Event Counts as Information Fractions

Why Count Events Rather Than Participants?

In an event-driven trial, interim analyses are often scheduled after specified numbers of outcomes have occurred. Those counts are convenient, but the number of participants enrolled is not automatically the amount of statistical information available. A participant who has not yet experienced the event may contribute little or no information to a time-to-event comparison, while an observed event contributes information through the treatment allocation among participants still at risk.

The key question is therefore not simply how many events have occurred, but whether each event contributes roughly the same amount of information. When it does, the fraction of planned events observed is a practical approximation to the fraction of final information accrued. That fraction can then be used as the information-time input to an alpha-spending function.

Definition: Let \(I_k\) be the statistical information available at look \(k\), and let \(I_K\) be the planned information at the final look \(K\). The information fraction at look \(k\) is \(t_k=I_k/I_K\), with \(0\leq t_k\leq 1\). If \(d_k\) is the planned cumulative event count at look \(k\), and \(d_K\) is the planned final event count, the event-count fraction is \(d_k/d_K\). It is an information-fraction approximation, not a definition of information.

In the usual group-sequential framework, information time indexes how much information has accumulated relative to the final analysis. The earlier material on computing information fractions from sample size gives the parallel idea for settings where information is proportional to sample size. In an event-driven survival analysis, the relevant quantity is generally the information in the event-time comparison, not the number randomized or the calendar time elapsed.

When the Event-Count Ratio Works

Consider a log-rank or Cox score analysis for a treatment effect expressed as a log hazard ratio. Under the null hypothesis, the score statistic is approximately normal, with variance equal to its information. In a simple event-driven design, the information can be represented as a sum of event-specific contributions. Write \(v_j\) for the information contributed by the \(j\)th event, so the information after \(D_k\) events is approximately

I_k=\sum_{j=1}^{D_k}v_j.

If each event contributes the same amount \(c\), then \(I_k=cD_k\), and the information fraction is exactly the event-count fraction:

t_k=\frac{I_k}{I_K} =\frac{cD_k}{cD_K} =\frac{D_k}{D_K}.

In practice, event contributions need not be identical. The count ratio is a useful approximation when their average is sufficiently stable across the planned follow-up. For a standard log-rank comparison with a stable treatment allocation among those at risk, each event has a similar expected contribution under the null. For example, with a constant allocation proportion \(r\), a simplified approximation gives information per event proportional to \(r(1-r)\). If that risk-set allocation remains stable, the common factor largely cancels in the ratio of interim to final information.

This justification is an approximation tied to the design and analysis. It is not enough that the trial has a fixed event target. The comparison also relies on broadly comparable event contributions over time, an analysis whose information is driven by the event process, and a planned final event count that corresponds to the intended final information. If the treatment allocation among the risk set changes substantially, or the event mix changes in ways that alter event-level information, counts alone may be inadequate.

Worked Example: Converting Planned Events to Information Fractions

Suppose a hypothetical trial plans its first interim analysis after 120 events, its second after 300 events, and its final analysis after 600 events. Assume that the average information per event is stable over follow-up, so that event counts are an appropriate proxy.

The planned event-count fractions are each look’s cumulative event target divided by the final target:

t_1=\frac{120}{600}=0.200,\qquad t_2=\frac{300}{600}=0.500,\qquad t_3=\frac{600}{600}=1.000.

Thus, under the stated approximation, the analyses are planned at 20%, 50%, and 100% of final information. The increments in event-count fraction are \(0.200\), \(0.300\), and \(0.500\); those are not the cumulative fractions, but the additional fractions accrued between successive looks. This distinction matters when interpreting a spending function, which is evaluated at cumulative information times.

The numbers of participants enrolled at these looks are not needed for this calculation. They may be useful for operational planning, but they do not replace the event counts as a proxy for information in this event-driven setup.

How the Fractions Enter an Alpha-Spending Rule

Let \(A(t)\) be a cumulative alpha-spending function, with \(A(0)=0\) and \(A(1)=\alpha\). At look \(k\), the spending rule is evaluated at the cumulative information fraction \(t_k\). The quantity \(A(t_k)\) is the cumulative amount of type I error allocated through that look under the design. The difference \(A(t_k)-A(t_{k-1})\) is the additional cumulative spending target since the previous look.

For a planned event-driven design, one uses the planned event-count fractions as approximations to \(t_k\) when the event-contribution assumptions are reasonable. This supplies the time inputs to the spending function. The resulting values are spending targets; they are not, by themselves, the critical Z-values or the probability of crossing a boundary at that look. Boundary calibration must account for the joint distribution of the sequential test statistics and the possibility of crossing at earlier looks.

Worked Example: Applying a Spending Function to Event Fractions

Suppose a hypothetical one-sided design has \(\alpha=0.025\), plans looks at 100, 250, and 500 events, and assumes stable information per event. For illustration, use the linear spending function \(A(t)=\alpha t\). This function spends the total level gradually in proportion to information time.

The planned information fractions are \(100/500=0.200\), \(250/500=0.500\), and \(500/500=1.000\). Evaluating the function gives cumulative spending targets:

$$ A(0.200)=0.025(0.200)=0.00500, $$

A(0.500)=0.025(0.500)=0.01250,\qquad A(1.000)=0.025(1.000)=0.02500.

