constraint: A proposed dose $d$ satisfies the constraint when $$ P(M<d\mid Y)\leq \gamma, $$ where $M$ is the posterior MTD and $\gamma$ is the chosen feasibility bound. Here $\gamma$ denotes a bound on posterior overdose probability; it is not the significance level $\alpha$ used for hypothesis tests.

The rule follows directly by writing the posterior probability as an integral over parameter values:

P(M<d\mid Y) = \int I\{M(\theta)<d\}\,\pi(\theta\mid Y)\,d\theta, \qquad \text{require } \int I\{M(\theta)<d\}\,\pi(\theta\mid Y)\,d\theta\leq\gamma.

Here $\pi(\theta\mid Y)$ is the posterior density and $I\{\cdot\}$ is an indicator, equal to 1 when its condition holds and 0 otherwise. Equivalently, if $F_M$ is the posterior cumulative distribution function of $M$, then $P(M<d\mid Y)=F_M(d^-)$, the probability mass strictly below $d$. When the posterior distribution is continuous at $d$, this equals $F_M(d)$.

This is the derivation: define the MTD as the target crossing; use monotonicity to identify overdose with $M<d$; then restrict the posterior probability of that event. The rule does not require the MTD to be known exactly. It controls the posterior mass of parameter values under which the proposed dose would exceed the target-defined MTD.

Worked Applications

Worked Example: A Continuous Posterior MTD

Suppose a hypothetical analysis gives a posterior distribution for the MTD that is uniform from 2 mg to 10 mg. The study team has specified $\gamma=0.25$. Calculate the posterior overdose probability for a proposed dose of 5 mg and assess the constraint.

For $2\leq d\leq10$, the fraction of the uniform posterior interval below $d$ is

P(M<d\mid Y)=\frac{d-2}{10-2}.

At $d=5$ mg,

P(M<5\text{ mg}\mid Y) = \frac{5-2}{10-2} = \frac{3}{8} = 0.375.

Since $0.375>0.25$, 5 mg does not satisfy the stated EWOC constraint. At 4 mg, the probability is $(4-2)/8=0.25$, so that dose meets the bound exactly. These calculations concern the posterior probability that the MTD is below each proposed dose, not a predicted fraction of patients who will experience toxicity.

Worked Example: A Discrete Posterior MTD

Suppose a hypothetical posterior assigns probabilities $0.10$, $0.25$, $0.40$, and $0.25$ to MTDs of 10, 20, 30, and 40 mg, respectively. Let $\gamma=0.30$. Evaluate proposed doses of 20 mg and 30 mg.

At 20 mg, only the posterior mass at an MTD of 10 mg lies strictly below the proposal:

P(M<20\text{ mg}\mid Y)=0.10.

This is at most $0.30$, so 20 mg satisfies the constraint. At 30 mg, the mass at 10 mg and 20 mg is below the proposed dose:

P(M<30\text{ mg}\mid Y)=0.10+0.25=0.35.

Because $0.35>0.30$, 30 mg does not satisfy the constraint. The posterior mass at exactly 30 mg is not included: the event is $M<d$, not $M\leq d$. This distinction is important whenever the posterior has point masses, as can happen when MTD values are represented by a finite set of possible doses.

Worked Example: Computing MTDs From a Logistic Model

Consider a hypothetical logistic dose-toxicity model $p(d;\beta)=\{1+\exp[-(-3+\beta d)]\}^{-1}$, with dose $d$ in mg, fixed intercept $-3$, and positive slope $\beta$ measured per mg. The target is $p_T=0.20$. Suppose the posterior has three equally likely values of $\beta$: 0.08, 0.10, and 0.12 per mg. Find the posterior overdose probability at 18 mg and compare it with $\gamma=0.30$.

For a fixed $\beta$, solve the target-crossing equation. Taking the logit of both sides gives

-3+\beta M=\log\left(\frac{0.20}{1-0.20}\right) =\log(0.25)=-1.3863, \qquad M=\frac{1.6137}{\beta}.

