What is Cohen's d?
Cohen's d is a standardized measure of effect size that expresses the difference between two group means in units of standard deviation. It is widely used in psychology, education, and medicine to quantify how large a difference is, independent of sample size. Unlike a p-value, which only tells you whether a difference is statistically significant, Cohen's d tells you how meaningful that difference is.
How to use this calculator
Enter the mean, standard deviation, and sample size for each of your two groups. The calculator computes the pooled standard deviation and divides the difference in means by it to give Cohen's d. It also classifies the result as negligible, small, medium, or large using Cohen's conventional benchmarks.
The formula explained
The effect size is $$d = \frac{\bar{x}_1 - \bar{x}_2}{s_p}.$$ The pooled standard deviation is $$s_p = \sqrt{\frac{(n_1-1)\cdot s_1^2 + (n_2-1)\cdot s_2^2}{n_1 + n_2 - 2}}.$$ This weights each group's variance by its degrees of freedom, giving a single combined spread used to standardize the mean difference.
Worked example
Suppose Group 1 has mean 100, SD 10, n 30, and Group 2 has mean 90, SD 12, n 30. The pooled variance is $$\frac{(29\cdot 100)+(29\cdot 144)}{58} = \frac{2900+4176}{58} = 122.$$ So \(s_p = \sqrt{122} \approx 11.0454\), and \(d = (100-90)/11.0454 \approx 0.905\) — a large effect.
Cohen's d Benchmarks and Overlap Reference
Cohen's d expresses a difference between two group means in pooled standard deviation units. Cohen (1988) suggested 0.2, 0.5 and 0.8 as rough labels for small, medium and large effects. Sawilowsky (2009) later added 1.2 (very large) and 2.0 (huge). The overlap figures below assume two normal distributions with equal variance and are computed from d.
| Cohen's d | Label | Distribution overlap (OVL) | Cohen's U3 (share of group 2 below the group 1 mean) | Probability of superiority |
|---|---|---|---|---|
| 0.2 | Small | ≈ 92.0% | ≈ 57.9% | ≈ 55.6% |
| 0.5 | Medium | ≈ 80.3% | ≈ 69.1% | ≈ 63.8% |
| 0.8 | Large | ≈ 68.9% | ≈ 78.8% | ≈ 71.4% |
| 1.2 | Very large | ≈ 54.9% | ≈ 88.5% | ≈ 80.2% |
| 2.0 | Huge | ≈ 31.7% | ≈ 97.7% | ≈ 92.1% |
The formulas behind the table are \(\text{OVL} = 2\,\Phi(-d/2)\), \(U_3 = \Phi(d)\) and probability of superiority \(= \Phi(d/\sqrt{2})\), where \(\Phi\) is the standard normal cumulative distribution function.
- Values of |d| below 0.2 are usually described as negligible.
- Cohen described these benchmarks as rough conventions for use when no better field-specific reference exists. They are not rigid cutoffs, and a d of 0.49 is not meaningfully different from 0.50.
Interpreting Your Cohen's d Result
d is measured in standard deviation units. A d of 0.5 means the two group means differ by half a pooled standard deviation. A d of 1.0 means they differ by one full pooled SD.
Sign versus magnitude. The sign only shows which group has the higher mean. With \(\text{Mean}_1 - \text{Mean}_2\), a positive d means group 1 is higher and a negative d means group 2 is higher. The size of the effect is the absolute value |d|, which is what the small/medium/large labels refer to.
Effect size is not statistical significance. A p-value depends heavily on sample size. A tiny d can be "significant" with thousands of participants, and a large d can be non-significant with a handful. Cohen's d does not grow with n, so it describes how big the difference is, not how certain you are that it exists. Reporting a confidence interval for d alongside it is good practice.
Context matters. The labels 0.2, 0.5 and 0.8 are generic. In some fields, such as education or medicine, a d of 0.2 can be important when the outcome is cheap to change or affects many people. In tightly controlled lab research, 0.2 may be unremarkable. Compare your d with effects reported in your own field and judge the practical consequences.
Small-sample bias. Cohen's d slightly overestimates the population effect when samples are small. Hedges' g corrects this approximately:
$$g \approx d \times \left(1 - \frac{3}{4(n_1 + n_2) - 9}\right)$$
For example, with \(n_1 + n_2 = 80\) the factor is \(1 - 3/311 \approx 0.990\), so a d of 0.53 becomes a g of about 0.53. The correction matters mostly when the total sample is below roughly 20 to 50.
Assumptions of the pooled SD.
- The two populations have roughly equal variances. If the SDs differ a lot, a pooled SD can be misleading, and alternatives such as Glass's Δ (using the control group SD) may fit better.
- The data are approximately normal. Heavy skew or outliers distort both means and SDs, and the overlap percentages in the reference table no longer hold exactly.
- The groups are independent. For paired or repeated-measures designs, a different effect-size formula is needed.
Scenario Comparison: How Inputs Change d
Each row uses the pooled SD \(s_p = \sqrt{\dfrac{(n_1-1)s_1^2 + (n_2-1)s_2^2}{n_1+n_2-2}}\) and \(d = (\text{Mean}_1 - \text{Mean}_2)/s_p\). The classification uses Cohen's conventions.
| Scenario | Group 1 (mean, SD, n) | Group 2 (mean, SD, n) | Pooled SD | d | Classification |
|---|---|---|---|---|---|
| A. Baseline | 105, 15, 30 | 100, 15, 30 | 15.00 | 0.33 | Small |
| B. Same difference (5), larger SD | 105, 25, 30 | 100, 25, 30 | 25.00 | 0.20 | Small |
| C. Same d as A, larger samples | 105, 15, 100 | 100, 15, 100 | 15.00 | 0.33 | Small |
| D. Unequal n and unequal SD | 82, 10, 20 | 75, 14, 60 | 13.14 | 0.53 | Medium |
| E. Larger gap, similar SDs | 78, 8, 25 | 70, 10, 25 | 9.06 | 0.88 | Large |
Worked checks
Scenario D: $$s_p = \sqrt{\frac{19(10^2) + 59(14^2)}{20 + 60 - 2}} = \sqrt{\frac{1900 + 11564}{78}} = \sqrt{172.62} \approx 13.14$$ $$d = \frac{82 - 75}{13.14} \approx 0.53$$
Scenario E: $$s_p = \sqrt{\frac{24(8^2) + 24(10^2)}{48}} = \sqrt{\frac{1536 + 2400}{48}} = \sqrt{82} \approx 9.06$$ $$d = \frac{78 - 70}{9.06} \approx 0.88$$
What the comparison shows
- A vs. B: the raw mean difference is 5 in both, but the larger spread in B shrinks d from 0.33 to 0.20.
- A vs. C: sample size does not change d. Only the means and SDs do. A larger n would make the estimate more precise and the result more likely to be statistically significant.
- D: with unequal n, the pooled SD is weighted toward the larger group's SD, so it lies closer to 14 than to 10.
FAQ
What counts as a large effect? By Cohen's rules of thumb, 0.2 is small, 0.5 medium, and 0.8 or above large.
Can d be negative? Yes — a negative value simply means Group 2's mean is higher than Group 1's. The magnitude (absolute value) is what matters for size.
Do the sample sizes need to be equal? No. The pooled formula weights by degrees of freedom, so unequal group sizes are handled correctly.