From the author:
V.A. Kiri and colleagues bring up three points regarding my recent paper on bias from unaccounted immortal time in observational studies of inhaled corticosteroids and mortality in chronic obstructive pulmonary disease (COPD) 1. Indeed, unlike my previous paper on misclassified (not “unaccounted” as stated by the authors) immortal time 2, the nature of the unaccounted immortal time bias presented in the current paper 1 is more insidious and I welcome this opportunity to clarify these points.
First, V.A. Kiri and colleagues note that, in their study, the 317 patients receiving fluticasone and salmeterol (F+S: the exposed group) had a duration of 74 days between the first and third prescriptions, compared with 87 days for the 3,620 patients receiving short-acting bronchodilators (SABA: the reference group). The fact that the study used a 180-day period from the date of the first of these three prescriptions before starting to count the deaths implies two periods of immortal time. The first is the span between the first and third prescriptions, which can be addressed by using the date of the third prescription as cohort entry. The second is the time between the third prescription and day 180. This period differs between the two groups, with a mean of 106 days for the F+S group and 93 days for the SABA reference group. This difference implies that a patient who dies 100 days after their third prescription will not be considered if they were in the F+S exposed group but will be counted if in the SABA reference group. To avoid such inconsistencies from differential classification and the resulting immortal time bias, the rule is straightforward: if regular use is defined by three prescriptions, simply use the date of the third prescription to define cohort entry.
The stepped care approach to COPD treatment, while appropriate, may lead to an inappropriate hierarchical definition of exposure, which results in the problem of “unaccounted immortal time bias”. To understand this principle, consider patient A diagnosed with COPD in 1995 who for the first time receives three prescriptions for a SABA within 6 months in 1996, subsequently receives for the first time three prescriptions for F+S in 1999 and dies in 2001. Under the hierarchical approach, this patient is included only in the F+S exposed group. The fact that the patient survived for 3 yrs after the reference SABA is never considered under this approach. This is incorrect because the identical patient B also diagnosed with COPD in 1995, who receives for the first time three prescriptions for a SABA within 6 months in 1996 but dies in 1997 is included in the SABA group. Thus, patient A, who as patient B would have contributed to the SABA reference group had they died before receiving F+S, was not counted because they survived. Here again, the rule is straightforward: the design and analysis must use all time accumulated after any of the exposure criteria are met. Moreover, the use of a nested case-control approach within an incorrectly defined cohort will only produce incorrect results.
Finally, we agree that patients switched to inhaled corticosteroids (ICS) (who should be the majority of patients on ICS because of the stepped care approach) are probably the sicker patients. Therefore, we would expect that the crude mortality rate of such patients on ICS to be higher than the rate for those on bronchodilators alone, i.e. confounding by indication. Surprisingly, the reverse was observed in the study of Soriano et al. 3. The crude 1-yr mortality rate of the F+S group was 3.8% compared with 11.6% for the SABA group. This anomalous absence of confounding by indication is simply due to the fact that the SABA group (3,620 patients) excluded a large number of patients (up to 1,045) with 1 yr of immortality, who survived to receive F+S. As a result, the rate of death in the SABA group is overestimated because its calculation is based on the person-yrs of the 3,620 patients only (the group where all the deaths occurred) instead of all 4,665 patients (that includes the 1,045 where no death could have occurred). In fact, this denominator is much larger because it should also include the contribution of immortal SABA time from patients subsequently put on ICS other than fluticasone, who were excluded from this study.
In all, observational database studies of drug effectiveness are tricky. It would be helpful if V.A. Kiri and colleagues could present a proper reanalysis of their data that avoids immortal time bias and the resulting false appearance of drug effectiveness.
- © ERS Journals Ltd