What is immortal time bias?

A recent JAMA paper by Yadav and Lewis (2021) provide the answer:

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

Yadav and Lewis (2021)

Sounds simple but how do these errors occur in practice? Here are some examples:

Death before intervention. One example would be to look at the effectiveness of heart transplant. Consider the case where researchers wanted to compare standard of care for patients with end stage heart failure to those who received a heart transplant. Consider the case where people are enrolled at the time when they are diagnosed with end stage heart disease. One seemingly sensible thing to do would be compare mortality rates for patients who did versus did not get a heart transplant. However, the time period between diagnosis and heart transplant is the “immortal time”. If a non-trivial number of patients die before getting a heart transplant, it may bias the results towards finding that heart transplants are favorable.

Drug duration studies. Consider the case where you want to compare Drug A to Drug B and its impact on mortality. This is a straightforward case and there is no immortal time bias. However, let’s say you want to compare whether using Drug A for a longer time period improves outcomes. You could divide the cohort into long-term users and short-term users. The problem would be that anyone who dies would be assigned to the short-term users. Thus, long-term users–by definition–have to live to use the drug past the median person (if that is how you are defining groups).

Exposure onset by age. A study by Honigberg et al. (2019) examined whether early menopause (i.e., menopause occurring before age 40) was associated with the development of cardiovascular disease (CVD). In this study, the authors restricted the sample to individuals who enrolled in the UK BioBank who were post-menopausal and had not yet experienced a CVD event. As Yadav and Lewis (2021) note in their recent JAMA paper, there are multiple problems with this study design. First, individuals who experienced a CVD event between menopause to BioBank enrollment were excluded from the study. Second, individuals who experienced menopause after BioBank enrollment were excluded.

So what is the solution? The best approach is to start from first principles. If you were designing a randomized controlled trial (RCT), how would you do it? Once you have this constructed in your mind, then go to your observational data and see in what ways your data generating process may differ from an RCT. If possible, use various statistical or study design approaches to address these limitations. Conduct a series of sensitivity analyses to test whether immortal time bias is an issue.

Immortal time bias may be more or less problematic depending on your study design, but it is one you should take into account whenever you are working with real-world data.