This JAMA Guide to Statistics and Medicine explains immortal time bias, an error in estimating the association between an exposure and an outcome that results from misclassification or exclusion of time intervals; explains how this misclassification or exclusion can occur; and presents approaches to minimize or avoid immortal time bias.
Observational studies are commonly used to evaluate the association between a risk factor or “exposure” and the time that elapses until an outcome of interest occurs (eg, smoking decreasing the time to cardiovascular death). This relationship may be assessed by analyzing natural variation in the exposure and patient outcomes. Such studies may be subject to immortal time bias, meaning that, during the period of observation, there is some interval during which the outcome event cannot occur. The research participants are “immortal” in that they must survive long enough to receive the intervention being studied. An example of immortal time bias can be found in a study by Honigberg et al1 that evaluated the hypothesis that menopause occurring before age 40 years is associated with the development of cardiovascular disease (CVD). In that study, the authors examined CVD outcomes for postmenopausal women aged 40 to 69 years using the UK BioBank and found an increased risk of CVD associated with both natural premature menopause (hazard ratio, 1.36 [95% CI, 1.19-1.56]) and surgical premature menopause (hazard ratio, 1.87 [1.36-2.58]), relative to women who had experienced normally timed menopause.
WHAT IS IMMORTAL TIME BIAS?
Bias from immortal time periods is the error in estimating the association between the exposure and the outcome that results from misclassification or exclusion of time intervals.2 Depending on how the exposed and unexposed groups are defined in an observational study, periods may be misclassified as time spent as exposed to an intervention when, in fact, an individual was not yet exposed (Figure 1). Similarly, the study enrollment process may have led to potential participants being excluded from the accounting of the time spent while either exposed or not exposed to the intervention.2
Examples of Immortal Time Intervals
A, Participants are included based on a qualifying medical event that starts the time interval. Those ultimately included in a drug treatment group may have a period of unexposed immortal time before starting the drug. Events in this period would be erroneously assigned to the drug that had not yet been received, and benefit associated with a drug would be overestimated. B, Participants are included based on UK BioBankenrollment, which sets the time interval, but all included participants may have varying periods of excluded immortal time before enrollment, beginning with the medical event of menopause. Two censored groups represent excluded immortal time bias: anyone undergoing a cardiovascular (CV) outcome prior to enrollment and participants undergoing menopause after enrollment (latter not shown).