Immortal Time Bias in Observational Studies

Esta obra está bajo una Licencia Creative Commons Atribución-NoComercial-CompartirIgual 4.0 Internacional.

Biases in the design of observational studies is one of the main sources of spurious association, in addition to chance errors or confounders. Immortal time bias is a type of bias that often occurs unnoticed by researches. This bias consists in the fact that there is a period, in the observation of the subjects, in which the event or endpoint of the study cannot occur. This is particularly important when comparing the incidence of an event between two groups: one with an exposure variable and others without it. Immortal time bias often occurs when, in the exposed group, there is a follow-up period between study entry and the determination of the exposure variable, a period in which the event or endpoint does not occur, by definition, resulting in a lower incidence of the event in the exposed group compared to the unexposed group (Figure 1). This difference in incidence is not real, but is biased by the study design that causes the immortal time bias, where subjects are protected for a period of time from developing the event in the exposed group. This situation occurs when cohort entry occurs occurs before the determination of the exposure variable. Thus, we call immortality time bias, when there is a period of time between cohort entry and the determination of the exposure variable in which the event cannot occur by definition.

Figure 1
Cohort study graph plotting the exposed and unexposed groups. The period between study entry and the determination of the exposure variable corresponds to the immortal period (event-free)

Sharif SZ clearly describes, from an article by Hemmelgarn et al., the immortal time bias.^{1, 2} In the aforementioned article, the authors report a 50% reduction in the risk of all-cause mortality for patients who had chronic kidney disease and attended multidisciplinary care clinics compared to those who had received usual care.² Mortality curves diverge rapidly between the group attending multidisciplinary care clinics compared to standard treatment. There is no biological plausibility in the fact that attending a multidisciplinary care clinic will result in an immediate and marked reduction in mortality. The reduction in mortality could occur over time (months or years), if patients evaluated in the multidisciplinary care clinic are accompanied by a more complete treatment and more complete treatment, with evidence-based and proven efficacy and/or associated with a better adherence to the interventions performed. The immediate reduction observed can only be explained by the immortal time bias. The Hemmelgarn's study was to examine the association between multidisciplinary care clinic and survival. It consisted of a retropective cohort of 187 clinical patients, who were exposed to a multidisciplinary care clinic, that were matched with 187 clinical control patients who were not part of a multidisciplinary care clinic (standard care). Control subjects were chosen on the basis of the propensity to match, so that individuals, in the control group, were similarly likely to be referred to a multidisciplinary care clinic as those in the multidisciplinary care clinic group. All patients were required to have an outpatient serum creatinine test performed between July 1 and December 31, 2001. The patients in the multidisciplinary care clinic group must also have attended a multidisciplinary care clinic between the July 1, 2001 and December 31, 2002. A Cox survival analysis was used between the multidisciplinary care clinic group and the standard group. Survival time was measured from the date of each patient's serum creatinine test. In other words, the date of each patient's serum creatinine represented the date they entered the cohort, or time 0. The patients were followed until the end of the evaluation (December 31, 2004) or until death, whichever came first. The difference in survival between the two groups was illustrated by the Kaplan-Meier survival curves (Figure 2). In this analysis, censorship occurred only at the end of the evaluation; therefore, the curves essentially represent the proportion of patients who were still alive at each time point during follow-up. The curves diverge immediately, which underlines the immortality time bias. After a year and a half, the two curves have a similar slope. Much of the early beneficial effect observed may be due to immortal time bias.

Figure 2
Kaplan-Meier survival curve, modified from the article of Hemmelgarn BR, which compares the survival of patients with renal failure who attended the multidisciplinary care clinic with the standard follow-up group²

There are two ways to prevent or correct the immortal time bias. One is at the time of the study design, which consists of "matching" in the selection of the subjects.³ Continuing with the example, an extra criterion is added in the selection of subjects in the control or standard group: they must be alive at the time they are selected subjects participating in the multidisciplinary care clinic. That is, the cohort begins to run as soon as the subjects enter the multidisciplinary care clinic and not when the serum creatinine sample is obtained. The other solution is to use time-dependent covariates, at the time of analysis.⁴ A time-dependent covariate is a predictor whose value can change over time. For example, a patient, in a multidisciplinary care clinic, would be considered unexposed from the date of study entry until they visit the multidisciplinary care clinic and are exposed from that point forward. Many statistical software packages can incorporate time-dependent covariates in survival analysis.

Correspondence: Aníbal Arias Email: anibal.arias@hospitalitaliano.org.ar