i squared statistic meta-analysis: Meaning and Interpretation

The i squared statistic meta-analysis readers often use to describe inconsistency is I². It estimates the proportion of observed variation in study effect estimates attributed to differences between studies rather than sampling error. It is a descriptive measure—not a measure of effect size, clinical similarity, or the number of studies that disagree.[1][2]

I² is most useful when interpreted alongside the individual study estimates, their confidence intervals, the pooled result, and the clinical and methodological context of the studies. It should not be treated as a stand-alone verdict on whether results can be combined or which statistical model to use.[1][2]

What does I² measure in meta-analysis?

I² describes the relative extent of inconsistency among study effect estimates, rather than the absolute size of the differences between them.[1][2]

The statistic is based on observed variation across studies and distinguishes variation attributed to between-study heterogeneity from variation expected through sampling error. Because it is relative, the same I² can accompany different absolute amounts of variability on the effect scale. It therefore cannot, by itself, tell you whether differences between studies are large enough to matter for the review question.[1]

I² is also distinct from treatment-effect size. A synthesis can have a large or small pooled effect alongside different levels of inconsistency; those are separate features of the results. Nor does I² indicate whether the studies are clinically similar. That judgment requires examining what the studies measured and how they were conducted.[1][2]

Cochran’s Q and I² are calculated from the same heterogeneity statistic, but answer different questions. Q tests compatibility with a model that assumes no heterogeneity, whereas I² describes the relative extent of inconsistency. A Q-test result is therefore not a substitute for considering the magnitude and implications of variability.[1][2]

How is I² calculated?

The conventional estimator is I² = 100% × (Q − df) / Q, where Q is Cochran’s heterogeneity statistic and df is usually the number of studies minus one; negative estimates are set to zero.[2]

The formula connects I² to Q and its degrees of freedom, but the result remains an estimate, not a direct measurement of why study results differ. The formula alone also does not show the absolute spread of true effects. For that, τ² estimates between-study variance on the effect scale; a prediction interval can describe the expected range for a similar new study, subject to its assumptions and possible instability when there are few studies.[1][2]

Interpretation should account for uncertainty. When the evidence base is small or estimates are imprecise, the available data may not clearly distinguish heterogeneity from sampling error. In particular, an I² of zero does not establish that the underlying study effects are identical.[1][2]

How should you interpret I² values and thresholds?

Interpret I² values as rough guidance about relative inconsistency, not as fixed categories that determine what the results mean.[1][2]

Cochrane’s ranges are overlapping interpretive guidance, not mutually exclusive categories or validated cutoffs. A value should not be assigned a definitive label without considering uncertainty, the clinical and methodological diversity of studies, and what differences in effects would mean for the review question.[1][2]

Avoid interpreting I² as the percentage of studies that disagree. It is an estimate of the proportion of observed variation attributed to between-study heterogeneity rather than sampling error; it does not count studies or indicate how many point in different directions.[1][2]

The uncertainty around the evidence matters as well. An I² of zero can arise even when study effects are not identical, particularly when few or imprecise studies make heterogeneity difficult to estimate. Conversely, a heterogeneity test can flag inconsistency without telling you whether its magnitude is important.[1][2]

There is no universal threshold at which inconsistency automatically makes a synthesis unusable. A more defensible interpretation asks whether the observed variation changes the answer to the review question, and whether differences across studies in populations, interventions, outcomes, or methods offer a plausible context for that variation.[1]

What can I² not tell you?

I² cannot show the absolute spread of true effects, explain why study estimates differ, or establish that a pooled result is clinically meaningful.[1][2]

Because it is relative, I² does not quantify heterogeneity on the effect scale. τ² provides an estimate of between-study variance on that scale, while a prediction interval can describe a range of effects expected in a similar new study. Prediction intervals depend on assumptions and may be unstable when few studies are available, so they also require cautious interpretation.[1]

I² cannot identify the cause of inconsistency. A random-effects analysis allows for differences between study effects under its model, but it does not explain those differences. Explanations require considering the review question and the studies’ characteristics, rather than treating a heterogeneity statistic as a diagnosis.[1][2]

Nor does I² tell you whether an individual result or the pooled estimate is precise. That requires examining confidence intervals and the estimates themselves. A confidence interval that crosses the null is not proof of no effect; it indicates that the interval includes the null value for the chosen effect measure.[1][3]

How should I² inform model choice and pooling?

