Why a Conditional Probability Can Guide Stopping
At an interim analysis, a trial may still have a substantial amount of information to collect, yet the data so far may make eventual success look unlikely. Stochastic curtailment uses this interim information to ask a forward-looking question: under a specified model and effect assumption, what is the probability that the trial will meet its success criterion if it continues? A small answer can trigger a prespecified recommendation to stop for futility.
The probability is calculated from the data observed so far and the planned future design. The cutoff used to decide whether to stop is a separate design choice. Keeping those quantities distinct is essential: a conditional probability of 0.04 is not itself a universal instruction to stop. It leads to a stopping decision only if it meets the rule’s prespecified threshold.
The rule is stochastic because the interim data, and therefore the conditional power, vary from trial to trial. It does not mean that investigators flip a random coin to decide whether to stop. A trial’s stopping decision follows from comparing its calculated conditional power with the fixed threshold. The rule is also conditional on choices that must be stated: the success event, the assumed effect and variance model, the remaining analysis plan, and the value of \(q\).
Assumptions and the Stopping Rule
A conditional probability is meaningful only relative to a model for future data. For the simple normal illustration used below, let \(t\) be the fraction of total information available at the interim analysis, and let \(Z(t)=z\) be the observed standardized statistic. Assume a one-sided final success criterion \(Z(1)\ge c\). Under an assumed standardized effect \(\theta\), suppose the cumulative score follows the canonical independent-increments model: \(Z(t)\) has mean \(\theta t\) and variance \(t\), and increments after the interim analysis are independent of the data observed so far, with mean \(\theta\) times the added information and variance equal to that added information.
Under these assumptions, the final statistic conditional on the interim result has mean \(\sqrt{t}\,z+\theta(1-t)\) and variance \(1-t\). Thus the conditional probability of meeting the final criterion is the upper-tail probability beyond \(c\). This particular formula applies to the stated one-final-look normal model; with multiple remaining looks, a different success event and corresponding conditional calculation are needed.
A practical stopping rule can be written compactly as
Here \(CP_k\) is a data-dependent probability, whereas \(q\) is the prespecified decision threshold. The inequality, not the probability alone, defines the stopping action. A design must also specify which looks are eligible and whether stopping is binding or nonbinding in the design’s operating-characteristic calculations. A nonbinding futility rule permits the monitoring committee to continue despite low conditional power; a binding rule requires stopping when the criterion is met.
Worked Decisions in a Normal Model
Worked Example: A Probability Below or Above the Cutoff
Suppose a hypothetical trial has one final analysis remaining. At an interim look, half the planned information has accrued, so \(t=0.50\), and the observed statistic is \(z=0.80\). Use a one-sided final success threshold \(c=1.645\), assume \(\theta=0.40\), and set the futility cutoff to \(q=0.10\). The normal independent-increments assumptions above are taken to hold for this illustration.
The conditional mean of the final statistic is \(\sqrt{0.50}(0.80)+0.40(1-0.50)=0.5657+0.2000=0.7657\). Its conditional standard deviation is \(\sqrt{1-0.50}=0.7071\). Therefore,
The estimated conditional probability of final success under \(\theta=0.40\) is about 10.7%. Since \(0.1068>q=0.10\), the stochastic-curtailment rule does not stop the trial at this look. The result is close to the cutoff, but the decision follows the prespecified inequality, not an informal judgment that 10.7% is “low.”
Worked Example: A Later Interim Result Meets the Rule
Consider a second hypothetical interim result under the same one-sided normal model and the same assumed effect, \(\theta=0.40\), with \(c=1.645\) and \(q=0.05\). Suppose \(t=0.75\) and the observed statistic is \(z=0.70\). Assume the trial has not already stopped and this is an eligible look.
The conditional mean is \(\sqrt{0.75}(0.70)+0.40(0.25)=0.6062+0.1000=0.7062\). The conditional standard deviation is \(\sqrt{0.25}=0.5000\), so
Because \(0.0302\le 0.05\), the rule calls for stopping for futility. The 3.02% value is the model-based conditional probability of meeting the final success criterion if the trial continues; 5% is the separate threshold that converts that probability into a decision. Neither number is the probability that the stopping decision itself is correct.
Worked Example: The Effect Assumption Changes the Decision
At a hypothetical interim analysis, take \(t=0.50\), \(z=0.40\), and \(c=1.645\), with futility cutoff \(q=0.05\). Use the same normal model, but compare two assumptions about the future effect: \(\theta=0\) and \(\theta=0.80\). This comparison checks how sensitive the stopping rule is to the effect used to calculate conditional power.
