Why Baseline Information Must Be Prespecified
An emulated trial needs a clear rule for what was known about each eligible person when treatment was assigned. A diagnosis or laboratory value in a database is not automatically baseline information: it may describe a period long before the treatment decision, or it may have been measured or recorded after that decision. Treating post-assignment information as baseline can make the analysis use knowledge that would not have been available at assignment.
The purpose of prespecification is to turn “adjust for clinical history” into an operational rule: which variables are needed, how each is defined, what period of history qualifies, and how values at or near time zero are classified. This does not establish that all relevant confounding has been measured. It makes the intended information set explicit so its timing and meaning can be assessed.
A protocol should list information at the variable level, not merely state “use baseline covariates.” Depending on the clinical question, the list may include demographics, prior diagnoses, disease severity, medication use, utilization, laboratory results, and relevant risk factors. These variables may help describe treatment selection and prognosis and may be needed to support the planned causal comparison. Their inclusion does not by itself guarantee adequate adjustment.
Define Time Zero and Each Measurement Window
First define \(t_0\) precisely: for example, the date and time of the first eligible treatment dispensing, or the timestamp at which a treatment decision is recorded. The definition must agree with the protocol’s eligibility, treatment strategies, and follow-up rules. If time zero is recorded only as a calendar date, the protocol should say how same-day records are handled; it should not silently assume an order that the data do not establish.
For a fixed lookback of \(k\) elapsed days, a useful convention is the half-open interval \([t_0-k\text{ days},t_0)\). The left endpoint is included and the right endpoint is excluded. Thus, an event exactly \(k\) days before time zero qualifies, but an event at or after time zero does not. State whether the window uses elapsed hours or calendar dates, and specify the time zone when timestamps matter. A phrase such as “the prior six months” is not interchangeable with “the prior 180 days.”
Different variables can reasonably have different windows. A recent laboratory result may represent current disease severity, whereas a longer history may be needed to identify chronic disease or prior medication use. Specify the rule for selecting among repeated measurements—for example, the most recent valid value before \(t_0\)—rather than choosing whichever value appears convenient after inspecting the results. Also distinguish “no qualifying record” from “known to have no condition”: a blank or absent code may reflect incomplete observation rather than a negative clinical history.
Worked Example: A 180-Day History Before Treatment Initiation
Consider a hypothetical new-user protocol in which treatment assignment occurs at 00:00 on 1 October 2025. The protocol requires prior diagnosis history and a recent laboratory value. It defines the diagnosis lookback as the 180 elapsed days immediately before time zero and selects the most recent valid laboratory result in the 30 days before time zero.
Diagnosis window. The 180-day window is \([4\text{ April 2025 at 00:00},1\text{ October 2025 at 00:00})\). The lower boundary is 180 days before 1 October 2025; it is included. The upper boundary is time zero; it is excluded. A qualifying diagnosis dated 4 April is included, while a diagnosis recorded at 00:00 on 1 October is not baseline under this rule.
Laboratory window. The 30-day window is \([1\text{ September 2025 at 00:00},1\text{ October 2025 at 00:00})\). Suppose a participant has valid results of 6.8 mmol/L on 10 September and 7.1 mmol/L on 25 September, with no result on or after time zero that qualifies. The protocol selects 7.1 mmol/L because it is the most recent valid result within the defined window. The value is reported in mmol/L, and the selection rule is reproducible.
These rules assume the dates or timestamps represent when the diagnosis or laboratory assessment occurred, not merely when an old result was imported into the database. If only an import date is available, the protocol cannot infer that the clinical measurement occurred within the window; that limitation must be flagged rather than concealed by assigning the import date as the event date.
Separate Event Time From Record Time
Clinical databases may contain several relevant dates: when a symptom began, when a specimen was collected, when a result became available to the clinician, and when the information was entered into the database. These dates answer different questions. A specimen collected before \(t_0\) but resulted afterward may describe pre-assignment biology, yet the result may not have been available to inform treatment selection. Whether it qualifies depends on the protocol’s purpose and the information actually available at assignment.
