What is a good I-squared value in a meta-analysis?

There is no single good value, only rough bands. As a guide, I-squared around 25 percent is low, around 50 percent is moderate, and 75 percent or more is high. These cut-offs are not absolute, so read I-squared alongside Cochran's Q, its p-value, and tau-squared, and consider the clinical and methodological diversity of your studies before deciding.

What is the difference between Cochran's Q and I-squared?

Cochran's Q is the weighted sum of squared deviations of each study from the pooled estimate, and it grows with the number of studies, so its size is hard to interpret on its own. I-squared rescales that same information onto a 0 to 100 percent scale that estimates the proportion of total variation due to genuine between-study heterogeneity rather than chance, which makes it easier to compare across analyses.

What does tau-squared tell me?

Tau-squared is the estimated variance of the true effects across studies, on the analysis scale. Unlike I-squared, which is a relative measure, tau-squared is an absolute measure of how much the underlying effects differ. It is the quantity a random-effects model adds to every study's variance, so a larger tau-squared widens the pooled confidence interval and pulls study weights closer together.

Why is the Q test sometimes not significant when I-squared is high?

Cochran's Q has low statistical power when there are few studies, so a high I-squared can sit beside a non-significant Q p-value, especially with a small number of trials. For that reason a non-significant Q should not be read as proof that heterogeneity is absent. Inspect I-squared and tau-squared together rather than relying on the Q p-value alone.

Does high heterogeneity mean I should not pool the studies at all?

Not automatically. High heterogeneity is a signal to investigate rather than an instant veto. If the variation can be explained by a clear factor, such as dose or population, a subgroup analysis or meta-regression may be more informative than one pooled number. If the studies are simply too diverse to share a meaningful average, a structured narrative synthesis is the honest alternative.

How does heterogeneity affect which pooling model I choose?

Low heterogeneity is consistent with a fixed-effect model, which assumes one shared true effect. Once there is appreciable between-study variance, the random-effects model is more appropriate because it builds tau-squared into every weight and into the pooled confidence interval. In practice most reviews default to random-effects because some genuine variation is almost always present.

Heterogeneity calculator (I-squared, Q, tau-squared)

Paste your included studies and see how much of the variation between them is real rather than chance, with the full set of heterogeneity statistics you need to choose a pooling model.

Heterogeneity is the genuine variation in true effects from one study to another, beyond what sampling error alone would produce. This calculator quantifies it with Cochran’s Q and its chi-square p-value, I-squared (the percentage of total variation that is due to real between-study variance rather than chance), the H statistic, and tau-squared, the estimated variance of the true effects. Together these tell you whether a random-effects model is warranted for your meta-analysis.

Effect measure

What you are entering

Confidence levelApplies to the intervals you enter and to every interval reported.

Drag and drop or click. CSV, TSV, Excel; header row with columns like study, label, author, trial, estimate.

Your studies, one per lineEach line: optional label, then estimate, lower limit, upper limit.

I-squared

30.8%

Moderate heterogeneity (around 50%). Favour a random-effects model and explore sources.

Heterogeneity statistics

Cochran's Q4.33

Degrees of freedom (df)3

Q p-value (chi-square)0.2276

H1.20

Tau-squared (between-study variance)0.0116

Number of studies4

Pooled context

Random-effects pooled estimate1.31 (95% CI 1.09 to 1.59)

Prediction interval (effect in a new study)0.70 to 2.45

Report-ready text

Across 4 studies, heterogeneity was moderate (I-squared = 30.8%, tau-squared = 0.0116, Q(3) = 4.33, p = 0.2276). The random-effects pooled estimate was 1.31 (95% CI 1.09 to 1.59). The 95% prediction interval for the effect in a new study ranged from 0.70 to 2.45.

What heterogeneity is and why it matters

When you place several studies side by side, their effect estimates never line up perfectly. Some of that scatter is pure chance: every study is a sample, and samples wobble around the truth. Heterogeneity is the part of the scatter that is not chance, the genuine variation in the underlying effect from one study to another because populations, interventions, settings, and outcome definitions differ. Quantifying it is the step that tells you whether a single pooled number is a fair summary or a convenient fiction, which is why it sits at the centre of any defensible meta-analysis workflow.

The four statistics this tool reports answer slightly different questions. Cochran’s Q tests whether the observed scatter exceeds what chance alone would produce. I-squared rescales that information onto a 0 to 100 per cent scale that is comparable across analyses. The H statistic expresses the same idea as a ratio relative to one. Tau-squared is the absolute variance of the true effects on the analysis scale, the quantity a random-effects model actually uses.

Reading the I-squared bands

The most widely cited guide treats I-squared near 25 per cent as low, near 50 per cent as moderate, and 75 per cent or above as high. Treat these as soft landmarks, not legal thresholds. An I-squared of 60 per cent in a set of clinically similar trials may be unremarkable, while 40 per cent across studies you expected to agree closely may deserve a hard look. Crucially, I-squared describes the proportion of variation that is real, not the amount; a large I-squared with a small absolute effect spread can matter less than a moderate I-squared with a wide one. That is why tau-squared and a prediction interval add information that a percentage cannot. The choice this all feeds into is covered in our guide to fixed-effect versus random-effects modelling.

