What a Blinded Update Can Tell You
A sample-size plan depends not only on the treatment effect the design aims to detect but also on features of the outcome data that affect precision or the number of participants who contribute usable information. These features are nuisance parameters: quantities needed for planning, but not the primary treatment effect of interest. Interim data can provide information about them even when treatment assignments remain masked.
The key distinction is between learning about the pooled data and learning about the difference between treatment groups. A blinded analysis may estimate an overall variability, event frequency, or attrition proportion. It does not, by itself, reveal which outcomes came from which treatment group or estimate the treatment contrast. Moreover, a pooled quantity is not automatically the same as the nuisance parameter required by a sample-size calculation. Whether it is usable depends on the outcome model, allocation, treatment effects, and the role that quantity plays in the design.
For example, a pooled standard deviation might be relevant if the design requires an estimate of outcome variability. A pooled event proportion might inform the expected number of events per enrolled participant. A pooled attrition proportion might inform the number who need to be enrolled to reach a target number of evaluable participants. Each interpretation requires its own assumptions; the word “pooled” does not establish that those assumptions hold.
What Masking Does—and Does Not—Identify
With treatment labels hidden, an analyst can calculate summaries across the observed participants without separating the data by group. Such summaries can be useful for operational planning. However, the masked data generally combine information from both treatment groups. If those groups have different outcome distributions, the pooled distribution reflects both within-group variation and differences between groups. Consequently, a pooled summary may not equal the group-specific nuisance parameter used in the original plan.
This is an identifiability issue, not merely a matter of choosing a more precise calculation. A pooled mean alone does not determine two group means or their contrast. Likewise, a pooled event proportion does not reveal the two group-specific event probabilities. Keeping labels masked protects against using interim treatment comparisons for the update, but it also limits what can be learned from the data.
A blinded update is most straightforward when the planning quantity has the same relevant value in both groups, the pooled statistic estimates the required target under the design assumptions, and interim observations represent the population expected at completion. If treatment changes the outcome distribution, follow-up, or probability of an event, the pooled estimate may instead be a mixture of group-specific values. It can still be relevant when the calculation is explicitly based on that mixture, but it should not be silently interpreted as a common value.
Three Types of Interim Information
| Quantity updated | What pooled data can summarize | Key question before using it |
|---|---|---|
| Outcome variance | Observed spread across participants with treatment labels masked | Does this spread estimate the variance required by the planned calculation, or is it affected by differences between groups? |
| Event rate | Overall observed fraction with the event | Does the design need an overall event rate, or a rate specific to one group or a group contrast? |
| Attrition rate | Overall observed fraction not retained or not evaluable by a defined time | Are attrition definitions, follow-up, and attrition probabilities sufficiently comparable across groups? |
These are not interchangeable quantities. An event rate describes an outcome occurring during follow-up; an attrition rate describes loss of usable follow-up or outcome information. A participant can experience the event and still complete follow-up, or leave follow-up without experiencing it. The protocol must define each quantity and its time window before interim estimation.
Worked Example: Updating a Planning Variance Without Estimating the Contrast
Hypothetical setting. A balanced, two-group trial plans to detect a fixed mean difference using a normal-outcome calculation. The original calculation uses a common within-group variance of \(20\) square units and requires \(160\) participants per group. At an interim look, masked outcomes give a pooled variance estimate of \(25\) square units. Suppose, for this illustration, that the pooled estimate is a suitable estimate of the common within-group variance and that the calculation’s other inputs remain fixed.
Under the usual fixed-effect normal approximation with allocation, target difference, significance level, and power held constant, the required number per group is proportional to the variance. Thus the revised planning target is
The estimated variance is \(25\) square units, which is \(25/20=1.25\) times the planned variance. The corresponding target is \(1.25(160)=200\) participants per group. This calculation updates a planning input; it does not calculate a treatment-effect estimate or use an observed treatment contrast.
The assumption is consequential. The example treats the pooled variance as estimating the common within-group variance needed by the calculation. If the treatment groups have different means, a pooled spread can also reflect between-group separation. Masking alone does not show that the pooled estimate is free of that contribution. In practice, the analyst must use the prespecified model and re-estimation rule to establish whether this estimate is appropriate; the arithmetic cannot validate the assumption.
Worked Example: Interpreting a Pooled Event Rate
Hypothetical setting. In a masked interim dataset, \(30\) of \(100\) participants have experienced a specified event by the planned assessment time. A design’s operational target is \(180\) observed events, and for this illustration its enrollment calculation uses one overall event probability that is assumed to remain stable over the remaining enrollment.
The pooled event-rate estimate is
Using that estimate as the assumed overall event probability, a simple event-count projection is
The projection is \(600\) participants under the stated assumptions. It does not estimate either treatment group’s event probability, nor does it estimate a treatment effect. It is appropriate only if the planning target is based on an overall rate and the interim rate is a reasonable forecast for the participants and follow-up that remain.
