I-Squared and Tau-Squared as Heterogeneity Measures

Two Measures, Two Questions

A meta-analysis combines estimates from several cohorts, but the estimates may differ by more than expected from sampling error alone. Cochran’s \(Q\) provides a test of that compatibility. The measures in this tutorial address a different need: summarizing the amount of between-study heterogeneity and its scale. I-squared (\(I^2\)) describes heterogeneity relative to the observed dispersion of estimates; tau-squared (\(\tau^2\)) estimates the variance of study-specific true effects. They are related, but they are not interchangeable.

For \(k\) independent studies, let \(y_i\) be the estimated effect in study \(i\), and let \(v_i\) be its within-study sampling variance. The fixed-effect inverse-variance weight is \(w_i=1/v_i\), and the fixed-effect weighted mean is \(\bar{y}_w=\sum_i w_i y_i/\sum_i w_i\). Cochran’s statistic is \(Q=\sum_i w_i(y_i-\bar{y}_w)^2\), with \(df=k-1\) under the usual common-effect model. The definitions and calculations below use this setup and the corresponding \(Q\).

The maximum with zero is important: the formulas do not permit negative heterogeneity. If \(Q\le df\), both estimates are set to zero. This boundary estimate means that these particular estimators do not detect excess dispersion; it does not establish that all study-specific effects are identical.

Calculating the Measures

I-squared uses \(Q\) and \(df\) only. Tau-squared also depends on the studies’ fixed-effect weights through \(C\). Therefore, even when two meta-analyses have the same \(Q\), degrees of freedom, and I-squared, their tau-squared estimates can differ. The units also differ: if study effects are log odds ratios, tau-squared is in squared log-odds-ratio units; if effects are measured in millimeters of mercury, it is in squared millimeters-of-mercury units.

Why the Tau-Squared Formula Has This Form

The DerSimonian–Laird estimator is a method-of-moments estimator. Its rationale is that, under a model with independent study estimates and a common distribution of study-specific effects, the expected fixed-effect \(Q\) is approximately its degrees of freedom plus a term proportional to the between-study variance.

More precisely, suppose the study-specific true effects \(\theta_i\) have common mean \(\mu\) and variance \(\tau^2\), and the observed estimates satisfy \(y_i=\theta_i+\varepsilon_i\), where the independent sampling errors have mean zero and variance \(v_i\). With \(W=\operatorname{diag}(w_1,\ldots,w_k)\), \(w=(w_1,\ldots,w_k)^\mathsf{T}\), and \(P=W-ww^\mathsf{T}/\sum_i w_i\), the statistic can be written \(Q=y^\mathsf{T}Py\). Because \(P\) removes a shared mean, its contribution from the common mean is zero. The covariance of \(y\) is \(\operatorname{diag}(v_i)+\tau^2 I\); for a quadratic form, its expectation is the trace of \(P\) times this covariance. The two trace terms are:

Consequently, \(E(Q)=df+\tau^2 C\) under this model. Solving the moment equation \(Q=df+\tau^2 C\) for \(\tau^2\) gives \((Q-df)/C\). Since a variance cannot be negative, the estimate is truncated at zero. This justification also clarifies the assumptions behind the calculation: estimates should be independent, within-study variances should be appropriate, and the effects should be comparable and described by a meaningful common mean and between-study variance. Dependence or incompatible effect scales can make the standard formulas misleading.

Relative Heterogeneity Versus Effect-Scale Variance

I-squared is a relative summary, not an absolute effect-size measure. Tau-squared is an absolute variance on the chosen effect scale, but its value depends on that scale and on how precisely the studies estimate their effects. Neither value alone answers whether differences between cohorts are scientifically important. For that judgment, consider the size of plausible effect differences in context, the outcome and population, the precision of the evidence, and the purpose of the analysis.

Assumptions, Diagnostics, and Alternatives

Before interpreting either measure, check that each study estimates the same defined association on a harmonized scale and that the effect directions are aligned. Allele coding, outcome definition, covariate adjustment, and study population can all affect comparability. The standard formulas also rely on independent study estimates; overlapping cohorts can create dependence, so the usual \(Q\), I-squared, and tau-squared calculations may not be valid without accounting for it.

Check the inputs and the underlying results rather than treating a single summary as a diagnostic on its own. Confirm that \(Q\ge0\), \(df=k-1\), and all \(v_i>0\), so the weights are defined. Inspect the study estimates and their standard errors: a large I-squared can occur when sampling errors are small even if differences in effects are modest, while a low I-squared can accompany substantial uncertainty when studies are imprecise. With few studies, both measures can be unstable; \(Q\) also has limited power to detect heterogeneity in that setting.

Different tau-squared estimators are available, and they can yield different estimates, especially with few studies or boundary results. The DerSimonian–Laird estimate is calculated from \(Q\) and the fixed-effect weights; it is a particular method-of-moments choice, not a uniquely determined property of the data. I-squared is also an estimate, not a precise known proportion. Report the estimator used, the scale and units, and the study count alongside the values. These summaries describe heterogeneity; they do not identify its cause.

Reporting the Measures Responsibly

A clear report gives \(Q\), its degrees of freedom, the I-squared estimate, the tau-squared estimator, and the effect scale. Interpret each in its own terms. For example: “Across four independent cohorts, \(Q=9.6\) with \(df=3\), \(I^2=68.75\%\), and the DerSimonian–Laird estimate was \(\hat{\tau}^2=0.1886\) log-odds-ratio units squared.” That statement reports both relative and scale-based summaries without claiming that either establishes the clinical or biological importance of the variation.

The choice between these measures is not a contest. I-squared helps describe how much observed dispersion exceeds the sampling-error expectation in relative terms. Tau-squared estimates the between-study variance in the actual effect units. Reading them together is more informative than treating either as a stand-alone verdict.