How Immortal Time Bias Artificially Slashes Mortality Rates in Observational Medical Research

When medical researchers evaluate a new therapy using observational health records, a subtle accounting error can make a completely ineffective drug look like a life-saving breakthrough. By crediting a treatment with the days a patient survived before actually taking it, analysts inadvertently manufacture a statistical illusion. This distortion routinely derails clinical guidelines.

The phenomenon is known as immortal time bias, and it plagues longitudinal cohort studies that rely on electronic health records rather than randomized trials. It occurs when there is a delay between a patient entering a study—such as the day of a cancer diagnosis—and the day they receive their first prescription.[2][4]

During that waiting period, the patient is effectively "immortal" regarding the study's outcome. To eventually be classified as a treated patient, they must survive long enough to pick up the prescription. A patient cannot initiate a new therapy after they have died.[2][8]

The fundamental definition of the problem is straightforward. As McGill University epidemiologist Samy Suissa wrote in a landmark 2008 methodological review, "Immortal time is a span of cohort follow-up during which, because of exposure definition, the outcome under study could not occur."[1]

Because every person in the treatment group successfully survived their waiting period, the group possesses a guaranteed block of event-free time. The bias arises when researchers misclassify that guaranteed survival period, counting it as time protected by the drug rather than time the patient survived on their own.[1][2]

The Anatomy of Immortal Time

In a randomized controlled trial, this misalignment cannot happen. Patients are assigned to a treatment or a placebo at "time zero," and the clock starts ticking for both groups simultaneously. Every day of survival is correctly attributed to the intervention they were assigned on day one.[8]

Observational studies, however, must observe treatment decisions as they unfold in the real world. A patient might be discharged from the hospital on a Monday but not fill their cardiac prescription until Friday. Those four days represent immortal time, because the patient had to live through them to reach the pharmacy.[2][6]

If a researcher looks back at the data and categorizes that patient as "treated" from the moment of hospital discharge, the accounting breaks down. The four days of survival are added to the drug's ledger. The treatment is credited for keeping the patient alive before it was even ingested.[4]

This misclassification exerts a double penalty on the untreated control group. Not only does the treated group receive artificial, guaranteed survival time, but that exact block of time is stolen from the untreated group's denominator, where it rightfully belongs.[2][4]

The Mathematical Illusion

To understand the scale of the distortion, consider a synthetic cohort where a drug has absolutely no biological effect. Suppose the untreated group experiences 20 deaths over 150 person-years of observation, yielding a mortality rate of 13.3 deaths per 100 person-years.[3]

Meanwhile, the patients who eventually take the useless drug spend 50 person-years waiting for their prescriptions, experiencing zero deaths during that immortal window. After starting the drug, they spend another 150 person-years on the therapy, experiencing 15 deaths. Their true mortality rate on the drug is 10.0 deaths per 100 person-years.[3]

If the researchers correctly classify the 50 immortal years as untreated time, the untreated denominator grows to 200 person-years, dropping its rate to 10.0 deaths per 100 person-years. The rate ratio between the two groups is exactly 1.0, accurately reflecting that the drug does nothing.[3][8]

But if the researchers misclassify those 50 immortal years as treated time, the math collapses. The treated group now claims 15 deaths across 200 person-years, dropping its apparent rate to 7.5. The untreated group is left with 20 deaths in 150 person-years, raising its rate to 13.3.[3]

Through bad accounting alone, the useless drug now appears to reduce mortality by 44 percent, boasting a rate ratio of 0.56. The researchers publish a paper claiming a massive survival benefit, and clinical practice shifts toward a therapy that provides no actual biological advantage.[3][8]

Real-World Medical Consequences

This is not a theoretical vulnerability. The bias was first formally identified in 1972 by researcher Mitchell Gail, who noticed it distorting the apparent survival benefits of heart transplantations. Patients who lived long enough to receive a donor heart were being compared unfairly to those who died waiting.[2][5]

