Why Outcome and Follow-Up Rules Need to Be Explicit
A report that says “we assessed hospitalization during follow-up” leaves several important questions unanswered. Does hospitalization mean any admission or only an admission for a specified diagnosis? Which records identify it? Does follow-up start on the cohort entry date, the treatment date, or a later date? What happens if a person leaves the health plan before the stated follow-up ends? Different answers can produce different outcome data even when analysts use the same underlying records.
Reporting outcomes is therefore not just naming a clinical concept. It requires a reproducible operational definition, a specified source and ascertainment schedule, and rules for the observation period. This tutorial assumes that the study’s population, time zero, and treatment strategies have already been defined. The outcome and its follow-up window must align with those choices.
For an event outcome, a useful specification says whether the analysis counts the first occurrence or all occurrences; whether it requires a particular diagnosis, procedure, or care setting; and how dates are assigned. For a continuous outcome, it states the instrument or laboratory test, units, assessment window, and rule for selecting a value if more than one is available. A clinically familiar label alone is not enough.
Define the Event, Source, and Assessment Time
A clear outcome description answers three linked questions. What qualifies? State the event criteria or measurement precisely enough that a second analyst could apply them. Where does the evidence come from? Name the database, clinical record, registry, survey, or other source, and say how records or measurements identify the outcome. When is it assessed? State whether it is assessed continuously from event dates, at scheduled visits, or in a defined window.
For coded outcomes, a reproducible definition can identify diagnosis or procedure codes, care settings, and whether a code must appear in a particular position or be confirmed in more than one record. Such rules should reflect the intended clinical event: a diagnosis code on an outpatient claim may not represent the same event as an inpatient admission. If an outcome algorithm was validated or adapted from a specified source, report that provenance and any modifications. Avoid implying that a code proves a clinical event when the code is only a recorded indicator.
Ascertainment can also differ across data sources. Claims may capture billed services during enrollment but not care outside the covered system. An electronic health record may contain detailed laboratory values but miss tests performed elsewhere. A registry may provide standardized event dates but cover only certain outcomes. State the relevant coverage and limitations so readers can judge what “no recorded outcome” means. It may mean the event did not occur, or it may mean the event was not observed in the available source.
Define the Window and Its End Rules
State the time origin explicitly. In many comparative studies it is the previously defined index date, but the report should still say that follow-up begins on that date or at a specified time relative to it. Define whether outcomes on the index date count. For date-based records, specify the calendar-day convention; if the exact order of treatment and outcome on the same day is not known, do not silently treat the outcome as occurring after treatment.
For a fixed-horizon outcome, state the horizon and its boundary rule, such as “from the day after index through day 30, inclusive.” Then say what happens if observation ends before that horizon. The end could be an administrative study end, disenrollment, death, or another stated event. These are not interchangeable. If a participant has no recorded outcome before leaving the data source on day 20, that record does not establish that the participant remained outcome-free through day 30.
For time-to-event outcomes, report the event date, maximum follow-up, and censoring rules. Censoring means that the event time is not observed beyond a known time; it is not evidence that the event never occurs. Also distinguish censoring from a competing event. For example, death before a nonfatal hospitalization prevents later observation of that hospitalization. Whether death is treated as a competing event or handled in another specified way depends on the outcome and analysis; it should not be described as ordinary loss to follow-up without explanation.
For repeated measurements, a window needs an assessment rule as well as start and end dates. “At six months” is ambiguous if actual visits vary. A report can specify an allowable range of days and how one value is selected when multiple measurements occur. It should also distinguish a missing measurement from a value indicating no change or no event.
A Short Justification for Recording Censoring Explicitly
Let \(T_i\) denote the time from the defined start of follow-up to the event for person \(i\), and let \(C_i\) denote the time to the end of observable follow-up. Let \(\tau\) be the maximum follow-up horizon. The observed time is the earliest of the event, the end of observation, and the horizon:
The event indicator is one only if the event is observed by that time, before censoring, and within the horizon:
If \(C_i < \tau\) and no event has been observed, the record establishes an event-free period only through \(C_i\), not through \(\tau\). This is why coding everyone without a recorded event as a fixed-horizon non-event can be wrong. A simple proportion of observed events over unequal follow-up periods also combines different opportunities for an event to occur. The window and censoring rules must therefore be reported before readers can interpret an outcome summary.
