Avoiding immortal time bias in observational studies

From the authors:

We thank A.J.N. Raymakers and co-workers for their letter, which clarifies the method of data analysis used in their study of the association between inhaled corticosteroids (ICS) and lung cancer incidence in COPD [1]. Indeed, as shown by the computing code they provide, the data analysis did consider ICS use as a time-dependent exposure, thus avoiding immortal time bias.

Our assumption that this bias affected their study arose mainly because of the scarcity in data presentation, in particular the crude rates. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines make extensive recommendations that allow to better understand and interpret data from observational studies, thus avoiding such misunderstandings [2]. In particular, it advocates that in “a cohort study with events as outcomes, report the numbers of events for each outcome of interest … the event rate per person-year of follow-up … presenting such information separately for participants in different exposure categories of interest … giving the unadjusted analyses together with the main data” [2]. A presentation of these data by A.J.N. Raymakers and co-workers, including information on the person-time and crude rates before and after ICS initiation, would have mitigated concerns about potential immortal time bias in the design and analysis of this observational study.

We agree that the availability of a cancer registry, such as the excellent one from British Columbia, with the confirmed diagnoses used in their study, provides more accurate outcome events than claims-based diagnoses. Nevertheless, validation studies in other databases have shown good accuracy for the diagnosis of lung cancer from claims data [3, 4]. This suggests that the resulting magnitude of the potential dilution of the hazard ratio towards the null in our study is likely marginal.

In all, it is reassuring to find out that the study of Raymakers et al. [1] used the proper time-dependent analysis that avoided immortal time bias. On the other hand, the relatively good accuracy of the diagnosis using claims data in our study likely does not fully explain the wide discrepancy in results between the two studies, namely a 30% reduction versus no reduction in lung cancer incidence with ICS use [5]. Thus, in view of our and other studies that also found no reduction in lung cancer incidence with ICS use, further research is needed to better understand this association, particularly if the suggestion of a randomised trial is contemplated [6].

Shareable PDF