Why Use an Information Fraction?
In a group sequential design, an interim analysis occurs before the planned maximum amount of information has accrued. The key question is not simply how many participants have enrolled, but how much information the available data provide about the treatment effect or other estimand. An information fraction expresses the current information relative to the information planned at the maximum analysis.
This ratio is useful because sequential analyses are organized around the statistical information available at each look. In the canonical model, the information fraction also indexes the joint distribution of the sequential test statistics. It is therefore a design quantity, not a measure of the observed treatment effect and not a probability.
The denominator is the planned maximum information, not automatically the planned maximum number of participants. The ratio is meaningful only after specifying how information is defined. For example, a design may use expected Fisher information under its model, or a specified estimate of information available at an interim analysis. The chosen definition must be applied consistently to the numerator and denominator.
This tutorial uses the standard group sequential convention that information accumulates over the study and is nondecreasing across planned looks. The canonical-model interpretation also relies on the usual assumptions of that framework: analyses concern the same estimand, information is evaluated on a consistent scale, and the sequential statistics have the canonical joint distribution described in the earlier material on the canonical joint distribution. Departures from these conditions can make the simple interpretation of \(t_k\) unreliable.
Connection to Sequential Test Statistics
In the canonical model, the information fraction locates an analysis within the planned information scale. In particular, for cumulative standardized statistics \(Z_i\) and \(Z_j\) at looks \(i\) and \(j\), where \(I_i\leq I_j\), the canonical covariance structure gives
The second equality follows directly from the definition: \(t_i/t_j=(I_i/I_{\max})/(I_j/I_{\max})=I_i/I_j\), because the same planned maximum information appears in both fractions. Thus the information fraction is not merely a convenient percentage. It preserves the relative positions of the looks on the information scale, which in turn determines their canonical correlation structure.
A short derivation clarifies why the information ratio is the relevant scale. In the canonical model, a cumulative score has variance proportional to its cumulative information. Standardizing that score divides it by the square root of its information. For nested analyses, the earlier score is part of the later score; their covariance is the earlier cumulative information. Consequently, standardization gives a correlation equal to the earlier information divided by the later information, under a square root. Expressing each cumulative information as a fraction of the same \(I_{\max}\) leaves that ratio unchanged. This argument depends on the canonical model assumptions; it is not a claim that any enrollment schedule automatically has this covariance structure.
Worked Example: Converting Interim Information to a Fraction
Suppose a hypothetical two-look trial plans to reach maximum information \(I_{\max}=120\) information units. At its interim analysis, the trial has accumulated \(I_1=48\) units. Calculate and interpret the interim information fraction.
Formula and substitution. $$ t_1=\frac{I_1}{I_{\max}}=\frac{48}{120}=0.400. $$
The interim analysis is at 40.0% of the planned maximum information. This does not mean that the observed effect is 40.0% of its eventual value, or that there is a 40.0% probability of a particular trial outcome. It describes the amount of information accrued relative to the specified maximum.
If the assumptions of the canonical model hold, this look has information time \(t_1=0.400\). The value is within the planned range because \(0\leq 48\leq 120\). The calculation itself does not establish that the model assumptions hold; those require separate design and data checks.
Information Fraction Is Not Enrollment Fraction
The enrollment fraction is the number enrolled at a given time divided by the planned maximum number enrolled. It describes recruitment progress. The information fraction describes statistical information progress. These ratios can be equal, but equality requires information to accrue in proportion to enrollment. That may be a reasonable approximation in some designs, but it is not a definition and should not be assumed without justification.
Participants can contribute different amounts of information. This may occur because their measurements have different precision, their observation times differ, or the relevant outcome data are not yet available for everyone enrolled. The information fraction therefore cannot generally be recovered from the participant count alone. The comparison below separates the quantities.
| Quantity | Numerator | Denominator | What it describes |
|---|---|---|---|
| Information fraction | Information available at the analysis | Planned maximum information | Progress on the statistical information scale |
| Enrollment fraction | Participants enrolled so far | Planned maximum enrollment | Progress on the recruitment scale |
Worked Example: Unequal Information per Participant
In a hypothetical trial, 120 participants are planned. At an interim look, 60 have enrolled, so the enrollment fraction is \(60/120\). The trial's prespecified information calculation gives 44 information units currently available and 100 units planned at maximum. Calculate both fractions and explain the difference.
Enrollment fraction. $$ \frac{60}{120}=0.500. $$
Information fraction. $$ t=\frac{44}{100}=0.440. $$
Half of the planned participants have enrolled, while 44.0% of the planned maximum information is available. In this hypothetical example, the two progress measures differ because the available data have not contributed information in direct proportion to enrollment. The information fraction should be reported as 0.440 on the stated information scale; the participant count alone would not justify replacing it with 0.500.
The calculation assumes the 44 and 100 units are measured for the same estimand and under the same information definition. If they were obtained using incompatible models or scales, their ratio would not be a valid information fraction.
