Evaluating Futility and Trial Feasibility

Futility Is Both a Statistical and an Operational Question

An interim efficacy result is not, by itself, a forecast of whether the trial can answer its objective. A committee considering futility asks a forward-looking question: given the evidence so far and realistic assumptions about what remains, is the trial still capable of meeting its objective? The answer depends on statistical evidence and on whether the necessary participants, treatment exposure, endpoint data, and follow-up can be obtained.

These are related but distinct assessments. A trial may have low conditional power under a particular effect assumption even though recruitment and follow-up are proceeding as planned. Conversely, the observed treatment contrast may remain promising while recruitment, retention, or endpoint collection make the planned analysis unattainable on the available timeline. Neither situation should be reduced to one number without examining its assumptions.

The efficacy boundary used in a conditional-power calculation must be the boundary appropriate to the planned design. In a group sequential trial, it may differ from a conventional single final-look cutoff because the design accounts for interim testing. The earlier material on reviewing efficacy evidence explains why an interim estimate should not be treated as a definitive effect estimate; here, that caution matters because projections can be especially sensitive to the effect assumed for future data.

Conditional Power: State the Model Before Calculating

Let \(Z_F\) denote the final test statistic, and let \(c\) be the prespecified final success boundary for the analysis in question. For a one-sided success criterion, success occurs when \(Z_F \ge c\). Suppose the conditional projection model is \(Z_F\mid\text{interim data}\sim N(m,v)\), where \(m\) is the conditional mean and \(v\) is the conditional variance. Then conditional power is the upper-tail probability beyond the boundary.

Here, \(\Phi\) is the standard normal cumulative distribution function. This calculation assumes the stated conditional normal model is an adequate approximation, that \(v>0\), and that the future-data assumptions used to obtain \(m\) and \(v\) are appropriate. Those assumptions include the planned information available at the final analysis and an explicit assumption about the treatment effect in the future data. The effect assumption might be the design alternative, a clinically important effect, or an extrapolation based on the interim estimate. These choices answer different questions and can give very different conditional powers.

The formula follows by standardizing the final statistic. Under the model, \((Z_F-m)/\sqrt v\) has a standard normal distribution. Therefore, \(P(Z_F\ge c)=P((Z_F-m)/\sqrt v\ge(c-m)/\sqrt v)\), which equals the upper-tail probability in the formula. This derivation also shows why a larger conditional mean raises conditional power, while a larger boundary lowers it, all else equal.

Conditional power is therefore a sensitivity analysis as much as a summary. If the assessment changes substantially when the assumed future effect changes from the design alternative to a smaller clinically relevant effect, the committee should see that dependence rather than receive a single apparently objective number. A projection using the observed interim effect can be particularly unstable when the interim sample is small or the estimate is noisy.

Operational Feasibility: Translate Plans Into Forecasts

Operational forecasts should be expressed in quantities that can be checked: participants still needed, recruitment rate, time remaining, expected endpoint availability, and the follow-up period required. A recruitment target alone is not enough. If the primary outcome is unavailable for many enrolled participants, the trial may not reach its planned information even if enrollment succeeds.

A simple recruitment projection is the current enrollment plus the expected additional enrollment over the remaining period. For planning purposes, a constant average rate can be used as a first calculation, but it is an assumption—not a guarantee. The rate should be examined by site and over time, and the forecast should account for site activation, seasonal variation, screening failures, and any known changes in capacity.

Adherence requires a separate interpretation. Lower-than-planned adherence can reduce the treatment contrast that a trial is capable of detecting, and it may affect how the result answers the clinical question. It does not automatically mean that nonadherent participants should be removed from the primary analysis or that randomized sample size should be discounted mechanically. The primary estimand and analysis plan determine what the treatment effect represents. An operational assessment should report adherence patterns and consider their implications without substituting an unplanned analysis for the planned one.

Combine the Evidence Without Collapsing It

A useful review lays out the statistical projection and operational forecast side by side, then asks whether their assumptions are compatible. For example, a conditional-power calculation might assume that the planned final information will be collected, while the recruitment and retention forecast suggests a substantial shortfall. In that case, the conditional-power number does not describe the likely trial as currently operating unless the projection is recalculated under an appropriate information assumption.

This separation is also useful when comparing alternatives. Conditional power summarizes the chance of meeting a statistical criterion under a specified future-data model. Operational forecasts summarize whether the data needed for that model can plausibly be collected. A forecast based on recruitment and retention cannot replace an efficacy analysis; a statistical projection cannot establish that sites can recruit or that participants will complete follow-up. Presenting the two views separately makes the assumptions and any tension between them visible.

Diagnostics, Failure Modes, and Limits

Before relying on a futility assessment, check the inputs and the data generating the projection. Unexpected changes in enrollment, incomplete follow-up, delayed endpoint adjudication, missingness, or inconsistent exposure can make apparently precise projections misleading. Compare current rates with the periods and sites on which the forecast is based; check whether the denominator is enrolled, randomized, or eligible participants; and verify that “evaluable outcome” is defined consistently with the planned analysis.

• One conditional-power number without its effect assumption. The correction is to state the assumed future effect and, when conclusions are sensitive, show more than one plausible scenario.
• Extrapolating a short-term recruitment rate as certain. The correction is to treat it as a forecast, examine site-level patterns and uncertainty, and describe what happens if the rate declines.
• Equating low adherence with a reduced randomized sample size. The correction is to assess its implications for the treatment contrast, estimand, and interpretability under the planned analysis.
• Counting enrollment as equivalent to complete information. The correction is to include retention, outcome availability, and follow-up timing in the forecast.
• Interpreting low conditional power as proof of no treatment effect. The correction is to recognize that it is conditional on the model and the interim data; it is not a test of whether the null is true.

The method has important limits. Normal approximations may be poor with sparse outcomes or unusual data patterns. Conditional power can be unstable when based on an interim effect estimate, and operational rates can change after the review. A projection also cannot resolve broader considerations such as the clinical value of the question or whether the available evidence supports a meaningful conclusion. It is an aid to a reasoned monitoring assessment, not a mechanical decision rule.