How this instrument works
A p-value tells you whether a difference between two groups is likely to be real rather than chance; it says nothing about how big that difference is. Cohen's d fills that gap: it expresses the difference between two sample means as a number of pooled standard deviations, so a d of 0.5 means the two group averages differ by half a standard deviation, regardless of what the original units were — test scores, milliseconds, milligrams, dollars.
Computing it takes two steps. First, the pooled standard deviation, sp, blends the spread of both samples into a single number, weighted by each sample's size so a larger sample's variance counts for more. Then d is simply the difference between the two group means divided by that pooled spread. The result is unitless, which is exactly what makes it comparable across different studies, different measurement scales, and different fields — a d of 0.8 in a psychology study and a d of 0.8 in a medical trial describe the same-sized effect, standardized.
Jacob Cohen, who introduced the measure, proposed rough benchmarks for interpreting it: d = 0.2 as a small effect, 0.5 as medium, and 0.8 as large. These are widely cited starting points, not strict cutoffs — a d = 0.3 difference can be practically important in one field and trivial in another, so context always matters more than the label alone.
- Enter each group's average into Sample 1 mean (x-bar-1) and Sample 2 mean (x-bar-2).
- Enter each group's spread into Sample 1 standard deviation (s1) and Sample 2 standard deviation (s2) — both must be greater than zero.
- Enter each group's size into Sample 1 size (n1) and Sample 2 size (n2) — both need at least 2 observations.
- Read Pooled standard deviation (sp) first — it's the combined spread the two means are measured against.
- Read Cohen's d (effect size) — its sign shows which group scored higher, and its magnitude shows how large that gap is in standard-deviation units.
Worked example — two groups of 25, means 105 and 100
Enter 105 into Sample 1 mean (x-bar-1), 100 into Sample 2 mean (x-bar-2), 15 into both Sample 1 standard deviation (s1) and Sample 2 standard deviation (s2), and 25 into both Sample 1 size (n1) and Sample 2 size (n2). The instrument first computes the pooled standard deviation: sp = √(((24×225)+(24×225))/48) = √(10800/48) = √225 = 15.0000.
Cohen's d (effect size) then reads 0.333333, computed as (105 − 100) / 15 = 5 / 15. By Cohen's own conventions (0.2 small, 0.5 medium, 0.8 large), that lands between small and medium — a real but modest difference between the two groups, well short of a large effect.
Questions
What does a Cohen's d of 0.33 actually mean?
It means the two group means differ by roughly a third of a pooled standard deviation — a small-to-medium effect by Cohen's own benchmarks (0.2 small, 0.5 medium, 0.8 large). It says nothing directly about statistical significance; a small sample can produce a large d that isn't statistically significant, and a huge sample can produce a tiny, highly significant d.
Why does Cohen's d use a pooled standard deviation instead of just one group's?
Because it needs a single, shared yardstick to express the mean difference in, and pooling both samples' variability — weighted by their sizes — gives a more stable estimate than relying on either group alone. Using only one group's standard deviation would make d sensitive to which group you happened to pick as the reference, which defeats the point of a standardized measure.
What's the difference between Cohen's d and a p-value?
A p-value answers 'is this difference likely to be real, or just chance?' Cohen's d answers 'how big is this difference, in standardized units?' They're complementary, not interchangeable — a tiny, practically meaningless difference can still be statistically significant with a large enough sample, while a genuinely large effect can fail to reach significance in a small one. Reporting both gives a fuller picture than either alone.
Can Cohen's d be negative?
Yes — the sign simply reflects which mean was entered first. If Sample 1 mean (x-bar-1) is lower than Sample 2 mean (x-bar-2), d comes out negative, of the same magnitude it would have with the two means swapped. The magnitude, not the sign, is what indicates the size of the effect; the sign just tracks direction.
Are Cohen's 0.2 / 0.5 / 0.8 benchmarks universal?
No — Cohen himself described them as rough, field-general starting points, not strict thresholds. In fields where effects are typically small (some areas of psychology or medicine), a d of 0.3 might be considered meaningfully large in context; in fields where large effects are common, the same 0.3 might be unremarkable. Always interpret d alongside domain knowledge of what a 'big' difference typically looks like in that specific area.
Why do both sample sizes need to be at least 2?
Because computing a standard deviation requires at least two data points to have any spread to measure at all — with a single observation, s1 or s2 would be undefined. The pooled standard deviation formula also divides by (n1+n2−2), which would be zero or negative if either sample had fewer than 2 observations.