Why Observation-Ending Rules Matter
A target-trial protocol must specify not only the outcome and follow-up horizon, but also what happens when the data stop revealing a participant’s outcome before that horizon. “No event recorded” can mean that the participant remained under observation without an event, that a measurement was not taken, or that the data source no longer captured the participant. Those situations are not interchangeable. In particular, an unobserved outcome is not automatically a non-event.
The protocol should define an observation-ending rule for each outcome and data source. The rule identifies when outcome ascertainment stops, what information can still establish the endpoint, and how an outcome that remains unknown is represented. These decisions should be made before examining outcome patterns by treatment group.
Four Different Ways Outcome Information Can Be Incomplete
The protocol should name the specific reason that observation ends or a measurement is unavailable. This makes the rule operational and prevents a single generic “censored” label from obscuring different data processes.
- Loss to follow-up occurs when the study can no longer ascertain a participant’s outcome through its planned contact or data-capture procedures, even though the participant remains eligible for follow-up. The protocol should define what constitutes loss—for example, a specified sequence of unsuccessful contacts—and the date assigned as the end of ascertainment. The last successful contact is often relevant, but it is not automatically the correct censoring date for every endpoint.
- Database disenrollment occurs when a participant leaves the database’s covered population or enrollment ends, so the data source no longer captures the participant’s outcomes. This is a data-capture boundary, not necessarily a change in the participant’s health or treatment. If an independent linked source continues to ascertain the endpoint, disenrollment from one database need not end outcome observation.
- A missing outcome measurement is an absent value at a scheduled or otherwise defined measurement occasion. A single missed visit does not, by itself, establish permanent loss to follow-up. Later measurements may still be observed. The protocol should state whether the endpoint requires that particular measurement, permits a defined measurement window, or can be ascertained from another source.
- Administrative censoring occurs when follow-up ends because the protocol or data extract reaches a prespecified calendar cutoff or maximum follow-up date. It is set by the study’s observation schedule, not by a participant-specific event such as leaving the database. Participants may therefore have different follow-up durations if they enter at different calendar dates.
These categories describe why information is unavailable; they do not decide how an outcome should be analyzed. A participant can have a missing visit but remain under observation, or leave a particular database while an outcome registry continues to provide complete follow-up. The protocol must connect the reason for incomplete information to the endpoint’s actual ascertainment process.
Writing the Observation-Ending Rule
A useful rule is specific enough that two analysts using the same records would assign the same end-of-observation date. It identifies the planned horizon, the sources that can establish the endpoint, the relevant date for each source, and the handling of events that are known before observation ends. The event definition and the follow-up horizon have already been specified as part of the target-trial protocol; the observation rule must be consistent with them.
Specify the time origin and the end of follow-up, such as 365 days after assignment. Keep this horizon distinct from the date the database extract was created.
List the data sources that can establish the outcome and which source governs if dates or records conflict. State whether outcome ascertainment continues after database disenrollment.
Specify how loss to follow-up is declared, the effective date of disenrollment, the treatment of missed measurement occasions, and the common administrative cutoff.
If the event is established before observation ends, record it as an event. If the endpoint is not established through the horizon, record it as unknown at that horizon rather than as event-free.
Confirm that the observation boundary does not silently change the outcome definition or follow-up horizon. State how an event that affects the endpoint, including a competing event, is classified under the prespecified outcome definition.
Worked Example: Assigning Event-Time Follow-Up Dates
Suppose a hypothetical protocol follows participants for up to 365 days after assignment and defines the outcome as the first hospitalization for a specified condition. The hospital record is the outcome source. The extract closes on day 300 for this participant. A participant has a hospitalization on day 80; a second participant has no hospitalization recorded and loses contact after day 120; a third remains observable through day 300 without hospitalization.
For the first participant, the event is established on day 80. The observed time is \(\min(80,300)=80\) days, and the event indicator is \(I(80\leq300)=1\). The hospitalization is counted even though observation later ends.
