To the Editor:— We appreciate the editorial1accompanying our article2and agree with the need to stress the limitations of propensity scores. A prospective randomized controlled trial (RCT), where possible, is always preferable to an observational study. However, RCTs have their own limitations, and prospective studies on blood transfusions based on hemoglobin thresholds are no exception. The exclusion of various diseases groups, such as patients with coronary artery disease, and the choice of treatment modality in the control group may challenge the applicability of the results of RCTs in everyday practice.3Indeed, Deans et al. 3recently highlighted the presence of coronary artery disease as a confounding factor in the RCT of Hébert et al. 4More specifically, a liberal blood transfusion strategy seemed to result in a higher mortality rate in younger patients with lower severity scores, but a lower mortality rate in the subgroup of patients with coronary artery disease.
Meticulous analyses, performed on large, unselected cohorts of critically ill patients, may provide useful additional information that can generate hypotheses and set the stage for subsequent RCTs. For example, a propensity analysis performed on a US database created a lot of turmoil when it raised serious concerns about the use of the pulmonary artery catheter,5but a similar analysis on the European Sepsis Occurrence in Acutely Ill Patients database using more meticulous adjustment did not yield similar findings.6
Thanks to the large database from the Sepsis Occurrence in Acutely Ill Patients study,7the propensity analysis in our recent article2included a large number of variables, allowing extensive and reliable adjustment for confounders. Unfortunately, the accompanying editorial by Drs. Nuttall and Houle1was misleading when it suggested that the time to survival was considered as from intensive care unit admission: As stated in the article,2the time to survival was counted from the day on which patients received a blood transfusion, and patients who were not transfused were censored at the time of intensive care unit discharge; hence, the time factor was taken into account in our analysis. Drs. Nuttall and Houle correctly stated that performance of a propensity analysis is weak when there are seven or fewer events per confounding variable, but this limitation does not apply when the number of patients is large, as in our study.8The editorial also described the statistical process as being opaque, simulating a “black box”; this could have been a relevant argument if covariate adjustment were performed using stratification according to propensity scores or the simple use of this score as a covariate in a multivariate analysis, where adjustment is not performed at the individual level and results cannot be examined for balance between treatment groups.9However, we performed case matching, with an excellent match between patients (table 6 of our article2).
Therefore, the editorialists are correct in underlining the potential limitations of a propensity analysis, but we believe our statistical methodology was strong enough to respond to most of these concerns.
*Erasme Hospital, Université libre de Bruxelles, Brussels, Belgium. email@example.com