Authors’ reply: Continuity corrections with Mantel–Haenszel estimators in Cochrane reviews

Abstract

Full text

Dear Editors,

We thank Ades and colleagues for their thoughtful letter regarding our paper on continuity corrections in Cochrane reviews.1 We fully agree with their concerns about the inappropriate implementation of continuity corrections with Mantel–Haenszel (MH) estimators, and appreciate their detailed historical perspective on this issue through the lens of Group B streptococcus prophylaxis reviews.

Their letter effectively illustrates how statistical methods that are known to be problematic can persist in widely-used software, potentially impacting clinical interpretations. The case study they present demonstrates how the default continuity correction in RevMan led to underestimation of intervention effects in reviews of intrapartum antibiotic prophylaxis, which aligns with our findings that approximately 30% of meta-analyses showed substantial differences in effect estimates when comparing methods with and without continuity correction.1

Cochrane states that it “exists to provide reliable evidence that people can use to make more informed health decisions.”2 In working toward this goal, we note that RevMan Web has recently implemented some methodological improvements, such as restricted maximum likelihood (REML) estimation for random effects models and Hartung–Knapp–Sidik–Jonkman methods for confidence intervals.3 However, the software still lacks the capacity to conduct MH meta-analysis without continuity corrections or to use exact methods such as conditional logistic regression for rare events meta-analysis.

We maintain our position that RevMan Web should incorporate more reliable statistical methods, including MH without continuity correction and logistic regression. Ades and colleagues raise an important point about statistical oversight in Cochrane’s editorial process. While organizational changes may help address this, we agree that having appropriate statistical methods readily available in the default software is crucial for supporting systematic reviewers in producing valid analyses.