For \(\theta=0\), the conditional mean is \(\sqrt{0.50}(0.40)+0(0.50)=0.2828\), and the standard deviation is \(0.7071\). Hence
This is below 0.05, so the rule stops for futility under the zero-effect assumption. For \(\theta=0.80\), the conditional mean is \(0.2828+0.80(0.50)=0.6828\), giving
This value is above 0.05, so the rule does not stop under the larger assumed effect. The observed data and threshold have not changed; the assumed effect has. A design should therefore state the effect assumption clearly and examine whether the stopping decision is sensitive to plausible alternatives.
Justification and Interpretation
The rule follows directly from two distinct steps. First, condition on the information already observed and calculate the chance of the specified future success event under the chosen model. Second, compare that chance with a prespecified threshold. In the one-final-look model, the independent-increments assumption gives the conditional normal distribution stated above, so standardizing the success boundary produces the normal-tail calculation. The stopping rule then applies the inequality \(CP_k\le q\); it does not alter the meaning of \(CP_k\).
The threshold is a policy choice, not a probability estimated from the interim data. Choosing a smaller \(q\) makes stopping for futility less likely, because fewer conditional-power values fall below it. A larger \(q\) makes the rule more willing to stop. The choice should be evaluated as part of the trial design, including its consequences for the chance of success under relevant effects and for stopping under the null. It should not be selected after seeing an interim result merely to justify a preferred action.
Diagnostics, Alternatives, and Limitations
Before applying the rule, check that the success event and remaining analyses used in the conditional-power calculation match the actual design. Check the information fraction, the interim statistic, the assumed effect, the variance treatment, and the implementation of any multiplicity-adjusted success boundaries. If the calculation treats an estimated variance as known, or uses an effect assumption without acknowledging its role, the reported conditional probability may be misleading. When the normal approximation is questionable, the calculation should reflect the trial’s actual outcome model and design rather than relying automatically on the normal formula shown here.
Stochastic curtailment differs from a deterministic futility boundary. A deterministic boundary specifies a cutoff directly on an observed statistic or another data summary. Stochastic curtailment first maps the interim data to conditional power, then compares that value with \(q\). The resulting boundary in the data can depend on the assumed effect, information remaining, and future analysis plan; it need not be a simple fixed cutoff on \(z\). A design with no futility stopping is another alternative: it avoids decisions based on a low conditional-power calculation but may continue a trial with little modeled chance of success. These approaches should be compared using the design’s operating characteristics, rather than by treating one rule as automatically preferable.
A futility stop is not proof that treatment has no effect, nor does a low conditional power give the probability that the null hypothesis is true. It is a conditional calculation under a specified model and effect assumption. Similarly, stopping when conditional power is above the cutoff does not guarantee success. Conditional power can be misspecified if the assumed effect or variance model is poorly chosen, if information accrual differs from plan, or if the success criterion is implemented incorrectly.
Finally, the rule does not by itself quantify how often a trial will stop for futility over repeated trials. That is a separate operating-characteristic question. Here the decision is defined for a particular observed interim history: calculate the conditional probability under the specified assumptions, then apply the prespecified threshold. This distinction prevents confusion between a probability of eventual success given current data and a probability of entering the stopping region across repeated trials.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A trial’s conditional power is 0.07 and its prespecified futility cutoff is 0.10. What does the rule recommend, and which quantity is the threshold?
- In the one-final-look normal model, why does the conditional standard deviation decrease as the information fraction \(t\) increases?
- At the same interim data, two assumed effects produce conditional powers of 0.03 and 0.08. With \(q=0.05\), explain why the resulting decisions differ.
- Does conditional power of 0.02 mean there is a 2% probability that the null hypothesis is true? Explain.
- Name two checks that should be made before relying on a stochastic-curtailment decision.
Solutions. (1) Since \(0.07\le0.10\), the rule recommends stopping for futility; 0.10 is the prespecified threshold, while 0.07 is the calculated conditional power. (2) Less information remains for future independent increments, so their variance \(1-t\) is smaller. (3) The first value is at or below \(q\) and triggers stopping; the second is above \(q\) and does not. The effect assumption changes the conditional probability. (4) No. It is the conditional probability of eventual success under the stated model and effect assumption, not a probability assigned to the null hypothesis. (5) For example, check that the success criterion and remaining design are represented correctly, and that the information fraction, variance model, and assumed effect are appropriate.