Statistics Glossary → Between-study variance (τ²) and heterogeneity (I²)
Between-study variance (τ²) and heterogeneity (I²)concept
Two measures in meta-analysis of how much the true effect sizes of the combined studies differ: τ² is the absolute size of that difference, and I² is the share of the observed spread that is due to it.
The random-effects model assumes that each study has its own true effect and that these effects come from a distribution with mean μ and variance τ². τ² is expressed in the units of the effect size (e.g. the log odds ratio or standardised mean difference), and its square root τ is the standard deviation of the true effects. It is estimated with methods such as DerSimonian-Laird, REML, Paule-Mandel, ML, Sidik-Jonkman and Hedges; simulation studies generally recommend REML or Paule-Mandel. τ² enters both the random-effects weights and the prediction interval.
I² = (Q − df)/Q × 100% (set to 0 if negative) is the percentage of the total observed variation that is not due to sampling error. Roughly, 25%, 50% and 75% are referred to as low, moderate and high heterogeneity. I² is a relative measure: as studies get larger their sampling error shrinks, and I² rises for the same τ². The answer to 'how much do the effects differ?' is therefore τ² and the prediction interval, not I². With few studies both estimates are very uncertain and should be reported with confidence intervals.
An everyday example
A meta-analysis of 12 studies of a training programme finds Q = 33 (df = 11). I² = (33 − 11)/33 ≈ 67%: about two-thirds of the observed spread comes from the studies genuinely finding different effects. A τ of 0.20 shows that even if the average effect is 0.40, the effect may be close to 0 in some settings and close to 0.8 in others.
In MerQur
In Meta-Analysis, the 'τ² estimator' option offers REML, DerSimonian-Laird, maximum likelihood (ML), Paule-Mandel, Sidik-Jonkman and Hedges (HE); REML or Paule-Mandel is recommended, and DL is the classical default.
Analyses where this concept appears
Related concepts
Fixed and random effectsKnapp-Hartung (HKSJ) adjustmentPrediction interval