What is the difference between Mantel-Haenszel and Inverse Variance pooling?

Inverse Variance pools study effect estimates by weighting each by the inverse of its variance (1/SE²). Mantel-Haenszel is specifically designed for binary 2×2 data and weights directly by study sample size distributions, making it more robust when individual studies have sparse events.

Why does Cochrane recommend Mantel-Haenszel for binary outcomes?

Cochrane recommends Mantel-Haenszel as the default fixed-effect method for binary outcomes (such as Odds Ratios and Risk Ratios) because it has better statistical properties and less estimation bias when trial sample sizes are moderate or event counts are small.

When should I use Peto's method instead?

Peto's odds ratio method is specifically recommended when events are very rare (less than 1% across arms), intervention effect sizes are moderate, and treatment arms are reasonably balanced. It avoids continuity corrections when one arm has zero events.

When should I switch from a fixed-effect to a random-effects model?

A random-effects model (such as DerSimonian-Laird) should be chosen when clinical or methodological diversity between studies suggests there is not a single true effect, or when statistical heterogeneity (I² > 50%) indicates substantial between-study variance (Tau² > 0).

STATISTICAL SYNTHESIS

Mantel-Haenszel vs Inverse Variance: Which Meta-Analysis Model Should You Use?

Choosing between Mantel-Haenszel, Inverse Variance, and Peto methods is one of the most critical statistical decisions when pooling clinical trials. This guide details the mathematical differences, handling of zero-event cells, and evidence-based rules for selecting the appropriate model.

Published October 12, 2026

Pool trial outcomes automatically in EvidenceFlow

Supports Mantel-Haenszel, Inverse Variance, Peto, and DerSimonian-Laird random effects.

Run Meta-Analysis →

1. Method Comparison at a Glance

Each pooling method assumes different statistical conditions and handles sampling variability differently:

Method	Outcome Types	Weighting Strategy	Sparse Data Performance	Best Use Case
Inverse Variance (IV)	Continuous (MD, SMD) and adjusted log ratios	1 / SE² (Variance)	Vulnerable to bias; requires +0.5 continuity correction	General across all effect measures
Mantel-Haenszel (MH)	Binary 2×2 counts (OR, RR, RD)	Cross-product sample weights	Robust; works with zero-event arms naturally	Gold standard for Cochrane binary outcomes
Peto Method	Binary Odds Ratio only	Variance of O - E statistic	Best for extremely rare events (< 1%)	Biased if treatment arms are highly imbalanced
DerSimonian-Laird (DL)	Any measure under random effects	1 / (SE² + Tau²)	Incorporates between-study variance	Accounts for clinical & statistical heterogeneity

2. Why Cochrane Recommends Mantel-Haenszel for Binary Data

In an Inverse Variance model, each study’s effect size (e.g., log Odds Ratio) is calculated first, along with its standard error. If a study has zero events in one arm, the standard error cannot be computed directly without adding an artificial continuity correction (usually +0.5 to each cell). This continuity correction introduces systematic distortion when event rates are low.

The Mantel-Haenszel method avoids this problem by calculating the pooled ratio directly from the 2×2 contingency table totals, weighting studies by their joint sample distribution. It provides superior statistical properties whenever study sample sizes are moderate or events are sparse.

3. When to Use Peto’s Odds Ratio Method

Peto’s method is a specialized variant of the fixed-effect model. It should be applied when:

Events are extremely rare (e.g., adverse events occurring in < 1% of participants).
The observed effect size is moderate (Peto odds ratios are biased when effects are huge).
Treatment and control arms are approximately equal in size (e.g., 1:1 randomization).

4. Fixed-Effect vs Random-Effects Decision Tree

A fixed-effect model assumes that all studies share a single identical true effect, with differences due only to sampling error. A random-effects model assumes that true effects follow a distribution across study populations:

Choose Fixed-Effect (Mantel-Haenszel / IV): When studies use identical patient populations, interventions, and protocols, and statistical heterogeneity is minimal ($I^2 < 25\%$).
Choose Random-Effects (DerSimonian-Laird): When clinical diversity exists between trial settings, dosages, or follow-up durations, or when $I^2 > 50\%$ indicates substantial between-study variance.

You can calculate your heterogeneity metrics with the free EvidenceFlow I² Calculator.

Automate your meta-analysis in EvidenceFlow

Seamlessly extract data, choose statistical models, and render publication-ready forest plots.

Start Free Project