I Squared (I²)

I squared (I²) is a statistic used in meta analysis to quantify the percentage of variability in effect estimates across studies that is attributable to true heterogeneity rather than sampling error. Values range from 0% (all variability is due to chance) to 100% (all variability reflects genuine differences between studies).

When combining results from multiple studies, understanding whether the individual estimates agree with each other is critical. If studies produce very different results, pooling them into a single summary estimate may be misleading. I² provides a standardized way to assess this inconsistency and is widely reported in published meta analyses. HTA agencies, clinical guideline panels, and Cochrane review groups use I² to judge the credibility of pooled estimates and to decide whether further investigation of heterogeneity is needed through subgroup analysis or meta regression.

I² is derived from the Cochran Q statistic, which tests the null hypothesis that all studies share a common effect size. The formula is I² = (Q minus df) divided by Q, multiplied by 100%, where df is the degrees of freedom (the number of studies minus one). If Q equals or falls below df, I² is set to zero. Higgins and colleagues (2003) tentatively labeled I² values of 25%, 50% and 75% as low, moderate and high. The Cochrane Handbook gives overlapping ranges instead: 0% to 40% might not be important, 30% to 60% may represent moderate heterogeneity, 50% to 90% may represent substantial heterogeneity, and 75% to 100% considerable heterogeneity. Both are rough guides rather than strict cutoffs: the Handbook says the importance of an I² value depends on the magnitude and direction of the effects and on the strength of the evidence for heterogeneity.

Several important properties of I² deserve attention. First, I² describes the proportion of variability due to heterogeneity but does not indicate its magnitude in clinically meaningful units. Two meta analyses can both report I² = 60% while having very different ranges of effect sizes across studies. Tau squared, the estimated between study variance in a random effects model, and its square root, tau, describe that magnitude on the scale of the effect measure. Second, I² is influenced by the precision of the included studies. Because I² compares the spread of the true effects with the sampling error within studies, the same spread gives a higher I² when the studies are large and precise: a meta analysis of large, precise studies can report a high I² even when the true effects differ only slightly, whereas a meta analysis of a few small studies may show I² = 0% simply because the studies lack precision to detect real variation. Third, I² should be interpreted alongside the prediction interval, which estimates the range within which the true effect of a future study is likely to fall, providing a more clinically interpretable picture of heterogeneity.

When I² is high, researchers should investigate the sources of heterogeneity. Common approaches include prespecified subgroup analyses (examining whether the effect differs by population, intervention dose, comparator, or study design), meta regression (testing whether continuous or categorical study characteristics explain variation in effects), and sensitivity analyses (removing outlying studies or restricting to studies at low risk of bias). Simply reporting a high I² without exploring its sources is considered incomplete.

I² should be calculated and reported in every meta analysis as part of the assessment of between study heterogeneity. It is useful for judging whether subgroup analysis is warranted and for contextualizing the pooled estimate, but it should not decide between a fixed effect and a random effects model: that choice should rest on whether a single common effect is plausible, judged from the clinical and methodological diversity of the studies, not on a statistical test or measure of heterogeneity. It is not appropriate as the sole indicator of heterogeneity; it should be accompanied by the Q test p value, a visual inspection of the forest plot, an estimate of tau squared, and, where possible, prediction intervals.

A frequent error is treating the conventional thresholds (25%, 50%, 75%) as rigid categories rather than rough guides, leading to statements like "heterogeneity was moderate" without further exploration. Another mistake is interpreting low I² as evidence that heterogeneity is absent, when, with few or small studies, it may simply reflect imprecision: I² is then estimated with wide uncertainty and can be 0% even when the true effects differ. Some researchers also conflate I² with the Q test p value; a nonsignificant Q test does not mean heterogeneity is absent, particularly when few studies are included. Finally, reporting I² without exploring its sources through subgroup analysis or meta regression represents a missed opportunity to understand why studies differ.

A meta analysis combines 18 RCTs examining the effect of mindfulness based stress reduction on anxiety scores. The pooled standardized mean difference is 0.52, but I² = 74% (Cochran's Q = 65.4 on 17 degrees of freedom), within the Cochrane Handbook's range for substantial heterogeneity. Subgroup analysis reveals that studies using 8 week programs show a larger and more consistent effect (SMD 0.69, I² = 32%) than studies using abbreviated 4 week programs (SMD 0.30, I² = 68%), suggesting program duration is one source of between study variation. It does not explain all of it, since the 4 week trials still differ considerably, and because the subgroups compare different studies rather than randomized groups, the finding is observational.