When several trials are combined, their effect sizes rarely line up perfectly. The degree to which they disagree is the heterogeneity of the meta-analysis.
It is captured by:
$$I^2 = \frac{Q - df}{Q}\times 100\%$$ where $$Q$$ is Cochran's statistic and $$df$$ the degrees of freedom. An $$I^2$$ near 0% means the studies essentially agree; values approaching 75% indicate substantial between-study variation, prompting use of a random-effects model.
Therefore heterogeneity is fundamentally a measure of variation between the included studies.
It does NOT measure:
• Publication bias - that is judged by funnel-plot asymmetry / Egger's test.
• Confounding - a within-study internal-validity problem.
• Single-study precision - reflected by each study's confidence interval and weight.
\[\boxed{\text{Evaluates variation between the included studies}}\]