Why Covariate Selection and Timing Need to Be Reported
A reported estimate from an observational comparison is difficult to assess if readers cannot tell which variables were adjusted for, how those variables were constructed, or whether they were measured before treatment began. A variable name such as “comorbidity” is not enough: it could mean a prior diagnosis, a diagnosis recorded at any point during follow-up, or a summary of several different sources. Those choices define different variables and can change the meaning of the analysis.
A transparent report connects each adjustment variable to the study question and to the timeline. It identifies the variable’s operational definition, data source, measurement window, and subject-matter rationale. It also states the treatment start or index time that separates baseline information from information that could have been affected by treatment. This documentation lets readers judge whether the selected variables plausibly address confounding without using information from after treatment began.
What to Document for Each Variable
For each variable in the intended adjustment set, report enough detail for a reader to understand what value was assigned to each participant. In many studies, this means reporting the following elements together rather than scattering them across several sections:
- Definition and construction: what characteristic is represented, including relevant categories, thresholds, lookback rules, or summary procedures.
- Source: which records or instruments supplied the information, such as enrollment forms, claims, electronic health records, or a named assessment.
- Measurement window: the calendar dates or interval relative to time zero during which the information was collected or used.
- Subject-matter rationale: why the variable was considered relevant to treatment assignment, the outcome, or both, and therefore to the intended adjustment.
- Temporal status: whether the value represents information available before treatment began, and how that was determined.
Use a precise anchor. If treatment can begin on different calendar dates for different participants, define the window relative to each participant’s time zero, not just by saying “in the year before the study.” Distinguish a variable’s reference period from the date on which a record was entered. For example, a visit entered after time zero might document a condition that existed earlier, but the report must explain how its pre-treatment status was established rather than silently treating every later entry as baseline information.
The rationale should be substantive, not merely that a variable was available or associated with an outcome in the observed sample. Explain the expected role of each variable based on clinical, behavioral, or other subject-matter knowledge. A variable may be included because it plausibly affects treatment choice and outcome, or because it is a well-established prognostic factor relevant to the comparison. The report should not imply that an observed association by itself proves a variable must be adjusted for.
Worked Example: Reporting a Claims-Based Baseline Variable
Suppose a hypothetical study compares adults who initiate a medication on January 1, 2025, with eligible adults who do not initiate it on that date. Investigators plan to adjust for prior myocardial infarction. They define time zero as January 1, 2025, and search claims during the 365 days before time zero.
The inclusive window is calculated as January 2, 2024, through December 31, 2024: January 2–31 contributes 30 days, February–December contributes 335 days, and \(30+335=365\). The report should say that prior myocardial infarction is an indicator for at least one qualifying diagnosis claim in that window, identify the claims source, and specify the diagnosis rule used to qualify a claim.
The subject-matter rationale might be that a prior myocardial infarction could influence the decision to initiate the medication and also predict the study outcome. That is a reasoned basis for considering the variable, not proof that all confounding has been removed. The dates also make the temporal claim checkable: the qualifying claims precede January 1, 2025. A complete methods description would state whether the same definition and window applied to both groups.
Why Time Zero Protects the Baseline Adjustment Set
The timing issue can be stated compactly. Let \(t_0\) denote time zero, and let \(X_j(t_0)\) be the value of adjustment variable \(j\) defined using information available before \(t_0\). For a lookback of \(L\) days, one common rule is to use records in the interval from \(t_0-L\) through the day before \(t_0\). This rule makes the boundary explicit:
Here \(g_j\) is the stated construction rule, such as “one or more qualifying diagnoses.” The notation is a reporting aid: it emphasizes that the variable is defined from a specified set of records, not from all records ever observed.
The justification for placing the window before time zero is temporal. A value recorded after treatment begins may reflect a consequence of treatment, a response to early symptoms, or a change in healthcare contact caused by treatment. In any of those cases it is not simply a pre-treatment characteristic. Restricting a baseline variable’s construction to pre-time-zero information prevents those later records from being inadvertently used as if they were baseline.
This restriction is necessary for a clean baseline definition, but it is not sufficient to guarantee an unbiased causal estimate. A pre-treatment variable is not automatically a confounder, and recording it before treatment does not establish that it was measured without error. The causal rationale and the timing rule answer different questions: the rationale explains why adjustment was considered; the timing rule explains which information was allowed to define the variable.
Worked Example: Checking a 90-Day Laboratory Window
In a hypothetical comparison with time zero on March 1, 2025, investigators define baseline kidney function using the latest serum creatinine result in the 90 days before treatment. They use laboratory records dated from December 1, 2024, through February 28, 2025. This is 90 days: December contributes 31 days, January 31, and February 28, so \(31+31+28=90\).
Suppose a participant has results dated November 20, February 10, and March 3. The November 20 result is outside the stated window. The February 10 result is inside it, and the March 3 result is after time zero. Under the stated rule, the baseline value is the February 10 result. If no eligible result exists in the window, the report should not imply that a later result was used as a baseline value; the handling of absent measurements must be described separately.
The methods should identify the laboratory source, the measurement and units, the “latest result” rule, and the exact window. A rationale might explain that kidney function could be relevant to treatment choice and to the outcome. Readers can then distinguish a measurement chosen for that rationale from one selected simply because it was the most recently available at any point in the records.
