Why Counting Participants Is Not Enough
In a group sequential study, an interim analysis is often described as occurring after a certain fraction of participants have enrolled. That description is not, by itself, a measure of how much statistical evidence has accrued. Participants can contribute different amounts of information, depending on their group, the variability of their outcomes, and whether usable outcomes are available. Therefore, the fraction enrolled and the information fraction can differ.
This distinction matters when locating an interim analysis on the planned information scale. If an analysis is described as occurring at 50% of planned enrollment, that does not establish that it occurs at 50% of planned information. The information fraction is the relevant quantity for interpreting precision and for placing an analysis in a design based on information.
Information is tied to the precision of the estimate for the parameter being tested, not simply to the number of records in a data set. For a scalar estimate \(\hat{\theta}\) with variance \(1/I\), information is \(I=1/\operatorname{Var}(\hat{\theta})\). Thus, greater information means lower variance and a smaller standard error. In practice, information may be derived from the model or calculated from the variance or standard error used in the analysis.
The information fraction framework in the earlier material on the canonical joint distribution uses cumulative information at each analysis. It is the information fraction, rather than a count fraction, that describes the analysis position on that scale. This tutorial focuses on why the two fractions need not match.
How Allocation and Variability Change Information per Participant
Consider two independent groups, A and B, with sample sizes \(n_A\) and \(n_B\). Suppose the outcome in each group has a fixed variance, \(\sigma_A^2\) and \(\sigma_B^2\), and the target is the difference in group means. Assume independent observations within and between groups, finite variances, and complete outcome data for the participants included in the analysis. Under this model, the variance of the difference in sample means is the sum of the two group-mean variances.
This formula makes the role of allocation explicit: adding participants to one group does not generally have the same effect as adding the same number to the other. It also shows why variability matters. A group with greater outcome variance contributes less information per participant than a group with lower variance, all else equal.
If group sizes in both arms grow in the same proportion and the variance model remains stable, information grows in that same proportion. For example, doubling both group sizes halves the variance of the difference and doubles its information. But when the interim allocation differs from the planned final allocation, a given total enrollment can correspond to a different information fraction.
Worked Example: A Half-Enrolled Study With Less Than Half the Information
Suppose a hypothetical two-group study plans to enroll 120 participants in each group. The outcome variance is assumed to be \(16\) units squared in both groups. At an interim analysis, 40 participants have been enrolled in group A and 80 in group B. Calculate the enrollment and information fractions.
The interim enrollment is \(40+80=120\), while planned final enrollment is \(120+120=240\). Thus the enrollment fraction is \(120/240=0.5000\), or 50.00%.
At the interim analysis, the variance of the difference in means is
The interim information is \(I_k=1/0.6000=1.667\) per squared outcome unit. At the planned final sample sizes, the variance is
Therefore, \(t_k=1.667/3.750=0.4444\). Although half the participants have enrolled, only 44.44% of the planned information has accrued. The interim allocation is less balanced than the final allocation, so its 120 participants provide less information than 60 participants in each group would under these assumptions.
When Group Variances Differ
If outcome variances differ across groups, allocation changes can have especially noticeable effects. The value of an additional participant depends partly on the variance in that participant’s group. A count-only comparison ignores this distinction.
Worked Example: Unequal Variability Changes the Information Fraction
In a hypothetical study, the planned final sample is 100 participants per group. The outcome variance is \(36\) units squared in the high-variability group and \(4\) units squared in the low-variability group. At an interim analysis, 80 participants have enrolled in the high-variability group and 20 in the low-variability group. Assume these variances apply at both analyses and outcomes are observed for everyone counted. Compare the information fraction with the enrollment fraction.
The interim enrollment fraction is \((80+20)/(100+100)=0.5000\). Its contrast variance is
At the planned final sample sizes, the contrast variance is
The information fraction is the ratio of final variance to interim variance:
Here, 50.00% of participants have enrolled, but 61.54% of planned information has accrued. The interim sample includes many participants from the more variable group, but it also has 20 participants in the more precise group, compared with 100 planned. With these particular variances and allocations, the combined information is nevertheless greater than half the planned amount. The direction of the discrepancy cannot be inferred from the enrollment fraction alone; it depends on both group sizes and variances.
This example does not imply that enrollment in a high-variance group is inherently more informative. Rather, it illustrates that information is a property of the contrast and its variance. A different allocation, variance pattern, or target parameter can produce a different information fraction at the same enrollment fraction.
Incomplete Observations and Usable Information
Enrollment counts include participants who have entered the study, but information at an analysis depends on which data are available and contribute to the analysis. If some participants have not yet provided an outcome, the interim analysis may have less information than the enrollment count suggests. The relevant sample sizes in a simple means model are the numbers with usable outcomes, not automatically the numbers enrolled.
