Korva is a scroll of the internet’s best ideas.
Immortal time bias inflates drug benefits
Immortal time bias
Picture a hospital database comparing people who took a drug with people who didn't. Many studies counted the time before a patient started the drug as part of the comparison group's follow-up, but counted only the time on the drug for the treated group. To be prescribed the drug at all, a person must already have lived to that point, so the treated group is immortal until the day treatment begins. They can't have died, or they wouldn't be in the group.
The bias affected many studies claiming metformin protects against cancer. When randomized trials followed, the conclusion for both metformin and vitamin D across various conditions was mostly that they don't work.
Confounding by selection
A second distortion is confounding by selection. People who receive a drug differ systematically from people who don't. Those with more advanced disease get certain drugs, healthier and more educated adults tend to get vaccines, and the rabies vaccine goes to people in certain jobs, such as vets. A randomized trial removes that by assigning treatment by chance.
How do researchers fix immortal time bias?
The standard fix is to treat exposure as something that changes over time. Each person's days before the first prescription count as unexposed time, and days afterward count as exposed time. Another method, called a landmark analysis, starts the clock at a fixed date and sorts people by what they had taken by then. A related approach, target trial emulation, designs the observational analysis to mimic the randomized trial that would answer the question.
Why did immortal time bias fool so many studies?
The error hides in how follow-up time is tallied rather than in the data itself, so it passes ordinary checks. Epidemiologist Samy Suissa of McGill University published influential analyses starting in the 2000s showing that it can produce large apparent protective effects for drugs. In a 2010 BMJ paper, Linda Lévesque and colleagues showed that an apparent protective effect of statins against diabetes progression disappeared when time was handled correctly.
How does randomization remove hidden differences?
Assigning treatment by chance balances characteristics between groups on average, including ones nobody measured or knew mattered. Statistical adjustment can only correct for factors that researchers recorded. Ronald Fisher developed the logic of randomization in agricultural experiments in the 1920s. The first randomized trial in medicine, of streptomycin for tuberculosis, was published by the UK Medical Research Council in 1948.
More from Saloni Dattani
