Overdose Control With Efficacy Requirements

Two Requirements, One Admissible-Dose Set

Dose finding must address two different questions: is the probability of unacceptable toxicity sufficiently low, and is the probability of efficacy high enough to make the dose worth considering? A dose can satisfy one condition and fail the other. An efficacy signal does not cancel an excessive toxicity risk, and apparent safety alone does not make an ineffective dose a useful choice.

The introductory material on posterior probabilities of acceptable doses treats posterior evidence about dose acceptability. Here the focus is the explicit combination of an overdose-control limit with a minimum-efficacy requirement, and how their intersection can constrain both escalation and final consideration. The numerical thresholds are design choices that must be specified before using the rule.

The strict inequality in the overdose event and the weak inequality in the efficacy event follow the definitions above; a protocol should use its chosen boundary conventions consistently. The thresholds \(\phi_T,\phi_E,\xi,\eta\) are not universal constants. They encode a clinical standard and a decision tolerance for uncertainty. For example, \(\phi_T=0.30\) means toxicity probabilities above 30% are considered excessive, while \(\xi=0.25\) allows at most 25% posterior probability that a dose exceeds that limit.

The two posterior checks are usually marginal: one concerns \(p_T(d)\), the other \(p_E(d)\). Passing both does not, by itself, imply that the posterior probability of a joint event involving both outcomes exceeds a specified value. If the design instead requires \(P\{p_T(d)\leq\phi_T,\ p_E(d)\geq\phi_E\mid D\}\) to exceed a threshold, that is a different rule and requires a joint posterior distribution. The dependence between efficacy and toxicity can matter for such a joint probability.

Posterior Checks Under a Beta-Binomial Model

For a transparent calculation, suppose that at each dose the toxicity outcomes are binary and conditionally independent across participants, with probability \(p_T(d)\). Assign \(p_T(d)\) a beta prior, and update it using the observed toxicity count. Use the analogous model for binary efficacy outcomes. With a \(\operatorname{Beta}(a,b)\) prior and \(x\) events among \(n\) participants, the posterior is \(\operatorname{Beta}(a+x,b+n-x)\). A uniform \(\operatorname{Beta}(1,1)\) prior therefore gives posterior \(\operatorname{Beta}(x+1,n-x+1)\).

These simple marginal models assume a consistent binary outcome definition, complete outcome ascertainment for the participants included, and the stated binomial sampling model at each dose. They do not require efficacy and toxicity outcomes to be independent within a participant for their marginal probabilities to be assessed separately. But a joint posterior probability cannot be calculated by multiplying marginal probabilities unless the posterior quantities are independent under an appropriate joint model.

When the posterior is beta with integer parameters, a beta-binomial identity can turn the efficacy probability into a finite binomial sum. If \(p\mid D\sim\operatorname{Beta}(a,b)\), \(a,b\) are positive integers, and \(X\sim\operatorname{Binomial}(a+b-1,q)\), then

This identity follows by differentiating the binomial sum on the right with respect to \(q\): its derivative is \(-q^{a-1}(1-q)^{b-1}/B(a,b)\), the negative beta density. Both sides equal 1 at \(q=0\), so integrating the matching derivatives establishes equality. It provides a direct way to evaluate the posterior efficacy condition using a binomial cumulative probability.

Using the Intersection to Constrain Decisions

The set of admissible doses is the intersection of the safety-passing set and the efficacy-passing set. This simple set operation has an important practical implication: a dose becomes eligible only when the evidence supports both conditions. If escalation is permitted only to an admissible dose, then the next dose must satisfy the overdose-control condition and the minimum-efficacy condition under the prespecified rule. This is a design choice, not a universal feature of escalation procedures; some protocols use efficacy chiefly to restrict final selection rather than to control every escalation.

A dose can fail either gate in either direction. High toxicity evidence can eliminate a dose even when efficacy evidence is compelling. Conversely, a dose with low posterior overdose probability can be excluded if evidence does not support the minimum efficacy requirement. If no dose passes both conditions, the rule does not identify an admissible dose; it does not justify relaxing a threshold after seeing the results.

Diagnostics, Alternatives, and Limitations

Before applying the gates, check that the outcome definitions, analysis populations, and assessment windows match those used to set the clinical boundaries. Delayed toxicity or efficacy outcomes can make a posterior calculation look more reassuring than the available evidence warrants if participants without mature outcomes are treated as though their outcomes were known. Likewise, small samples can leave broad posterior distributions; a point estimate below \(\phi_T\) is not a substitute for checking the posterior probability above \(\phi_T\).

A useful diagnostic is to examine the posterior probabilities across all doses, not just the dose currently under consideration. Unexpected reversals across ordered doses may reflect sampling variability, data problems, or a model that does not capture the dose-response pattern. Do not silently force monotonicity or smooth the estimates after observing results. If a model imposes a dose-response structure, assess whether its assumptions are plausible and report how that structure affects the gates. Also inspect sensitivity to reasonable prior choices, especially when data are sparse and the decision is close to a threshold.

Compared with a rule based only on toxicity, the combined rule prevents a dose from being considered merely because it appears tolerable when efficacy is unlikely to reach a meaningful minimum. Compared with a rule based only on efficacy, it prevents a promising efficacy estimate from overriding an unacceptable posterior overdose probability. A single utility score is another possible decision framework, but it allows different benefit-risk tradeoffs and is not equivalent to requiring both separate gates. The thresholds make the present rule interpretable, while also making clear which doses are excluded and why.

These rules depend on the clinical relevance of the cutoffs, the adequacy of the outcome models, and the timing and completeness of observed data. Separate marginal gates also do not summarize the probability that both true dose-level conditions hold simultaneously. If a design needs that joint statement, it must specify and fit a suitable joint model rather than infer it from two marginal checks.