Why the Q p-value can mislead

Cochran’s Q has low power when there are few studies, so with five or six trials it can return a non-significant p-value even when real heterogeneity is substantial. The reverse happens with dozens of large studies, where Q turns significant on trivial, clinically irrelevant differences. For this reason a non-significant Q is never proof that the studies agree. Always read Q, its p-value, I-squared, and tau-squared as a set. A deeper treatment lives in our article on interpreting and explaining heterogeneity.

A worked example

Imagine seven trials with a Cochran’s Q of 18.2 on six degrees of freedom. The expected value of Q under no heterogeneity equals its degrees of freedom, so 18.2 against an expected 6 already signals excess variation. I-squared works out at about (18.2 minus 6) divided by 18.2, which is around 67 per cent, in the moderate-to-high zone. Suppose tau-squared comes back at 0.04 on the log scale. Together these say roughly two thirds of the observed variation is real, and the spread of true log effects has a standard deviation near 0.2. You would report a random-effects pooled estimate, flag the heterogeneity openly, and try to explain it through subgroup analysis or meta-regression.

Common mistakes

Reading I-squared as an error rate. It is not the chance you are wrong; it is the share of total variation attributable to genuine between-study differences.
Trusting a non-significant Q with few studies. Low power means absence of significance is not evidence of homogeneity.
Switching to fixed-effect to make the interval narrower. A tighter interval bought by ignoring real variation is false precision, not a better result.
Stopping at the number. High heterogeneity is the start of an investigation into why effects differ, not a value to report and move past.
Confusing heterogeneity with publication bias. They are different problems; a funnel plot and the issues in our forest plot reading guide help separate genuine variation from missing studies.

How it works

Each study is placed on the analysis scale and given an inverse-variance weight. For ratio measures (odds ratios, risk ratios, hazard ratios) the estimate and interval are logged first, because heterogeneity is assessed on the log scale. The weight for each study is the inverse of its variance:

w_i = 1 / SE_i^2

Cochran’s Q is the weighted sum of squared deviations of each study from the inverse-variance pooled mean, with degrees of freedom one less than the number of studies:

Q = sum( w_i * (y_i - y_pooled)^2 )

df = k - 1

From Q the calculator derives the remaining statistics, where I-squared and H describe relative heterogeneity and tau-squared (DerSimonian-Laird method of moments) is the absolute between-study variance:

I^2 = max(0, (Q - df) / Q) * 100%

H = sqrt( max(1, Q / df) )

C = sum(w_i) - sum(w_i^2) / sum(w_i)

tau^2 = max(0, (Q - df) / C)

The I-squared interpretation band (low near 25 percent, moderate near 50 percent, high near 75 percent) follows common practice but the cut-offs are guides, not absolute thresholds.

Frequently asked questions

What is a good I-squared value in a meta-analysis?: There is no single good value, only rough bands. As a guide, I-squared around 25 percent is low, around 50 percent is moderate, and 75 percent or more is high. These cut-offs are not absolute, so read I-squared alongside Cochran's Q, its p-value, and tau-squared, and consider the clinical and methodological diversity of your studies before deciding.
What is the difference between Cochran's Q and I-squared?: Cochran's Q is the weighted sum of squared deviations of each study from the pooled estimate, and it grows with the number of studies, so its size is hard to interpret on its own. I-squared rescales that same information onto a 0 to 100 percent scale that estimates the proportion of total variation due to genuine between-study heterogeneity rather than chance, which makes it easier to compare across analyses.
What does tau-squared tell me?: Tau-squared is the estimated variance of the true effects across studies, on the analysis scale. Unlike I-squared, which is a relative measure, tau-squared is an absolute measure of how much the underlying effects differ. It is the quantity a random-effects model adds to every study's variance, so a larger tau-squared widens the pooled confidence interval and pulls study weights closer together.
Why is the Q test sometimes not significant when I-squared is high?: Cochran's Q has low statistical power when there are few studies, so a high I-squared can sit beside a non-significant Q p-value, especially with a small number of trials. For that reason a non-significant Q should not be read as proof that heterogeneity is absent. Inspect I-squared and tau-squared together rather than relying on the Q p-value alone.
Does high heterogeneity mean I should not pool the studies at all?: Not automatically. High heterogeneity is a signal to investigate rather than an instant veto. If the variation can be explained by a clear factor, such as dose or population, a subgroup analysis or meta-regression may be more informative than one pooled number. If the studies are simply too diverse to share a meaningful average, a structured narrative synthesis is the honest alternative.
How does heterogeneity affect which pooling model I choose?: Low heterogeneity is consistent with a fixed-effect model, which assumes one shared true effect. Once there is appreciable between-study variance, the random-effects model is more appropriate because it builds tau-squared into every weight and into the pooled confidence interval. In practice most reviews default to random-effects because some genuine variation is almost always present.

The guides behind this tool

Understanding heterogeneity and I-squared Fixed-effect vs random-effects models How to do a meta-analysis How to read a forest plot

More review and meta-analysis tools

Meta-analysisPool study effect estimates with fixed-effect and random-effects models, with the heterogeneity statistics shown.Open tool Odds ratio / risk ratioTurn a 2x2 table of events into an odds ratio, risk ratio, and risk difference, each with a 95% confidence interval.Open tool Standardized mean differenceCompute Cohen's d and Hedges' g from two groups' means, standard deviations, and sample sizes.Open tool