If the two groups have different event probabilities, the pooled proportion reflects the mix of participants in the masked interim data. A change in allocation or in the mix of groups over time can therefore make that pooled proportion a poor forecast of the overall rate at completion. If the calculation instead requires a group-specific event probability, the masked pooled estimate cannot supply it without further assumptions or information.
Worked Example: Adjusting Enrollment for Interim Attrition
Hypothetical setting. A trial needs \(200\) evaluable participants at a specified assessment. By an interim review, \(12\) of \(120\) participants have met the protocol’s prespecified definition of attrition before that assessment. Assume for this illustration that the masked overall attrition estimate applies to both groups, that future participants have the same attrition probability, and that attrition is not associated with a feature that changes the chance of reaching the assessment.
The estimated attrition proportion is
If \(N\) participants are enrolled and \(90\%\) are expected to be evaluable, the expected evaluable count is \(0.90N\). Setting this equal to the target gives
Since enrollment must be a whole number and rounding down would give an expected count below \(200\), the projected enrollment is \(223\) participants. The update concerns the number expected to provide evaluable data; it does not indicate whether the treatment works. If attrition differs by group, time, or prognostic characteristics, the pooled proportion may not provide the appropriate enrollment adjustment.
Why the Assumptions Matter
A masked estimate is usable for recalculation only when it answers the planning question the design actually poses. Before applying an update, check the following conditions.
- Target match. Define whether the calculation needs a common variance, an overall event rate, a group-specific rate, or an evaluability proportion. Do not substitute a pooled estimate for a different target without justification.
- Comparable interim and future data. Check that the outcome definition, measurement method, follow-up time, eligibility criteria, and data-collection procedures have not changed in ways that alter the estimated quantity.
- Relevant group structure. Consider whether treatment could change the pooled quantity. A common-variance or common-attrition assumption must be justified, not inferred from the fact that treatment labels are hidden.
- Adequate information. Early estimates can be unstable, especially when few participants or events have accrued. Report the denominator and assess whether the estimate is precise enough to support the proposed change.
- Prespecified and bounded action. Define how the estimate will change the target, and any permitted minimum or maximum, before looking at interim results. The update rule should not become an informal response to whichever estimate is observed.
Useful diagnostics include checking interim data completeness, comparing follow-up duration with the plan, reviewing whether event and attrition definitions were applied consistently, and examining whether the estimate changes sharply when a small number of observations are removed. These checks do not recover masked group-specific information. They help reveal whether the pooled estimate is based on comparable, sufficiently complete data.
Alternatives and Limitations
A blinded update is not the only way to address uncertainty in planning quantities. A fixed design can use a conservative planning value chosen before the trial. An external pilot can inform a nuisance parameter, but its estimate may not transport to the current setting. An unblinded interim re-estimation can use group-specific information, but it involves access to treatment comparisons and therefore requires design-specific procedures and safeguards.
The advantage of a blinded update is that it can respond to observed pooled features without directly estimating the treatment contrast. Its limitation is the same masking that provides that separation: it may not distinguish within-group variation from between-group differences, or reveal group-specific event and attrition patterns. A pooled estimate also has sampling uncertainty, and a single interim estimate may be a noisy forecast of the final data.
Finally, changing a sample-size target does not, by itself, establish that the study’s type I error is controlled. The statistical validity of a particular re-estimation rule depends on the design and the complete analysis procedure, including how interim information is used. A blinded update should therefore be implemented under a prespecified design rule whose operating characteristics have been assessed, not treated as an automatic guarantee of validity.
Common Pitfall
Correction: Masking prevents the update from directly using separated treatment-group results; it does not prove that treatment has no effect on the pooled data. First identify the target quantity required by the calculation, then justify why the pooled estimate informs that target under the design assumptions.
Check Your Understanding
Answer each question by identifying both what the pooled estimate says and what assumptions are needed to use it.
- A masked interim analysis reports a pooled outcome variance. Does that estimate, by itself, show that the treatment effect is zero?
- A trial’s sample-size calculation needs a control-group event rate. Can a masked pooled event rate generally provide that quantity?
- A masked estimate suggests \(15\%\) attrition. Name two assumptions needed before using it to inflate enrollment.
- A planning variance is \(16\) square units, and a suitable interim estimate is \(20\) square units. If the target scales proportionally with variance and all other inputs are fixed, by what factor does the target change?
- Give one diagnostic that could reveal that an interim pooled rate is a poor forecast of the remaining trial.
Solutions. (1) No. The pooled variance describes spread in the masked data; without group separation it does not estimate the treatment contrast or establish that the contrast is zero. (2) Not generally. The pooled rate combines participants from both groups and need not equal the control-group rate; using it requires additional assumptions or a design that needs the overall rate. (3) For example, future participants must have a similar attrition probability, and the attrition definition and follow-up period must be consistent. Comparable attrition across groups is also needed if the update uses one pooled rate for both. (4) The factor is \(20/16=1.25\), so the target increases by \(25\%\). (5) Examples include finding that interim participants have substantially shorter follow-up than planned, or that the event definition or data-collection process changed. Either could make the observed rate unrepresentative of the remaining trial.