MECIR and Cochrane Handbook Chapter 10: The Meta-Analysis Conduct Standards, Item by Item

• Cochrane Handbook Chapter 10 explains how to analyse data and run meta-analyses; MECIR standards C61–C73 set the conduct rules for synthesis.
• Nine of the 13 synthesis standards are mandatory; effect measure, model and pooling method choices are Handbook guidance, not MECIR standards.
• Since 2024 the Handbook prefers REML over DerSimonian-Laird, suggests HKSJ intervals and recommends prediction intervals with about five or more studies.

Cochrane Handbook Chapter 10, Analysing data and undertaking meta-analyses, sets out how Cochrane reviews combine study results. The MECIR standards (Methodological Expectations of Cochrane Intervention Reviews) turn part of that guidance into conduct standards, each classed as mandatory or highly desirable12.

This guide works through the Chapter 10 topics item by item, gives the Handbook section and, where one exists, the matching MECIR standard code and status, so review teams can audit their analysis plan before and after data synthesis12.

Versions used: Cochrane Handbook version 6.5 (2024) and MECIR Version August 2023. Check the MECIR manual for the current wording before quoting a standard in a protocol2.

Chapter 10 explains methods and their rationale; MECIR states what a Cochrane review must or should do. Not every Chapter 10 topic has a MECIR standard: the choice of effect measure, model and pooling method is Handbook guidance, not a numbered standard12.

Since the August 2023 version, MECIR covers conduct only. The former protocol and reporting standards were retired, and Cochrane now endorses PRISMA 2020 for reporting24. Many non-Cochrane review teams also use MECIR and Chapter 10 as a benchmark of methodological quality.

The table pairs each Chapter 10 topic with its Handbook section and, where one applies, the MECIR synthesis standard (C61–C73) and its status12. Record these codes in your audit trail.

Chapter 10 links the effect measure to the outcome type (10.4.3, 10.5.1, 10.9)1:

• Dichotomous outcomes: risk ratio (RR), odds ratio (OR) or risk difference (RD). Ratio measures tend to be more consistent across studies; RD is easier to interpret in absolute terms.
• Continuous outcomes: mean difference (MD) when studies use the same scale; standardised mean difference (SMD) when they measure the same construct on different scales.
• Time-to-event outcomes: hazard ratios, pooled using the generic inverse-variance method or from O−E and variance statistics.

A fixed-effect model assumes one common true effect; a random-effects model assumes true effects vary between studies and estimates their average (10.3.1, 10.3.2)1. The DerSimonian and Laird method has been widely used3, but its τ² estimate can be biased1.

Since version 6.5 (2024), Chapter 10 prefers estimators such as restricted maximum likelihood (REML), now the RevMan default. It suggests considering the Hartung-Knapp-Sidik-Jonkman (HKSJ) confidence interval when τ² is above zero and there are more than two studies, and a prediction interval when there are about five or more studies (10.10.4)1.

Chapter 10 describes three main approaches for combining studies1:

• Inverse-variance method (10.3): weights each study by the inverse of its variance; used for continuous and generic data.
• Mantel-Haenszel method (10.4.1): the usual choice for dichotomous data, especially when events are few or studies small.
• Peto method (10.4.2, 10.4.4): for odds ratios when events are rare (around 1% or less), effects are not very large and group sizes are balanced.

Assess statistical heterogeneity with the χ² (Q) test, I² and τ² (C63), and take it into account when interpreting results rather than using it alone to choose a model (C69)12. Pre-specify a small number of subgroup analyses with a rationale, compare subgroups with a formal test for differences (C67) and interpret them cautiously (C68). Meta-regression is generally not advised with fewer than 10 studies (10.11.4)1.

Use sensitivity analysis to check whether conclusions depend on decisions such as study inclusion, risk of bias or imputation of missing data (C71, 10.14)12. A forest plot shows the result of each of these decisions at a glance.

Use this checklist to audit the analysis section of a review24:

• Synthesis methods match the registered protocol, and deviations are explained.
• Each outcome has an appropriate, justified effect measure, with scale direction aligned (C61).
• The decision to pool considers clinical and methodological diversity (C62).
• Multi-arm, cluster and cross-over trials are analysed without unit-of-analysis errors (C66, C70).
• The model, τ² estimator, confidence-interval method and meta-analysis method are reported.
• Heterogeneity statistics are reported and interpreted (C63, C69).
• Subgroup and sensitivity analyses are pre-specified or labelled as post hoc (C67, C68, C71).
• Missing outcome data, skew and non-reporting bias are addressed (C64, C65, C73).
• Conclusions do not read “no evidence of effect” as “evidence of no effect” (C72).

Cochrane Handbook Chapter 10 and the MECIR standards together define what rigorous meta-analysis looks like. Mapping each analysis decision, from effect measure and model to heterogeneity and sensitivity analysis, against them produces a transparent, defensible synthesis whether or not the review is published by Cochrane. For the wider process, see the meta-analysis workflow and our meta-analysis questions and answers; for hands-on help, see meta-analysis support.