Immortal Time Bias and Confounding by Indication

Suppose you are conducting a post-marketing drug safety study for a particular drug using administrative health databases. There could be multiple medications for the same disease condition, or there could be multiple dose categories depending on the severity of the disease condition. Over the course of follow-up, the patient may switch to another medication for the same condition, change dosage, or discontinue medication altogether. In all these cases, the patient stays in one medication or dose category for a varying length of time. The epidemiologist must account for this dynamic nature of treatment course because not doing so results in misclassification of exposure person-time.

Suppose you are comparing the safety of drug X relative to no exposure in relation to the risk of the onset of disease Z, and you have included all the patients who were diagnosed with a condition Y. Some may start drug X immediately after being diagnosed with Y, whereas others may be prescribed X only after certain length of time. Some others may be diagnosed with Y but never initiate X throughout the study period. Classification of exposure categories is simple for those who have always been on the medication X or those who have never initiated X. When we come across patients who initiated X later in the follow-up, we have to avoid the mistake of classifying these patients simply as having been exposed to X for the entire follow-up period. In this case, there is immortal time bias (Figure 1) [1]. That is, there is misclassification of exposure person-time, in which the person-time that should be attributed to the ‘no exposure’ group is attributed to the ‘drug X’ group. We see from the present that the patient (who now has been prescribed drug X) in the past unexposed period could not have had the onset of disease Z. When this block of time is attributed to drug X without incidence of disease Z, the risk of Z associated with X becomes lower. Drug X would then appear to have a protective effect against Z even if it would not be the case in reality.

Figure 1: Immortal Time Bias Illustration

To avoid immortal time bias altogether, we come back to the notion of time. As opposed to time-fixed variables, which do not change in value over the course of the study, we can use time-varying variables, whose values are updated over time. In the above example, the epidemiologist could create a time-varying exposure variable, which changes in value as the patient changes his or her treatment regimen. If we return to the patient with condition Y, who does not initiate drug X in the beginning of the study period but initiates it in the middle, the period before initiation of X could be classified as being ‘unexposed’ in the time-varying exposure variable. Similarly, the epidemiologist can design a time-varying exposure to capture changes in dosage, simply by classifying the dosage according to pre-specified dose categories.

Let’s consider another scenario. In randomized controlled trials, we commonly see placebo controls. In observational studies, especially those utilizing pharmacy claims database, we cannot have placebo controls because the database does not capture prescription of “nothing”. Suppose you want to conduct an observational study to evaluate the safety of drug A. If you compare drug A with no exposure, patients in the two comparison groups may be different with regards to clinical characteristics and disease severity. Now, we assume that all patients in the study population consists of individuals who were diagnosed of a specific disease. Prescription of the medication likely indicates that that patient’s disease status is more severe than the disease status of the patient who was not prescribed the medication. When the patient’s health status is worse, the risk of mortality or adverse health outcome likely remains higher. So, when we revisit the example of drug A, the risk of the outcome would be biased away from the null. In other words, the risk of the outcome associated with drug A would appear to be higher than the truth. To avoid confounding by indication as in the above example, the selection of an active comparator remains critical. The active comparator is simply another medication or intervention that is prescribed to the patients with a degree of disease severity that is similar to that which would require prescription of the main drug of interest.

References

1. Suissa, S., Immortal time bias in pharmacoepidemiology. Am J Epidemiol, 2008. 167(4): p. 492–9.