Heterogeneity (I²)

There are two kinds. Clinical and methodological heterogeneity is variation in populations, doses, comparators, instruments and follow-up length, judged by reading the studies rather than by computing anything. Statistical heterogeneity is variation among the estimates larger than sampling error explains. Cochran's Q is the weighted sum of squared deviations of each study's estimate from the pooled one; I-squared rescales Q as a percentage, Q minus its degrees of freedom divided by Q, floored at zero. Tau-squared is the estimated variance of the true effects.

Cochrane offers deliberately overlapping bands rather than thresholds: zero to forty percent possibly unimportant, thirty to sixty moderate, fifty to ninety substantial, seventy-five upward considerable. I-squared is a ratio, not a magnitude. Pool several very large trials and sampling error is tiny, so a clinically negligible difference produces a high I-squared; pool a handful of small ones and real, important differences can produce an I-squared of zero.

High heterogeneity is a prompt to explain rather than to abandon. Subgroup analysis, meta-regression on dose or duration and restriction to one instrument all attempt this, though such comparisons are observational across studies. It also argues for a prediction interval, which uses tau-squared to say where the effect in a new setting would plausibly fall and is far wider than the interval around the pooled mean.

The everyday misuse is an I-squared of zero from four small trials, quoted as evidence that they agree; Q has very low power with few studies, so a non-significant Q is weak reassurance rather than a demonstration of consistency. The reverse error is discarding a review because I-squared was eighty percent, when trials that all point the same way and differ only in magnitude are not in the same situation as trials pointing opposite ways.