The additional targets are \(0.00500\) through the first look, \(0.01250-0.00500=0.00750\) between the first and second looks, and \(0.02500-0.01250=0.01250\) between the second and final looks. These increments sum to \(0.02500\). They describe the spending schedule, not three independent per-look significance levels. In particular, the second-look boundary cannot be obtained by treating \(0.00750\) as an ordinary single-look alpha without accounting for the first look.

When Event Counts Misrepresent Information

The count ratio can be misleading when average information per event differs across periods. For example, changes in the proportion of treated participants remaining at risk, changing allocation within risk sets, or differences in event composition can alter event contributions. The direction of the discrepancy depends on which period has larger contributions.

If later events contribute more information per event than earlier events, the interim events contribute less information per event than the overall average. The event-count fraction can then overstate the information fraction at an interim look. If later events contribute less information per event than earlier events, the event-count fraction can understate the information fraction.

Worked Example: Unequal Information Per Event

Consider a hypothetical design with 100 events by an interim look and 400 events planned overall. Suppose, purely for illustration, that each of the first 100 events contributes 0.20 information units, while each of the next 300 contributes 0.10 information units. These contributions are assumed known for this calculation.

The event-count fraction is \(100/400=0.250\). But the interim information is \(100(0.20)=20\) units, and final information is \(100(0.20)+300(0.10)=50\) units. Therefore the information fraction is \(20/50=0.400\). In this constructed example, the event-count ratio understates the interim information fraction because the early events contribute more information per event than the later events.

Reverse the contributions: let the first 100 events each contribute 0.10 units and the next 300 each contribute 0.20 units. The interim information is then \(100(0.10)=10\) units, while final information is \(10+300(0.20)=70\) units. The information fraction is \(10/70\approx0.143\), below the event-count fraction of 0.250. Here later events contribute more information per event, so the event-count ratio overstates interim information.

These examples isolate the direction of the effect. In an actual design, event-level information may not be known in advance; the practical implication is to question the count approximation when the risk-set composition or other features affecting information are expected to change.

Planning Checks and Alternatives

Before using event fractions as spending-function inputs, check that the event targets refer to cumulative events analyzed at each look, that the final target is the event count associated with the intended final analysis, and that the planned looks occur in increasing order. Then consider whether information per event is plausibly stable. For a log-rank analysis, the treatment allocation among participants at risk is an important diagnostic; substantial departures from its planned pattern can change the information contribution of events.

If the design can estimate information more directly, using \(I_k/I_K\) is preferable to relying on a count proxy. For example, information estimates from the planned score or likelihood analysis can reflect changing risk sets better than a simple ratio. A sample-size fraction is another possible proxy when information is expected to scale with participants rather than events, but in a genuinely event-driven survival design, enrollment fractions can be particularly misleading when follow-up or event rates differ over time.

The event-count approach is limited by its assumptions: it does not make event counts and information identical in every survival analysis, and it does not correct a misspecified analysis model or an inaccurate event forecast. Moreover, planned event fractions describe the intended timing of looks. If realized information differs materially from the plan, the planned spending inputs may no longer describe the realized information times. The distinction between planned and realized timing should therefore be retained in design documentation and monitoring.

Common Mistakes

Using participants randomized in place of events. In an event-driven time-to-event design, an enrolled participant who has not yet had an event may not contribute the same information as a participant with an observed event. Use the event ratio only when it is a credible proxy for information.
Calling the event-count ratio the information fraction without qualification. It is exact only under equal information per event and otherwise is an approximation. State the assumption or use an information estimate.
Reversing the direction of bias. If later events contribute more information per event, early events are below the overall average, so the event-count fraction overstates the interim information fraction. If later events contribute less, it understates it.
Interpreting spending increments as boundary levels. \(A(t_k)-A(t_{k-1})\) is an increment in cumulative spending, not automatically the nominal significance level or crossing probability for look \(k\).

Key takeaway: Planned event-count fractions are useful spending-function inputs when information per event is reasonably stable. The ratio \(d_k/d_K\) approximates \(I_k/I_K\); it does not replace the information definition or the later calibration of sequential boundaries.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A trial plans analyses after 80 and 200 events, with 400 events planned at the final analysis. Under stable information per event, what are the two interim information-fraction approximations?
If the first 100 events contribute more information per event than the later events, does the event-count fraction tend to overstate or understate the information fraction at the first look? Explain.
A design uses \(A(t)=0.025t\) and has a planned event fraction of 0.40. What is its cumulative spending target at that look?
Why is the spending increment between two looks not automatically the nominal significance level for the later look?

Solutions. (1) The fractions are \(80/400=0.200\) and \(200/400=0.500\). (2) It tends to overstate the information fraction: the early events contribute more information per event than the later events, so their information share exceeds their share of the total event count. (3) \(A(0.40)=0.025(0.40)=0.0100\), a cumulative target through that look. (4) Sequential test statistics are dependent and a boundary crossing may already have occurred at an earlier look, so boundary calibration must account for the full sequence rather than equating an increment with a single-look alpha.