Dr. Manouchehr Hessabi
← All writing
7 min readresearch methods · epidemiology · observational studies

Immortal time bias: when one group gets free survival

Immortal time bias explained: how counting follow-up before an exposure begins can make a drug or an honor look protective, and how epidemiologists avoid it.

By Manouchehr Hessabi, MD, MPH

Immortal time is a stretch of follow-up in a study during which, because of how exposure was defined, the outcome being studied could not have happened. When that time is credited to the exposed group, the exposure can look protective even if it does nothing at all. The term and its formal treatment are set out in a widely cited review by Samy Suissa in the American Journal of Epidemiology (Suissa, 2008).

What do cohort study, follow-up and time zero mean?

A few terms first. A cohort study (compared with other designs in the explainer on case-control versus cohort study design) follows a group of people forward through time and compares how often an outcome, such as death or hospitalization, occurs among people with and without some exposure. Follow-up is the period during which each person is observed. Time zero is the moment the clock starts for each person: the point from which their outcome-free time is counted.

Immortal time bias is, at bottom, a problem with time zero. If the clock starts for everyone at the same moment, but some people only become "exposed" later, then the period between time zero and their exposure is time they were guaranteed to survive. They had to be alive, and free of the outcome, to become exposed in the first place.

How can an Oscar make an actor live longer?

The cleanest published illustration comes from Hollywood. In 2001, Redelmeier and Singh reported in the Annals of Internal Medicine that Academy Award-winning actors and actresses lived almost four years longer than their less successful peers, a statistically significant difference.

In 2006, Marie-Pierre Sylvestre, Ella Huszti and James Hanley reanalysed the evidence in the same journal (Sylvestre, Huszti and Hanley, 2006). They pointed out that the statistical method in the original study gave winners an unfair advantage: it credited an Oscar winner's years of life before winning toward survival after winning. When they reanalysed the data with methods that avoided this immortal time, the survival advantage was closer to one year and was not statistically significant.

The intuition is simple once it is named. To win an Oscar, an actor must be alive on the night of the ceremony. An actor who wins at 60 contributed decades of guaranteed survival to the "winner" group, years during which that person could not have died and still become a winner. An actor who died at 45 could never have entered that group. Classifying people by something that happens later in life, then counting their whole lives as if they belonged to that group from the start, quietly tilts the comparison.

The 2006 authors were explicit that the problem is not limited to awards. They noted that the same type of bias can, and often does, occur in nonexperimental studies of the life-extending or time-extending benefits of medical interventions.

Why does this show up in drug studies?

According to Suissa's review, bias from immortal time was first identified in the 1970s, in cohort studies of the survival benefit of heart transplantation. The structure is easy to see: a patient has to survive the wait for a donor heart to be counted as a transplant recipient, so the waiting time is guaranteed survival for that group.

The same review describes how the bias resurfaced in pharmaco-epidemiology, the study of how medicines perform in large populations, usually using routinely collected data such as prescription and hospital records. Several observational studies had reported that various medications were extremely effective at reducing illness and death. Suissa showed that these studies, while using different cohort designs, all involved some form of immortal time and the corresponding bias.

A common version looks like this:

  • A researcher assembles a cohort of patients from the date of a diagnosis or hospital discharge. That date is time zero.
  • Anyone who fills a prescription for the drug of interest at any point during follow-up is classified as a "user" from time zero onward.
  • A patient whose first prescription comes eight months after time zero is therefore counted as exposed for those eight months, even though the patient was not taking the drug, and had to survive those months in order to fill the prescription at all.

The comparison group, meanwhile, includes people who died or had the outcome early, before they ever had the chance to start the drug. The result can be a striking apparent benefit that the drug did nothing to produce.

For readers who want a worked example in a specific clinical setting, Lévesque, Hanley, Kezouh and Suissa used statins and the progression of diabetes to demonstrate the problem in the BMJ (Lévesque and colleagues, 2010). The point of that paper, as its title states, is the method: how immortal time arises in a cohort study and what it does to the estimate. It is not a verdict on whether statins work.

How big can the bias get?

There is no single answer, because the size depends on the study. What Suissa's analysis did establish is how it behaves. For time-based, event-based and exposure-based ways of defining a cohort, the bias in the rate ratio (the outcome rate in the exposed group divided by the rate in the unexposed group) grows in proportion to the duration of the immortal time (Suissa, 2008). The more guaranteed survival a design hands the exposed group, the more protective the exposure appears.

The same paper showed that the bias is more pronounced when the hazard of the outcome decreases over time. The hazard is, roughly, the instantaneous risk of the outcome at a given moment among people who have not yet had it. Many outcomes are most likely soon after a diagnosis or a hospital stay and become less likely later. In that situation, the early months are exactly when the most events occur, so misplacing those months does the most damage to the comparison.

The Oscar example shows the practical stakes: an apparent advantage of almost four years shrank to about one year, and the difference was no longer statistically significant, once the time was counted correctly.

How do epidemiologists prevent it?

The core fix is to count each person's time in the right place.

Classify exposure as it actually happens over time. Rather than labeling a person a "user" from time zero because of something that happens later, the analysis can treat exposure as time-varying: a person's follow-up before the first prescription counts as unexposed time, and only the time after it counts as exposed. Suissa's review frames the bias as coming from immortal time that is misclassified or excluded; assigning that time to the period it actually belongs to removes the guaranteed survival from the exposed group.

Align time zero, eligibility and treatment assignment, as a trial would. In a randomized trial these three things happen at the same moment by design: a person is found eligible, is assigned to a treatment, and starts follow-up, all at once. Hernán and colleagues describe immortal time bias, and several related errors, as the result of a failure to align the start of follow-up, the specification of eligibility, and treatment assignment (Hernán and colleagues, 2016). Their proposed remedy is to write down the target trial: the randomized trial the observational study is trying to emulate, including who would be eligible, what strategies would be compared and when follow-up would begin. Many analyses of observational data are, in their words, attempts to emulate such a trial, and specifying it explicitly makes misaligned time much harder to overlook.

The target trial framework does not make observational data equivalent to randomization. Confounding, the distortion caused by a third factor linked to both exposure and outcome, still has to be addressed separately, as does selection bias, which arises from who enters or goes missing from a sample rather than from how time is counted. What the framework does is prevent a class of errors that the researchers introduced through the design itself.

What should a careful reader check?

A study's headline rarely mentions when the clock started. A few questions help:

  • When did follow-up begin for each group? If the exposed group's clock starts before the exposure did, look for immortal time.
  • Was anyone classified by something that happened after time zero? Becoming a winner, receiving a transplant, or filling a later prescription are all events a person must survive to reach.
  • Is the benefit implausibly large? Suissa's review noted that the studies affected by this bias often reported medications to be extremely effective. A very large protective effect in observational data is a reason to read the methods closely.
  • Does the paper describe exposure as time-varying, or name the trial it emulates? Either is a sign the authors have thought about time zero.

Recognising immortal time bias is a skill for reading evidence, not a judgment on any particular treatment. Some exposures studied this way are genuinely beneficial; the point is that a flawed design cannot tell the difference.

Readers interested in how study design shapes what research can show will find related work among Dr. Hessabi's peer-reviewed publications.

About the author. Dr. Manouchehr Hessabi is a physician-epidemiologist and Senior Research Scientist at the BERD core of UTHealth Houston's Center for Clinical and Translational Sciences. See his peer-reviewed publications or research programs.