A Single-Arm Oncology Response Analysis

Why a Single-Arm Response Analysis Needs a Decision Rule

In an early-phase oncology trial, a response rate is often compared with a prespecified benchmark representing a level of activity judged too low to justify further development. The analysis must answer two distinct questions: what response rates are plausible given the prior information and trial data, and whether the evidence meets the rule chosen for a go/no-go decision. A posterior probability can summarize the first question and inform the second, but the rule and its clinical threshold should be set before outcomes are examined.

The analysis below is hypothetical. It treats response as a binary outcome for each patient in a defined analysis population. It does not establish that treatment caused responses: in a single-arm study, the response rate is observed under treatment, without a concurrent randomized comparison.

Definition: Let \(p\) be the probability of response for a patient in the target population under the treatment and response-assessment procedures being studied. For \(n\) patients with \(x\) responses, a beta-binomial analysis specifies a beta prior for \(p\) and a binomial likelihood for \(x\). A go/no-go rule may be written as: go if \(P(p>p_0\mid x)\geq c\), where \(p_0\) is a prespecified response-rate threshold and \(c\) is a prespecified posterior-probability cutoff.

Model, Assumptions, and Updating

The basic sampling model is \(X\mid p\sim\operatorname{Binomial}(n,p)\). Conditional on the common response probability \(p\), each patient's response is modeled as a Bernoulli outcome, and the outcomes are conditionally independent. The number \(n\) is fixed for the analysis, and every patient counted in \(n\) has a response status defined under the trial's prespecified assessment and follow-up rules. The model does not account automatically for missing outcomes, varying follow-up, patient-level response probabilities, or changes in assessment practice.

A beta prior \(p\sim\operatorname{Beta}(a,b)\), with \(a>0\) and \(b>0\), is conjugate to this binomial likelihood. Here \(a\) and \(b\) are prior shape parameters, not necessarily literal counts of prior patients. After observing \(x\) responses and \(n-x\) nonresponses, the posterior distribution is \(\operatorname{Beta}(a+x,b+n-x)\). The posterior probability above a threshold is the area of this distribution above that threshold.

\begin{aligned} P(X=x\mid p) &\propto p^x(1-p)^{n-x},\\ \pi(p\mid x) &\propto p^{a-1}(1-p)^{b-1}p^x(1-p)^{n-x}\\ &=p^{a+x-1}(1-p)^{b+n-x-1},\\ p\mid x &\sim \operatorname{Beta}(a+x,b+n-x). \end{aligned}

The last line follows because the kernel is that of a beta density with updated shape parameters; its normalizing constant is the beta-function constant for those parameters. Thus the observed responses add \(x\) to the first shape parameter, while nonresponses add \(n-x\) to the second.

Assumptions to check: The response definition and analysis population are fixed; response status is sufficiently ascertained for the included patients; the binomial model's common-\(p\), conditional-independence approximation is reasonable; and the prior is defensible for this population and treatment setting. If enrollment, follow-up, or stopping depends on accumulating outcomes, the likelihood and decision analysis must reflect that design rather than treating the data as a simple fixed-\(n\) sample.

Worked Analysis: A Prespecified Efficacy Threshold

Suppose a hypothetical single-arm phase 2 study has 33 evaluable patients, of whom 11 have a confirmed response. The trial team prespecifies \(p_0=0.20\) as the minimum response rate of clinical interest, uses a uniform prior \(\operatorname{Beta}(1,1)\), and sets the go criterion at \(P(p>0.20\mid x)\geq0.90\). The analysis population, response criteria, and treatment of unevaluable patients would need to be specified in the protocol; here, the stated 33 patients are assumed to meet that definition.

Worked Example: Posterior Probability of Exceeding the Benchmark

Update the prior. There are \(x=11\) responses and \(n-x=33-11=22\) nonresponses. Therefore,

p\mid x=11\sim\operatorname{Beta}(1+11,1+22) =\operatorname{Beta}(12,23).

