Why the Calendar Is Not an Information Scale
A group sequential design may schedule an interim analysis for a particular month, but the month itself does not determine how precise the interim estimate will be. Participants may enter at different rates over the study, and outcomes may become available only after follow-up. Even when outcomes are available, they may contribute different amounts of information. Thus, two studies that have each used half their planned calendar duration need not have accumulated half their planned information.
The distinction matters when assessing whether an interim analysis has reached its intended statistical information target. Calendar time is a measure of elapsed duration. Information is a measure of precision for the parameter being estimated. The introductory material on information fractions defines the latter as the information available at an interim analysis divided by the information planned at the final analysis. Here, the focus is why that ratio need not equal the fraction of calendar time elapsed.
A familiar special case helps establish the distinction. Suppose independent observations \(Y_i\) follow \(Y_i\sim N(\theta,\sigma^2)\), with a common mean \(\theta\) and known variance \(\sigma^2\). Each observation contributes \(1/\sigma^2\) units of Fisher information about \(\theta\). With \(n(t)\) observations available by time \(t\), the total information is \(I(t)=n(t)/\sigma^2\). If the same variance applies throughout, the information fraction is \(n(t)/n(T)\): it is the fraction of planned observations available, not automatically the fraction of time elapsed.
Calendar time and information can track each other if information accrues at a constant rate over the planned period. But steady accrual is an assumption about the study's data-availability process, not a property of the calendar. Recruitment may accelerate or slow, and an enrolled participant may not yet have a usable outcome. We can see the consequences in concrete timelines.
Recruitment That Changes Over Time
Consider a design in which each participant provides an independent outcome with the same variance, all outcomes are available immediately, and the parameter of interest and measurement process do not change over time. Under those assumptions, information is proportional to the number of participants with outcomes. If recruitment is faster in one part of the study than another, the participant fraction can differ from the calendar-time fraction, and so can the information fraction.
Worked Example: Recruitment Accelerates After a Slow Start
Suppose a hypothetical study plans to recruit 200 participants over 12 months. Recruitment is expected to be 10 participants per month in months 1–4 and 20 per month in months 5–12. Each participant contributes an independent outcome with known variance \(\sigma^2=100\) in squared outcome units, and the outcome is available at enrollment. Compare calendar time and information at month 6.
Check the model. The schedule yields \(4(10)+8(20)=200\) participants by month 12, as planned. Under the stated common-variance normal model, information is the number of available outcomes divided by 100. Immediate outcome availability and unchanged measurement assumptions mean completed outcomes equal recruited participants.
The calendar-time fraction at month 6 is \(6/12=0.500\). The information at month 6 is \(I(6)=80/100=0.800\), while planned final information is \(I(12)=200/100=2.00\). Therefore,
The study has used 50.0% of its planned calendar duration but has accumulated 40.0% of its planned information. The mismatch occurs because the first six months recruit fewer participants per month than the study's average rate of \(200/12\approx16.67\) participants per month. It would be incorrect to infer the information fraction from the elapsed time alone.
The direction of the discrepancy is not fixed. If recruitment starts quickly and later slows, information can be ahead of calendar time. Also, a recruitment curve is not necessarily the same as an outcome-availability curve. If follow-up is required, participants can be enrolled without yet contributing their planned outcome. That distinction can widen or reverse the gap.
Outcome Completion Can Lag Behind Enrollment
For an analysis based on completed outcomes, counting enrolled participants as though all their outcomes were available overstates the information currently in hand. The appropriate count depends on the analysis and the information each available observation contributes. A fixed follow-up period creates a particularly clear lag: an outcome from a recently enrolled participant is not available until that follow-up is complete.
Worked Example: A Fixed Follow-Up Delay
Suppose a hypothetical study plans to recruit 240 participants, at 20 per month, from month 1 through month 12. Each participant's outcome is measured after three months of follow-up. The final analysis is planned for month 15, when the last recruited participants can complete follow-up. Assume independent outcomes with common known variance \(\sigma^2=64\), and assume each completed outcome contributes information about the same mean parameter. Compare the fractions at month 9.
Check the timeline and model. By month 9, the study has enrolled \(9(20)=180\) participants. With a three-month delay, only participants enrolled by month 6 have completed the measurement, so there are \(6(20)=120\) outcomes available. By month 15, all \(12(20)=240\) planned outcomes are complete. This calculation assumes no loss to follow-up, no outcome availability before the scheduled measurement, and a common outcome distribution over the recruitment period.
For the normal mean with known variance, \(I=n/64\). Hence the information fraction is
At month 9, 60.0% of the planned calendar duration has elapsed, but 50.0% of planned information is available. Counting the 180 enrolled participants as 180 completed outcomes would incorrectly report an information fraction of \(180/240=0.750\). Enrollment is not the same as outcome completion.
A follow-up delay does not by itself guarantee a particular relationship between the two fractions. For example, a study might recruit much faster early on, making many outcomes available despite a delay; a slow recruitment period can have the opposite effect. Moreover, an interim dataset may include partial or repeated measurements rather than one completed outcome per participant. In that case, information should be computed for the data actually used in the analysis, rather than inferred from a participant count.
Information Can Differ Even for the Same Number of Outcomes
The number of available outcomes is a useful proxy for information only when their contributions are comparable. For independent normal observations about a common mean with known variances \(\sigma_i^2\), observation \(i\) contributes \(1/\sigma_i^2\) units of information. More variable observations contribute less. If the outcome precision changes across cohorts or settings, an interim count fraction can differ from the information fraction even when outcomes are complete.
