Medical research methods · Cohort timing
Time zero in retrospective cohort studies: aligning eligibility, exposure and follow-up
Turn unclear timing definitions into a plan paragraph for discussion Requested outputs / delivery goal: A study-timing definitions table, a time-zero plan paragraph and a question list.
Define the question first, then who becomes eligible when, when exposure is classified and when outcomes start being observed. Do not retrospectively turn exposure occurring later in follow-up into a baseline group. A more elaborate model does not replace a check of the time structure.
This guide concerns medical cohorts using existing records, not treatment advice or automatic causal inference. “Retrospective” does not specify actual collection, measurement and follow-up dates. STROBE items 5 and 6(a) require relevant periods, eligibility and follow-up methods to be reported.
The feasibility guide asks whether a study can proceed; the participant-flow guide reconciles counts and denominators. This page addresses timing. Treatment-strategy comparisons can use the target-trial framework to check alignment, but not every descriptive, association or prediction cohort is a target-trial emulation.
A concrete task with RigorMed
Turn unclear timing definitions into a plan paragraph for discussion
These are suggested task inputs and outputs to request, not a new execution, a customer outcome or a guarantee of completion.
Materials to provide: The research question, eligibility, exposure and follow-up definitions, field documentation and lawful aggregate illustrations.
Requested outputs / delivery goal: A study-timing definitions table, a time-zero plan paragraph and a question list.
An example request
Map eligibility, exposure classification, baseline measurement and follow-up origin individually to dates and evidence. List risks of grouping with later information, then draft a time-zero definition and questions requiring further evidence.
What still needs review
Time alignment does not certify causality. Do not backdate future exposures, invent follow-up or let a model replace statistical design decisions.
Separate four kinds of time
Different questions may require different origins. The rule should be reproducible and relevant, not necessarily a common calendar date for everyone.
| Time | Specify at least | Common misuse |
|---|---|---|
| Eligibility | When all criteria are met and which existing records establish them | Selecting the baseline population by future survival or treatment completion |
| Exposure classification | Measurement, initiation and changes; when classification information becomes available | Treating any exposure during follow-up as exposure from baseline |
| Baseline variables | Measurement window relative to the origin and value-selection rule | Calling post-exposure or post-outcome information baseline |
| Outcome and follow-up | Origin, event, last observable time and administrative end | Using database closure as everyone’s last observed date |
Prescription, dispensing, actual use and record-entry dates need not coincide. Verify field meaning and recording delays. Retain unknown dates rather than asking AI to invent them.
Five checks before analysis
Freeze question and unit
Specify population, exposure or prediction task, outcome and horizon. Distinguish people, encounters and repeated measurements. Decide the estimand before seeing results, not the origin after seeing them.
Define eligibility and time zero
Provide each person’s eligibility rule and available records. For simple treatment-initiation comparisons, check alignment of eligibility, strategy classification and follow-up start. Grace periods or dynamic strategies need their own suitable design and analysis, not a mechanical date shift.
Check what was knowable then
Mark measurement time, recording time and relation to baseline for each field. Prediction uses information actually available at prediction time; treatment comparisons must not classify baseline groups from future exposure or survival.
Specify follow-up and exit rules
Predefine event ascertainment, loss to follow-up, administrative closure and the role of competing events such as death. Censoring does not establish absence of an outcome or automatically satisfy an independent-censoring assumption.
Review a small timeline and retain changes
Use lawful public or synthetic materials to test same-day ordering, repeated eligibility, exposure changes and missing dates. Actual patient records belong only in authorized private controlled processing. Resolve conflicts through documented protocol changes.
Synthetic timeline: starting on day 10 is not exposure on day 0
Suppose eligibility is met on day 0. A begins an exposure on day 10; B dies on day 5. If baseline groups are defined by whether exposure ever occurs during follow-up, A must survive until day 10 to enter the exposed group, whereas B cannot. Counting A’s earlier time as exposed can introduce immortal time and related selection or classification bias.
These dates are synthetic, with no real study, treatment-effect numbers or statistical calculation. Reassigning time with one click is not a general repair. Time-varying exposure, landmark and target-trial designs concern different populations, estimands and assumptions. Appropriate methodological review must determine the plan before analysis and state the revised question.
Aligned time is not a causal certificate
Confounding, selection, measurement error, missingness, incomplete follow-up and transportability remain possible even with aligned timing. Exposure preceding outcome alone does not justify “causes”; a completed STROBE checklist does not certify unbiased estimates or ethical compliance.
If dates cannot be verified, event order is unresolved or eligibility depends on future information, mark what is not identifiable and revise the question or source. Do not fabricate dates, discard early events to improve curves or transfer a treatment-comparison design directly to cross-sectional data.
Checklist before statistical work
Keep definitions and evidence, not just ticks:
- Question, study unit and horizon are explicit.
- Eligibility, exposure and follow-up-origin rules are reproducible.
- Baseline windows exclude information not knowable at that time.
- Events, loss to follow-up, administrative end and competing events have predefined rules.
- Missing dates, same-day ordering, exposure changes and protocol revisions are recorded.
This page neither executes statistics nor validates patient data, and guarantees neither significance, causality nor submission readiness.
Common questions
Must time zero be the same calendar day for everyone?
No. It is usually defined relative to individual eligibility or a specified initiating event. Still report calendar periods for recruitment, exposure, follow-up and collection.
Will more covariates or a propensity score remove immortal time?
Not as a guarantee. Incorrect grouping or allocation of time is a design problem. Correct the study definition first, then choose methods and assumptions matching the estimand.
Can first record-entry date replace a missing event date?
Not by default. Only a verified field representing the required event time, with acceptable error and disclosed limitations, can potentially serve that purpose. Otherwise retain the unknown.
Verified sources
The official sources below were checked on 2026-10-12 / RigorMed: 2026-10-13. Access conditions and update dates are rechecked before publication.
Start with a specific task
Turn unclear timing definitions into a plan paragraph for discussion
Bring legally usable materials and this page's example request to the workspace. Check the pre-execution material and credit notices before confirming, then review the resulting outputs.