Find Outcome-Free Time Hidden in Exposure Definitions

Search for information unavailable at baseline

An exposure definition can require a person to remain alive or event-free long enough to qualify for the exposed group. If earlier follow-up is then attributed to that group, the comparison may include a period during which the defining event could not yet have occurred. Review exposure classification and person-time allocation together.

Hernán and colleagues’ time-zero paper explains how explicit trial specification helps prevent these design errors. The problem is not repaired by adding more baseline covariates to a model whose exposure groups or follow-up periods were constructed incorrectly.

Trace several synthetic records

Create a timeline containing eligibility, first treatment, any required repeat treatment, outcome, loss to follow-up and administrative end. Ask what information would have been available at each date. Flag labels such as ever treated, completed treatment or received a minimum number of doses when they are assigned at the beginning of observation.

Check whether people with early events are excluded from one group but retained in the other. Look for inconsistent lookback requirements and different opportunities to accumulate exposure. Preserve the original date variables so the classification can be reconstructed from events rather than from a final binary label alone.

Examine a hypothetical delay

Suppose a person becomes eligible on day 0, starts treatment on day 30 and survives to day 100. An analysis labels them treated from day 0 because they eventually initiated treatment. If initiation requires survival to day 30, the first 30 days are guaranteed survival time among people classified this way.

A different person who dies on day 10 can never satisfy that future initiation rule and may be placed in the untreated group. Comparing the resulting groups from day 0 can favor treatment even without a true benefit. The example isolates the classification problem; it does not specify a complete correction for every treatment question.

Choose a design that answers the intended question

Depending on the scientific target, an appropriate solution may involve a baseline initiation comparison, a time-varying exposure analysis, a landmark design or a carefully specified grace-period emulation. These approaches answer different questions and bring different assumptions. Do not select one solely because it removes a warning or changes the estimate favorably.

After revising the design, repeat the synthetic-timeline tests and reconcile total follow-up across groups. Report the start of follow-up, exposure-assignment rule and handling of events during any initiation window. A clear timing diagram and a few tested edge cases can reveal more about this bias than a long list of adjusted covariates.

Source verification: 6 November 2026. Examples are hypothetical.