These definitions describe what was observed; they do not, by themselves, establish causal effects. In analyses that use time-to-event methods, assumptions about censoring may be needed for valid inference. Reporting should make the observation and censoring process transparent rather than assume that incomplete observation is harmless.
Worked Example: A 30-Day Claims-Based Outcome
Suppose a hypothetical study follows 100 people from an index date of January 1, 2025, using linked health-insurance claims. The outcome is a new inpatient admission for acute myocardial infarction. The study’s coding algorithm requires a qualifying acute myocardial infarction diagnosis in the principal discharge-diagnosis field of an inpatient claim.
Operational definition and source. The outcome is the first qualifying inpatient admission identified in the claims data. The diagnosis-field requirement and inpatient setting are part of the definition; an outpatient diagnosis alone does not count. The report should name the code set and any algorithm version used, rather than leave “acute myocardial infarction” as an unspecified label.
Window. Let January 1 be day 0. Count qualifying admissions dated from January 2 through January 31, inclusive: the 30 calendar days after the index date. An event recorded on January 1 is excluded by this date-based rule because the claims data do not establish that it occurred after the index event. The end of follow-up for a person with continuous coverage is day 30.
Application to the hypothetical data. Assume all 100 people have claims coverage through day 30, and eight have a qualifying admission during the window. The observed fixed-window proportion is
Thus, 8% of this hypothetical cohort had a recorded qualifying admission within the specified 30-day window. This summary is interpretable because the example stipulates complete claims observation through day 30 for everyone. If a person instead disenrolled on day 20 without an observed admission, that person’s claims would not establish the absence of an admission through day 30. The report would need to describe that shortened observation and the analysis rule for it.
Worked Example: First Nonfatal Event With Earlier Death
Consider a hypothetical cohort followed for up to 180 days from March 1. The outcome is first hospitalization for a specified nonfatal condition, identified from inpatient claims. Use event time in days after the index date, and count admissions after index through day 180. For illustration, assume the records show the following:
| Person | Observed record | Follow-up interpretation |
|---|---|---|
| A | Qualifying admission on day 40 | Event observed at day 40 |
| B | Death on day 70; no qualifying admission before death | Competing event at day 70 |
| C | No admission; coverage ends on day 100 | Censored at day 100 |
| D | No admission; covered through day 180 | Observed event-free through day 180 |
For person A, \(T_A=40\), so the recorded event time is \(\tilde{T}_A=\min(40,C_A,180)=40\), and the event indicator is 1. For person C, no event is seen before coverage ends, so the data support event-free observation only through day 100; they do not establish that the person remained event-free through day 180.
Person B illustrates why a report should not automatically call every earlier end of event-free observation “censoring.” Death prevents a later nonfatal hospitalization and is a competing event for this outcome. The report should state how death is represented in the analysis. Person D, in contrast, has no qualifying event during the full specified horizon, so the data support an event-free record through day 180. The example does not treat death as if it were loss of claims coverage.
Worked Example: A Repeated Laboratory Measurement
Suppose a hypothetical study compares change in glycated hemoglobin, HbA1c, among adults with diabetes. Laboratory results come from the electronic health record, and HbA1c is measured in percentage points. The protocol defines baseline as the result closest to the index date in the 30 days before index, excluding the index date. It defines follow-up as the result closest to day 180 within days 150 through 210, inclusive. If two results are equally distant from day 180, the earlier result is selected.
