In the early 2000s, a study published in the Annals of Internal Medicine reported that Oscar winners had a longer life expectancy compared to their less successful counterparts. The conclusion was that an elevated social status conferred a survival advantage of about four years to the winners. This study received wide attention in the general press, even making its way to an Oscar winner’s acceptance speech.
Desirable as they were, these results were not valid. The problem was that winners had an unfair advantage over non-winners from the beginning of time. How? Subjects in the study were first classified as winners or non-winners and then observed from their day of birth. Therefore, unlike non-winners, winners had to live long enough to win.
Such illusive findings—a product of immortal time bias—are also common in non-randomized studies of health care interventions. Awareness about this bias in pharmacoepidemiology was first raised when there was a growth in the use of real-world data (RWD).
Early studies using RWD reported certain medications like inhaled corticosteroids being effective at reducing all-cause mortality. These findings invited further scrutiny at the time due to a lack of plausible mechanism of action by which these medications can protect one from all-cause mortality. They were found to be prey to immortal time bias.
Immortal time refers to a period of time during observation when a subject cannot experience an outcome of interest. In the Oscar winners study, time from birth until the receipt of an Oscar is immortal for winners. Further, when treatment assignment is based on an event that occurs in the follow-up, those in the treated group by definition have follow-up long enough for the event to occur. Exclusion or misclassification of immortal time produces biased findings.
Researchers must be aware of when their studies may be influenced by immortal time bias and how to eliminate it. Central to the problem with immortal time is the specification of ‘time zero’—or put simply, when does the clock start?
Failure to align the start of follow-up can result in bias. Examples of study design choices that have the potential to introduce immortal time bias include: definition of exposure based on events occurring after cohort entry, hierarchical definition of treatment groups, or drawing contrasts between treated and untreated subjects. Such choices may be unavoidable in certain circumstances but care must be taken to align the start of follow-up and appropriately classify person-time.
Going back to the study on Oscar winners, a reanalysis using appropriate methods found no significant survival advantage. Alas, the problem of immortal time bias however remains prevalent in the literature as evidenced by recent appraisals demonstrating that some conclusions drawn about medications in cardiometabolic, rheumatologic and gastrointestinal therapeutic areas, among others, may have been fallible.
No wonder immortal time bias has been considered the fallacy that does not die.