Quantifying variation beyond random chance
I-squared (I2) measures what fraction of observed variation between studies reflects true differences in effect rather than random sampling noise. I2 = 0% means all variation is sampling error and the estimate is homogeneous; I2 = 100% means all variation is real difference. Values between signal graded heterogeneity.
Rough benchmarks label I2 < 25% as low, 25-50% as moderate, 50-75% as substantial, and > 75% as considerable heterogeneity. A large meta-analysis with I2 = 20% may still harbor important subgroup differences, while a small analysis with I2 = 60% might show substantial scatter from limited data. The context (number of studies, their sample sizes) shapes interpretation.
Responding to heterogeneity with investigation
Moderate to substantial heterogeneity (I2 > 50%) usually prompts subgroup or meta-regression analysis to identify study characteristics (patient age, intervention dose, follow-up length) explaining variation. Exploring heterogeneity is more informative than ignoring it; consensus estimates mask important nuance when true variation exists.
Some fields accept high I2 as normal (treatment effects genuinely differ by context), while others view it as problematic aggregation. The pooled effect becomes less interpretable as I2 climbs; in highly heterogeneous analyses, narrative synthesis or separate estimates by subgroup may communicate findings more honestly than a single-effect estimate with weak heterogeneity adjustment.