Why Follow-Up Must Be Defined Before Comparing Outcomes
A trial patient observed for one year and an external-control patient observed for six weeks have not had equal opportunity for an outcome to be detected. Comparing their outcome counts as if observation were equally complete can misrepresent the contrast, even when eligibility, treatment strategies, and time zero have been aligned. Follow-up rules specify when outcome observation begins, how long it is intended to continue, and what ends reliable observation.
The objective is not to force every patient to have the same observed duration. It is to define a common clinical question and represent each patient’s actual observation faithfully. This requires separating the clinical event of interest from the end of the period in which the event could be ascertained. In an external-control comparison, that separation is especially important: data-system enrollment, visits, and record availability do not always track a patient’s underlying clinical status.
The earlier material on setting time zero and aligning calendar time provides the basis for choosing a common origin and addressing when patients entered observation. Here, the emphasis is on what happens after that origin. For example, if the endpoint is first hospitalization for a specified condition within 180 days, the endpoint definition must state which hospitalizations qualify, and follow-up must establish whether such an event could be detected through the relevant period. “No event recorded” is not automatically equivalent to “known to remain event-free.”
Observed Follow-Up and Censoring
Let \(T\) be the time from time zero to the event of interest, and let \(C\) be the time from time zero to the end of reliable event ascertainment, if ascertainment ends before the event occurs. For a right-censored observation, the recorded follow-up time is \(Y=\min(T,C)\). The event indicator \(\delta\) equals 1 when the event is observed by the end of follow-up and 0 when observation ends first. These definitions assume the event time and observation-ending time can be placed on a consistent time scale and that the event is recorded if it occurs while reliable ascertainment continues.
When \(\delta=1\), the observed time \(Y\) is the event time. When \(\delta=0\), the data establish that the patient was event-free through time \(Y\), but do not establish what happened afterward. A patient censored at 60 days is therefore not known to be event-free through day 180. Censoring is a statement about the available observation, not a clinical outcome.
Typical reasons for ending ascertainment include a prespecified database cutoff, the end of a study’s observation period, loss of a link to outcome records, or a patient leaving a health system whose records are the only source for the endpoint. The reason and date should be recorded where possible. A patient may also reach the planned horizon without an event; for an analysis limited to \(\tau\), this is administrative end of follow-up at the horizon, rather than evidence about time beyond it.
Worked Example: Translating Patient Timelines Into Follow-Up
Suppose the endpoint is first qualifying hospitalization within 180 days after time zero. A hypothetical trial participant has a qualifying hospitalization on day 94. A hypothetical external-control patient has no recorded qualifying hospitalization, and the data source can ascertain the endpoint through day 121 but not afterward.
For the trial participant, \(T=94\), so the observed follow-up is \(Y=94\) days and \(\delta=1\). The hospitalization is an observed endpoint event within the 180-day horizon.
For the external-control patient, outcome ascertainment ends at \(C=121\) days before any qualifying event is observed. The record therefore gives \(Y=121\) days and \(\delta=0\). The patient is known to have no recorded event through day 121, but the data do not establish event-free status for the remaining 59 days. Treating this patient as event-free through day 180 would overstate what the source supports.
The two records are not equivalent observations of 180-day outcome status. They can be represented as an event at 94 days and right censoring at 121 days, respectively. Whether an analysis can use the shorter observed follow-up depends on its assumptions; simply comparing the endpoint counts without accounting for the observation periods is not justified.
Data-Source Exit Is Not the Same as Clinical Follow-Up
A patient can leave one data source while remaining under observation for the clinical endpoint, or remain listed in a data source while meaningful outcome ascertainment has stopped. The relevant question is not whether the patient had another visit or remained enrolled in a particular file. It is whether events meeting the endpoint definition could still be detected after the apparent exit.
For instance, departure from an insurer’s claims database may end ascertainment of hospitalizations if no other linked records are available. But if a validated national hospital registry continues to capture those hospitalizations, insurer departure alone does not end outcome follow-up. Conversely, a patient’s last clinic visit is not necessarily the censoring date if later events are captured through comprehensive claims or registry data. A gap between encounters is not, by itself, proof of either continued event-free status or loss of outcome ascertainment.