For the second participant, the protocol defines loss to follow-up as occurring at the last date on which the outcome source can establish event status, day 120. The observed time is \(\min(T,120)=120\) days if no earlier event is found, and the event indicator is 0 through day 120. This does not establish that the participant remained event-free through day 365. The 365-day event status is unknown unless another source establishes it.
For the third participant, observation ends administratively on day 300. The participant is known to be event-free through day 300, but the data do not establish 365-day status. The assigned administrative end date is not day 365 merely because that was the intended horizon.
The rule is internally consistent only if the extract cutoff truly limits the available outcome information for this endpoint. If the hospital record remains linkable beyond day 300, the data-extract date alone need not end the participant’s outcome observation.
Fixed-Horizon Outcomes and Missing Measurements
For a fixed-horizon outcome, the key question is whether the outcome can be determined at the specified horizon, not simply whether every planned visit occurred. A missed intermediate measurement may have no effect on a later endpoint that is observed as planned. Conversely, a completed earlier visit cannot establish that a participant remained event-free for the rest of the horizon.
Worked Example: A Missed Visit Is Not Automatically Censoring
In a hypothetical study, the endpoint is a symptom score measured at 12 months, with a permitted window from day 350 through day 380. A participant completes the baseline and 6-month assessments, misses the 12-month appointment, and completes a score on day 365 through a remote assessment. Another participant misses the 6-month assessment but completes the day-365 assessment.
For the first participant, the day-365 score falls within the prespecified 12-month window. The 12-month outcome is observed; the missed appointment does not make the endpoint missing. The protocol should say whether the remote assessment is an accepted source and how the value is assigned to the 12-month endpoint.
For the second participant, the missing 6-month score does not make the 12-month score unavailable. The endpoint is observed at day 365, although the intermediate measurement is missing. That intermediate value must not be filled with an assumed score of zero or carried forward unless the protocol has explicitly defined and justified such a procedure.
If a third participant has no score in the day-350-to-day-380 window and no valid alternative source, the 12-month measurement is missing. That fact alone does not say whether the participant was lost to follow-up, disenrolled, or simply missed the visit. The reason and the endpoint status should be recorded separately.
What Censoring Assumptions Do—and Do Not—Say
A protocol can define censoring precisely without making censoring harmless for estimation. When outcome ascertainment stops before the horizon, the outcomes after that time are unknown. An analysis of observed follow-up generally needs assumptions about how that loss of information relates to the outcome, as well as sufficient continued observation across relevant participant histories. For example, a conditional independent-censoring assumption says that, conditional on specified observed information, censoring does not additionally depend on the future event time. Defining a censoring date does not establish that assumption.
The consequence can be seen for a fixed-horizon binary outcome. Let \(R=1\) indicate that the outcome is observed through \(\tau\), and let \(p=P(Y(\tau)=1)\) be the target risk. The risk among participants with observed outcomes is
If \(R\) and \(Y(\tau)\) are independent, this expression equals \(p\). But if participants who experience the event are less likely to have an observed endpoint, the observed risk can be lower than the target risk. For example, suppose the target risk is \(0.20\), \(60\%\) of participants with an event are observed through the horizon, and \(90\%\) of those without an event are observed. Then
The observed risk is about \(14.3\%\), rather than the target \(20\%\). This calculation is not an adjustment method; it shows why treating only observed outcomes as representative requires justification. In a causal comparison, differences in outcome ascertainment between treatment groups can also compromise the intended contrast.
Worked Example: Database Disenrollment With Continued Outcome Ascertainment
Consider a hypothetical 180-day study in which the primary endpoint is death, ascertained from a linked vital-status registry. A participant’s insurance database enrollment ends on day 100, but registry linkage is available through day 180. The participant has no death record through day 180.
If the protocol specifies that the linked registry establishes vital status regardless of insurance enrollment, day 100 is not the end of outcome observation for this endpoint. The endpoint is observed through day 180, and the participant is classified as having no recorded death by that horizon, subject to the registry’s defined coverage and reporting rules.
