Part of our Statistical paradoxes series · Immortal time bias
In May 2001, two researchers, Donald Redelmeier and Sheldon Singh, published a study in the Annals of Internal Medicine with the kind of finding that writes its own headline. They had taken every actor and actress ever nominated for a leading or supporting Academy Award — 762 performers — and paired each one with a same-sex cast member from the same film, born in the same era. Then they compared how long everyone lived. The Oscar winners came out ahead: a life expectancy of 79.7 years against 75.8 for their less-celebrated peers. Nearly four extra years of life, with a p-value of 0.003. The press loved it, and the explanation practically suggested itself: status, recognition, and a sense of control over your life are good for your health.
Five years later, three statisticians at McGill — Marie-Pierre Sylvestre, Ella Huszti, and James Hanley — went back to the same question and found that most of those four years came from how the clock was started. Analyzed in a way that only gave winners credit for the years they had actually lived as winners, the advantage shrank to roughly one year, and it was no longer statistically significant. The actors hadn't changed. The bookkeeping had.
The original headline comparison classified each performer by whether they ever won an Oscar, and measured survival from birth. That sounds harmless until you notice what it quietly assumes. To be counted as a winner, an actor has to survive long enough to win. Anyone who died young, before their winning performance, could only ever be counted in the other group. So every year a future winner lived before the ceremony was guaranteed, risk-free survival time, and it was all credited to "winning," even though the Oscar hadn't happened yet.
Epidemiologists call those guaranteed years immortal time: a stretch of follow-up during which, by the way the groups were defined, the outcome could not occur in one of them. Immortal time bias always flatters the group whose membership depends on something happening later. It doesn't need a wrong dataset or a sloppy calculation. Every number can be correct and the comparison can still be rigged, purely by when you start counting.
The fix is conceptually simple. Treat "winner" as something a person becomes at a specific date, not something they always were. Before the ceremony, a future winner's years count in the not-yet-a-winner column. From the ceremony on, they count as a winner. That is what the McGill reanalysis did, and most of the advantage went away.
It's worth being honest that the debate didn't end there. Redelmeier and Singh returned to the question in 2022 with longer follow-up and a larger sample, and reported a survival advantage for winners again. Whether Oscars are good for you is still argued over. The methodological lesson isn't: if the groups are defined by something that happens partway through follow-up, the analysis has to respect when it happened, or the headline number is partly an artifact.
Whenever someone compares outcomes between people who did something and people who didn't, ask one question: when does the clock start for each group, and could anyone in the "did it" group have had the outcome before they did it? If the answer to the second half is no, there's immortal time in the comparison. The remedy is to count each person in the group they actually belonged to at each point in time — or to start everyone's clock at a common moment, like the date of exposure or a landmark date — rather than sorting people by how their story eventually turned out.
If your study, evaluation, or product analysis compares people by something that happened partway through follow-up, talk to us before the head start becomes your headline.
One email whenever we publish something new. No spam, unsubscribe anytime.
{{ subscribeErrorMsg }}