In the decades since, immortal time bias has repeatedly surfaced in pharmacoepidemiology. In the early 2000s, multiple observational studies claimed that inhaled corticosteroids dramatically reduced the risk of death in patients with chronic obstructive pulmonary disease. Later re-analyses revealed that the survival benefit was largely a statistical mirage.[1][2]

The distortion also heavily impacted early pandemic research. During the first waves of COVID-19, several high-profile observational studies reported that hydroxychloroquine and azithromycin significantly reduced mortality in hospitalized patients. These studies classified patients as treated even if they received the drugs days into their admission.[3]

When epidemiologists re-analyzed the COVID-19 data using proper time-varying methods, the apparent protective effect vanished. In one major cohort, a primary analysis showing a 35 percent reduction in mortality flipped to show an 83 percent increase in the rate of death once the immortal time was correctly categorized.[3]

The Danger of Duration Thresholds

The bias can also be introduced through more subtle study design choices, such as requiring a minimum duration of therapy. Researchers often want to compare "long-term users" of a medication against non-users to see if sustained exposure provides a cumulative benefit.[4][7]

If a study defines a long-term user as someone who takes a statin for at least six months, every patient in that group is guaranteed to survive their first six months of follow-up. If they die in month five, they are either excluded or dumped into the non-user control group.[4]

This creates a massive block of immortal time. The long-term users are granted a six-month survival advantage by definition, artificially slashing their observed mortality rate. The control group, meanwhile, absorbs all the early deaths, making the lack of treatment look far more dangerous than it actually is.[2][7]

The Impact of Decreasing Hazards

The magnitude of immortal time bias is not uniform across all diseases. It becomes particularly severe when the risk of the outcome is highest immediately after diagnosis and tapers off over time—a pattern statisticians call a decreasing hazard function.[1]

Hospital discharges often follow this pattern. The risk of readmission or death is acute in the first few days after a patient goes home, but stabilizes if they survive the first week. If a treatment is typically started during that high-risk window, the bias is amplified.[1][6]

By classifying the high-risk, pre-treatment days as immortal time, the analysis shields the treated group from the most dangerous period of the disease trajectory. The untreated group is left to absorb the concentrated wave of early events, drastically exaggerating the drug's apparent efficacy.[1]

Conversely, in diseases with a constant hazard—where the risk of death remains steady month after month—the bias grows linearly with the length of the delay. A six-month wait for a prescription will distort the findings twice as much as a three-month wait, regardless of when the clock started.[1][8]

Target Trials and Time-Varying Fixes

Epidemiologists have developed strict methodological frameworks to prevent this distortion. The most robust approach is target trial emulation, which forces researchers to design their observational study exactly as they would a randomized trial, strictly aligning the start of follow-up with the moment treatment is assigned.[8]

When treatment is delayed, analysts must use time-varying covariates. Under this method, a patient's observation period is split into segments. The days before they pick up their prescription are strictly tallied in the untreated column, and they only cross over to the treated column on the day the drug is ingested.[2][4]

Alternatively, researchers can use a landmark analysis. This involves setting a fixed point in time—such as 30 days after diagnosis—and only analyzing patients who survived to that landmark. Treatment status is defined strictly by what happened before the landmark, and follow-up begins only after it.[2][8]

While landmark analyses discard some early data, they successfully eliminate the guaranteed survival advantage. By ensuring that no patient is credited with time they had to survive just to enter a group, the method restores a level playing field between the exposed and unexposed cohorts.[2]

As electronic health records become the primary engine for medical research, policing these methodological boundaries is critical. Algorithms can process millions of patient records in seconds, but they cannot inherently recognize when a study's design has granted one group the mathematical gift of immortality.[6][8]

Without rigorous alignment of time zero and treatment status, the sheer volume of modern medical data will simply generate highly precise, entirely spurious correlations. Ensuring that survival is earned biologically, rather than granted statistically, remains the foundation of reliable observational science.[8]