Continuity corrections with Mantel–Haenszel estimators in Cochrane reviews

The paper by Tsujimoto et al. 1 raises serious questions about the conduct of Cochrane reviews.They found that 64% of 885 meta-analyses that included studies with zero cells implemented the Mantel-Haenszel (MH) method in RevMan software, which by default implements the standard continuity correction (CC), in which 0.5 is added to every cell of a 2 2 table with a zero cell.The CC is unnecessary when calculating MH statistics, and it biases estimates towards the null.Tsujimoto et al. reported that in about 30% of cases, point estimates of the odds ratio were biased by 25% or more.Readers will surely support their conclusion that 'future updates of RevMan should incorporate less biased methods,' and their call for MH without CC to be incorporated.However, 'bias' may be too mild a word for a procedure that is, and has always been, recognized as frankly incorrect. 2Readers might have expected a call for biased methods to be entirely removed: they might also be wondering how the Cochrane organization could have allowed this blunder to persist for so long.In 2016, we raised RevMan's CC issue with the Cochrane Neonatal group, following the 2014 publication of their review of intrapartum anti-bacterial prophylaxis (IAP) to prevent neonatal earlyonset Group B streptococcus (EOGBS) infection. 3The headline estimate (0.17, 95%CI 0.04-0.74) of the efficacy of IAP to prevent EOGBS disease included a CC.This underestimates the efficacy by a factor of nearly 2 (Table 1).Our communication with the Neonatal Group and the authors' response was recorded as 'Feedback' on pages 32 and 33 of the Cochrane review.The authors replied: 'We are not statisticians, but assume that the statistical methods used in RevMan 5.2 are correct'.They went on to quote the 2008 Cochrane Handbook section 9.2.2.2 to the effect that ORs and RR cannot be calculated if there are zero cells, and that RevMan 'automatically makes the [continuity] correction when necessary'. 4 Almost identical wording appears in section 6.4.1.2 of the 2nd edition. 5he author's response is understandable, and it is hard to attach any blame to the distinguished neonatologist at Toronto's Hospital for Sick Children who led the review.Of course, a later section (16.9 in the 2008 Handbook, 10.4.4.1 in the 2019 edition) explains that the correction is not necessary when the MH estimators are used, and also that it causes bias.However, even if the Handbook had given explicit advice not to implement the CC with MH estimators, this would have made no difference: until recently, the use of RevMan was mandatory in Cochrane reviews (Handbook section 2.3.5).