© 2006 American Thoracic Society
Observational Studies of Inhaled Corticosteroids in Chronic Obstructive Pulmonary Disease: Misconstrued Immortal Time BiasFrom the Authors:We thank Dr. Suissa for his interest in our article, in which we used two different designs to analyze observational data on chronic obstructive pulmonary disease (COPD) and the use of inhaled corticosteroids (ICS) (1). Our designs were chosen because of recent concerns that some earlier analyses could have been subject to bias due to unbalanced baseline characteristics or by failure to account for differing wait times before the onset of treatment. In the case of our nested case-control analysis, which is not subject to immortal time bias, Suissa is concerned about an information bias. He suggests that patients with COPD who died might have had earlier hospitalizations during which ICS were prescribed unknown to their general practitioners (GPs). However, in METHODS, we explained how we designed our analysis to avoid this possibility. To enter the nested case-control study, the patient had to have a first COPD-related hospitalization. The period subsequently analyzed for treatment was drawn from between 30 and 365 days after discharge when the patient was under the clinical care of the GP. If the patient was rehospitalized within these 365 days, the retrospective analysis started from the index day of readmission. To allow for cases in which duration between discharge and index days was less than 6 months, control subjects were matched for duration of follow-up of each such case. Indeed, we defined an outcome event as the first occurrence of either COPD-related rehospitalization or death. Hence, no adjustment is required to our published data. Our second analysis, using propensity scores to improve matching of baseline characteristics, is, as stated in our article, subject to wait time bias, which we sought to avoid by ensuring both groups were determined on the day of discharge. Because of the great difficulty in handling data from patients in whom treatment was switched during the period of follow-up, inevitably we had to take a "peek into the future." Indeed, our ability to construct the whole cohort of patients for subsequent analysis by both methods requires that we know that suitable records exist for 1 year of follow-up or until an earlier outcome. Regarding our propensity scores–based analysis, Suissa's recommendation ignored the reasons we gave against an application of what was essentially the "treatment switching" methodology. As referenced in our article, standard texts and publications on biostatistics and epidemiology have warned against use of this methodology unless the strong assumption on which it is based can be satisfied—namely, that the reason for the change of exposure status is unrelated to the risk of subsequent event (2–5). For our study, this translates to an assumption that the reason why a GP subsequently prescribed ICS to a patient for the first time was unrelated to the patient's disease severity status—an assumption that clearly violates both existing guidelines (6) and the regular judgment involved in clinical practice. This problem cannot be avoided by censoring patients at first exposure to ICS as Suissa suggests, since such censoring will be informative to the risk of an event, making the analysis still inappropriate. We note that Suissa has not offered any explanation on how his recommended approach satisfied this fundamental assumption. The problem of confounding by indication was more influential than immortal time bias in our data. Consequently, adoption of Suissa's suggestion would result in both ignoring the confounding bias and incorrectly accounting for the much smaller immortal time bias. Not surprisingly, Suissa's calculation indicated ICS as possibly associated with an elevation of risk of COPD hospitalization or death. If ICS were not effective, an association should be absent even if exposed patients were only compared with the subpopulation of patients who were never exposed to ICS. In our article, we went further to address this issue of exposure misclassification and confounding by severity by restricting our comparison to exposed patients with the same propensity for ICS prescription as unexposed patients at time of discharge to reduce the more influential bias due to absence of randomization. We do, however, agree with Samet that, while the contributions of observational data are important, more definitive conclusions can be drawn from controlled trials (7).
GlaxoSmithKline, Greenford, Middlesex, United Kingdom
Fundación Caubet-Cimera, Mallorca, Spain
Imperial College, London, United Kingdom
Wythenshawe Hospital, Manchester, United Kingdom FOOTNOTES Conflict of Interest Statement: V.A.K. is an employee of GlaxoSmithKline (GSK). J.B.S. is an employee of GSK. N.B.P. in the past has been a consultant on COPD for GSK and is currently on an advisory committee for a long-term trial of treatment of COPD sponsored by GSK, for which he has received a total of $6,500 since 2002. J.V. received $8,000 in 2002, $10,000 in 2003, and $10,000 in 2004 for speaking at conferences organized by GSK. He received $4,000 in 2002, $7,000 in 2003, and $2,000 in 2004 for speaking at conferences organized by AstraZeneca, and $1,000 in 2004 for speaking at a conference organized by Boehringer-Ingelheim. He has served on advisory boards for GSK, receiving $2,000 annually in 2002–2004. His previous department (Department of Respiratory Medicine, Hvidovre University Hospital, Copenhagen, Denmark) received a research grant for approximately $400,000 during 2001–2003 to support research into the genetic basis of COPD. His wife (Inge Vestbo) has been an employee of GSK from 1988 to January 2004. Neither J.V. nor his wife have shares or options in GSK, or any other pharmaceutical company. REFERENCES
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
