In Reply

We would like to thank Drs. Warner, Qiu, and colleagues for their valuable inputs regarding our systematic review.¹ 

Dr. Warner rightly points out that CIs crossing the equality line correspond to nonsignificant results and suggests that the wording of our findings may have failed to reflect this lack of statistical significance. Although we agree that “borderline” results (i.e., where one end of the CI just overlaps the null value) should be interpreted with caution, it is worth to note that the Cochrane Collaboration discourages formulations such as “nonsignificant” or “not statistically significant,” since these terms are commonly misinterpreted as an indication that “the intervention has no effect.”²  Although some authors would describe such findings as a “tendency” or a “trend” toward an effect, we opted for a more moderate wording (i.e., using formulations such as “seemed to” or “possible increase”), as suggested elsewhere.³  As for the interpretation of borderline findings, it might help to remember that the true effect is more likely to lie around the point estimate (i.e., around the risk ratio) than at the margins of the CI.³  The traditionally significant P < 0.05 may well be suitable for testing efficacy, but CIs rather than hypothesis testing are preferred when testing safety, equivalence, or noninferiority.⁴ 

A second concern of Dr. Warner’s is that our analysis did not include transfusion-related pulmonary complications, which may have resulted in an underestimation of potential harmful effects associated with liberal transfusion strategies. The rationale behind the exclusion of pulmonary complications was mainly related to the quality of the reported data in the original trials: in most studies, there was no distinction between transfusion-related events (such as acute lung injury or pulmonary edema due to circulatory overload) and events secondary to inadequate oxygen supply, such as left-sided heart failure due to myocardial infarction. Including outcomes with opposite etiologies could have resulted in a dilution of the intervention effects.

Qui et al. highlight a potential issue encountered in trials addressing transfusion strategies, i.e., the fact that heterogeneity in hemoglobin levels within individual treatment groups may potentially dilute treatment effects. Their concern is based on the assumption that patients assigned to a restrictive strategy who received blood transfusions would eventually have the same (posttransfusion) hemoglobin levels as those from the liberal group. A similar issue may occur if some patients assigned to a restrictive strategy never developed anemia (i.e., perioperative hemoglobin levels maintained in the range of the liberal group). This could indeed lead to an underestimation of adverse events, since only a small fraction of patients assigned to a restrictive strategy would truly be at risk of developing anemia-related complications. To address this potential source of heterogeneity, Qui et al. propose to stratify the analysis according to hemoglobin levels (see table 1, which provides a detailed description of hemoglobin levels across studies). Although the idea is very elegant, such exploratory analyses should be carried out with caution, since the quality of the reported data remains limited (data not extractable, heterogeneity in the frequency or duration of hemoglobin measurements, or use of inadequate statistics [e.g., Student’s t test for data correlated over time]). It is also worth noting that the randomized design used in the original studies tends to protect from bias and residual confounding. We certainly agree that large, well-designed randomized controlled trials are still needed to fully explore the effects of transfusion strategies and eagerly await the results of the ongoing Transfusion Requirements in Cardiac Surgery-III trial (NCT 02042898).