What Is Immortal Time Bias in Observational Studies?

Immortal time bias is a flaw in the design or analysis of observational studies that makes a treatment, exposure, or experience look more beneficial than it actually is. It gets its name from a period of follow-up during which certain participants, by the very structure of the study, cannot experience the outcome being measured. That guaranteed-survival window then gets incorrectly counted toward the treatment group’s results, inflating how effective something appears. The bias has been found lurking in studies of medications, surgical procedures, lifestyle factors, and even Academy Award wins, and correcting for it has reversed or eliminated findings that once seemed solid.

What “Immortal Time” Actually Means

Imagine a study that compares patients who received a new drug to patients who did not. In a typical observational setup using medical records, researchers look back and sort people into groups based on who eventually started the drug and who never did. The catch is that starting a drug takes time. A patient might enter the hospital on day one but not begin the medication until day ten. Those first nine days are “immortal time” because the patient had to survive them in order to join the treated group at all. A patient who died on day five never had the chance to start the drug and therefore gets counted as untreated.

When those guaranteed-survival days are credited to the drug group, the drug appears to keep people alive longer, even if it does nothing. The treated group gets a head start of risk-free time baked into its results. The untreated group, meanwhile, includes people who died early and never had the opportunity to receive treatment. The comparison is rigged from the start.

The bias occurs whenever the start of formal observation begins before the exposure of interest, and people who experience the outcome before they experience the exposure get excluded from the exposed group.¹ This is not a rare edge case in study design. It is a structural feature of many observational studies, and if it goes unrecognized, it quietly distorts the results.

The Oscar Winners Who Did Not Actually Live Longer

One of the most memorable demonstrations of immortal time bias comes from outside medicine entirely. In 2001, a study published in a major medical journal concluded that actors and actresses who won Academy Awards lived almost four years longer than their peers who were nominated but never won. The finding made headlines: winning an Oscar, it seemed, was literally life-extending.

The problem was subtle but devastating. The analysis credited all of an Oscar winner’s years of life, including the decades before they won, toward their post-win survival. An actor who won at age 40 and died at 80 was counted as having 80 years of life associated with the “winner” group, even though 40 of those years happened before the win. Meanwhile, actors who never won had no such pre-award period inflating their totals. When researchers later reanalyzed the data using methods that avoided this bias, the survival advantage shrank to about one year and was no longer statistically meaningful.²

The Oscar example is useful because it strips away the complexity of medical data and makes the logic error visible. You cannot credit time before an event to the group defined by that event. Yet this is exactly what many studies do, often without the researchers realizing it.

Why Drug Studies Are Especially Vulnerable

The bias shows up most frequently, and most consequentially, in observational studies of drug effectiveness. These studies typically rely on health databases to compare people who took a medication to people who did not. Because treatment decisions happen over time rather than at a single starting gun, the window between entering the study and actually starting the drug creates immortal time.

A review of the problem found that several high-profile observational studies claiming drugs were remarkably effective at reducing sickness and death had used a flawed approach that introduced immortal time bias, generating what amounted to an illusion of treatment effectiveness.³ The drugs in question may have had some real benefit, but the size of that benefit was artificially inflated by the bias.

An illustrative case involved oseltamivir (the antiviral drug sold as Tamiflu) in critically ill influenza patients. Using standard statistical models, the drug appeared to protect against death. But when researchers accounted for the time-dependent nature of when patients actually received the drug, there was no longer evidence of a protective effect.⁴ The apparent benefit had been entirely an artifact of the analytical method. Patients who survived long enough to receive the drug were being compared against patients who died before they could get it, and the drug was getting credit for their survival.

How Large Can the Distortion Be?

The magnitude of the bias is not trivial. In simulation studies designed to quantify the problem, mishandling immortal time led to overestimates of treatment benefit by as much as 96 percent. In other scenarios using different analytical methods, the bias inflated or deflated the measured effect by well over 100 percent in either direction.⁵ That means a drug with zero real effect could appear to cut the risk of death nearly in half, purely because of how time was handled in the analysis.