Worked Example: Interpreting an Apparent Data-Source Exit
Consider a hypothetical patient with a 180-day endpoint horizon. Insurance claims end on day 75. A linked hospitalization registry is confirmed to capture qualifying hospitalizations through day 150, and no qualifying hospitalization is recorded through that date. There is no reliable endpoint source after day 150.
If the endpoint can be ascertained through the linked registry, day 75 is not the end of outcome follow-up. The patient has reliable outcome-free observation through day 150, so \(Y=150\) days and \(\delta=0\). The observation ends at day 150, 30 days short of the horizon.
If the registry is not actually linked for this patient, and claims are the only source capable of capturing the endpoint, ascertainment ends at day 75. Then \(Y=75\) days and \(\delta=0\). The source exit is relevant because it coincides with loss of endpoint capture, not because leaving insurance is itself a clinical event.
In neither case does “no event recorded” mean no event occurred after the last date on which the endpoint could be reliably observed. The data lineage and linkage coverage determine the defensible censoring date.
Choosing a Comparable Follow-Up Rule
A common follow-up rule should define the same duration and endpoint opportunity for both cohorts. One practical specification is: start at the aligned time zero; ascertain the defined outcome until the first of the outcome, the horizon \(\tau\), or the end of reliable ascertainment; and record why observation ended. The horizon should be clinically meaningful and supported by available data. Calendar-time alignment also matters: a common duration after time zero does not guarantee comparability if one cohort’s later outcomes fall beyond its data cutoff.
The following simple construction makes the rule explicit. Suppose a patient’s first qualifying event occurs at day \(T\), reliable data end at day \(C\), and follow-up is capped at \(\tau\). The effective observation end is the earliest of those times. An event is counted only if it occurs no later than both the ascertainment end and the horizon.
If no event occurs before observation ends, the patient contributes outcome-free observation only through \(Y\). If the patient is still reliably observed and event-free at day \(\tau\), follow-up for this question ends at the horizon. The definition presumes the same event-date conventions across cohorts, including how to handle an event recorded on the same day as a data cutoff. That convention should be stated and applied consistently.
Worked Example: Applying a Common 120-Day Horizon
Suppose a hypothetical analysis specifies \(\tau=120\) days. Patient A has a qualifying event on day 80, with reliable data through day 120. Patient B has no event recorded, but reliable data end on day 103. Patient C remains event-free with reliable data through day 120.
For A, \(T=80\) and \(C\geq120\), so \(Y=\min(80,C,120)=80\) days and \(\delta=1\). The event occurs within the horizon.
For B, no event is observed before ascertainment ends at day 103. Thus \(Y=\min(T,103,120)=103\) days and \(\delta=0\), where \(T\) is later than 103 or unobserved. This record establishes event-free observation only through day 103.
For C, there is no event through the horizon and observation is reliable through day 120. Follow-up for the specified question is \(Y=120\) days and \(\delta=0\). C has completed the horizon; B has not. Both have \(\delta=0\), but their records provide different amounts of observed event-free time.
Why Censoring Assumptions Matter
Censoring is not harmless merely because it is recorded correctly. Many time-to-event methods rely on an assumption that, within relevant groups or conditional on measured predictors, the event time is independent of the censoring time. In plain terms, among patients with the same relevant observed information, the people whose outcome follow-up ends earlier should not systematically have different unobserved event prospects from those who remain observed. This is often called non-informative censoring conditional on the variables used.
The assumption can fail if patients leave a source because their health is worsening, if care moves elsewhere in a way related to prognosis, or if a data system captures outcomes less completely for patients with particular characteristics. For external controls, the mechanism that determines observation can differ from that in the trial. Merely setting the same nominal horizon does not resolve that difference.
A useful justification for the standard right-censoring likelihood shows the role of the assumption. Let \(f_T(y)\) be the event-time density and \(S_T(y)=P(T>y)\) the probability of remaining event-free beyond \(y\). For an observed event at \(y\), the event-time contribution is \(f_T(y)\); for censoring at \(y\), the contribution is \(S_T(y)\), because the patient is known to remain event-free through that time. If censoring is independent of the event time, the joint contribution factors into an event-process term and a censoring-process term. The latter can be treated as a nuisance contribution when the target is the event-time distribution. If censoring depends on unobserved event risk, this factorization does not justify ignoring the censoring process.
