When a meta-analysis pools many studies into one number, the pooled result is only as trustworthy as the agreement behind it. Statistical heterogeneity is the measure of that agreement: it describes studies that vary more than chance alone would explain, often because they used different patient groups, doses, or outcome measures. If heterogeneity is low, the single number is a fair summary. If it is high, the average may be hiding studies that point in genuinely different directions.
The most common yardstick is the I-squared index, expressed as a percentage. According to ScienceInsights, an I-squared of 0% means all the variation across studies is likely random, while an I-squared of 75% means three-quarters of the observed variation comes from real differences between studies rather than chance. That distinction changes how much weight a pooled figure deserves, and it is the first thing to check after the headline result.
This article explains what heterogeneity means in a meta-analysis, how I-squared and Cochran's Q work, and when disagreement between studies should make you read the pooled number more carefully. It is information, not medical or insurance advice.
What does heterogeneity mean in a meta-analysis?
In plain terms, heterogeneity means the parts of a group differ from each other. The word applies everywhere in science, from chemistry to genetics, but in research reviews it has a specific job. As Wikipedia's overview of homogeneity and heterogeneity puts it, study heterogeneity in meta-analysis is when multiple studies of an effect are measuring somewhat different effects, due to differences in the subject population, the intervention, the choice of analysis, or the experimental design.
That matters because pooling assumes the studies are estimating roughly the same thing. If one trial tested a drug in young healthy adults and another tested it in older patients with several conditions, the two results can differ for real reasons. Averaging them produces a number that may describe neither group.
Heterogeneity is partly a matter of perspective. A set of trials can be homogeneous on one dimension, say age, and heterogeneous on another, say baseline health status. Researchers try to control some sources of difference while measuring others, so "the same" always depends on what the comparison is.
How do I-squared and Cochran's Q measure it?
Two tools do most of the work. Cochran's Q test asks a yes-or-no question: is the variation across studies larger than random sampling would produce? It gives a p-value, and like any p-value it answers a narrow question. Our guide to what a p-value can and cannot tell you covers why a single threshold number rarely settles anything on its own.
The I-squared index goes further. Instead of a yes-or-no answer, it estimates what share of the observed variation reflects real differences between studies rather than chance. ScienceInsights describes the scale directly: 0% means the variation is probably all random; 75% means three-quarters of it is genuine difference, perhaps because studies used different patient populations, dosages, or outcome measures.
Two cautions apply. First, I-squared depends on how much the studies disagree relative to their own precision, not on an absolute standard of "good" or "bad." Second, a single percentage cannot tell you why studies differ. It flags the disagreement; it does not explain it.
When is a pooled result still reliable despite high heterogeneity?
High I-squared does not automatically invalidate a meta-analysis. It means the average deserves less weight on its own and more attention should go to why the studies disagree. A pooled estimate built on studies that measured the same intervention the same way in similar populations can survive a moderate I-squared. A pooled estimate that mixes different interventions, different comparison groups, and different outcome definitions is fragile no matter what the percentage says.
Our earlier guide, How to Read a Meta-Analysis: What Pooling Studies Can and Cannot Fix, walks through the other checks worth running before you trust a pooled figure, including how studies were selected and weighted. We covered a connected angle in Linked to Lower Risk: How to Read an Observational Study Before It Becomes a Headline.
A practical reading order: check what the studies pooled (design, population, intervention), then check I-squared, then check whether the authors explored the disagreement. If a headline reports only the average and never mentions heterogeneity, that omission is itself information.
What is subgroup analysis, and when does it help?
When studies disagree, researchers often split them into subgroups and re-pool each group separately. The question becomes: do the results line up once studies are grouped by, for example, dose, population, or outcome measure? If the subgroups show consistent results within themselves, the disagreement has an identifiable source, and each subgroup estimate is more informative than the overall average.
ScienceInsights makes the related clinical point that a modest average benefit can be misleading when it is actually a mix of large benefits for one subgroup and no benefit, or harm, for others. That is heterogeneity of treatment effects: real differences among patients in how they respond to a drug and how vulnerable they are to side effects. The same logic applies at the study level in a meta-analysis.
Subgroup analysis has limits. The more subgroups researchers test, the more likely some split looks meaningful by chance alone. A subgroup finding is a hypothesis about why studies differ, not a confirmed result, and it carries the same burden of evidence as any other claim.
What this means for reading health headlines
Our analysis: heterogeneity is the honesty check on pooling. A meta-analysis with low heterogeneity and a clear, consistent set of studies earns more confidence than one with a high I-squared and no exploration of why. Neither number replaces reading what the studies actually did.
Three habits cover most of it:
- Look for the I-squared figure alongside the pooled result. If it is high, ask what the authors say about why.
- Treat subgroup findings as leads, not conclusions, especially when many subgroups were tested.
- Remember that the same diagnosis or intervention can look very different across studies because the studies themselves differ, not because one is fraudulent.
For the wider toolkit, our Research section collects guides on trial design, risk figures, and how findings move from study to headline. And if the pooled result traces back to observational studies rather than trials, the design behind the finding changes what the claim can say; our piece on cohort studies versus randomized trials explains why. For related coverage, see Cohort Study or Randomized Trial? Why the Design Behind a Finding Changes the Claim.
What the evidence establishes is this: heterogeneity measures disagreement, I-squared quantifies it as a share of variation, and disagreement traced to a real source can be more informative than the average that once hid it. What remains unknown in any single review is which difference explains the disagreement, and that answer requires the subgroup work, or new studies, to be done.
