Why Interpretive Precision Matters
A predictive monitoring result is a probability about a possible future trial outcome, not a report of what has already happened. Confusing it with the current posterior probability of efficacy can produce a misleading interim interpretation; ignoring how many outcomes remain can make an impossible outcome appear attainable. Even a correctly calculated predictive probability is not a guarantee: it averages over the uncertainty in future outcomes.
The earlier material on predictive monitoring for binary outcomes establishes the Beta-binomial calculation used here. The focus now is interpretation: which probability answers which question, what a threshold licenses the monitoring rule to do, and what it does not establish.
State the Rule and the Remaining Outcomes
For clarity, suppose the prior is \(p\sim\operatorname{Beta}(a,b)\), and individual binary outcomes are conditionally independent given \(p\), with each patient having the same response probability \(p\). If \(x\) responses and \(n-x\) nonresponses have been observed, the posterior is \(\operatorname{Beta}(a+x,b+n-x)\). This model assumes a common response probability over the patients being monitored and that the observed outcomes are recorded correctly.
Let \(N\) be the planned total number of patients with outcomes available for the terminal analysis, and let \(m=N-n\) outcomes remain. Write \(K\) for the number of responses among those \(m\) patients. Given the current posterior \(\operatorname{Beta}(A,B)\), where \(A=a+x\) and \(B=b+n-x\), \(K\) has a beta-binomial predictive distribution:
Here \(B(\cdot,\cdot)\) is the beta function. If terminal success is defined by a posterior rule such as \(P(p>p_0\mid\text{final data})\geq\gamma\), then the predictive probability of terminal success is the sum of the predictive probabilities for those \(k\) values that make the final data satisfy that rule. It is not generally the current posterior probability \(P(p>p_0\mid x,n)\), nor is it simply the probability of any response among the remaining patients.
A Short Derivation: The Probability of All Remaining Patients Responding
A useful special case occurs when terminal success requires every remaining patient to respond. Given the current posterior \(\operatorname{Beta}(A,B)\), the probability of \(m\) responses in a row is obtained by averaging the conditional probability \(p^m\) over uncertainty in \(p\):
The integral is the beta-function ratio because multiplying the beta density by \(p^m\) changes the exponent of \(p\) from \(A-1\) to \(A+m-1\). The product form follows by expanding the beta-function ratio. This calculation shows why a point estimate plugged into a binomial probability is not generally the predictive probability: the predictive calculation averages over the entire posterior distribution of \(p\).
Worked Example: A High Current Posterior Probability, but a Lower Predictive Probability
Suppose the prior is \(p\sim\operatorname{Beta}(1,1)\). A trial plans \(N=5\) patients, and terminal success requires \(P(p>0.5\mid\text{final data})\geq0.90\). At an interim look, two patients have completed outcomes and both responded. Calculate the current posterior success probability and the predictive probability of terminal success.
The current posterior is \(\operatorname{Beta}(1+2,1+0)=\operatorname{Beta}(3,1)\). Its probability above 0.5 is \(P(p>0.5\mid x=2,n=2)=1-(0.5)^3=0.875\). This is evidence about the response rate based on the two observed outcomes; it is not the probability that the trial will meet its final rule.
To identify the terminal boundary, consider the final data. With four responses among five patients, the posterior is \(\operatorname{Beta}(5,2)\), for which \(P(p>0.5)=0.890625\), below 0.90. With five responses, the posterior is \(\operatorname{Beta}(6,1)\), and \(P(p>0.5)=1-(0.5)^6=0.984375\). Thus terminal success requires all five responses. Since two have already responded, all three remaining patients must respond.
Using the all-responses formula with \(A=3\), \(B=1\), and \(m=3\), \(P(K=3\mid x=2,n=2)=3(4)(5)/[4(5)(6)]=0.5000\). The predictive probability of terminal success is therefore 0.5000, despite the current posterior success probability of 0.875. The former concerns the final decision after three uncertain outcomes; the latter concerns \(p\) given the data observed so far.
Reachability Is Part of the Calculation
A terminal rule may be mathematically defined but unreachable from a particular interim state. Before calculating a predictive probability, compare the number of responses already observed with the largest possible final count, \(x+m\). If no possible value of \(K\) can satisfy the terminal rule, the predictive probability of terminal success is zero. This is a structural consequence of the decision rule and remaining sample size, not evidence that the current response rate is zero.
Worked Example: A Terminal Boundary That Cannot Be Reached
Let \(p\sim\operatorname{Beta}(1,2)\), with \(N=4\). Define terminal success as \(P(p>0.5\mid\text{final data})\geq0.80\). At an interim analysis, one of two completed outcomes is a response, so \(x=1\), \(n=2\), and \(m=2\). Find the terminal response boundary and the predictive probability of terminal success.
With three responses at the end, the posterior would be \(\operatorname{Beta}(1+3,2+1)=\operatorname{Beta}(4,3)\), and \(P(p>0.5)=0.65625\), below 0.80. With four responses, the posterior would be \(\operatorname{Beta}(5,2)\), and \(P(p>0.5)=0.890625\). Therefore, terminal success requires four responses.
Only two outcomes remain, so even if both are responses the final total is \(1+2=3\), short of the required four. Consequently, no value of \(K\in\{0,1,2\}\) meets the terminal rule, and \(P(\text{terminal success}\mid x=1,n=2)=0\). For comparison, the current posterior is \(\operatorname{Beta}(2,3)\), with \(P(p>0.5)=0.3125\). Neither that current probability nor the probability that both future patients respond changes the reachability calculation: two future responses still do not produce four total responses.