For covariates intended to represent the information available to the treatment decision, a strong rule is to require both a pre-\(t_0\) measurement time and pre-\(t_0\) availability. If the target trial would use a specimen value regardless of when the result was returned, that is a different operational definition and should be stated explicitly. When the data lack the timing detail needed to decide, classify the timing as ambiguous; do not automatically label the value baseline.
Worked Example: Records Close to Time Zero
Suppose time zero is 10:00 on 15 July 2025. A protocol asks for a blood-pressure measurement within the preceding 48 elapsed hours, using \([t_0-48\text{ hours},t_0)\). The window begins at 10:00 on 13 July and ends immediately before 10:00 on 15 July.
A measurement at 10:00 on 13 July is exactly at the included left endpoint and qualifies. A measurement at 09:50 on 15 July also qualifies. A measurement at exactly 10:00 on 15 July is at the excluded right endpoint and does not qualify; a measurement at 10:12 is post-time-zero and does not qualify.
Now suppose the database gives only “15 July” for a measurement and time zero, with no time of day. The records alone do not establish whether the measurement occurred before or after assignment. The protocol should flag this as unresolved timing, apply a prespecified conservative rule if one is justified, or exclude the value from the baseline measure. It should not assume that the blood pressure was measured before treatment merely because both events share a date.
The interval convention is checked directly: qualifying times satisfy \(t_0-48\text{ hours}\leq t<t_0\). That inequality includes the first boundary and excludes time zero, so the classification of all four boundary examples follows from the same rule.
Why Post-Time-Zero Values Are Not Baseline Covariates
A variable may be clinically important yet still be unsuitable as a baseline covariate if its recorded value can occur after assignment. Treatment may affect subsequent symptoms, testing, diagnosis, or clinical contact. Adjusting for such a post-assignment value can therefore condition on information influenced by treatment, rather than describe the state on which treatment was assigned. The protocol should flag variables whose measurement window could cross \(t_0\), even when the underlying condition may have existed earlier.
A short justification follows from the information available at assignment. Let \(\mathcal{H}(t_0^-)\) denote the clinical history available strictly before \(t_0\). A quantity used as baseline must be defined from that history. If a recorded value depends on an event after \(t_0\), it is not guaranteed to be a function of \(\mathcal{H}(t_0^-)\). Therefore it cannot be treated as baseline solely because it appears in the participant’s record. The distinction is about the timing and definition of the information, not just the date on which an analyst extracts the database.
Worked Example: A Baseline Diagnosis or a Follow-Up Diagnosis?
In a hypothetical study, treatment is initiated at 09:00 on 3 February. Investigators want to characterize disease severity using a diagnosis code and a laboratory value. The diagnosis code is entered at 14:00 on 3 February, and the laboratory specimen is collected at 08:30 that morning but the result becomes available at 11:00.
The diagnosis code’s entry time is after time zero. Unless a reliable clinical event time or prior record establishes that the diagnosis was known before 09:00, it cannot be assumed to describe baseline information. The code may reflect a later assessment, and the database provides no basis for moving it backward in time.
The specimen was collected before time zero, but its result became available afterward. If the protocol’s baseline variable means “laboratory result available to the treating clinician at assignment,” this result does not qualify. If the protocol instead defines baseline physiology by specimen collection before assignment, it may qualify, but the protocol must state that estimand-relevant definition and retain the result-availability flag. These are not interchangeable rules.
The conclusion is not that either value is clinically useless. It is that neither may be silently assigned baseline status without a timing rule. The example also shows why event date, result date, and entry date should be separate fields whenever the source data permit.