Recognizing Post-Treatment Information
A variable measured after treatment begins is not automatically unsuitable for every possible analysis, but it should not be presented as a baseline adjustment variable without a clear justification. If treatment can affect the variable, adjusting for it may change the question being answered and can introduce bias. For instance, conditioning on a post-treatment variable can create bias if that variable is affected by treatment and also shares causes with the outcome. The direction and size of such bias are not determined merely by knowing that the measurement occurred later.
Use causal language precisely. A variable lies on a pathway from treatment to outcome only if it is on a directed causal path that connects treatment to outcome. For example, in the chain treatment → symptoms → dose adjustment → hospitalization, dose adjustment is on that pathway only if the specified links include an effect of dose adjustment on hospitalization. If that link has not been established or assumed, do not claim that adjusting for dose adjustment blocks that mediated pathway. It remains important to flag the variable as post-treatment and explain why conditioning on it may be problematic, without asserting an unsupported pathway.
When a variable’s timing is ambiguous, report the rule that resolves the ambiguity. Say whether the date refers to the service date, specimen collection, encounter date, or data-entry date where those differ. Explain whether information recorded after time zero was excluded, whether historical information was accepted only when it documented a pre-time-zero event, and how the same rule was applied across treatment groups.
Worked Example: Avoiding an Unsupported Mediation Claim
Consider a hypothetical study of a treatment initiated at week 0 and hospitalization during the next 12 weeks. Symptoms are assessed at week 2, and clinicians may adjust the dose at week 3. Both observations occur after treatment begins. The investigators should report their dates and should not include them in the baseline adjustment set as though they preceded treatment.
Suppose the proposed causal relationships are treatment → symptoms and symptoms → dose adjustment. Those links alone do not establish that dose adjustment lies on a path to hospitalization. To make that claim, a further link from dose adjustment to hospitalization must be part of the assumed causal structure; for example, treatment → symptoms → dose adjustment → hospitalization. Without that link, the correct report is that dose adjustment is post-treatment and that conditioning on it may introduce bias, not that it blocks a treatment-to-hospitalization pathway.
The reporting decision is therefore to define baseline covariates using information before week 0, identify symptoms and dose adjustment as later measurements, and state the reason for any separate analysis that uses them. This preserves the distinction between a timing fact that can be checked and a causal claim that requires an explicit rationale.
Diagnostics, Alternatives, and Limitations
Before finalizing a report, check the covariate list against the study timeline and data extraction rules. Warning signs include a covariate described only by a vague label, a lookback window with no reference point, a baseline value taken from a record after treatment initiation, or different rules for treated and comparator participants without an explanation. Also check that the stated rationale matches the variable actually constructed. A “prior diagnosis” derived from any diagnosis during follow-up is not a prior diagnosis.
Subject-matter selection and automated or significance-based selection are not interchangeable reporting strategies. A prespecified, subject-matter rationale makes the intended adjustment set interpretable and allows readers to assess whether it is plausible. Selecting variables because they differ between treatment groups or have small outcome-association p-values can omit relevant variables and make the analysis dependent on the observed sample. Automated feature selection can be useful in some settings, but it does not remove the need to explain the included variables, define their timing, and assess whether post-treatment information entered the candidate set.
A detailed covariate table is often the clearest presentation. It can give the variable, operational definition, source, pre-time-zero window, and rationale in adjacent columns. The methods text should still define the time-zero convention and explain any shared rule that applies to multiple variables. This reporting does not prove that confounding is absent: unmeasured or poorly measured causes can remain, and choices about variable definitions can still affect the analysis. Its value is that readers can see and evaluate those choices instead of having to infer them.
| Reporting item | Question a reader should be able to answer |
|---|---|
| Definition | What exactly counts as the variable being present or having a particular value? |
| Source | Which record system, assessment, or instrument supplied the information? |
| Timing | What dates or relative window define the value, and where is time zero? |
| Rationale | Why was this variable considered relevant to treatment assignment or outcome? |
| Post-treatment check | Could any included information have been observed after treatment began? |
Check Your Understanding
For each question, give your answer before reading the solution that follows it.
- A study defines time zero as June 1 and describes a covariate as “any diagnosis recorded during the study.” What reporting problem should be addressed?
Solution. The definition does not distinguish pre-treatment from post-treatment records. Specify a pre-time-zero measurement window and the qualifying diagnosis rule, or explain explicitly why later information is used in a different analysis. A baseline variable should not silently include diagnoses recorded after June 1.
- A 30-day lookback ends on the day before an October 15 time zero. What dates should the window cover, and what should the report state?
Solution. The 30 days are September 15 through October 14, inclusive: September 15–30 contributes 16 days and October 1–14 contributes 14 days, for \(16+14=30\). Report those relative dates or the equivalent rule, the data source, and how the variable is constructed from eligible records.
- Investigators include baseline disease status because it may influence treatment choice and predict the outcome. Does this rationale prove that the analysis is free of confounding?
Solution. No. It gives a subject-matter rationale for considering the variable, but it does not establish that all relevant confounders were measured or correctly controlled. Report the rationale and the operational definition without claiming that it guarantees an unbiased estimate.
- Treatment affects a post-treatment symptom, and the symptom affects a medication change. May the report say that adjusting for the medication change blocks a pathway to the outcome?
Solution. Not from those links alone. The medication change lies on a treatment-to-outcome pathway only if it also affects the outcome in the specified causal structure. Without that link, report that the medication change is post-treatment and that conditioning on it may be problematic, but do not claim it blocks a mediated pathway.