Choosing the Interim Information Fraction

Why the Interim Time Matters

A sample-size re-estimation rule acts on information available at an interim analysis. If that analysis occurs early, the effect estimate is usually imprecise, so small changes in the observed data can change whether the rule recommends an increase. A later analysis usually gives a more precise estimate, but leaves less time to recruit participants and complete follow-up. Choosing the interim time therefore involves both statistical and operational considerations.

The key quantity is the interim information fraction. It measures how much information has accrued at the interim analysis relative to the information planned at the original final analysis. In a simple model where every observation contributes the same amount of information, this fraction is the interim sample size divided by the originally planned final sample size. In other designs, such as event-driven studies or settings with unequal information per participant, the information fraction need not equal the fraction of participants enrolled.

To isolate the effect of timing, suppose observations follow a normal model \(X_i\sim N(\theta,\sigma^2)\), with independent observations, a common known variance \(\sigma^2\), and a stable outcome measurement process. The interim estimate is \(\hat{\theta}_1=\bar{x}_1\), and its standard error is \(SE(\hat{\theta}_1)=\sigma/\sqrt{n_1}\). These assumptions make the precision comparison transparent; in a real design, the appropriate information measure and variance model must match the endpoint and allocation.

Precision Changes With the Information Fraction

If \(n_1=tN_0\), the standard error at the interim analysis is

Thus, relative to the standard error at the planned final sample size, an interim estimate has standard error \(1/\sqrt{t}\) times as large. Precision does not increase linearly with sample size: quadrupling the information halves the standard error. Under the normal model, a two-sided 95% interval for \(\theta\) at the interim has the form \(\hat{\theta}_1\pm1.96SE(\hat{\theta}_1)\). Its width contracts as information accrues, although the estimate's realized value can move in either direction.

Timing Changes the Chance of Triggering an Increase

Precision matters because a re-estimation rule is applied to an observed interim result. Consider a deliberately simple rule that increases the planned final sample size from \(N_0\) to \(N_{\max}\) when the estimate falls in a specified region \(A\). If the true effect is \(\theta\), then under the normal model

This probability depends on both the information fraction and the location of the region relative to the true effect. A more precise estimate concentrates more tightly around \(\theta\); it does not, by itself, imply that an increase becomes more or less likely. For a real promising-zone rule, \(A\) may be defined through a conditional-power or effect-estimate criterion. The earlier material on conditional power and promising-zone rules supplies those decision constructs; here the focus is how their operating probabilities can change with timing.

For a binary rule that either retains \(N_0\) or increases to \(N_{\max}\), expected final sample size follows directly by weighting the two possible final sizes by their probabilities:

What Timing Does—and Does Not—Change About Sample-Size Range

In the examples, the possible final targets are 200 and 280 observations at either interim time. The range of possible increases over the original plan is therefore zero to 80 observations, independent of \(t\), because the rule's permitted targets are fixed. Timing changes how likely each target is under a given true effect, and therefore changes expected sample size; it does not automatically change the design's stated maximum.

Operationally, however, timing changes how much recruitment remains. At \(t=0.25\), 50 observations have accrued: reaching 200 requires 150 more, while reaching 280 requires 230 more. At \(t=0.75\), 150 have accrued: reaching 200 requires 50 more, and reaching 280 requires 130 more. The extra 80 observations beyond the original plan are the same in both cases, but the later decision leaves less calendar time to obtain them and complete follow-up. If a protocol also imposes a fixed recruitment window or a limit on the number of participants that can still be enrolled, some nominally permitted final targets may no longer be feasible.

Keep these two meanings separate: the statistical range is the set of final sample sizes allowed by the rule and its cap; the operationally feasible range is the subset that can still be recruited and followed on schedule. State explicitly whether a cap refers to final total sample size or to additional recruitment after the interim. Confusing the two can make a design appear to permit an increase that cannot actually be delivered.

Choosing and Checking a Timing

There is no universally best information fraction. An earlier interim analysis provides more opportunity to act on a recommendation, but its effect estimate is noisier. A later analysis gives a more precise estimate, but can make a permitted increase impractical. The choice should be evaluated against the planned decision rule, recruitment and follow-up timelines, and the effects and variances considered plausible for the endpoint.

A useful comparison evaluates, at several candidate values of \(t\), both the probability of each sample-size decision and the expected final sample size across a range of true effects. For a rule with more than two possible final sizes, use the general calculation

where \(N_j\) are the permitted final sample sizes. The probabilities must reflect the actual interim estimator distribution and the prespecified rule, including any rounding or bounds. A single expected sample size can hide very different decision probabilities across effects, so it should not be considered in isolation.

Several checks help reveal a problematic timing choice:

• Precision check: calculate the interim standard error at each candidate information fraction. If the estimate is too noisy to distinguish regions that lead to different actions, the rule may be unstable.
• Feasibility check: verify that recruitment and outcome ascertainment can reach every permitted target after the proposed look. Account for the time needed to act on the recommendation.
• Operating-characteristic check: calculate decision probabilities and expected sample size across plausible true effects, not only at the planning effect.
• Model check: confirm that the assumed information accrual, variance, and estimator distribution are reasonable. Unequal information per participant, missing outcomes, or a changing variance can make \(n_1/N_0\) a poor substitute for an information fraction.

A fixed-sample design avoids an interim adaptation decision and has a simpler sample-size profile, but cannot respond to interim evidence. A rule based on conditional power or a promising zone can respond to interim information, yet its decision probabilities depend on the look's timing and the rule's assumptions. Neither approach is automatically preferable. In particular, changing sample size based on unblinded interim estimates can affect the final test's operating characteristics; those characteristics must be evaluated for the actual adaptive design rather than assumed to be unchanged.

Common Pitfall and Limitations

Wrong reasoning: “A later interim analysis is more precise, so it must make a sample-size increase less likely.” Precision describes the spread of the estimate, not the position of the increase region. If the true effect lies inside that region, added information may increase the chance of an increase; if the true effect lies elsewhere, the probability can move differently. The correction is to calculate \(P(\hat{\theta}_1\in A\mid\theta,t)\) under the model for each relevant \(\theta\) and \(t\).

The normal known-variance examples are deliberately simple. Actual designs may use a treatment-effect estimator with a different variance, unequal allocation, binary outcomes, or event-driven information. Interim estimates may also be affected by missing data, delayed outcomes, or variance estimation. The information-fraction comparison remains useful, but its numerical calculations must use the design's correct estimator distribution. Finally, expected sample size is an average over repeated studies under specified assumptions; it does not guarantee that a particular study will use that many participants.