Immortal Time Bias: Cause and Correction

Immortal time bias occurs when a period of follow-up during which the outcome cannot, by design, occur is misclassified as time under the exposure being studied — and because that period is guaranteed to be event-free, it makes the exposed group look protected even when the drug or intervention does nothing at all. The bias doesn’t come from a coding mistake or a small sample; it comes from a specific, common design choice: defining who counts as “exposed” using information that only becomes available after a subject has already survived part of follow-up, then crediting that survival back to the exposure as if it had been present from the start.

The mechanism: how “cannot-yet-have-the-outcome” time gets misclassified

Every cohort study needs two things fixed at the start: a time zero (when follow-up begins) and an exposure status. Immortal time bias appears when those two things are not actually aligned — when exposure status is really determined at some point after time zero, but the analysis treats the subject as exposed for the whole follow-up period anyway.

Concretely: a cohort enters at some index date (hospital discharge, diagnosis, cohort enrollment). Some subjects go on to receive a drug, but not on day one — they have to survive long enough, and be well enough, to have a prescription written and filled. The interval between time zero and the moment exposure actually starts is immortal for the exposed group specifically: a subject who died during that interval could never have gone on to fill the prescription, so by construction every subject counted in the exposed group survived through it. If the analysis assigns that pre-treatment interval to the exposed group’s person-time — which happens automatically in a naive time-fixed comparison — the exposed group is credited with survival it did nothing to earn, and the unexposed group is denied the chance to have had that same guaranteed-survival period, because by definition no unexposed subject’s time gets protected this way.

The result is not a small distortion. Because the misclassified time is, by construction, 100% event-free, it drags the exposed group’s apparent event rate down regardless of whether the drug has any real effect — and the longer the average gap between time zero and treatment initiation, the larger the spurious benefit.

A worked pharmacoepidemiology example

The cohort below is an illustrative composite built to walk through the arithmetic of the bias, not a real published study — the pattern it follows (discharge cohort, prescription-defined exposure, all-cause mortality) matches the structure of real inhaled-corticosteroid-in-COPD and statin-adherence literature that has been specifically scrutinized for this bias, but the cohort size and event counts here are round numbers chosen for clarity, not data from any specific paper.

Take a cohort of 1,000 patients discharged alive after hospitalization for a chronic respiratory condition, followed for one year for all-cause mortality. The exposure of interest is a maintenance drug; a patient is classified as “exposed” if they fill at least one prescription for it at any point during the year, and “unexposed” otherwise. Suppose 400 patients eventually fill a prescription, at a mean of 60 days after discharge, and suppose the drug has no true effect on mortality whatsoever — true one-year mortality is 10% in both groups if you could observe it under identical conditions.

In a naive analysis, every day between discharge and the actual prescription fill — on average 60 days per exposed patient, roughly 24,000 patient-days across the 400 eventually-exposed patients — gets counted as exposed person-time. But a patient cannot fill a prescription after they have died, so none of those 24,000 patient-days could have contributed a death to the exposed group; any death occurring before a patient’s first fill is, by definition, counted in the unexposed group instead (or excluded from being “exposed” entirely, depending on how crudely the study is done). The exposed group’s denominator has been inflated with guaranteed-survival time it never risked losing, so its computed mortality rate comes out lower than the unexposed group’s — not because the drug helped, but because the group definition itself filtered out early deaths by construction. The direction is always the same in this pattern: the more of an exposed patient’s early follow-up gets absorbed by the immortal interval, the larger the apparent (and spurious) survival benefit, independent of anything the drug actually does.

Why the name “immortal” time

The earliest documented example of this exact error is usually credited to Mitchell Gail’s 1972 analysis of heart-transplant survival, where transplant recipients had, by definition, survived long enough on the waiting list to receive a transplant — time that got folded into the “transplanted” group’s survival time and made transplantation look far more protective than it was. The term itself, and its systematic treatment as a recurring problem specific to drug-effectiveness studies, comes from a 2007 review by Samy Suissa in Pharmacoepidemiology and Drug Safety, which catalogued roughly twenty published pharmacoepidemiology studies carrying the same structural flaw — overwhelmingly studies where exposure was defined by prescription fills or refills observed sometime after cohort entry. The name reflects the mechanism directly: during the misclassified interval, the exposed group is functionally immortal in the analysis, because no death occurring in that interval could ever be attributed to them.

The fix: define exposure as time-varying, not fixed at baseline

The standard correction is to stop treating exposure as a single fixed label attached to a subject for their entire follow-up, and instead let it change value at the moment exposure actually starts — a time-varying (time-dependent) exposure in a Cox proportional-hazards model or an equivalent survival-analysis framework. Each subject contributes unexposed person-time from time zero until their first prescription fill, then switches to contributing exposed person-time from that point forward. A subject who dies before ever filling a prescription contributes only unexposed person-time and an unexposed death — they are never counted as exposed at all, which is the correct treatment, since they never were.