If the registry link were unavailable after day 100, then day 100 would end ascertainment from the available sources. The absence of a death record through day 100 would not establish 180-day survival. The protocol must therefore distinguish database disenrollment from loss of endpoint ascertainment, rather than automatically equating the two dates.
Diagnostics, Alternatives, and Limitations
A protocol’s observation rule should produce records that can be audited. Before analysis, summarize the number and timing of each type of observation ending, separately by treatment group and relevant follow-up interval. Check for impossible dates, such as an event recorded after the stated censoring date without a source that supports later ascertainment. Confirm that participants with a known event before censoring are retained as events, and that a missing scheduled measurement has not been silently converted to a non-event.
Useful checks include comparing the recorded last-contact date with the disenrollment date; reviewing whether endpoint sources cover the full administrative period; and checking that follow-up windows are applied consistently. Large differences in loss, disenrollment, or missingness across treatment groups do not by themselves prove bias, but they are a warning that a simple comparison of observed outcomes may not represent the target contrast.
| Protocol choice | What it means | Important limitation |
|---|---|---|
| Censor at last verified outcome information | Use the last date through which the endpoint is actually ascertainable. | Does not make future outcome status known or guarantee censoring is unrelated to future outcomes. |
| Continue ascertainment through a linked source | Use another specified source after contact or database enrollment ends. | Requires a clear source hierarchy and a defined period of coverage. |
| Define a measurement window for a fixed-horizon endpoint | Accept measurements in a prespecified range around the target time. | Does not recover a value when no valid measurement is available in the window. |
| Stop everyone at a common administrative date | Apply a shared calendar cutoff to outcome collection. | Participants entering later may have less potential follow-up before the cutoff. |
These choices answer different questions and should not be conflated. In particular, changing the horizon to the last observed date for each participant would make follow-up duration participant-dependent; it would not recover the originally specified fixed-horizon outcome. A competing event also requires the endpoint-specific classification already set by the protocol: it should not be relabelled as ordinary loss to follow-up merely because it prevents a later event of interest.
Common Mistake: “No Record Means No Event”
Wrong reasoning: “There is no event code after day 120, so the participant had no event through day 365.” This treats absence of data as evidence of absence of the outcome.
Correction: First establish whether the endpoint source could have recorded the event through day 365. If ascertainment ended on day 120, the participant’s 365-day status is unknown unless another valid source resolves it. Record the event-free period only through the last date supported by the data, and apply the prespecified observation-ending rule.
Check Your Understanding
For each question, identify the observation rule and distinguish an observed outcome from an unknown outcome.
- A study has a 90-day binary event endpoint. A participant has no event record through day 40, when contact ends, and no alternative source is available. Is the 90-day outcome 0, 1, or unknown?
- A participant misses the 6-month visit but completes the protocol-defined 12-month outcome assessment. Does the missed visit, by itself, censor the participant before 12 months?
- A data extract ends on day 250 in a study with a 365-day horizon. A participant has no event recorded by day 250. What can be concluded about the 365-day endpoint?
- In the risk calculation above, suppose \(P(Y(\tau)=1)=0.20\), \(P(R=1\mid Y(\tau)=1)=0.60\), and \(P(R=1\mid Y(\tau)=0)=0.90\). What is the risk among participants with observed outcomes?
Solutions. (1) Unknown. No event was established through day 40, but the data do not determine whether an event occurred from day 41 through day 90. (2) No. The endpoint is observed if the day-365 assessment meets the protocol’s measurement rule; the missed intermediate assessment remains a separate missing measurement. (3) The participant is known to have no recorded event through day 250, but the 365-day endpoint is unknown unless another source establishes status after day 250. (4) The observed risk is \(0.60(0.20)/[0.60(0.20)+0.90(0.80)]=0.12/0.84=0.1429\), or about \(14.3\%\).