Confounders in randomized controlled trials (RCTs) reporting significant effects on mortality in critically ill patients using non-surgical techniques have not been systematically explored. We aimed to identify factors unrelated to the reported intervention that might have affected the findings and robustness of such trials.Methods:
We searched Pubmed/MEDLINE for all RCTs on any non-surgical interventions reporting an effect on unadjusted mortality in critically ill patients between 1/1/2000 and 1/12/2015. We assessed: the number needed to treat/harm (NNT or NNH), sample size, trial design (blinded/unblinded, single or multinational, single or multicenter (sRCT or mRCT)), intention to treat (ITT) analysis, and countries of origin.Results:
Almost half of RCTs were sRCTs. Median sample size was small, and 1/3 were not analyzed according to ITT principle. Lack of ITT analysis was associated with greater effect size (p = 0.0028). Harm was more likely in mRCTs (p = 0.002) and/or in blinded RCTs (p = 0.003). Blinded RCTs had double sample size (p = 0.007) and an increased NNT/NNH (p = 0.002). Finally, mRCTs had higher NNT (p = 0.005) and NNH (p = 0.02), and harm was only detected in studies from Western countries (p = 0.007).Conclusions:
These observations imply that major systematic biases exist and affect trial findings irrespective of the intervention being studied.