The direction of the distortion is almost always the same: immortal time bias favors the treated group. This matters because observational studies are the backbone of evidence for many medical decisions, particularly when randomized trials are impractical or unethical. If a meaningful share of those studies contain this bias, clinicians and policymakers are making decisions based on inflated evidence.

How Common Is It in the Published Literature?

Depressingly common. A study examining high-impact cardiovascular journals found that about 13 percent of published observational cohort studies contained immortal time bias. Most of the affected studies did not use an active comparator group, a design choice that makes the bias more likely to go undetected.⁶ The majority of the bias in those papers came from misclassifying when patients were exposed to treatment.

A broader analysis across 25 different research topics, covering 182 studies, found that 44 percent of eligible studies were affected by immortal time bias.⁷ That figure is strikingly high and suggests the problem is not confined to a handful of careless studies but is a widespread feature of observational research.

A separate evaluation of 162 cohort studies using routinely collected health data classified about 22 percent as being at high risk for the bias, while roughly 8 percent lacked enough information to even assess whether immortal time was a concern.⁸ These numbers vary depending on the field and the stringency of the assessment, but the overall picture is clear: a substantial minority of published observational studies contain a bias that makes their findings look better than they should.

Three Ways Researchers Fix It

Recognizing the problem is half the battle. Several analytical strategies can correct for or prevent immortal time bias, and each works by aligning the clock so that time is counted fairly for both groups.

Time-varying exposure: Instead of permanently labeling someone as “treated” from the start of the study, this approach lets a person’s treatment status change over time. A patient is counted as untreated until the day they start the drug, then switches to the treated group. This prevents the pre-treatment survival window from being credited to the drug. Simulation studies show that this method consistently provides unbiased treatment effect estimates, while the standard approach leads to bias in favor of treatment.⁹
Landmark analysis: This method picks a fixed time point after study entry and only includes patients who are still alive at that point. Groups are then defined by whether they received the treatment before the landmark. Outcomes are counted only after the landmark, ensuring that no one gets credit for surviving a period they had to survive by definition.¹⁰ When the treatment effect is small or absent, landmark analysis performs well at eliminating the bias.¹¹
Target trial emulation: This is a broader framework where researchers design their observational analysis to mimic the structure of a randomized trial. By explicitly defining when follow-up starts, when treatment is assigned, and when eligibility is established, the framework forces alignment of these time points and prevents the mismatches that create immortal time.¹² The approach can be extended to emulate a sequence of trials when few patients start treatment at any single time point.¹³

Among these approaches, time-varying exposure modeling has been called the most appropriate method for certain types of studies, particularly when analyzing cumulative dose effects.¹⁴ No single correction method is universally best, though, and the right choice depends on the study question and data structure. What matters most is that researchers acknowledge the issue and use at least one of these tools.

Does Correcting the Bias Change Published Conclusions?

Often enough to matter. When researchers went back and applied corrections to 16 studies identified as being at high risk for immortal time bias, a third of them showed meaningfully different results when using time-varying exposure models, and a quarter showed different conclusions under landmark analysis.¹⁵ In some cases, findings that appeared statistically significant became non-significant. In at least one case, the direction of the effect flipped entirely: what looked like a benefit turned into a harm.

The Oscar study is the cleanest public example, but the same pattern plays out in clinical research. A drug or intervention that appeared to reduce mortality in the original analysis turns out to have a weaker effect, no effect, or occasionally the opposite effect once the immortal time is properly accounted for. These are not minor academic quibbles. They are the difference between a treatment guideline and a retraction.

Electronic Health Records Make the Problem Worse

The rise of large electronic health record databases has been a boon for observational research, letting scientists study millions of patients without running expensive trials. But these databases also make immortal time bias easier to introduce and harder to spot. Because diagnoses, prescriptions, and outcomes are recorded at different times across a patient’s history, the gap between when someone enters the dataset and when their exposure begins can be large and irregular.