The resulting MTDs are approximately $20.17$ mg when $\beta=0.08$, $16.14$ mg when $\beta=0.10$, and $13.45$ mg when $\beta=0.12$. The MTD lies below 18 mg for the latter two posterior values, so

P(M<18\text{ mg}\mid Y)=\frac{2}{3}=0.667.

Since $0.667>0.30$, 18 mg fails the constraint. As a check on the event interpretation, at 18 mg the toxicity probabilities for the three slope values are approximately 0.174, 0.231, and 0.302. Exactly the latter two exceed the target 0.20, matching the two of three MTDs below 18 mg. The calculation relies on positive slopes and the logistic curve’s strict increase in dose, which ensure a unique target crossing.

Assumptions, Diagnostics, and Interpretation

The event equivalence and its clinical meaning depend on the model and the definition of the MTD. In particular, the derivation assumes that the toxicity probability increases with dose over the relevant range, that the target crossing exists, and that the crossing is unique. It also assumes that the posterior distribution being used reasonably represents uncertainty in the dose-toxicity parameters, conditional on the accrued data and the chosen model and prior.

Useful checks include verifying that posterior parameter draws produce finite MTDs within the model’s meaningful dose range, and confirming that the dose-toxicity curve is increasing over that range. If posterior draws imply no target crossing, multiple crossings, or implausible MTD values, then the MTD event may not be defined as assumed. The constraint can also be sensitive to model specification or prior information when data are sparse. Posterior sensitivity checks and the earlier material on skeleton and prior sensitivity are relevant for judging how stable the calculation is.

The rule is distinct from selecting a dose just because it is closest to a point estimate of the MTD. A point estimate, such as the posterior mean, does not by itself reveal how much posterior probability lies below a candidate dose. Two candidate doses can be on the same side of that estimate yet have quite different posterior overdose probabilities. Conversely, the EWOC constraint is not a guarantee that a trial will have no toxicities: it limits a model-based posterior probability about the target-defined MTD, not the outcome of each future patient.

Common pitfall: “The proposed dose has a 20% toxicity probability, so its posterior overdose probability must be 20%.” These are different quantities. The first is a toxicity probability at a dose under a specified parameter value (or a posterior predictive calculation, if uncertainty is integrated appropriately); the second is $P(M<d\mid Y)$, the posterior probability that the target-defined MTD is below the proposed dose. To apply the derived EWOC rule, calculate the latter from the posterior distribution of $M$.

If the model’s monotonicity or target-crossing assumptions fail, the simple event identity needs reconsideration rather than mechanical application. Likewise, if the posterior has discrete mass at the proposed dose, the strict-event convention must be implemented consistently. These are not merely numerical details: they determine exactly which parameter values count as overdosing.

Check Your Understanding

Answer each question using the posterior MTD constraint, then compare your reasoning with the solutions.

A continuous posterior MTD distribution is uniform from 4 mg to 12 mg. What is $P(M<7\text{ mg}\mid Y)$? Does 7 mg satisfy $\gamma=0.40$?
A discrete posterior assigns probabilities 0.2, 0.5, and 0.3 to MTDs of 10, 20, and 30 mg. What is $P(M<20\text{ mg}\mid Y)$, and why is the mass at 20 mg excluded?
State the assumption that lets the event $p(d;\theta)>p_T$ be rewritten as $M(\theta)<d$.
In one sentence, distinguish posterior overdose probability from the probability that a patient treated at dose $d$ experiences toxicity.

Solutions. (1) $(7-4)/(12-4)=3/8=0.375$, so 7 mg satisfies the bound because $0.375\leq0.40$. (2) The probability is 0.2; the mass at 20 mg is excluded because the event is strictly $M<20$ mg. (3) The modelled dose-toxicity curve must increase strictly with dose and have a unique crossing of the target. (4) Posterior overdose probability is the probability, given the data, that the MTD lies below the proposed dose, whereas patient toxicity probability concerns a toxicity outcome at that dose.

Deriving EWOC From a Posterior MTD Constraint

What You'll Learn