Loss To Follow-Up And Information Loss

Why Loss to Follow-Up Matters

In a time-to-event trial, a participant lost to follow-up stops contributing to later risk sets. That participant is not counted as having had the event: the event time is unknown after the last contact. Consequently, loss to follow-up can reduce the number of later events observed and can alter the allocation of participants who remain at risk. Both changes matter for a group sequential logrank analysis, but they are not the same as a change in the event rate.

The distinction is important. A trial may accrue fewer events because less follow-up time is observed, even if the event rate among participants under observation is unchanged. Conversely, selective loss may change the mix of participants still at risk and hence the observed event rate. The latter can also challenge the assumptions needed for valid logrank inference.

Risk Sets and Event-Level Information

At a given event time, let $Y_1$ and $Y_0$ be the numbers at risk immediately before the event in the treatment and control groups, and let $Y=Y_1+Y_0$. For a single event, the logrank score contribution for treatment is the observed treatment event indicator minus its expected value under $H_0$. Under the null, conditional on one event occurring in this risk set, the probability that it occurs in treatment is $Y_1/Y$.

Definition: For a single event at a time with participants at risk in both groups, the logrank information contribution is the conditional variance of the treatment event indicator under $H_0$: $$ \Delta I = \frac{Y_1}{Y}\left(1-\frac{Y_1}{Y}\right) =\frac{Y_1Y_0}{Y^2}. $$ The total information over the observed event times is the sum of these contributions. A loss-to-follow-up censoring time is not an event time and contributes no event-level score or information.

This definition shows why information is not simply the number of participants enrolled or the number remaining under observation. With one event, the risk-set information contribution is largest when the groups are equally represented, and smaller when the risk set is imbalanced. In particular, the total information depends both on how many events occur and on the treatment allocation in the risk sets at those events.

Assumptions for Interpreting the Information

The logrank calculation treats censoring as independent of the event process, at least conditional on treatment group and any covariates or strata used in the analysis. In practical terms, after conditioning on that information, participants censored at a particular time should not have a different subsequent event experience from comparable participants who remain under observation. Randomization supports treatment-group comparisons, but does not by itself guarantee this censoring condition.

The interpretation also depends on using correct event and censoring times, applying the same event definition across groups, and forming each risk set from participants known to be event-free and under observation just before the event. The calculation below uses single, distinct event times. Tied event times require the tie-handling convention used in the analysis; they should not be treated as a sequence of arbitrary single events.

Worked Example: Fewer Events, the Same Observed Event Rate

Hypothetical illustration. Suppose that without loss to follow-up, balanced treatment and control risk sets produce 60 distinct events during 1,000 person-years of observed follow-up. With loss to follow-up, suppose there are 40 distinct events during 667 person-years. Assume the risk sets remain approximately balanced at the event times in both scenarios.

For balanced risk sets, each event contributes $$ \Delta I = 0.5(1-0.5)=0.25. $$ Thus the first scenario has approximately $60(0.25)=15.0$ units of information, while the second has approximately $40(0.25)=10.0$ units.

The observed event rates are $$ \frac{60}{1000}=0.0600 \text{ events per person-year} \quad\text{and}\quad \frac{40}{667}=0.05997 \text{ events per person-year}. $$ Both are approximately 0.0600 events per person-year. In this illustration, fewer observed events and less information accompany reduced follow-up time, not a lower observed event rate. The rate comparison alone does not establish that the event processes are identical; it only shows that event count and observed rate answer different questions.

How Loss to Follow-Up Changes Risk-Set Information

If loss to follow-up occurs in similar proportions in both groups, the treatment allocation among those still at risk may remain close to its original ratio. Then the information contribution per event may change little, while the total information falls because fewer events are observed by a given time. If loss differs by group, the risk-set allocation can shift. Even for the same number of events, this change can reduce information per event.

For example, with a 50:50 risk set, the contribution is $0.5(0.5)=0.25$. With a 70:30 risk set, it is $0.7(0.3)=0.21$, not 0.09. Thus, if 50 event times each had a risk-set allocation of 70:30, the summed information would be $50(0.21)=10.5$. For 50 balanced event times it would be $50(0.25)=12.5$. The imbalanced risk sets supply $2.00$ fewer information units, a 16.0% reduction relative to 12.5.

Worked Example: Unequal Loss and a Changed Risk-Set Allocation

Hypothetical illustration. At each of 50 distinct event times, suppose 70% of the participants at risk are in treatment and 30% are in control. Compare this with 50 event times at which the risk sets are balanced. Assume all these event times have both groups represented.

For each imbalanced risk set, $$ \Delta I=0.70(0.30)=0.2100. $$ The total is $50(0.2100)=10.50$ information units. For each balanced risk set, $\Delta I=0.50(0.50)=0.2500$, giving $50(0.2500)=12.50$ units. The difference is $12.50-10.50=2.00$ units, and the relative reduction is $2.00/12.50=0.160$, or 16.0%.