This expression is a justification for why an event and a censored observation contribute differently; it is not a guarantee that the assumption holds. In practice, inspect reasons and timing of censoring by cohort, compare the available follow-up distributions, and examine whether source exit is associated with measured characteristics or clinical history. These checks can reveal concerning patterns, but cannot prove that censoring is independent of unobserved prognosis.
Distinguishing Censoring From Other Endings
The endpoint definition determines whether a recorded event ends follow-up as the outcome, as a competing event, or not at all. For an endpoint of first nonfatal hospitalization, death before hospitalization prevents a later hospitalization from being observed. Death is not ordinary loss to follow-up: it is a clinical event that can alter the probability of the outcome. Whether it is treated as a competing event, included in a composite endpoint, or addressed under another estimand depends on the clinical question. Simply coding death as ordinary censoring can answer a different question and may bias estimates if the relevant assumptions do not hold.
Other distinctions matter as well. If the event is known only to have occurred between two assessments, its timing is interval-censored rather than exactly observed at the date of the later visit. If a patient has a treatment change, that change is not automatically censoring; its role depends on the treatment strategy and estimand. These are endpoint and estimand decisions, not clerical conventions to be made after seeing the results.
| Record situation | What is known | Follow-up interpretation |
|---|---|---|
| Qualifying endpoint observed | The defined clinical event occurred at a known time | Record an event at that time |
| Outcome source ends before the event | Event-free status is known only until the last reliable observation time | Right-censor at that time, subject to the censoring assumptions |
| Patient leaves one source but a linked source continues | Endpoint ascertainment may continue despite the original source exit | Use the end of reliable outcome capture, not automatically the original source exit |
| Death before a nonfatal endpoint | A clinical event changes the possibility of observing the endpoint | Apply the prespecified competing-event or composite-endpoint definition |
Common Pitfalls and Practical Checks
- Calling every last encounter a censoring date. A last encounter is not necessarily the last time an endpoint could be found. Check the actual coverage of claims, registries, and linked records.
- Calling a patient event-free through the horizon after early source exit. No recorded event after data capture ends is not evidence of continued outcome-free follow-up.
- Using different horizons for the two cohorts without justification. Different horizons change the clinical question and opportunity to detect an event. State a common target horizon and assess whether data support it in both cohorts.
- Treating every end of observation as independent censoring. If source exit is related to prognosis, the usual censoring assumption may fail. Describe the mechanism and avoid interpreting a coding choice as proof of validity.
- Ignoring event-date uncertainty or competing events. Apply consistent date rules and preserve the distinction between loss of observation and clinical events that change the outcome process.
A practical audit should verify that time zero and the horizon are consistently applied; each endpoint and censoring date is supported by a source; patients with no recorded event have a documented last reliable observation date; and censoring patterns and reasons are summarized by cohort. When one cohort has much shorter or less complete ascertainment, the comparison may require a narrower horizon or additional assumptions and analyses. If those choices cannot support the intended estimand, the limitation should be explicit rather than hidden by coding all unobserved periods as event-free.
Check Your Understanding
Answer each question by distinguishing the clinical event from the observation process. Brief solutions follow the questions.
- A patient is event-free in the record through day 88, when the only outcome database ends. The horizon is 150 days. What are \(Y\) and \(\delta\), and what is not established?
- A patient leaves an insurer on day 60, but a linked registry reliably captures the endpoint through day 130. What date should define the end of ascertainment if no event is recorded?
- Why is a patient censored at day 40 not interchangeable with a patient known to be event-free through a 90-day horizon?
- Give one example of a reason censoring might be informative, and state the concern it raises.
- For a nonfatal outcome, why might death before that outcome not be treated as ordinary censoring?
Solutions. (1) \(Y=88\) days and \(\delta=0\); the record does not establish event-free status from day 89 through day 150. (2) Day 130, because the linked registry continues reliable endpoint ascertainment after insurer exit. (3) The first patient is known to be event-free only through day 40; the second is observed event-free through the full 90-day horizon. (4) For example, patients may leave the source when their health worsens; those censored may then have different unobserved event risk, challenging conditional independence. (5) Death changes whether the nonfatal outcome can subsequently occur, so it is a clinical competing event rather than simply a loss of outcome observation; the endpoint strategy must specify how to handle it.