A Predictive Threshold Is Not a Promise
Monitoring rules may compare the predictive probability of terminal success with a threshold to guide an interim action. For example, a protocol might say to stop early for efficacy if this predictive probability is at least a specified value. Crossing such a threshold means that the prespecified action rule is met, conditional on the model and data. It does not mean that terminal success is certain, or even that the predictive probability is close to one.
The distinction is especially important when future outcomes are variable. A predictive probability of 0.60 means that, under the model and current information, the terminal success event has probability 0.60. It leaves probability 0.40 for terminal non-success. The chosen threshold determines an action; it does not alter the probability distribution of future outcomes.
Worked Example: Meeting a Predictive Threshold Does Not Guarantee Success
Suppose \(p\sim\operatorname{Beta}(1,1)\), the planned total is \(N=4\), and terminal success requires \(P(p>0.5\mid\text{final data})\geq0.90\). Two of the first two patients responded. A monitoring rule recommends early efficacy stopping when the predictive probability of terminal success is at least 0.50. Does the rule recommend stopping, and does that guarantee terminal success?
At the final analysis, three responses give posterior \(\operatorname{Beta}(4,2)\), with \(P(p>0.5)=0.8125\), below 0.90. Four responses give \(\operatorname{Beta}(5,1)\), with \(P(p>0.5)=0.96875\). Thus success requires four responses, so both remaining patients must respond.
The current posterior is \(\operatorname{Beta}(3,1)\). The predictive probability that both remaining patients respond is \(E[p^2\mid x=2,n=2]=3(4)/[4(5)]=0.6000\). The predictive probabilities for zero, one, and two future responses are 0.1000, 0.3000, and 0.6000, respectively; only the last outcome meets the terminal rule. Because 0.6000 exceeds the monitoring threshold of 0.50, the specified rule recommends stopping for efficacy. But under the model there remains probability \(1-0.6000=0.4000\) of not meeting the terminal rule. The threshold triggers an action; it does not guarantee the final outcome.
Diagnostics, Alternatives, and Common Errors
A sound interpretation can be checked systematically. First, state the terminal success rule in terms of final data. Second, identify how many outcomes remain, including outcomes that are pending if they will contribute to the terminal analysis. Third, list which possible future response counts satisfy the rule. Finally, sum their predictive probabilities. These checks expose both calculation errors and mismatches between the probability reported and the question being asked.
- Reporting the wrong probability. \(P(p>p_0\mid x,n)\) is a posterior probability about the response rate. The predictive probability is about future data meeting the terminal rule. Always name the event to which the reported probability refers.
- Ignoring remaining sample size. A high current posterior probability cannot make an unreachable terminal boundary attainable. Check \(x+m\) against the response counts that satisfy the rule.
- Equating a threshold with certainty. A predictive threshold determines an action under a specified monitoring rule. It does not set the probability of terminal success to one.
- Using the wrong future event. “At least one future response” is not interchangeable with “terminal success.” Derive the final response counts required by the terminal rule before calculating any probability.
- Plugging in a point estimate. A binomial probability calculated at \(\hat p=x/n\) ignores posterior uncertainty. The Beta-binomial predictive probability integrates over the posterior for \(p\).
Exact enumeration of the possible values of \(K\) is often the clearest approach for binary outcomes: it makes the terminal event and each included probability visible. Simulation can approximate the same predictive probability when the future-data model is more complicated, but simulation error must be considered and rare events may need many draws. A current posterior probability is a useful summary of present evidence, but it answers a different question. Neither summary replaces the prespecified terminal rule.
These interpretations depend on the model and the definition of success. A common-\(p\) Beta-binomial model may not represent changing response rates, dependence among outcomes, or heterogeneous patient populations. A predictive probability is conditional on the assumptions used to generate it; it is not a guarantee that those assumptions are adequate. If pending outcomes are treated as future outcomes, the analysis must also make clear that their eventual results are being predicted rather than already observed.
Check Your Understanding
Answer each question using the distinction between current posterior evidence and future terminal outcomes. Solutions are provided below.
- In the first worked example, what does 0.875 describe, and what does 0.500 describe?
- In the second worked example, why is the predictive probability of terminal success zero even if both remaining patients respond?
- In the third worked example, the predictive probability is 0.600 and the stopping threshold is 0.50. What does crossing the threshold establish, and what does it not establish?
- Suppose a terminal rule requires at least four responses, and there are two observed responses with one outcome remaining. What is the maximum final response count, and can terminal success occur?
- Why is \(P(K=m\mid x,n)\) generally not equal to the probability obtained by substituting \(\hat p=x/n\) into a binomial formula?
Solutions. (1) 0.875 is the current posterior probability that \(p>0.5\) after two responses; 0.500 is the predictive probability that all three remaining patients respond and the terminal rule is met. (2) The final total would be \(1+2=3\), below the required four responses, so the success event is impossible. (3) The specified monitoring rule recommends stopping for efficacy; the threshold does not guarantee terminal success, which still has predictive probability 0.600 rather than 1. (4) The maximum is \(2+1=3\), so success cannot occur and its predictive probability is zero. (5) The plug-in calculation treats the estimated response probability as fixed, whereas the predictive probability averages future-outcome probabilities over the posterior uncertainty in \(p\).