Calculate the posterior probability. The beta-binomial identity gives \(P(p>t)\) for \(p\sim\operatorname{Beta}(a,b)\) with positive integer \(a,b\) as \(P\{\operatorname{Binomial}(a+b-1,t)\leq a-1\}\). Here \(a+b-1=34\), so

\begin{aligned} P(p>0.20\mid x=11) &=\sum_{k=0}^{11}{34\choose k}(0.20)^k(0.80)^{34-k}\\ &\approx 0.9723344. \end{aligned}

Thus the posterior probability that the response rate exceeds 20% is about 97.23%. Because \(0.9723\geq0.90\), this hypothetical analysis meets its prespecified go criterion. The result is evidence, under this model and prior, that \(p\) exceeds the benchmark; it is not a 97.23% probability that the treatment works, nor does it establish a causal effect relative to an untreated or active-control group.

As a useful descriptive quantity, the posterior mean is \(12/(12+23)=12/35\approx0.3429\), or 34.29%. This mean is not the decision statistic: the decision rule uses the posterior probability above 20%, which accounts for the full posterior distribution.

Why the Tail Probability Has a Binomial Form

The identity used above can be justified directly for integer beta parameters. If \(p\sim\operatorname{Beta}(a,b)\), its density is proportional to \(p^{a-1}(1-p)^{b-1}\). For positive integers \(a,b\), the beta cumulative distribution at \(t\) equals the upper tail probability \(P\{Y\geq a\}\) for \(Y\sim\operatorname{Binomial}(a+b-1,t)\). Taking the complement gives \(P(p>t)=P(Y\leq a-1)\). In the example, \(a=12\), \(b=23\), and \(t=0.20\), producing the sum from \(k=0\) through 11 with 34 binomial trials. This is a computational identity, not a claim that the clinical trial enrolled 34 patients; its binomial variable is a device for evaluating the beta tail.

What the Rule Does—and Does Not—Establish

A go/no-go rule is a decision convention, not a universal statistical law. A cutoff of 0.90 means that the team proceeds when the posterior probability of exceeding the stated benchmark reaches at least 90%. Its practical meaning depends on the clinical consequences of advancing an ineffective treatment, abandoning a useful one, and the costs and design of the next study. Those considerations motivate the benchmark, cutoff, prior, and trial size. They should not be chosen after seeing the observed response count simply to obtain a desired decision.

The benchmark may come from historical experience or a clinical judgment about minimum activity. Either way, uncertainty about that benchmark and differences between historical and current patients are not represented by this simple model: \(p_0\) is treated as fixed. A single-arm response rate may also be affected by patient selection, assessment frequency, supportive care, and changes in standard practice. Consequently, exceeding a historical threshold supports a development decision under stated assumptions; it does not provide the same evidence as a randomized comparison.

The prior matters particularly when the sample is small. A uniform \(\operatorname{Beta}(1,1)\) prior assigns equal density to all response rates from 0 to 1. It is a transparent reference choice, not a statement that all rates are equally likely based on substantive oncology knowledge. A clinically informed prior could alter the posterior tail probability, so reporting the prior and checking reasonable alternatives are part of a credible analysis. This tutorial's numerical decision is conditional on the specified uniform prior.

Worked Example: A Trial With No Observed Responses

Consider a separate hypothetical cohort with \(n=33\) and \(x=0\), analyzed with the same \(\operatorname{Beta}(1,1)\) prior and benchmark \(p_0=0.20\). The posterior is

p\mid x=0\sim\operatorname{Beta}(1,34).

For a \(\operatorname{Beta}(1,b)\) distribution, the probability above \(t\) is \((1-t)^b\). Therefore,

P(p>0.20\mid x=0)=(1-0.20)^{34}=0.80^{34}\approx0.0005071.

This is about 0.05071%, far below the 0.90 go cutoff, so this hypothetical result would not meet the go criterion. The conclusion is tied to the prespecified rule; it is not a claim that the response rate is exactly zero.

Worked Example: Predicting Responses in a Future Cohort

Return to the posterior \(\operatorname{Beta}(12,23)\) from the 11-of-33 analysis. Let \(R\) be the number of responses among five future patients whose response outcomes follow the same model. Conditional on \(p\), \(R\sim\operatorname{Binomial}(5,p)\); integrating over the posterior gives a beta-binomial predictive distribution. In particular, the probability of no responses is

\begin{aligned} P(R=0\mid x=11) &=E\{(1-p)^5\mid x=11\}\\ &=\frac{23\cdot24\cdot25\cdot26\cdot27} {35\cdot36\cdot37\cdot38\cdot39}\\ &=\frac{9{,}687{,}600}{69{,}090{,}840} \approx0.1402. \end{aligned}

Thus the model assigns about a 14.02% posterior predictive probability to observing no responses in the next five patients. This prediction averages over uncertainty in \(p\); it is not the probability that \(p=0\), and it assumes that the future patients and response assessment are sufficiently comparable to those in the analyzed cohort.

