Article Text


Biases in the observational study of β blockers in COPD
  1. S Suissa,
  2. P Ernst
  1. Center for Clinical Epidemiology and McGill Pharmacoepidemiology Research Unit, Jewish General Hospital, and Departments of Epidemiology and Biostatistics and of Medicine, McGill University, Montreal, QC Canada
  1. Dr S Suissa, Center for Clinical Epidemiology, Jewish General Hospital, 3755 Côte-Sainte-Catherine Road, Montréal, Québec, Canada H3T 1E2; samy.suissa{at}

Statistics from

Dransfield and colleagues1 reported, using an observational study design, that inpatient use of β blockers by patients hospitalised for COPD is associated with a surprisingly important 61% reduction in mortality.1 They also reported an astonishing 92% reduction in mortality associated with short acting β agonist use. Several biases introduced in the design and analysis of this observational study must be considered when interpreting these startling mortality reductions in patients with COPD.

Firstly, immortal time bias was introduced by defining exposure to β blockers or to short acting β agonists by billings occurring at any time during the hospitalisation.2 Indeed, the exposed patients necessarily had some initial period with no exposure before they received these drugs during the course of their hospitalisation. This period is “immortal”: a patient whose hospitalisation lasted 8 days and who received a β blocker on day 5 had an immortal period of 5 days during which they could not die. Indeed, had the patient died on day 4, they would have been classified as a non-user of β blockers. Thus by defining exposure in this way, the immortal period conferred a guaranteed survival advantage to the users of β blockers and an apparently longer survival. This is suggested by the mean length of stay of 7.8 days for users of β blockers versus 5.3 days for non-users. This bias unquestionably explains the phenomenal 92% reduction in mortality associated with short acting β agonists as over 95% of subjects used the agents and the magnitude of the bias is directly proportional to the frequency of exposure.3

Secondly, bias was introduced if β blockers are less likely to be used in the fatal hospitalisation of a patient with COPD who is in the final stages of the disease. Indeed, if these drugs are withheld in the context of palliative care, the rate of death in patients exposed to β blockers will be underestimated, which will make β blockers appear protective.

Thirdly, selection bias was likely introduced by the way the cohort was defined. The cohort of 825 subjects was formed using the last hospitalisation for a COPD exacerbation that occurred during the period 1999–2006. There were, however, approximately 2120 hospitalisations that occurred during this period (calculated from table 1 of the paper). By selecting the last hospitalisation, the cohort necessarily overrepresented the hospitalisations resulting in death. Basic tenets of epidemiology propose instead to use either the first hospitalisation to define the cohort, or to use all hospitalisations, albeit with a data analysis complicated by the correlated nature of hospitalisations occurring in an individual patient. Selection bias is amplified if β blockers are likely to be withheld in fatal hospitalisations.

Another important source of selection bias was introduced by identifying study subjects according to death summaries citing COPD as the probable cause of death. As death from cardiovascular causes is frequent in patients with COPD,4 and as patients prescribed a β blocker, and therefore with cardiovascular disease, are less likely to have COPD listed as the cause of death,5 subjects with COPD receiving a β blocker who died were systematically less likely to be included. As a result, a significant number of deaths exposed to β blockers was likely left out, leaving only eight such subjects in the study, thus leading to the appearance of a protective effect of β blockers. The presence of this bias is further suggested by the trend towards a protective effect of calcium channel blockers (odds ratio 0.76).

Observational studies are essential to complement information from randomised controlled trials. However, when such studies suggest astounding benefits that are inconsistent with trial data and use methods that are known to introduce well recognised biases, their results regrettably must be considered unfounded.


View Abstract


  • Competing interests: None declared.

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Linked Articles

  • Letters
    M T Dransfield S M Rowe J Johnson W Bailey L Gerald
  • Letters
    M T Dransfield S M Rowe J Johnson W Bailey L Gerald