An effect size is a standardized measure of the magnitude of a phenomenon, such as how far apart two group means are or how strongly two variables relate, expressed on a common scale that does not depend on the sample size. Because it strips out the units of the original measurement and the number of participants, an effect size lets you compare a result from a small reading-comprehension trial against a large blood-pressure study and ask the same question of both: how big is the difference, really? The most widely reported effect size for comparing two means is Cohen's d, the standardized mean difference between two groups divided by their pooled standard deviation. It tells you, in standard-deviation units, how far the average of one group sits from the average of the other.
Why the effect size matters more than the p-value
A common mistake is to treat a small p-value as proof of an important result. It is not. A p-value answers a narrow question: assuming there is no true effect, how surprising is the data we observed? With a large enough sample, even a trivial difference becomes statistically significant, because significance depends heavily on sample size. The effect size answers the question that actually matters to a clinician, teacher, or policymaker: how large is the difference, and is it big enough to care about? Two studies can share an identical p-value while describing effects that differ tenfold in real-world magnitude, so reporting the effect size alongside its confidence interval is now expected by most journals and reporting guidelines. Significance tells you whether an effect is detectable; the effect size tells you whether it is meaningful.
The standardized mean difference and Cohen's d
When an outcome is measured on a continuous scale, the natural building block is the standardized mean difference. You take the difference between the two group means and divide it by a measure of spread, which converts a difference expressed in the study's original units (millimetres of mercury, test points, kilograms) into a unitless number. Cohen's d is the best-known version. Its numerator is the difference between the treatment mean and the control mean, and its denominator is the pooled standard deviation, a weighted average of the two groups' spread. A Cohen's d of 1.0 means the two means lie one full standard deviation apart; a value of 0.5 means they differ by half a standard deviation. Because it is unitless, the same value carries the same interpretation whether the original outcome was measured in points, millilitres, or seconds.
Several related measures share this logic. Glass's delta divides by the control group's standard deviation alone, which is preferred when an intervention is expected to change the spread as well as the mean. The correlation coefficient r and the variance-explained measures such as eta squared and R squared are effect sizes for relationships and for analysis of variance. All of them serve the same purpose: turning a raw result into a portable, comparable quantity. If you have group means and standard deviations to hand, our standardized mean difference calculator will compute the value and its confidence interval for you.
Cohen's benchmarks, and why they are only a starting point
Jacob Cohen offered rough benchmarks to help researchers interpret a standardized mean difference when no field-specific yardstick exists: roughly 0.2 is small, 0.5 is medium, and 0.8 is large. These thresholds are genuinely useful as a first orientation, but Cohen himself warned that they were arbitrary conventions, not laws of nature. Their meaning is context-dependent. In a mature drug trial, a Cohen's d of 0.2 can represent a clinically valuable benefit worth millions of healthy life-years across a population, while in an early laboratory study an effect of 0.8 might be unremarkable. A sensible interpretation always asks what counts as a meaningful difference in the specific field, compares the value to effects reported in comparable studies, and reads the confidence interval rather than fixating on the point estimate. Treat the benchmarks as a vocabulary, not a verdict.
Hedges' g and the small-sample correction
Cohen's d has a known flaw: in small samples it is biased upward, meaning it systematically overestimates the true effect because the pooled standard deviation is estimated with limited data. The smaller the study, the larger the inflation. Hedges' g fixes this by multiplying Cohen's d by a correction factor (often written as J) that shrinks the estimate slightly toward zero, removing most of the bias. The correction depends on the degrees of freedom, so it is strong for tiny studies and negligible once samples grow into the dozens, where Hedges' g and Cohen's d become almost identical. Because meta-analyses routinely include small studies, Hedges' g is the default standardized mean difference in most modern meta-analysis software and methods guidance.
Why meta-analysis pools effect sizes, not p-values
The decisive reason effect sizes dominate evidence synthesis is that they, not p-values, are what a meta-analysis actually combines. A p-value carries no information about direction or magnitude and cannot be averaged meaningfully across studies. An effect size can: each study contributes a standardized estimate and a measure of its precision, usually the inverse of its variance, and the pooled result is a precision-weighted average in which larger, more precise studies count for more. This is the engine behind every forest plot you have seen, where each row is one study's effect size and confidence interval and the diamond at the bottom is the pooled estimate. The same machinery powers our pooled effect-size calculator, which weights and combines the standardized differences you enter.
Once effects are on a common scale, the remaining choices are about how to combine them. The overview of how a meta-analysis synthesises evidence explains why standardization comes first, and the comparison of fixed-effect and random-effects models determines how much weight to give large studies and how to handle differences between them. In both cases the unit being pooled is the effect size, which is exactly why getting it defined, corrected, and reported correctly is the foundation of a defensible synthesis.