The variance calculation alone does not establish that an analysis of observed outcomes is unbiased or otherwise valid. The example below assumes that the observed outcomes support a valid analysis under the study’s prespecified approach. In an actual study, the reasons for incomplete observations and the analysis assumptions must also be considered.
Worked Example: Half Enrolled but Only One-Third of Planned Information
A hypothetical study plans to enroll 100 participants in each of two groups. Both group outcome variances are assumed to be \(25\) units squared. At interim, 50 participants have enrolled in each group, but usable outcomes are available for only 40 participants in group A and 30 in group B. Assume these available outcomes can be analyzed validly under the study’s prespecified method, and that the planned final analysis will have outcomes for all 100 participants per group.
The enrollment fraction is \((50+50)/(100+100)=0.5000\). For the interim information calculation, use the observed counts with usable outcomes:
The planned final contrast variance is
Thus, the information fraction is \(0.5000/1.4583=0.3429\), or 34.29%. The enrollment fraction is 50.00%, but the information fraction is only 34.29%. Counting the 20 and 20 enrolled participants without usable outcomes as though they contributed full outcome information would overstate the information available at interim.
Checking the Fraction and Choosing a Comparison
A practical check is to calculate the information fraction from the variance or standard error of the target estimate, then compare it with the count fraction. If the same scalar estimate and compatible variance calculations are used at both analyses, the standard-error relationship is
For example, if the interim standard error is \(1.5\) times the planned final standard error, the corresponding information fraction is \(1/1.5^2=0.4444\). A disagreement between this calculation and the fraction based on enrolled participants is not automatically an error; it is a prompt to identify what drives the difference.
Useful diagnostics include checking that the interim and final calculations refer to the same parameter and model, confirming the group-specific counts used, and verifying whether counts represent enrollment or participants with usable data. Also check that the variance assumptions are appropriate and that the final-information denominator matches the planned design. If the variance estimate is unstable or the analysis model changes, an apparent information fraction can be uncertain or not directly comparable.
| Quantity | What it measures | When it can be a useful proxy |
|---|---|---|
| Enrollment fraction | Participants enrolled by interim divided by planned enrollment | When information per participant is approximately constant and data availability tracks enrollment |
| Information fraction | Interim information divided by planned final information for the target parameter | When comparing analysis position on the information scale |
| Standard-error ratio | Precision at interim relative to planned final precision | When standard errors are based on compatible models and the same estimand |
Enrollment fraction is attractive because it is simple, observable, and often available before outcomes are complete. It is a reasonable approximation when allocation stays close to plan, outcome variability is stable, and participants contribute comparable amounts of usable data. Its limitation is that it cannot account for changes in the precision of the target estimate. Information fraction addresses that limitation, but it depends on a correctly specified and comparable information calculation.
Common Mistake and Key Takeaway
- Wrong reasoning: “Half the planned participants have enrolled, so the interim analysis must have half the planned information.”
- Correction: This conclusion requires information to accumulate proportionally to enrollment. Check allocation, group-specific variability, and outcome availability, then calculate \(I_k/I_K\) for the target analysis.
- Another warning: A small standard error does not by itself guarantee that the analysis is unbiased or that its assumptions hold. Precision and validity are related but distinct questions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study has planned sample sizes of 80 per group and interim sample sizes of 40 per group. Under stable variances and complete outcomes, what is the enrollment fraction, and what is the information fraction for a difference in means?
- Why can an interim allocation of 70 participants in group A and 30 in group B provide a different information fraction from an allocation of 50 and 50, even though both have 100 participants?
- A hypothetical study plans 60 participants per group, with variance \(9\) in each group. At interim, 30 per group have usable outcomes. Calculate the enrollment and information fractions if 40 per group have enrolled.
- State two checks to make when the information fraction differs substantially from the enrollment fraction.
Solutions. (1) The enrollment fraction is \(80/160=0.5000\). Since both group sizes are half their planned values and variances are stable, the contrast variance doubles, so information is halved; the information fraction is also \(0.5000\). (2) The variance of the difference is \(\sigma_A^2/n_A+\sigma_B^2/n_B\), so changing the allocation changes the two variance contributions, even when the total count is fixed. (3) The enrollment fraction is \(80/120=0.6667\). Interim variance is \(9/30+9/30=0.6000\); final variance is \(9/60+9/60=0.3000\). Therefore \(t_k=0.3000/0.6000=0.5000\). (4) Check whether the counts are enrolled or have usable outcomes, whether allocation and variances match the planned assumptions, and whether the interim and final calculations use the same parameter and model.