Author + information
- Alberto Bouzas-Mosquera, MD⁎ (, )
- Jesús Peteiro, MD, PhD and
- Nemesio Álvarez-García, MD
- ↵⁎Department of Cardiology, Juan Canalejo Hospital, As Xubias, 84, 15006 A Coruña, Spain
We read with great interest the article by Fonarow et al. (1) evaluating the effect of continuation or withdrawal of beta-blocker drugs on outcomes in patients hospitalized with heart failure. The authors performed an analysis of 2,374 patients admitted with decompensated heart failure and concluded that withdrawal of beta-blocker therapy in these patients was associated with higher mortality.
There are several baseline characteristics that substantially differ among treatment groups. In addition to differences in the prevalence of coronary risk factors and coronary artery disease, patients who were withdrawn from beta-blocker drugs had lower left ventricular ejection fraction and higher expected post-discharge mortality risk. The authors performed a propensity score analysis to adjust for potential treatment selection bias.
Propensity scores represent the conditional probability of being assigned to a treatment group given a set of potential confounders (2,3). The bias and variance of the estimated effect of the treatment under study depend on the covariates selected for propensity score estimation. The authors claim that the propensity scores in their study were calculated with the set of all possible covariates that were related to the probability of receiving beta-blocker therapy; the inclusion of all information regarding the factors that might affect the selection of the treatment is, in fact, an important mainstay of propensity score analysis. However, the authors did not indicate which variables they used for estimation of the propensity scores. Furthermore, the reasons for beta-blocker withdrawal during hospital stay were not collected; this information is essential, because beta-blocker continuation or withdrawal might depend strongly on the clinical evolution of the patient during hospital stay, which in turn might be associated with outcome. Unfortunately, the authors also failed to provide information on the accuracy of the propensity scores for predicting treatment assignment, which might be assessed by the area under the receiver operating characteristic curve of the logistic regression model.
The presence of unmeasured variables that both affect the choice of the treatment and the outcome and the generation of propensity scores from potentially inaccurate models might preclude an adequate comparison among the different groups, which might compromise the validity of the estimated effect of the intervention.
- American College of Cardiology Foundation
- Fonarow G.C.,
- Abraham W.T.,
- Albert N.M.,
- et al.,
- OPTIMIZE-HF Investigators and Coordinators
- D'Agostino R.B. Jr.