I-Squared

A statistic quantifying the percentage of total variation across studies in a meta-analysis attributable to genuine heterogeneity rather than chance.

Source: Higgins et al. 2003

The I-squared statistic expresses, as a percentage, how much of the total variation among study results reflects genuine differences in the true effect rather than chance. A value near zero means the scatter is about what sampling error alone would produce, while a high value means real heterogeneity dominates, so the studies are estimating meaningfully different effects. Reported as a single number, it offers a quick gauge of how consistent a body of studies is. The share of variation that is real is what it measures. Higgins and colleagues (2019) describe this statistic.

Source: Higgins et al. 2019

I-squared is calculated from a measure of heterogeneity, relating the observed variation across studies to the variation expected by chance, and expressing the excess as a percentage of the total. It ranges from zero, where all variation is due to chance, to one hundred per cent, where the variation reflects genuine differences. Rough conventions describe low, moderate, and high heterogeneity, though interpretation depends on context. So I-squared is calculated from the heterogeneity relative to chance and interpreted as the percentage of variation among studies due to real heterogeneity, guiding how consistent the studies are and how to combine them.

Source: Higgins et al. 2003

I-squared is useful because it provides an interpretable, scale-free summary of heterogeneity that does not depend directly on the number of studies or the metric, unlike the Cochran Q test, making it easier to compare across meta-analyses. It conveys how much of the variation among studies is genuine, which informs whether pooling is appropriate and how to interpret the combined estimate. So I-squared is valued for summarising heterogeneity in a standardised, comparable way, helping reviewers judge the consistency of the evidence and decide on the synthesis approach, which is why it is widely reported in meta-analyses.

Source: DerSimonian & Laird 1986

The limitations of I-squared include that it depends on the precision of the studies, so with large studies even small genuine differences can produce a high value, while with small studies real heterogeneity may give a low value, meaning it should not be interpreted mechanically; that it does not indicate the magnitude or direction of the heterogeneity; and that conventional thresholds for low, moderate, and high are rough. So I-squared is interpreted alongside other information, such as the estimate of the between-study variance and the clinical context, rather than relied upon alone, and its value is understood in light of the studies' sizes.

Source: Higgins et al. 2003

I-squared relates to other heterogeneity measures in that it is derived from the Cochran Q statistic, which tests for the presence of heterogeneity, by expressing the excess variation as a percentage; and it complements tau-squared, which estimates the actual between-study variance in the effect. Q tests whether heterogeneity is present, I-squared summarises its proportion of total variation, and tau-squared quantifies its magnitude on the effect scale. So I-squared is one of several related measures, providing a standardised proportion of variation due to heterogeneity, and is interpreted alongside Q and tau-squared for a fuller picture of the variation among studies.

Source: Higgins et al. 2003