I² is the percentage of the variability among the effect sizes in your meta-analysis that is not explained by sampling error, but by real differences between studies. If I² = 0%, all the spread you see in the forest plot fits within chance; if I² = 75%, three-quarters of that spread reflects the fact that the studies are not estimating the same thing. The tentative thresholds from Higgins and Thompson (2002) are 25% (low), 50% (moderate) and 75% (high), but keep this in mind: I² does not tell you how much heterogeneity there is in effect size units, only what proportion of the total variation is heterogeneity. That is why an I² of 80% across very large studies can be practically irrelevant, and why you should never report it on its own, always alongside τ² and the prediction interval.
Heterogeneity is possibly the most important and most poorly managed concept in the practice of meta-analysis. When we combine the results of several studies investigating the same question, the first issue we must consider is not whether the combined effect is significant, but whether it makes sense to combine those studies in the first place. If the individual studies show very disparate results (some find large effects and others find no effect, or even find effects in the opposite direction), calculating a mean effect may obscure more than it reveals.
The I-squared statistic (I²) is the most widely used measure of heterogeneity today, and its popularity is partly due to its apparent interpretive simplicity. I² expresses the percentage of total variability in effect sizes that is due to real heterogeneity between studies, rather than sampling error. An I² of 0% would mean that all variability between studies is explained by chance. An I² of 75% would mean that three-quarters of the variability reflects real differences between studies. Higgins and Thompson (2002) proposed tentative cutoff points: 25% for low heterogeneity, 50% for moderate, and 75% for high (you can use our heterogeneity calculator to obtain these values automatically).
Limitations of I² that few people know about
However, I² has important limitations that many users are unaware of. The most fundamental is that I² depends on the precision of the individual studies, not only on the magnitude of real heterogeneity. With very large and precise studies, even small amounts of real heterogeneity produce very high I² values. With small and imprecise studies, substantial amounts of real heterogeneity can produce moderate I² values. This means that the same I² value can mean very different things depending on the size and precision of the studies included in the meta-analysis.
Borenstein et al. (2017) have argued convincingly that I² should not be used as the sole measure of heterogeneity, and that researchers should supplement it with other statistics that provide different information. The Q statistic (Cochran's test) evaluates whether statistically significant heterogeneity exists, but it has low power when the number of studies is small (which is typical) and detects trivial heterogeneity when the number of studies is large. Tau-squared (τ²) estimates the true between-study variance on the effect size scale, allowing assessment of the absolute magnitude of dispersion. And the prediction interval, which many meta-analyses do not report, indicates the range within which the effect of a future study would be expected to fall, and is probably the most informative measure for evaluating the practical implications of heterogeneity.
What to do when heterogeneity is high
When I² is high, the answer is not simply to report the figure and move on. High heterogeneity is a signal that there are moderators, that is, characteristics of the studies that explain why some find larger effects than others. Identifying these moderators is, in many ways, more informative than calculating the overall combined effect. Moderators can be methodological (type of design, study quality, measurement instrument used), population-related (age, gender, clinical severity of the sample), or contextual (country, type of intervention, duration of treatment).
Moderation analyses in meta-analysis are conducted through meta-regression (when the moderator is continuous) or through subgroup analyses (when the moderator is categorical). Meta-regression is conceptually similar to a weighted regression where studies are the observations and effect size is the dependent variable. The practical problem is that most meta-analyses include a small number of studies (frequently fewer than 20), which severely limits the number of moderators that can be examined simultaneously without overfitting the model. A practical rule of thumb is not to include more than one moderator for every ten studies.
And before you go hunting for moderators, check something simpler: that the heterogeneity is not coming from one or two outlying studies. A Galbraith plot shows you this at a glance, and a sensitivity analysis dropping those studies will tell you how much they really weigh on the pooled result. If excluding a single study makes I² drop from 78% to 30%, you do not have a heterogeneity problem: you have one odd study that needs explaining in the discussion.
Fixed-effects vs. random-effects models
The choice between a fixed-effects model and a random-effects model is directly related to heterogeneity. The fixed-effects model assumes that all studies are estimating exactly the same population effect, and that the only source of variation between estimates is sampling error. The random-effects model assumes that studies are estimating effects that vary from one another, and that the variation reflects both sampling error and real differences between studies.
In practice, the random-effects model is almost always more appropriate, whether you work in health, education, biology or economics, because studies differ in their samples, procedures, instruments, and contexts, and it would be naive to assume that all are estimating exactly the same value. The fixed-effects model is only defensible when the studies are near-exact replications of the same protocol, which rarely occurs. Furthermore, with high heterogeneity, the confidence intervals of the fixed-effects model are artificially narrow, which can give a false sense of precision (a problem compounded when publication bias is also present). The random-effects model produces wider and more honest intervals that reflect the additional uncertainty caused by between-study variability.
How to get I², Q, τ² and the prediction interval in R
The good news is that all four come out of the same object, the same rma() you later use to draw your forest plot in R with metafor. If your data already contain each study's effect size (yi) and its variance (vi), this is all you need:
library(metafor)
# dat: yi = effect size, vi = variance of each study
res <- rma(yi, vi, data = dat, method = "REML")
summary(res)
res$I2 # % of variability that is not sampling error
res$H2 # total variability / variability expected by chance (1 = homogeneity)
res$tau2 # true between-study variance, on the effect size scale
res$QE # Cochran's Q
res$QEp # its p value
# Where the effect of a future study would fall
predict(res, digits = 3)H² is the poor relative of I² and deserves more attention: it is the ratio between the total observed variability and the variability you would expect from sampling error alone. It equals 1 under homogeneity and grows without an upper bound, so it separates a moderately heterogeneous meta-analysis from a disastrous one, two situations where I² would look similar because it is capped at 100%. The defensible minimum to carry into the manuscript is: k, the pooled effect with its 95% CI, Q with its degrees of freedom and p value, I², τ (better than τ², because it lives on the effect size scale and reads just like a Cohen's d) and the prediction interval.
That prediction interval is what really changes the conversation. You can have a pooled effect of d = 0.45 with a 95% CI of [0.30, 0.60], which looks solid, and a prediction interval of [-0.20, 1.10], which is telling you the next study could come out in the opposite direction. Reporting the first and staying quiet about the second is the most common way of selling more certainty than you actually have.
Before you submit: poorly justified heterogeneity is one of the things that most often stalls a meta-analysis in review, and it is almost always the same story: a high I² reported with no τ², no prediction interval and no attempt to explain where it comes from. If your manuscript is already written, run it through the Q1 Paper Reviewer, a free Reviewer 2 style pre-review that tells you what they will push back on before the journal does.