Why Loss to Follow-Up Needs Special Attention
A participant who is lost to follow-up stops contributing observed event information at the last time their status is known. The record is right-censored at that time, but the reason for stopping observation matters. A scheduled study end can censor participants without regard to their health. By contrast, a participant may stop attending because their health is worsening, because they feel well and see no reason to return, or for reasons unrelated to prognosis. These patterns have different implications for inference.
The central question is not simply whether some participants were lost. It is whether, among people still under observation and free of the event at a given time, the chance of loss is related to their future event experience. Participant characteristics and health measurements recorded before loss can reveal such relationships. They are evidence to investigate, not a direct observation of what would have happened after follow-up stopped.
This definition concerns the relation between censoring and the event process, not whether loss is predictable from a measured variable in isolation. A characteristic may predict loss but, after conditioning on that characteristic, censoring may still be independent of future event time. Conversely, loss can depend on unrecorded health changes even when measured baseline variables show little association with loss.
In competing-risks data, an observed competing event is an event, not a loss-to-follow-up censoring time. A participant who experiences a competing event is no longer at risk of the event of interest under the observed event process. The assessment here concerns people whose follow-up ends without an observed event because they cannot be reached or no longer provide data.
What Can Be Learned From Observed Characteristics?
Start by describing who is lost, when loss occurs, and what was known about each participant before it occurred. Baseline characteristics can include age, disease severity, treatment group, or access to care. Evolving health status can include symptoms, laboratory results, or functional measures collected during follow-up. For an evolving measure, its value must have been recorded before the loss being examined; a measurement taken after the person has already disappeared cannot explain that earlier loss.
A useful first comparison is the proportion lost over a defined period in groups formed by a characteristic. With unequal follow-up, a single overall proportion can be misleading: participants observed longer have more time in which loss can occur. Comparisons should therefore use a clearly defined time window or intervals among participants still under observation and event-free at each interval. The denominator and the treatment of observed events must be stated.
For evolving health status, compare loss over a subsequent interval among participants who were still being followed and event-free when that health status was measured. This ordering supports a temporal interpretation: the recorded status preceded the observed loss. It does not show that the status caused the loss, nor does it reveal the event outcomes that would have occurred after loss.
Worked Example: Baseline Severity and One-Year Loss
Hypothetical data. At enrollment, 80 participants have severe disease and 120 have mild disease. By 12 months, 24 severe-disease participants and 12 mild-disease participants have been lost to follow-up before an observed event. Suppose status through 12 months is known for everyone, and the comparison is the cumulative proportion lost before an event in that period.
The observed loss proportions are \(24/80=0.300\) for severe disease and \(12/120=0.100\) for mild disease. The risk difference is \(0.300-0.100=0.200\), or 20 percentage points. The risk ratio is \(0.300/0.100=3.00\). In this hypothetical cohort, loss before an observed event was three times as common by 12 months in the severe-disease group.
For an approximate 95% confidence interval for the difference in two independent proportions, use the estimated standard error \(SE=\sqrt{\hat p_1(1-\hat p_1)/n_1+\hat p_0(1-\hat p_0)/n_0}\). Here, \(SE=\sqrt{0.300(0.700)/80+0.100(0.900)/120}=\sqrt{0.003375}=0.05809\). The interval is \(0.200\pm1.96(0.05809)\), or approximately \((0.0861,0.314)\). Under the assumptions for this large-sample interval, the data are compatible with a 12-month difference of about 8.6 to 31.4 percentage points. This interval describes uncertainty in the loss-proportion difference; it does not quantify the effect of loss on an event estimate.
The comparison is a warning sign if disease severity also predicts future events: the people lost may have a different prognosis from those remaining under observation. It is not proof of informative censoring. For example, the difference could be explained by recorded severity if the analysis conditions appropriately on severity and censoring is independent of future events within severity groups.
Check the Timing of Changing Health Status
A health measurement made during follow-up can be more informative about the circumstances of loss than a baseline measurement, but it also requires careful construction. At each assessment time, define the eligible group as participants still observed and free of the event. Classify their health status using information available at that time, then examine loss during a later, specified interval. Do not classify participants using measurements recorded after that interval begins or after loss.