Build and Audit the Information Inventory
A compact inventory makes the requirements reviewable. It should link each concept to a specific variable definition and timing rule. Separate “required for the intended baseline description or adjustment” from “available in the database”; a protocol requirement does not establish that the source data contain a valid measurement.
| Protocol item | Specification to record | Timing concern to flag |
|---|---|---|
| Prior diagnosis | Code set or clinical definition; lookback interval; rule for repeated codes | Diagnosis date may differ from entry date; absence of a code is not necessarily absence of disease |
| Laboratory value | Test, units, valid range, measurement window, and selection among repeats | Specimen, result-availability, and database-entry times may differ |
| Medication history | Medication definition; exposure window; rule for active use at time zero | Dispensing or order date may not equal actual use; records may straddle time zero |
| Clinical utilization | Encounter types and count or summary period | An encounter after assignment may reflect treatment response or follow-up |
| Demographic or fixed characteristic | Source field and reference date | Some fields are updated retrospectively or may have uncertain effective dates |
Useful diagnostics include checking whether timestamps have sufficient precision to order measurement and assignment; whether values fall outside the specified window; whether units and coding definitions are consistent; and how often records have missing or ambiguous dates. Compare the frequency of measurement across treatment groups as well: a variable may appear more often in one group because of different clinical monitoring, not because the underlying characteristic is more common. Such checks reveal problems with the information definition and its capture, though they do not prove that the remaining baseline set is sufficient.
The main alternatives are not universally right or wrong. A narrow recent window may represent current status better but leave more people without a qualifying value. A longer window increases coverage but may use stale information. Selecting the latest value is transparent and often clinically interpretable; averaging repeated values answers a different question and can give more weight to people measured more often. The protocol should choose according to the clinical meaning and the intended trial, then report the choice and its consequences. It should not search among windows or summaries for whichever gives a preferred treatment estimate.
Common Pitfalls and Limitations
- Calling every pre-follow-up record “baseline.” A record can be dated before follow-up but still fall outside the prespecified window or have an uncertain event time. Apply the variable-specific rule.
- Using calendar dates as if they gave an exact order. Same-day records without times cannot establish whether a measurement preceded assignment. Mark that ambiguity explicitly.
- Confusing no record with a negative history. An absent diagnosis or measurement may indicate incomplete capture. Define how absence is represented rather than coding it automatically as “no.”
- Allowing a window to cross time zero. A rolling window that includes follow-up measurements can place post-assignment information in a purported baseline covariate. Use an explicit endpoint and inspect it.
- Assuming measured covariates eliminate confounding. Precise timing and definitions improve protocol clarity; they do not establish that all relevant causes of treatment and outcome are observed or correctly measured.
Prespecifying requirements also cannot repair inaccurate clinical timestamps, inconsistent coding, or unobserved history. It makes those limitations visible and gives the analyst a defensible basis for deciding whether a record qualifies as baseline, is ambiguous, or must be left out of that baseline definition.
Check Your Understanding
Answer each question using the stated interval and timing rules.
- A 90-day lookback ends at 12:00 on 1 December. State the interval and say whether a record exactly at its left endpoint qualifies.
- Time zero is 08:00 on 20 May. A lab result is collected at 07:30 and becomes available at 09:00. Does it qualify if baseline means information available to the clinician at assignment?
- Why is a diagnosis code entered after time zero not automatically a baseline diagnosis, even if the illness may have started earlier?
- A protocol uses the most recent valid result in the 30 days before time zero. A participant has results 35 days and 4 days before time zero. Which is selected, and why?
Solutions. (1) The interval is \([t_0-90\text{ days},t_0)\), or \([2\text{ September at 12:00},1\text{ December at 12:00})\) for elapsed 24-hour days. The left endpoint is included, so a record exactly there qualifies. (2) No. Although collection preceded assignment, the result was not available until 09:00, after the 08:00 assignment; it fails the stated availability definition. (3) The entry time does not establish when the illness began or when it was known. Without reliable pre-time-zero evidence, treating the code as baseline would infer timing the record does not provide. (4) The result from 4 days before time zero is selected because it is within the 30-day window and is more recent; the result from 35 days before time zero is outside the window.