Worked Example: Different Precision Across Recruitment Periods
Imagine a hypothetical 10-month study of a common mean. By month 5, 40 outcomes are complete; the remaining 60 are completed between months 5 and 10. Assume independent normal outcomes with known variance 4 for each of the first 40 participants and known variance 9 for each of the remaining 60. The design and scientific interpretation of the mean are assumed comparable across participants. Compare the calendar-time, outcome-count, and information fractions at month 5.
Check the model. The assumed observations are independent and share the same target mean, while their known variances may differ. For this model, the information contributions add, with a contribution of \(1/\sigma_i^2\) from each observation.
The calendar-time fraction is \(5/10=0.500\), and the outcome-count fraction is \(40/100=0.400\). The information fraction is
Thus 40.0% of outcomes provide about 60.0% of planned information in this example, because the first group of outcomes is more precise. The example assumes known variances and a common mean; if the groups actually target different means, simply adding these contributions would not answer the same scientific question. The relevant model and estimand must be checked before treating a numerical information calculation as meaningful.
Why the Information Fraction Has This Form
The ratio \(I(t)/I(T)\) is not an adjustment to the calendar. It follows from comparing the information in the data available now with the information planned for the final analysis. A short derivation in the simple normal model makes the distinction explicit.
The simplification depends on the same variance applying to the observations and on the interim data being part of the planned final data. If variance differs, information need not be proportional to the count. If information is accumulated at a constant rate, then \(I(t)=I(T)t/T\) and the two fractions are equal. But a constant calendar-time rate is an additional assumption; it does not follow from the definition of information or from the use of a group sequential design.
Comparing the Three Fractions
These quantities answer different questions. Calendar-time fraction describes how much planned duration has elapsed. Enrollment fraction describes how much of the planned recruitment has occurred. Information fraction describes how much statistical information is available for the analysis. They coincide only under particular accrual, completion, and precision patterns.
| Quantity | What it measures | What can make it differ from the others? |
|---|---|---|
| Calendar-time fraction | Elapsed time divided by planned final-analysis time | It advances even during slow recruitment or follow-up delays. |
| Enrollment fraction | Participants enrolled divided by planned enrollment | Enrolled participants may not yet have usable outcomes. |
| Information fraction | Information available divided by planned final information | It depends on outcome availability, precision, and the analysis model. |
Calendar time is valuable for planning staffing, monitoring study progress, and coordinating operational activities. Enrollment is useful for tracking recruitment. Neither quantity, by itself, establishes that an analysis has reached a planned information target. When an interim analysis is intended to occur at a specific information fraction, the information available under the analysis model is the relevant quantity.
Diagnostics, Failure Modes, and Limits
A mismatch between calendar and information fractions is not inherently a problem; it may simply reflect the actual accrual pattern. The problem is treating them as interchangeable without checking whether the planned data-availability assumptions still hold. Useful checks include:
- Compare the expected and observed timelines. Review recruitment and outcome-completion counts by period. A steep rise or long flat segment in completed outcomes shows why a linear calendar-time assumption may be implausible.
- Use the analysis dataset, not only the enrollment log. Confirm which outcomes are available and eligible for the interim analysis. Enrollment counts cannot substitute for outcome completion when follow-up is pending.
- Check the model that translates data into information. Common-variance formulas require a common variance; a count-based calculation can mislead when precision differs. Confirm that the outcome, estimand, and analysis remain comparable over time.
- Distinguish planned from estimated information. A planned curve is a forecast based on assumptions. The information available at an interim analysis should be assessed using the data and model appropriate to that analysis. Uncertainty in variance or other nuisance quantities can make the estimated fraction uncertain.
A simple monotonicity check is also useful: with nested datasets and a fixed model, adding independent observations with positive information contributions should not reduce total information. If a reported information total falls as data accrue, investigate whether the model, dataset, parameterization, or information calculation has changed. Estimated standard errors can fluctuate, so a fluctuating standard error does not automatically mean that the underlying accumulated Fisher information has decreased.
There are limits to what a fraction alone can tell you. Two studies can have the same information fraction but different calendars, recruitment histories, or data-quality concerns. A correctly computed fraction also cannot repair a misspecified model or make outcomes available sooner. Finally, if participants entering at different times have systematically different outcome distributions, combining their information under a common-parameter model may be inappropriate. The assumed statistical model and the study's scientific target must be considered together.
Check Your Understanding
Use the definitions and assumptions in each question. The solutions follow the questions.
- A study plans 120 equally precise, independent outcomes over 10 months. At month 4, 36 outcomes are complete. What are the calendar-time and information fractions?
- A trial has enrolled 150 of 200 planned participants, but only 100 have completed the outcome measurement. Under a common-variance model where each completed participant contributes one equal information unit, which count determines the information fraction?
- In the known-variance normal mean model, why does the information fraction equal the completed-outcome fraction when all outcomes have the same variance?
- A study has reached 70% of its planned calendar duration, but recruitment has been unusually slow and follow-up is delayed. Is it justified to report 70% of planned information? Explain.
- Suppose interim observations are more precise than later observations, and both groups estimate the same mean under the model. Could the information fraction exceed the outcome-count fraction? Explain briefly.
Solutions. (1) The calendar-time fraction is \(4/10=0.400\). Equal precision makes information proportional to completed outcomes, so the information fraction is \(36/120=0.300\). (2) Use the 100 completed outcomes, not the 150 enrolled participants; the information fraction is \(100/200=0.500\). (3) Each outcome contributes \(1/\sigma^2\), so the interim and final totals are \(n(t)/\sigma^2\) and \(n(T)/\sigma^2\); their ratio is \(n(t)/n(T)\). (4) No. Elapsed duration does not establish how many outcomes are available or how much information they contain; calculate information using the data and analysis model. (5) Yes. More precise interim observations contribute more information per outcome, so their information share can be greater than their share of the total outcome count.