Heterogeneity in Meta-Analysis: Understanding I²

Heterogeneity is one of the first things a careful reader checks in a meta-analysis, because it answers a basic question: are these studies actually measuring a consistent effect, or are we pooling together things that don't belong together?

What heterogeneity means

Heterogeneity refers to the variability in effect estimates across the studies included in a meta-analysis, beyond what you'd expect from random chance (sampling error) alone. Some variability is normal — different studies have different samples, settings, and designs. The question is whether that variability is small enough to make pooling a single summary estimate meaningful.

The Q-statistic and I²

Cochran's Q is a statistical test for heterogeneity, but it has low power when there are few studies and can flag heterogeneity as "not significant" even when real variability exists. I² (I-squared) addresses this by describing the percentage of total variability across studies that's due to heterogeneity rather than chance, which makes it easier to interpret and compare across meta-analyses regardless of study count.

Interpreting I² values

A commonly used rough guide (from the Cochrane Handbook) is:

• 0–40%: might not be important.
• 30–60%: may represent moderate heterogeneity.
• 50–90%: may represent substantial heterogeneity.
• 75–100%: considerable heterogeneity.

These ranges overlap deliberately — the guidance is meant to be interpreted alongside the direction and size of effects, not read as a strict cutoff.

What to do when heterogeneity is high

• Switch to a random-effects model if you were using fixed-effects, since random-effects models account for between-study variability directly.
• Run a subgroup analysis to see whether a specific study characteristic (population, intervention dose, study design) explains the variation.
• Try meta-regression to test whether a continuous study-level variable is associated with the effect size.
• Consider whether pooling is appropriate at all — sometimes high heterogeneity is a sign that a narrative synthesis is more honest than a forced statistical pool.