The Bias That Makes Every Treatment Look Good

Almost every surgical comparison you have ever read has one thing in common. Nobody could randomize it. You cannot assign a patient to open surgery versus minimally invasive by coin flip in most real world settings, so researchers reach for registry data instead, compare whoever got treatment A to whoever got treatment B, adjust for a list of variables, and call the result evidence.

Hidden inside a large share of these comparisons is a specific, well documented error that silently makes almost any treatment look protective, even when it does nothing at all. It is called immortal time bias, and once you learn to see it, you will notice it everywhere.

Here is the mechanism, stripped to its simplest form.

Say you want to know whether starting a certain medication after surgery improves survival. You take your cohort, label everyone who eventually started the medication as treated, and everyone who did not as untreated. Then you start the clock for both groups at the same point, say the date of surgery.

The problem is this. A patient can only be labeled treated once they actually receive the medication, which might be weeks after surgery. But by counting them as treated from the date of surgery, you have quietly given the treated group a period of guaranteed survival before the treatment even started, since they had to live long enough to receive it. That guaranteed survival window is immortal time, and it is not a real treatment effect. It is a bookkeeping error that inflates the apparent benefit of almost anything you study this way.

The treatment did not make the patients live longer. Living longer is what let them receive the treatment. The direction of causality got quietly reversed.

This is not a rare or theoretical problem. It has been documented across oncology, cardiology, critical care, and increasingly in surgical outcomes research, and it has reversed the conclusions of published studies once corrected. A 2025 editorial in the British Journal of Surgery specifically made the case for this framework in surgical and perioperative research, where randomized trials are especially hard to run.

The fix has a name: target trial emulation. Before touching the data, you write out the randomized trial you wish you could run. Who would be eligible. What exactly counts as the treatment strategy. When follow up starts for everyone, which must be the same moment eligibility is confirmed, not the moment treatment happens to occur. What the outcome is and how long you follow patients. Only once that hypothetical trial is fully specified do you go into the observational data and force your analysis to emulate it exactly.

Done properly, this single discipline eliminates immortal time bias and several related errors at once, because it forces time zero, the moment your clock starts, to be the same for every patient regardless of when they actually happened to start treatment.

20%

You do not need to become a causal inference specialist to benefit from this. You need to ask one question before starting your next observational comparison, and ask it before you touch the data, not after.

If you are currently designing a comparison from registry or cohort data, this is worth working through before the analysis begins. AskDimitri can help you specify the target trial before you touch the dataset.

Specify your target trial →

Randomization is not the only thing that makes a comparison trustworthy. Careful design of the clock is what makes an observational study earn the right to be believed.

Editor's note: in September 2026 the figures in this letter were checked against the sources listed below, and the text was corrected where the letter as sent differed from them.