Assessment of Risk of Bias
The overall risk of bias was low for most studies, except that the study by Thome (2006) 16 had a high overall risk of bias for BPD and NDI (Supplemental Table B). In the study by Thome (2006), the mortality in the permissive hypercapnia group was higher than that in the normocapnia group (36% vs. 19%). This can lead to a bias in the rates of BPD and NDI because those who died before the assessment of BPD or NDI could not have these outcomes. Owing to the characteristics of the intervention (targeting PaCO2), care providers were not blinded in any of the included trials. BPD diagnosis that was made without an objective algorithm (e.g., oxygen reduction test) might be influenced by the outcome assessors’ knowledge of the intervention assignment.20 Therefore, the measurement outcome of BPD was considered to be concerning, to some extent, except for that reported in the study by Carlo (2002) 15, which used an algorithm for BPD diagnosis (Supplemental Table B).
Of note, the study by Carlo (2002) was terminated early because the infants treated with dexamethasone (another intervention of the RCT with a 2 Ă— 2 factorial design) had high rates of NEC. Because the reason for early termination was not related to the comparison between permissive hypercapnia and normocapnia, we did not consider that it increased the risk of bias. For the cohort study of Hagen (2008) 19, the overall risk of bias was judged to be serious for IVH and NDI owing to concerns about remaining confounding and measurement outcomes (Supplemental Table C).