Research specifically examining this issue in electronic health records found that treating immortal time as exposure time led to an overestimate of life expectancy in the exposed group and could lead to fundamentally wrong conclusions about survival.¹⁶ The problem is especially tricky for lifelong conditions like disabilities or congenital diseases, where the “exposure” has been present since birth but may not appear in the medical record until much later.

Step-by-step causal analysis frameworks for health record data emphasize that improper alignment of the event that defines study inclusion and the time of intervention is a major source of bias. When follow-up starts before the intervention occurs, the treated group is automatically biased toward patients who were still alive at the time of treatment assignment, inflating the estimated effect.¹⁷ As health systems increasingly turn to automated analysis of patient records to guide clinical decisions, this problem will only grow more consequential.

Beyond Drug Studies

While medications and clinical outcomes get the most attention, immortal time bias can affect any study comparing groups defined by something that happens over time. Surgical outcomes studies are vulnerable: patients who die before they can get an operation are counted in the non-surgical group, making surgery look more effective. Cancer biomarker studies face the same issue when a patient’s test result at some later time point is treated as though it were known at the start of the study. In circulating tumor DNA research, for example, standard survival analysis methods can introduce immortal time bias if a time-dependent biomarker result is analyzed as a fixed baseline variable instead of being aligned with when the test was actually performed.¹⁸

Studies of lifestyle factors are vulnerable too. Research examining whether people who exercise regularly live longer can introduce immortal time bias if participants must survive long enough to be classified as regular exercisers. Vaccination studies can suffer from it when unvaccinated person-time before someone receives a shot gets counted against the unvaccinated group. Essentially any time-to-event study where the exposure group is defined retrospectively and the exposure does not happen at a fixed, common starting point is at risk.

How to Spot It When You Are Reading a Study

You do not need to be a statistician to notice the warning signs. When reading about an observational study claiming a drug, procedure, or behavior dramatically reduces the risk of death or disease, a few questions can help you gauge whether immortal time bias might be lurking.

First, ask when the clock started. If the study compared “users” versus “non-users” of a drug starting from hospital admission or study enrollment rather than from the moment the drug was actually started, the gap between enrollment and first dose is immortal time. Second, check whether the study mentions using any of the correction methods: time-varying covariates, landmark analysis, or target trial emulation. Among studies that handled the issue properly, the most common approaches were an active comparator new-user design and time-varying exposure definitions.¹⁹ If the methods section does not mention any of these, the study may be at risk.

Third, consider the size of the claimed benefit. When an observational study reports that a common, inexpensive medication cuts the risk of death by 40 or 50 percent, that is an extraordinary claim. Randomized trials of expensive, heavily researched drugs often find benefits in the range of 10 to 20 percent. An observational finding that greatly exceeds what trials show for similar interventions deserves extra scrutiny. Immortal time bias is one of the most common explanations for such implausibly large effects.

Finally, pay attention to whether the study has an active comparator group, meaning it compares users of one treatment to users of a different treatment rather than to people who received nothing. Studies without an active comparator were far more likely to contain the bias in at least one review of cardiovascular research.²⁰ An active comparator design naturally avoids some of the time-misalignment problems because both groups have a treatment start date that can serve as a common starting point.

Why Peer Review Does Not Always Catch It

Given how well-documented this bias is, you might expect journals to screen for it before publishing. In practice, the bias continues to slip through peer review for several reasons. The flaw is not in the data but in how time is structured in the analysis, and it can be invisible to reviewers who focus on the clinical question rather than the analytical architecture. The immortal period is often short relative to the total follow-up, making it easy to assume the distortion is negligible even when it is not. And the bias reliably produces positive findings, which are exactly the kind of results that journals are most eager to publish.

The fact that roughly one in eight cardiovascular cohort studies in top journals contained the bias suggests that awareness alone has not solved the problem. Some researchers have called for checklists and reporting standards that require authors to explicitly state how they handled pre-exposure follow-up time. Others have argued that reviewers and editors should be trained to look for the telltale design features. For now, the burden falls partly on the reader to ask the right questions when confronted with an observational study claiming dramatic results.