This single change removes the bias at its source: the guaranteed-survival interval is now correctly attributed to the unexposed side of the comparison (where it belongs, since it occurred before treatment), rather than silently inflating the exposed side. See the Cox proportional hazards model for how the model itself is specified and its assumptions checked, and how to interpret a hazard ratio for what the resulting time-varying-adjusted estimate does and doesn’t tell you once the correction is applied.

Software implementation requires the dataset to be restructured into multiple rows per subject (one row per exposure interval, a “start-stop” or counting-process format) rather than one row per subject with a single exposure label — this is a data-management step, not just a modeling-option toggle, and it’s the part most often skipped when a naive baseline-exposure analysis is run out of convenience rather than by deliberate choice.

Two more corrections worth knowing

Landmark analysis. Choose a fixed landmark time after cohort entry (for example, 60 or 90 days), restrict the analysis to subjects who survived to that landmark, and classify exposure based on status at the landmark rather than at any point across all of follow-up. This sidesteps immortal time bias entirely, at the cost of discarding early events (deaths before the landmark) and requiring a defensible, pre-specified choice of landmark time — too early and many exposed subjects haven’t started treatment yet; too late and real early events are thrown away. A sensitivity analysis across several plausible landmark times is standard practice for this reason.

Prescription time-distribution matching (PTDM). Proposed alongside the time-varying fix in the same pharmacoepidemiology literature, PTDM assigns each unexposed (comparison) subject an artificial “start of exposure” time drawn from the observed distribution of actual treatment-initiation delays in the exposed group, then measures both groups’ outcomes from that assigned point forward. It is less commonly used than the time-varying Cox approach but serves the same purpose: making the immortal interval symmetric across both groups instead of asymmetric.

All three corrections share the same underlying logic: the guaranteed-survival interval has to be treated identically for exposed and unexposed subjects, or excluded from the comparison altogether — it can never be allowed to count only in the exposed group’s favor.

How to spot it when you’re reading a paper

A few structural tells are enough to raise the question before checking the methods section closely:

Exposure is defined by something observed during follow-up (a prescription fill, a lab result crossing a threshold, a procedure date) rather than by baseline status at cohort entry.
The methods section doesn’t explicitly say whether the survival-analysis model treats exposure as time-varying, or describes exposure with a single yes/no column with no mention of when exposure started relative to time zero.
The reported effect size is unusually large and unusually protective for a drug class where the underlying pharmacology gives no strong prior reason to expect that scale of benefit — immortal time bias tends to produce dramatic, almost too-good hazard ratios rather than modest ones.
Time zero (cohort entry) and the point exposure was assessed are described separately, with a gap between them that the methods section doesn’t explain how it handled.

None of these confirm the bias on their own, but the combination — post-baseline exposure definition plus no explicit mention of time-varying handling — is the single most common signature of an uncorrected immortal-time problem in the observational pharmacoepidemiology literature. This same category of design flaw is one reason target trial emulation has become the preferred framework for observational comparative-effectiveness work: specifying the hypothetical trial protocol in advance forces time zero and exposure assignment to be aligned before the analysis is run, rather than corrected for after the fact.

Frequently asked questions

Is immortal time bias the same as lead-time bias?

No, though they’re both survival-time distortions and get confused. Lead-time bias comes from earlier detection of a condition (e.g. through screening) making survival-from-diagnosis look longer without actually delaying death — it’s about when the clock starts relative to disease onset. Immortal time bias comes from misclassifying which group a guaranteed-survival interval belongs to. They can co-occur in the same study but arise from different design errors.

Can immortal time bias make an exposure look harmful instead of protective?

Yes, if the misclassified interval is assigned to the comparison (unexposed) group instead — this is less common but happens in designs where, for example, “exposed” is the baseline default and subjects are reclassified as “unexposed” only after some later event. The direction of the bias follows whichever group absorbs the guaranteed-survival time, not a fixed rule that it always favors treatment.

Does excluding early deaths fix the problem?

Not by itself, and it introduces a new bias (informative censoring) if those early deaths are excluded unevenly across groups. Landmark analysis excludes early events in a structured, pre-specified, symmetric way, which is different from ad hoc exclusion of “inconvenient” early deaths after looking at the data.

Does this only apply to drug studies?

No — the same structural error appears anywhere exposure or treatment status is determined by a post-entry event: surgical procedures, device implantation, transplant receipt (the original Gail 1972 example), even employment or behavioral exposures identified partway through a cohort’s follow-up. Pharmacoepidemiology is simply where the pattern has been most systematically documented, because prescription-fill data makes the post-baseline exposure definition especially common.

How big a problem is it in practice?

Suissa’s 2007 review found the bias present, uncorrected, across a substantial share of the pharmacoepidemiology studies it examined, with apparent protective effects that shrank or disappeared once a time-varying correction was applied — it is not a rare or theoretical concern in observational drug-effectiveness research.

Immortal Time Bias: How It Inflates Drug Effects and How to Correct It