An interval-based comparison is descriptive. It asks whether the observed frequency of loss in the next interval differs by the health status recorded at the interval’s start. Events observed during the interval must be distinguished from loss; they are not simply treated as participants lost to follow-up. If the groups differ in their follow-up time or opportunity to be lost, a raw proportion may not be comparable.
Worked Example: Worsening Health Before Loss
Hypothetical data. At a scheduled visit, 150 participants are still under observation and event-free. Sixty have a worsening symptom score since the prior visit; 18 are lost to follow-up before an observed event during the next three months. Of the 90 whose score is stable, 9 are lost in that interval. Assume each participant’s status is ascertained through the interval or until an observed event or loss, and that the same three-month window is used for both groups.
The three-month loss proportions are \(18/60=0.300\) for worsening symptoms and \(9/90=0.100\) for stable symptoms. Thus the risk difference is \(0.300-0.100=0.200\), and the risk ratio is \(0.300/0.100=3.00\). In this example, loss before an observed event was more frequent in the group whose symptoms had worsened before the interval.
Because symptoms were measured before the interval in which loss was counted, the ordering is consistent with worsening status being a predictor of later loss. If worsening symptoms also predict the event, this pattern suggests that the people who leave observation may differ in prognosis. It still cannot establish their unobserved event times after loss. The comparison also relies on the groups being defined at the same landmark and on losses and observed events being classified consistently.
Separate Crude Associations From Conditional Evidence
A crude association can arise because a characteristic is related to another measured prognostic factor. For example, symptom worsening may be more common in people with severe disease, and severity may itself predict loss. Comparing loss by symptom change without considering severity can therefore give a different picture from comparing within severity groups. The question relevant to a conditional independent-censoring assumption is whether censoring remains related to future event experience after conditioning on the observed variables and history that the analysis uses.
Worked Example: A Crude Association That Disappears Within Severity Groups
Hypothetical data. At a landmark visit, participants are classified by disease severity and whether symptoms worsened before the next interval. The numbers lost in that interval are:
| Severity | Symptoms worsened | Symptoms stable |
|---|---|---|
| Severe | 16 lost of 80 | 4 lost of 20 |
| Mild | 1 lost of 20 | 4 lost of 80 |
Ignoring severity, the worsening group has \(17/100=0.170\) lost, while the stable group has \(8/100=0.080\) lost. The crude risk difference is \(0.170-0.080=0.090\), or 9 percentage points, and the crude risk ratio is \(0.170/0.080=2.125\).
Within the severe group, the proportions are \(16/80=0.200\) and \(4/20=0.200\). Within the mild group, they are \(1/20=0.050\) and \(4/80=0.050\). Thus, in this constructed example, symptom worsening is not associated with loss within either severity group; the crude difference reflects the different severity composition of the worsening and stable groups. Severity, however, is associated with loss: the loss proportion is 20% in the severe group and 5% in the mild group.
This pattern illustrates why an unadjusted comparison is not enough to assess conditional censoring. If severity is measured adequately and is included in the information conditioned on, the symptom-worsening association may add no evidence beyond severity. Whether censoring is independent of the future event process within severity groups remains a separate assumption; this table alone cannot verify it.
Evidence, Diagnostics, and Limits
Evidence that loss may be informative is strongest when several observations align: loss is more frequent among people with known poor prognosis; a health change recorded before loss predicts subsequent loss; the pattern persists within relevant measured groups; and the timing of loss is clinically plausible as a response to health. Check these patterns at more than one follow-up interval when possible. A single end-of-study loss proportion can conceal when the divergence began.
Useful descriptive checks include:
- Tabulating counts and proportions lost by baseline characteristics, treatment group, and disease severity.
- Comparing loss in later intervals by health status measured before each interval, among participants still at risk at its start.
- Displaying the timing of loss alongside observed event timing, rather than reporting only a total number lost.
- Checking whether comparisons use the same time window and appropriately account for observed events that remove participants from the event-free risk set.
