Immortal Time Bias

Immortal time bias is a form of selection bias in observational studies that occurs when the period between cohort entry and the start of exposure includes a span of time during which the outcome of interest could not have occurred. Misclassifying or ignoring this "immortal" period inflates the apparent benefit of the exposure, producing spurious protective associations.

Immortal time bias is one of the most common and consequential methodological flaws in pharmacoepidemiological and outcomes research. Studies affected by this bias typically overestimate the benefits of treatments, leading to misleading conclusions that can influence prescribing practices and policy decisions. Peer reviewers, journal editors, and HTA assessors increasingly look for immortal time bias when evaluating observational evidence, and its presence can be grounds for discounting or rejecting a study's findings. Understanding this bias is essential for any researcher conducting observational analyses of treatment effects.

The mechanism underlying immortal time bias involves the incorrect handling of time between cohort entry and the moment a patient receives the treatment or exposure of interest. Consider a study examining whether statin use after a heart attack reduces mortality. Patients are classified as "statin users" if they fill a statin prescription at any point after discharge. However, to fill a prescription, the patient must survive long enough to do so. The period between discharge and the prescription fill date is "immortal time" because death during this period would have resulted in the patient being classified in the unexposed group instead. If this immortal time is attributed to the statin group (as person time at risk), or simply dropped from the statin group while the comparison group's follow up still starts at discharge, the comparison is biased in the statin group's favor.

Three analytic strategies can address immortal time bias. Time dependent exposure classification treats exposure status as a time varying covariate, so that each patient contributes person time to the unexposed group until the moment they receive the treatment and to the exposed group thereafter. Landmark analysis defines a fixed time point after cohort entry (for example, 90 days after discharge) and classifies patients based on whether they received the exposure by that landmark, excluding patients who died or were lost to follow up before the landmark. Matching on the timing of treatment initiation, as used in the prevalent new user design, pairs exposed and comparison individuals who were alive, under follow up and at similar points in their disease trajectory when exposure began, and starts follow up for both at that time.

Awareness of immortal time bias is necessary whenever an observational study classifies patients into exposed and unexposed groups based on an event that occurs after cohort entry, such as filling a prescription, receiving a procedure, or being admitted to a program. It is particularly relevant in database studies using administrative claims, electronic health records, or registry data. If exposure is defined at baseline (for example, treatment assigned at randomization in a trial), immortal time bias does not apply.

A frequent mistake is defining exposure based on any use of a treatment during the follow up period and then counting survival from cohort entry, which guarantees that exposed patients have survived at least until the time of their first exposure. A second error is conducting a landmark analysis but choosing an arbitrary landmark time without clinical justification, which can either introduce residual bias (if the landmark is too early) or exclude too many patients (if it is too late). Some researchers attempt to address immortal time bias by excluding the immortal period from both groups, but this approach can introduce other forms of selection bias if the reasons for exclusion differ between groups. Finally, failing to conduct sensitivity analyses that vary the exposure definition and analytic approach makes it impossible to assess the influence of immortal time on the results.

A database study examines whether enrollment in a cardiac rehabilitation program after myocardial infarction reduces one year mortality. Patients are classified as "enrolled" if they attend at least one session at any time during the year. The naive analysis reports a 40% reduction in mortality among enrolled patients. However, the median time from discharge to first session is 45 days: each enrolled patient was guaranteed to survive until their first session, and half of them for at least 45 days. When the researchers reanalyze the data using time dependent exposure classification, the apparent mortality reduction drops to 15%, and after applying a 90 day landmark analysis, it drops further to 12%, demonstrating that much of the original estimate was attributable to immortal time bias.