For one hypothetical participant, the baseline results are 8.2% on day -24 and 8.0% on day -8. The day -8 result is closer to index, so baseline HbA1c is 8.0%. In the follow-up window, results are 7.4% on day 183 and 7.8% on day 205. Their distances from day 180 are 3 and 25 days, respectively, so the selected follow-up value is 7.4%. The recorded change is
The participant’s recorded HbA1c decreased by 0.6 percentage points between the selected baseline and follow-up measurements. Another hypothetical participant has no HbA1c result between days 150 and 210. Under this definition, that participant has no observed follow-up measurement; the absence of a result is not a value of 0%, nor does it establish that HbA1c was unchanged. Reporting the allowed window and selection rule makes clear which observed value represents the follow-up assessment.
Choosing and Reporting an Outcome Approach
The right reporting format depends on what the outcome represents. A fixed-window binary outcome answers whether a qualifying event occurred by a stated horizon. A time-to-event outcome uses event timing and can account for different observed follow-up durations, provided its censoring and competing-event rules are specified. A repeated-measure outcome describes a value at a scheduled or windowed assessment. These approaches answer different questions and should not be substituted for one another just because the same records are available.
| Outcome approach | What to specify | Important limitation |
|---|---|---|
| Fixed-window event | Event algorithm, source, horizon, boundary dates, and handling of incomplete observation | Unobserved time before the horizon cannot be treated as confirmed event-free time |
| Time to first event | Event date rule, maximum follow-up, censoring dates, and competing events | Results depend on the event and censoring definitions; a competing event is not simply absence of the event |
| Repeated measurement | Measure and units, assessment window, source, and rule for choosing among multiple values | Measurements taken at different times may not represent the same assessment point |
A concise report can follow a consistent pattern: name the outcome; give its operational criteria and data source; state when it is assessed; define time zero, window boundaries, and the maximum horizon; then explain how earlier end of observation and competing events are handled. Use the same definitions for each comparison group unless a justified design feature requires otherwise. If ascertainment differs by group, describe that difference because it may affect how recorded outcomes are interpreted.
Diagnostics and Common Reporting Pitfalls
Before finalizing a report, check that the outcome definition can be implemented using the named source, that its dates align with the stated time origin, and that the follow-up end rule is unambiguous. Review whether every participant could in principle be observed for the stated horizon, and distinguish full follow-up from an earlier end to data coverage. For repeated measures, check whether the stated window and selection rule yield one defined value, or explicitly report when no value is available. These checks improve reproducibility; they do not establish that the source captures every event.
- Vague event label. “Cardiovascular event” is not reproducible without criteria for which events qualify and how they are identified. Name the components and the source or algorithm.
- Unspecified time origin. “Within 30 days” is ambiguous unless the report states what starts the clock and which dates count.
- Calling no record a non-event. A person with only 20 days of observation has not necessarily been event-free through a 30-day horizon. Report the shortened observation and the rule used.
- Calling a competing event ordinary censoring. Death before a nonfatal outcome differs from disenrollment that ends data observation. State how each is handled.
- Reporting a visit target without a window. “Six-month HbA1c” does not explain which actual measurement was selected. Give the permitted days and the selection rule.
Check Your Understanding
Answer each question using the definitions and rules in this tutorial. Brief solutions follow the questions.
- A study says “stroke within 90 days” but does not define the data source, time origin, or boundary dates. Name three details that should be added.
- A participant has no recorded event and leaves the data source on day 45 of a 60-day window. What does the record establish about event-free observation?
- For a nonfatal outcome, a participant dies on day 20 without first experiencing the outcome. Why should the report distinguish this from data coverage ending on day 20?
- A protocol defines a day-180 laboratory value using a window from days 150 through 210. A participant has results on days 170 and 200. Which value is selected if the rule is closest to day 180?
Solutions. (1) Specify the source and operational event criteria, identify the event’s time origin, and state whether the index date or day 90 is included; also describe early observation ends. (2) The record supports no recorded event only through day 45. It does not establish event-free status through day 60. (3) Death prevents later observation of the nonfatal outcome and is a competing event, whereas the end of data coverage is a censoring or observation-ending event. The analysis rule should state how each is handled. (4) Day 170 is 10 days from day 180, while day 200 is 20 days away, so the day-170 result is selected.