Tracing a Go-or-No-Go Decision From Question to Residual Risk

From a Clinical Question to a Decision

A development decision should be traceable: a reviewer should be able to follow the path from the clinical question to the evidence, model, decision threshold, recommendation, and remaining uncertainty. A go decision does not mean that a drug is effective or safe in every relevant setting. It means that, under stated assumptions and criteria, the current evidence supports continuing a specified development plan.

Consider a wholly hypothetical early oncology study of a new treatment in patients with a defined tumor type after standard therapy. The immediate question is whether the response and tolerability signals justify a larger, controlled study. This is not the same as asking whether the treatment improves survival compared with a control. The available single-arm data can inform a development decision, but they cannot by themselves establish comparative benefit.

To make the question operational, define \(p\) as the response probability in the intended study population and \(q\) as the probability of a prespecified serious treatment-related toxicity over the study’s defined observation period. Suppose the development team considers a response probability above 0.20 promising enough to justify further study, while a serious-toxicity probability of 0.20 or greater is unacceptable for this decision. These are hypothetical decision benchmarks, not universal oncology standards.

A prespecified rule can then require strong evidence for both conditions: \(p>0.20\) and \(q<0.20\). For this case, define “go” as a posterior probability of at least 0.85 that both conditions hold. The numerical threshold reflects the consequences of continuing versus stopping, the cost and feasibility of the next study, and the team’s tolerance for false progression. It must be chosen before interpreting the observed data, rather than selected to make the result look favorable.

Evidence and Model Assumptions

Suppose 9 of 31 evaluable patients had a confirmed response, and 0 of 13 patients in a separately summarized safety cohort experienced the specified serious toxicity during the observation period. The efficacy and safety cohorts need not have identical follow-up or analysis sets; their definitions must be checked before combining their results.

For a transparent efficacy summary, suppose each patient’s response is a binary outcome, patients are independent, and all 31 responses are assessed consistently. Let \(R\) be the number of responders. Conditional on a common response probability \(p\), use \(R\sim\operatorname{Binomial}(31,p)\). These assumptions can fail if response ascertainment differs, patients have materially different response probabilities, or dependence arises from shared sites or other structure.

For the decision calculation, use a uniform prior \(p\sim\operatorname{Beta}(1,1)\). With a binomial likelihood, the posterior is \(p\mid\text{data}\sim\operatorname{Beta}(1+R,1+n-R)\), where \(n\) is the number assessed. The prior is part of the analysis, not a fact supplied by the study; a different justified prior can change the resulting probability.

For safety, suppose a separate, prespecified Bayesian summary uses a beta-binomial model. In this hypothetical case, the prior is \(q\sim\operatorname{Beta}(1,1.721)\), representing an explicitly elicited prior, and 0 toxicities are observed in 13 patients. Under the assumptions of a common toxicity probability and independent binary outcomes, the posterior is \(q\mid\text{data}\sim\operatorname{Beta}(1,14.721)\). This unusual prior is stated openly because hiding its influence would make the safety summary difficult to assess.

Combining Evidence and Applying the Rule

The decision concerns both endpoints at once. Even if separate analyses provide posterior probabilities for efficacy and safety, their marginal probabilities cannot simply be multiplied unless posterior independence is justified. A conservative lower bound for the probability that both requirements hold follows from the union bound:

The bound does not require independence between \(p\) and \(q\). It is conservative: the actual joint posterior probability may be higher, but cannot be inferred from the marginals alone without further information about their dependence. The bound is approximately 0.8727, above the prespecified go threshold of 0.85.

Justification, Diagnostics, and Alternatives

The joint lower bound can be justified by considering the two ways the desired joint event can fail. Let \(A=\{p>0.20\}\) and \(B=\{q<0.20\}\). The complement of \(A\cap B\) is \(A^c\cup B^c\), so:

The inequality follows because the probability of a union is no greater than the sum of the probabilities of its component events. It remains valid when the efficacy and safety estimates are dependent. If a defensible joint model is available, directly calculating the posterior probability of \(A\cap B\) may be less conservative; the joint model’s assumptions then need their own scrutiny.

Before relying on the calculation, check whether response assessments are complete and consistently defined, whether the patients represent the population in the proposed next study, and whether the safety observation period is long enough to detect the toxicity of concern. Review event counts and patient-level patterns for evidence of site effects, heterogeneity, or outcome dependence that conflicts with the common-probability binomial models. Check prior sensitivity, especially for the safety probability informed by no observed events. A result that changes materially under plausible prior choices should be reported as fragile, not as a precise decision signal.

The main alternative is a frequentist analysis of the same binary outcomes, such as estimating response and toxicity rates with confidence intervals and comparing them with benchmarks. Such intervals describe repeated-sampling properties under their assumptions; they do not directly give the posterior probability that the decision conditions hold. Another alternative is a joint model that represents response and toxicity together, potentially accounting for dependence. It may use information more efficiently, but its structure can be harder to validate. The choice should follow the decision question and the data, not the method that produces the most favorable recommendation.

A go/no-go framework also has limits beyond statistical uncertainty. The benchmark response rate and toxicity ceiling are decision inputs, not biological constants. A go decision may be inappropriate if the study population, endpoint definition, follow-up, or treatment context differs from that assumed in the criteria. A model cannot repair weak endpoint ascertainment or make a single-arm comparison causal. The next study should be designed to resolve the uncertainties that matter most, including comparative benefit and the precision of safety estimates.

Common Pitfall

A common error is to say, “There is a 92.5% chance the drug works, so the treatment is proven.” The 92.5% value is a posterior probability, under a specific model and prior, that the response probability exceeds a chosen benchmark. It does not establish comparative benefit, guarantee future results, or remove uncertainty about toxicity. The correction is to state the event precisely, include its assumptions, apply the joint decision rule, and name the important uncertainties that remain.