Choosing and Maintaining the Denominator
The planned maximum information is a design reference. It must be specified clearly enough that the interim quantity can be compared with it. In a fixed design, \(I_{\max}\) is the information target for the planned maximum analysis. If the planned maximum changes, the information fraction changes even if the data already collected do not. A report should therefore identify which maximum-information target is being used.
For example, if current information is 80 units, its fraction of a 160-unit target is 0.500, whereas its fraction of a revised 200-unit target is 0.400. Neither division changes the current data. They answer different questions about progress toward different maximum targets. A fraction without its reference target can be ambiguous when the target has been revised.
A practical calculation can be organized as follows:
Use the same statistical target and model-based information convention at the interim and maximum analyses.
Record the value that defines the denominator, including any applicable design revision.
Use the prespecified interim calculation, rather than substituting the number enrolled unless the design explicitly justifies that approximation.
Calculate \(t_k=I_k/I_{\max}\), then confirm the numerator, denominator, and units are consistent and the ratio is plausible for the planned information range.
Worked Example: A Revised Maximum Target
A hypothetical study has current information \(I_k=72\) units. Its original maximum-information target was 144 units, but the planned target is subsequently revised to 180 units. Calculate the information fraction relative to each target and interpret the results.
Relative to the original target: $$ t_k^{\mathrm{original}}=\frac{72}{144}=0.500. $$
Relative to the revised target: $$ t_k^{\mathrm{revised}}=\frac{72}{180}=0.400. $$
The same current information is 50.0% of the original target and 40.0% of the revised target. The appropriate fraction depends on which maximum target is relevant to the analysis being described. A report should not state only “the information fraction is 0.400” if the target was recently changed; it should make clear that 0.400 is relative to the revised 180-unit maximum.
Diagnostics, Alternatives, and Limitations
Several checks help detect a calculation or interpretation that is going wrong:
- Check the scale. The current and maximum information must refer to the same estimand and information definition. Dividing information values from different models does not produce an interpretable fraction.
- Check the reference target. A changed maximum target changes the ratio. Record whether the denominator is the original or revised planned maximum.
- Investigate a value above 1. If \(I_k>I_{\max}\), then \(t_k>1\). This may indicate that the study has exceeded its planned maximum information, that the denominator is outdated, or that the quantities were calculated inconsistently. It is not a value to silently truncate to 1.
- Check the information estimate. If the method for estimating current information is unstable or poorly matched to the design assumptions, the fraction may also be unstable. The ratio is only as reliable as its inputs.
- Do not infer the fraction from enrollment automatically. Similar enrollment and information fractions may be useful evidence in a design where contributions are sufficiently uniform, but equality must be justified rather than presumed.
The main alternatives answer different questions. Enrollment fraction summarizes how much recruitment has occurred. Calendar-time fraction summarizes elapsed time relative to a planned duration. An event fraction summarizes observed events relative to a target number of events. Each can be operationally useful, but none is automatically the information fraction. In a design where the relevant information is not proportional to one of these counts, using that count as a substitute can misstate the position of an interim analysis on the information scale.
Information fraction also has limits. It does not measure the direction or magnitude of an estimated treatment effect, determine whether a stopping boundary has been crossed, or guarantee that the assumed canonical model is adequate. Its interpretation depends on the specified information measure, its planned maximum, and the assumptions connecting the sequential statistics to that measure. The introductory material on statistical information and precision provides the general motivation: more information ordinarily corresponds to greater precision, but the fraction alone does not supply an effect estimate or a decision.
Common Mistake
A useful interpretation always names the reference: “At this interim analysis, 44.0% of the planned maximum information is available.” That statement is precise about what the ratio measures and avoids implying that the same percentage of participants must have enrolled.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A design has current information of 36 units and planned maximum information of 90 units. What is the information fraction, and what does it mean?
- A trial has enrolled 40% of its planned participants. Can its information fraction be assumed to equal 0.40? Explain.
- Current information is 75 units. The planned maximum is revised from 150 to 200 units. What are the two corresponding information fractions?
- An interim calculation gives \(I_k=105\) units when the recorded maximum is \(I_{\max}=100\) units. What should be checked before interpreting the resulting ratio?
Solutions. (1) \(t=36/90=0.400\). The analysis has accrued 40.0% of the planned maximum information. (2) No. Forty percent enrolled is an enrollment fraction; information may not accrue in proportion to participant count, so current information and the planned maximum must be compared directly. (3) Relative to the original target, \(75/150=0.500\); relative to the revised target, \(75/200=0.375\). The denominator must be identified when reporting the fraction. (4) The direct ratio is \(105/100=1.050\), but first check whether the study has exceeded its planned maximum, whether the target is outdated, and whether numerator and denominator use the same information definition. Do not cap the ratio at 1 without resolving the discrepancy.