- Reviewing completeness and timing of health measurements. If measurements are missing more often among people approaching loss, a comparison based only on recorded values may itself be selected.
These checks detect patterns in observed data. They cannot establish independence from unobserved future outcomes, because those outcomes are precisely what is unknown after loss. An absence of an observed association is not proof that censoring is non-informative: relevant health status may be unmeasured, measured with error, or recorded too infrequently to capture a change preceding loss. Conversely, a clear association between a measured covariate and loss does not automatically invalidate an analysis if the association is adequately explained by variables on which the censoring assumption is conditioned.
A practical comparison is between administrative censoring and loss to follow-up. Administrative censoring often occurs at a planned boundary, such as a fixed study end; its timing may be explained by entry date and calendar time. Loss to follow-up is a departure from the planned observation process and may depend on participant circumstances or health. Neither label by itself guarantees independent censoring: assumptions depend on the actual mechanism and on what is conditioned upon. In the earlier material on right-censoring mechanisms and conditional independent censoring, the key distinction is whether censoring is independent marginally or after conditioning on observed information.
The main limitation is fundamental: observed predictors of loss are not the same as the unobserved outcomes after loss. Descriptive comparisons help determine whether a simple independent-censoring assumption is questionable and which observed information may matter. They do not, by themselves, recover missing event histories or prove that a particular analysis is unbiased. Methods that model censoring or adjust estimates for it require additional assumptions and are separate procedures.
A Short Justification for the Diagnostic Logic
The reason a loss association matters can be expressed using conditional probabilities. Let \(X\) denote the observed information on which independence is assumed, and let \(I(T>t)\) indicate that the event has not occurred by time \(t\). If censoring by \(t\) is independent of the event process conditional on \(X\), then among participants with the same \(X\), selection by remaining uncensored does not change the event-free probability:
The equality requires \(P(C\ge t\mid X)>0\) and conditional independence of the event and censoring processes given \(X\). If people with the same recorded \(X\) who remain under observation have different event prospects from those censored, this equality fails. A measured health status associated with later loss is a reason to examine whether it belongs in the relevant conditioning information. But even after adding measured variables, unobserved common causes can remain, so observed data cannot certify the assumption.
Common Mistakes
- Calling every censoring event informative. Loss can be independent of future event experience; the fact that follow-up ended does not establish dependence.
- Treating an association as proof. A measured predictor of loss raises a question about the censoring assumption but does not reveal post-loss event outcomes or prove bias.
- Using health status measured after loss. A predictor must be recorded before the loss interval being studied; otherwise temporal order is reversed or unavailable.
- Ignoring the risk set. Comparing people with different opportunities for loss, or treating observed competing events as losses, can produce misleading proportions.
- Concluding that no association means no problem. Unmeasured or poorly timed health changes can still link loss to prognosis.
Check Your Understanding
Answer each question using the definitions and reasoning above. Brief solutions follow the questions.
- In a hypothetical cohort, 15 of 50 participants with a poor-prognosis baseline characteristic are lost by six months, compared with 10 of 100 without it. Calculate the two loss proportions, risk difference, and risk ratio.
- Why should a symptom measure used to assess later loss be recorded before the interval in which loss is counted?
- A crude association between symptom worsening and loss disappears after comparing participants within disease-severity groups. What does this show, and what does it not show?
- Why is an observed competing event not counted as loss to follow-up?
- If no measured baseline characteristic predicts loss, can one conclude that censoring is independent of future event time? Explain.
Solutions. (1) The proportions are \(15/50=0.300\) and \(10/100=0.100\). The risk difference is \(0.300-0.100=0.200\), or 20 percentage points, and the risk ratio is \(0.300/0.100=3.00\). (2) Recording the symptom status first establishes that it was observed before the later loss; a post-loss value cannot serve as a prior predictor. (3) The crude association may reflect different severity composition, so worsening adds no association within the displayed severity groups. This does not establish independence from future event time within those groups. (4) A competing event is an observed event in the event process; the participant’s event status is known rather than lost. (5) No. Unmeasured or poorly timed health changes could link loss to future event outcomes, so absence of measured baseline associations is not proof of independent censoring.