This comparison isolates risk-set allocation: the event count is the same in both scenarios. It does not show that unequal loss necessarily changes the observed event rate. Nor does a 70:30 risk set alone establish that censoring is informative; it shows that each event carries less logrank information than an event in a balanced risk set.

Reduced Information Is Not Automatically a Changed Event Rate

Information is a property of event contributions conditional on risk-set composition. An observed event rate is a summary of events relative to observed time or another explicitly defined denominator. A reduction in the number of events can occur simply because participants contribute less observation time. A rate can remain similar while total information falls; a rate can also change because the participants who remain under observation differ in their event risk.

The second possibility is especially concerning when loss to follow-up is related to prognosis after accounting for the analysis variables. For example, if participants with worsening health are more likely to leave observation in one group, the observed risk set may become selectively healthier. A logrank test then compares those still observed, and its usual interpretation may not describe the randomized groups’ event experience. A change in the observed rate should therefore prompt questions about who was lost, when, and why—not an automatic conclusion about treatment effect.

Worked Example: Selective Loss Can Distort an Observed Comparison

Hypothetical illustration. Suppose 100 participants per group are divided equally into higher-risk and lower-risk participants. Under a null treatment effect, imagine that over a specified period 20% of higher-risk participants and 4% of lower-risk participants experience the event. In control, all 100 remain observed, with 10 events among the 50 higher-risk participants and 2 among the 50 lower-risk participants: 12 events in total.

Suppose that in treatment, 25 higher-risk participants are lost before their outcomes for the period are known. Among the 75 retained participants, imagine 5 events among the 25 higher-risk participants and 2 among the 50 lower-risk participants. The observed event proportion among those retained is $7/75=0.09333$, compared with $12/100=0.1200$ in control.

That difference cannot be read as evidence of a treatment benefit. The groups have different observed risk composition, and the outcomes of the lost higher-risk participants are unknown. If their loss is related to their subsequent event risk even after conditioning on variables in the analysis, the independent-censoring assumption is not supported. The logrank statistic and its information describe the observed risk sets; they cannot recover unobserved outcomes merely by counting the participants who remain.

Checking for Information Loss and Failure Modes

Useful monitoring summaries include cumulative events by group, numbers at risk by group at event times, the number and timing of losses to follow-up, observed follow-up time, and reasons for loss when available. A drop in event accrual alongside a drop in observed person-time is consistent with reduced observation, but is not enough to establish why it happened. Marked differences in loss between groups or losses concentrated among participants with poor prognosis raise concern about the censoring assumption.

These summaries diagnose potential problems; they do not prove independent censoring. The logrank information calculation is not a correction for informative loss, and more precise information arithmetic cannot compensate for systematic missing outcomes. When the censoring assumption is doubtful, the analysis plan may need sensitivity analyses or other methods appropriate to the missingness mechanism, rather than a simple interpretation of the ordinary logrank result.

A common mistake is to say, “There are fewer events, so the event rate must have fallen.” That conclusion confuses an event count with a rate. The correction is to examine observed follow-up and state the denominator used for any rate, then separately calculate information from event-time risk sets. Another mistake is to count a participant lost to follow-up as an event or to treat the participant as remaining at risk indefinitely. The correction is to censor at the last known event-free time, subject to the study’s prespecified data rules.

Compared with using enrolled sample size as a measure of progress, event-based information accounts for whether events occurred and how the groups were represented in their risk sets. Compared with using only the event count, it also reflects risk-set imbalance. Neither measure, however, is a substitute for examining follow-up completeness and censoring assumptions. The earlier material on logrank information from risk-set variances gives the general risk-set basis for this distinction.

Key takeaway: Loss to follow-up removes future observation and can reduce event accumulation. Its effect on logrank information depends on both the resulting event count and treatment allocation in the risk sets. Neither fewer events nor lower information alone proves a lower observed event rate or a treatment effect.

Check Your Understanding

Answer each question, then compare your reasoning with the solution provided.

A single event occurs when $Y_1=40$ and $Y_0=60$. What is its logrank information contribution? Solution: $Y=100$, so $\Delta I=(40/100)(60/100)=0.2400$ information units.
There are 30 events, each with balanced risk sets. What is the total information? Solution: Each contributes 0.25, so $I=30(0.25)=7.50$ units.
Does a lower event count necessarily imply a lower observed event rate? Solution: No. The event rate depends on its denominator, such as observed person-time. Fewer events with proportionally less observed time can yield a similar rate.
At an event time, 70% of the risk set is in treatment. What is the information contribution, and why is it not 0.09? Solution: It is $0.70(0.30)=0.2100$. The control proportion is 0.30, so the product is 0.21, not 0.09.
Why can selective loss of higher-risk participants threaten a logrank interpretation? Solution: If loss is related to subsequent event risk after conditioning on the analysis variables, censoring is informative. The remaining risk sets may no longer represent comparable event experience, and the ordinary logrank calculation does not recover the missing outcomes.