Immortal Time Bias in Cohort Studies: A Concept Simply Explained

Abstract

Immortal time bias is one of many types of bias that may flaw cohort studies. This bias occurs when formal observation of patients starts before they experience the exposure of interest, and when, by design, those who experience the outcome of interest before they experience the exposure are ineligible to be included in the exposure arm of the study. Consequently, patients in the exposure arm, by design, are immortal to the outcome between the start of observation and point of the exposure. If this period of immortality is included in analysis, the result is immortal time bias. This concept is explained with one hypothetical example, related to selective serotonin reuptake inhibitor exposure and prevention of ischemic heart disease events, and one published example, related to the use of suvorexant to prevent delirium during intensive care unit admission. Also explained is how immortal time bias can be avoided, and why an awareness of the concept is important when reading studies that employ observational designs.

Many types of bias are described in research; these include selection bias, information bias, confounding, and other forms of bias, and subcategories thereof. ¹ Earlier articles in this column discussed confounding and lead time biases.^2,3 This article explains immortal time bias, and why it is important for readers to understand this bias in the context of cohort studies.

Immortal Time Bias

We are interested in studying the effect of an exposure on an outcome. The exposure can be a risk factor or a treatment. The outcome can be a favorable event or an unfavorable event. As an example, we know that selective serotonin reuptake inhibitors (SSRIs) inhibit platelet aggregation and may, therefore, protect against first-onset ischemic heart disease (IHD) events. ⁴ So, we extract data from healthcare records to conduct a retrospective observational study in which we examine the occurrence of a first-onset IHD event in depressed patients who initiate treatment with an SSRI drug. The comparison group comprises all other depressed patients, whether initiated on treatment or not, and, if initiated on treatment, regardless of what the (non-SSRI) treatment was.

In this study, we define study baseline as the date of diagnosis of depression. This is reasonable because IHD risk expectedly increases with the onset of depression and depression-related behaviors ⁴ ; because we cannot accurately know the date of onset of depression, we use the date of diagnosis of depression as a proxy.

If follow-up for all patients starts from this baseline, then occurrence of an IHD event at any time after baseline is eligible to be recorded as the target outcome in the comparison group. However, in the SSRI group, occurrence of an IHD event is eligible to be recorded as the target outcome only after the SSRI is started. So, if there is an interval between baseline and SSRI initiation (e.g., because of a long interval of no treatment, or because another treatment was prescribed first), only patients who do not experience an IHD event during this interval are eligible for inclusion in the SSRI arm of the study. That is, the SSRI patients are immortal to the study outcome in the interval between baseline and SSRI initiation. The longer the interval, the greater the immortal time bias where time to event is the outcome of interest.

In summary, immortal time bias occurs when formal observation of patients starts before patients experience the exposure, and when, by design, patients who experience the outcome of interest before they experience the exposure of interest are ineligible to be included in the exposure arm of the study. In effect, patients in the exposure arm, by design, are immortal to the outcome between start of observation and start of the exposure. If this period of immortality is included in analysis, the result is immortal time bias. ⁵

Side Note

The longer the interval, the greater the immortal time bias even when occurrence of the event, rather than time to event, is the outcome of interest. This is because with longer immortal time, patients with poorer prognosis may experience the event early and so, by default, become patients who experienced the event during the comparison condition. Because these patients are now ineligible for inclusion in the exposure arm (e.g., SSRI arm, in the example above), the exposure arm will experience a selection bias for patients with good prognosis. This is an example of how the same circumstances may include more than one type of bias; in this case, immortal time bias and selection bias.

Avoiding Immortal Time Bias

Immortal time bias can be avoided by synchronizing onset of follow-up such that immortal time is similar in the groups being compared; or, such that experience of the exposure (exposure group) or eligibility to experience the exposure (comparison group) is synchronized. ⁶ As a simple example, immortal time bias in the SSRI-IHD study described above can be avoided by redefining the SSRI group to include only those patients who were started on an SSRI at the time of diagnosis of depression. Consequently, at study baseline, itself, patients fall into either SSRI or comparison groups and there is zero immortal time in each group and so no immortal time bias.

Suvorexant for the Prevention of Delirium

Suvorexant is a dual orexin receptor antagonist that addresses insomnia by reducing alertness. Suvorexant may reduce pathological wakefulness, improve sleep, regulate the sleep-wakefulness cycle, and through these and other mechanisms reduce the risk of delirium in patients admitted to intensive care units (ICUs). In this context, Shiotsuka et al. ⁷ illustrated immortal time bias in a retrospective observational study of ICU patients.

Among 2,054 patients, those who received suvorexant at any time during their ICU stay were at significantly lower risk of new-onset delirium than those with no exposure to suvorexant (adjusted hazard ratio (aHR), 0.21; 95% confidence interval (CI), 0.16–0.27). However, when an attempt was made to limit immortal time and synchronize baselines, among 1,255 patients, those who received suvorexant within 72 hours of ICU admission were at similar risk of delirium relative to those with no exposure to suvorexant within 72 hours of ICU admission (aHR, 1.02; 95% CI, 0.65–1.59). ⁷

General Notes

Concerns about immortal time bias most commonly arise in the context of cohort studies. However, this bias can also occur in case–control studies and randomized controlled trials, when the timing of exposure is misaligned between the study groups. Avoidance of immortal time bias should be considered at the time of study design, itself.

Immortal time bias favors the exposure group when the outcome is an undesirable event. Thus, immortal time bias will favor the SSRI group when the outcome is an IHD event, or the suvorexant group when the outcome is delirium.

Parting Notes

Observational studies are common in medical and neuropsychiatric research. Readers should, therefore, be aware of immortal time bias as an explanation for study findings when the study follow-up starts before exposure (thereby creating immortal time) and when baseline is not synchronized between exposed and unexposed groups (thereby creating the bias).