Diagnostics, Alternatives, and Limitations

Before interpreting the posterior, check the data and design that feed it. Confirm the response count against the prespecified endpoint definition, identify whether any patients lack adequate assessment, and verify the analysis denominator. Review whether patients share a sufficiently stable response probability. Strong clinical heterogeneity, dependence between outcomes, or time-varying response ascertainment can make a single binomial probability inadequate. Examine the prior's implications on the response-rate scale and calculate whether plausible alternative priors would change the decision. A decision that changes under modest prior variations should be reported as sensitive, rather than presented as an unqualified go.

The main alternative for a single-arm benchmark question is a frequentist one-sample binomial procedure, such as a test or confidence interval for a response rate. That approach is framed through long-run sampling properties under specified hypotheses or coverage; the beta-binomial approach directly summarizes posterior uncertainty conditional on the model and prior. Neither removes the need to justify the benchmark, response definition, or single-arm design. A randomized treatment comparison addresses a different question by comparing arms, and is not interchangeable with exceeding a fixed single-arm benchmark.

The beta-binomial model is compact and analytically convenient, but its simplicity is also a limitation. It treats the response probability as common across patients, uses a fixed benchmark, and does not by itself model delayed responses, censoring, missingness, competing events, or treatment-effect comparisons. If these features are material, a model and decision rule that represent them are needed. The posterior probability should therefore be reported alongside the model, prior, threshold, cutoff, observed response count, and relevant design limitations.

Confusing the posterior probability with a probability that the null is true. The value \(P(p>p_0\mid x)\) is probability over parameter values conditional on data and assumptions; it is not a posterior probability for a hypothesis unless a separate model assigns probability to that hypothesis.
Changing the rule after observing the data. Selecting \(p_0\) or the cutoff to make the observed trial pass undermines the interpretation of the stated decision rule.
Treating the historical benchmark as a randomized control. A single-arm comparison cannot eliminate differences in patient mix, response assessment, or time period.
Ignoring the denominator. Reporting only “11 responses” is not enough: the analysis requires the number assessed and a clear account of patients not included.

Key takeaway: A beta-binomial single-arm analysis updates a prior for the response rate using response and nonresponse counts. Its posterior tail probability can support a prespecified development decision, but the conclusion remains conditional on the model, prior, response definition, benchmark, and limitations of a nonrandomized study.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

With prior \(\operatorname{Beta}(1,1)\), 8 responses among 20 patients produce what posterior distribution?
In the main example, why does the beta-tail calculation use a binomial variable with 34 trials even though the trial had 33 evaluable patients?
If the posterior probability above \(p_0\) is 0.88 and the prespecified cutoff is 0.90, what is the decision under that rule?
Does a posterior probability of 0.9723 that \(p>0.20\) mean there is a 97.23% probability the treatment is effective compared with no treatment? Explain.
Name two assumptions or design features that should be checked before relying on the simple binomial likelihood.

Solutions. (1) The posterior is \(\operatorname{Beta}(1+8,1+20-8)=\operatorname{Beta}(9,13)\). (2) The identity maps a \(\operatorname{Beta}(a,b)\) tail to a binomial tail with \(a+b-1\) trials; for \(\operatorname{Beta}(12,23)\), that number is 34. It is a computational identity, not the study sample size. (3) The go criterion is not met because \(0.88<0.90\); under the stated rule, this is no-go. (4) No. It is the posterior probability that the modeled response rate exceeds the fixed 20% benchmark, conditional on the prior, data, and model; it is not a randomized treatment comparison or a probability of causal efficacy. (5) Examples include adequate ascertainment of response status, a fixed and clearly defined analysis population, conditional independence of patient outcomes, and a sufficiently common response probability across patients.