Reply to Letter: Association of Postoperative Transfusions With Adverse Outcomes After Noncardiac Surgery

    loading  Checking for direct PDF access through Ovid


We greatly appreciate the interest and comments from Xue and colleagues on our paper entitled “Variation in Transfusion Practices and the Effect on Outcomes after Noncardiac Surgery” published in Annals of Surgery.1 In our propensity score-matched observational study of nearly 50,000 patients undergoing noncardiac surgery across 52 hospitals in the state of Michigan, we demonstrate that patients receiving postoperative packed red blood cells transfusions had increased risks for adverse postoperative outcomes with the exception of myocardial infarction; there is wide hospital-level variation in transfusion practices; and hospitals that are more liberal to packed red blood cells transfusion practice are almost twice as likely to transfuse the average patient and had higher risk-adjusted mortality.
Our dataset indeed has some information on preoperative β-blocker and statin use. These variables were not originally included in our models as we do not have data on their postoperative use, timing, dosage changes, or the specific drugs used, which may introduce noise or bias. Nonetheless, we checked the distribution of these variables before and after matching. Table 1 shows that even though these 2 variables were not included in our propensity score models, they are fairly well balanced in the matched cohort. This speaks to the success of the matching algorithm in minimizing “unmeasured” confounding. Although we certainly agree that these medications may help prevent perioperative deaths, our models were well calibrated, had good discrimination, and the score was well specified as discussed in our study, such that adding extra variables in the risk-adjustment models should not significantly alter the results.2
Regarding the comments on anesthesia techniques and intraoperative management approaches, Xue and colleagues bring about important perspectives that we cannot address with our dataset, as we do not have granular data on specific induction agents or the other parameters mentioned, and across the 52 hospitals we expect variable practice patterns between anesthesiologists. We do, however, have data on intraoperative vasopressor use. Again, this crude “unmeasured” confounder was well balanced after matching, thus we do not believe that adding it to our risk-adjustment models would dramatically change the results. On the other hand, we feel that the decision to transfuse postoperatively would not necessarily be influenced by whether etomidate or another agent was used, and thus it would not be of value in estimating a propensity score.
Finally, we would like to stress that our study remains observational in design and propensity score matching is one method to minimize confounding by indication, as previously discussed in the paper. The possibility of residual unmeasured confounding is always present, although it seems that we achieved a well-balanced matched cohort even on variables not used for matching. In the absence of well-designed randomized controlled trials, researchers have to rely on advanced statistical techniques, which inherently come with a set of assumptions that we have to entertain to draw robust and meaningful conclusions.

Related Topics

    loading  Loading Related Articles