Use I² as one piece of evidence, not as an automatic rule for pooling or selecting a fixed-effect or random-effects model.[1][2]

The model should reflect the review question and the assumptions appropriate to the studies being synthesized. Neither I² nor the Q-test result should, by itself, decide whether pooling is appropriate. A random-effects model does not resolve clinical differences between studies or explain why their effects vary.[1][2]

A practical interpretation starts with the purpose of the synthesis: what effect is being estimated, and which studies does that estimate represent? Then consider the individual effects, their precision, the pooled estimate, and the clinical and methodological diversity of the evidence. Heterogeneity statistics can inform that assessment, but should not replace it.[1]

Automation can support reproducible calculation and make heterogeneity signals easier to review. Human reviewers still need to decide whether the studies address a coherent question, what variability means in context, and whether a summary estimate answers the review question. Those judgments should be reported transparently.[1][3]

How do you read I² alongside a forest plot?

Read the forest plot’s effect measure and null value first, then the study estimates and confidence intervals, the pooled result and its interval, and the heterogeneity information.[1][3]

Start by identifying the effect measure, its direction, and the value representing no effect. Then compare the study estimates and intervals: notice their spread, their precision, and whether their directions are consistent. Next assess the pooled estimate and confidence interval, without treating an interval that crosses the null as proof that there is no effect.[1][3]

Interpret I² after looking at the plotted results. It describes relative inconsistency, but does not show whether the variation is clinically important or explain its source. A forest plot provides a visual context for that judgment; it does not make the judgment on its own.[1][2]

What should a transparent I² report include?

A transparent report should state the model, effect measure, number of studies, pooled estimate and confidence interval, I² and τ² where relevant, and how inconsistency was assessed.[1][3]

Explain how the reported inconsistency relates to the review question and the clinical and methodological diversity of the studies. Avoid relying on an I² label alone, and do not present the statistic as a decision rule for pooling. Reporting both the calculation’s context and the interpretation helps readers distinguish the statistical summary from the review team’s judgment.[1][3]

Key takeaways

• I² estimates the relative contribution of between-study heterogeneity to observed variation, not effect size or clinical similarity.[1][2]
• The conventional I² estimator is derived from Q and its degrees of freedom, with negative estimates set to zero.[2]
• I² ranges are rough, overlapping guidance rather than validated cutoffs for deciding whether to pool.[1][2]
• An I² of zero does not prove that study effects are identical, especially when evidence is sparse or imprecise.[1][2]
• Interpret I² alongside study estimates, confidence intervals, τ², and the clinical and methodological context.[1]
• Report the model, effect measure, pooled result, heterogeneity statistics, and interpretation transparently.[1][3]

FAQ

What is the difference between Q and I²?

Both derive from the same heterogeneity statistic, but Q tests compatibility with a no-heterogeneity model, whereas I² describes the relative extent of inconsistency.[1][2] Neither statistic alone explains the cause or practical importance of differences between study results.

Does I² show how many studies disagree?

No. I² is not a count or percentage of studies that disagree; it estimates the proportion of observed variation attributed to between-study heterogeneity rather than sampling error.[1][2] Examine the study estimates and their intervals to assess the pattern of results.

What is the difference between τ² and I²?

I² describes heterogeneity relative to observed variation, while τ² estimates between-study variance on the effect scale.[1] They provide different information and should not be treated as interchangeable.

What does a prediction interval add?

A prediction interval can describe the expected range of effects for a similar new study, complementing the pooled estimate and its confidence interval.[1] Its interpretation depends on assumptions, and it may be unstable when few studies are available.