Tau-Squared

A statistical parameter representing the estimated variance of the true treatment effect across studies in a random effects meta-analysis.

Source: DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188. doi:10.1016/0197-2456(86)90046-2.

Tau-squared represents the estimated variance of the true effect across studies in a random-effects meta-analysis, that is, how much the underlying effect genuinely differs from one study to another. A larger value means the studies are estimating more widely spread effects, which widens the confidence interval around the pooled result and gives smaller studies relatively more weight. Unlike a percentage such as I-squared, it is expressed on the scale of the effect itself. Quantifying the spread of true effects is its role. Borenstein and colleagues (2009) describe this parameter.

Source: Borenstein et al. 2009

Tau-squared is estimated within a random-effects meta-analysis from the observed variation among the study effects relative to the variation expected from sampling error alone, attributing the excess to between-study heterogeneity. Several estimators exist, such as the method of moments approach of DerSimonian and Laird and likelihood-based methods, which can give somewhat different values. The estimate carries its own uncertainty, especially with few studies. So tau-squared is estimated by separating the genuine between-study variation from the within-study sampling variation, using one of several available estimators, to quantify the heterogeneity in the true effects across the studies in the meta-analysis.

Source: Higgins et al. 2003

Tau-squared affects a random-effects meta-analysis by entering the study weights and widening the confidence interval of the pooled estimate: a larger tau-squared makes the study weights more similar and increases the uncertainty around the mean effect, reflecting the genuine variation in the true effects. When tau-squared is zero, the random-effects analysis reduces to the fixed-effect one. So tau-squared shapes both the weighting and the precision of a random-effects meta-analysis, with greater heterogeneity producing wider intervals and a more evenly weighted pooled estimate, which is why estimating it is central to the random-effects approach to combining studies.

Source: DerSimonian & Laird 1986

Tau-squared and I-squared both concern heterogeneity but express it differently: tau-squared estimates the absolute between-study variance in the true effects, on the scale of the effect measure, while I-squared expresses heterogeneity as the percentage of total variation across studies that is due to genuine differences rather than chance. Tau-squared gives the magnitude of the variation, whereas I-squared gives a relative, scale-free proportion. So the two differ in that tau-squared measures the size of the between-study variance directly and I-squared its share of the total variation, and they are often reported together to describe heterogeneity fully.

Source: Higgins et al. 2003

Tau-squared is important because it quantifies the actual magnitude of between-study heterogeneity, which determines how much the true effects vary and thus how the studies should be combined and how uncertain the pooled estimate is. Knowing tau-squared informs the width of prediction intervals for the effect in a new setting and the interpretation of the mean effect. So tau-squared matters for understanding and conveying the variability in the true effects across studies, supporting appropriate weighting, honest uncertainty, and prediction, which is why it is a key output of a random-effects meta-analysis rather than a purely technical detail.

Source: DerSimonian & Laird 1986