Written and maintained by CASRAI Editorial Board
Published
On this page
Cohen’s h is an effect size for the difference between two proportions. It converts each proportion with the arcsine square-root transformation, φ = 2·arcsin(√p), and subtracts the results: h = φ₁ − φ₂. Cohen’s conventional labels read |h| = 0.20 as small, 0.50 as medium and 0.80 as large, but Cohen himself cautioned against leaning on them when theory or experience gives better values.
Definition: what Cohen’s h measures
Cohen’s h describes how far apart two proportions are on a scale where equal values of h are meant to be equally detectable. A plain difference in percentage points does not have that property: a gap between 5% and 1% is much harder to detect, and means something different, than a gap between 55% and 51%. The arcsine square-root transformation stretches the scale near 0 and 1 so that proportions at the extremes are treated with the same sensitivity as proportions near 0.5 (the variance-stabilising idea). The measure comes from Jacob Cohen’s power-analysis work, where it is used to size studies that compare two proportions, and it is also reported after the fact to separate statistical significance from practical size.
It belongs with the other families of effect size measures: Cohen’s d for mean differences, Cohen’s h for proportions, and Cramér’s V for association in a contingency table. For the wider picture of why effect sizes are reported at all, see effect size in research.
Cohen’s h formula, with every symbol explained
For two proportions p₁ and p₂, each between 0 and 1:
- φ = 2·arcsin(√p): the arcsine square-root transformation of a proportion. arcsin is the inverse sine function and the result is in radians, so φ runs from 0 (when p = 0) to π ≈ 3.1416 (when p = 1).
- h = φ₁ − φ₂ = 2·arcsin(√p₁) − 2·arcsin(√p₂): the directional form. Its sign tells you which proportion is larger.
- |h| = |φ₁ − φ₂|: the nondirectional form, used when only the size of the gap matters.
- p₁, p₂: the two observed (or hypothesised) proportions, written as decimals, for example 0.50 and 0.40, not 50 and 40.
The R function ES.h in the pwr package documents exactly this definition: “The effect size is 2*asin(sqrt(p1))-2*asin(sqrt(p2))”, and cites Cohen (1988), Statistical Power Analysis for the Behavioral Sciences (2nd ed.).
Worked example: 50% versus 40%
Suppose 50% of participants in one group and 40% in another show an outcome (p₁ = 0.50, p₂ = 0.40). The arithmetic:
- √0.50 = 0.707107, and arcsin(0.707107) = 0.785398 radians (π/4). Doubling gives φ₁ = 1.570796.
- √0.40 = 0.632456, and arcsin(0.632456) = 0.684719 radians. Doubling gives φ₂ = 1.369438.
- h = 1.570796 − 1.369438 = 0.2014.
The pwr documentation’s own example uses the same pair (ES.h(0.5, 0.4), taken from Cohen’s Exercise 6.5), and the value 0.2014 agrees with the arithmetic above. By the conventional labels, a ten-percentage-point gap at this baseline is just over a small effect.
Why the same percentage-point gap gives different values of h
The data table below applies the formula to six pairs of proportions. Three of them differ by ten percentage points yet give three different values of h, because the same gap is a larger step on the transformed scale when the proportions sit nearer 0 or 1. The pairs 20% vs 10% and 90% vs 80% give identical h because the transformation is symmetric around 0.5.
Interpreting Cohen’s h: reading guidance
| |h| | Conventional label (Cohen) | How to read it |
|---|---|---|
| Below 0.20 | Below the “small” convention | The two proportions are close on the transformed scale; look at the confidence interval before saying anything about practical size. |
| 0.20 | Small | Cohen’s lower benchmark. The 50% vs 40% example sits just above it at 0.2014. |
| 0.50 | Medium | Cohen’s mid benchmark. For orientation: 55% vs 35% gives 0.4049, still short of it. |
| 0.80 | Large | Cohen’s upper benchmark. |
| π ≈ 3.14 | Maximum possible | Reached only when one proportion is 1 and the other is 0. |
Treat these labels as a starting vocabulary rather than a verdict. Cohen advised using exact values from theory or experience in preference to the conventions where possible, so the defensible reading of an h is relative to what a meaningful difference is for the outcome in your field: a drop in infection rate, a change in response rate, a difference in retention.
Second worked example: a larger gap near the middle
Take p₁ = 0.55 and p₂ = 0.35, a twenty-point gap. φ₁ = 2·arcsin(√0.55) = 1.670964 and φ₂ = 2·arcsin(√0.35) = 1.266104, so h = 1.670964 − 1.266104 = 0.4049. Doubling the percentage-point gap from ten to twenty did not double h relative to the 50% vs 40% case (0.2014 to 0.4049 is close, but not exactly twice), which is again the transformation at work. Reversing the order, p₁ = 0.35 and p₂ = 0.55, gives h = −0.4049: the magnitude is unchanged and only the sign flips, which is why a paper should say which group comes first or report |h|.
Two-sample and one-sample uses
The same h feeds two different tests. In a two-sample design you compare the proportions observed in two groups (the pwr function pwr.2p.test, or pwr.2p2n.test when group sizes differ). In a one-sample design you compare one observed proportion with a hypothesised value (pwr.p.test); the pwr documentation’s example takes h from ES.h(0.5, 0.4) and runs pwr.p.test with n = 60 and a two-sided α of 0.05. Decide before data collection which design you have, because the sample-size arithmetic differs between them.
Assumptions and common mistakes
- Use proportions, not percentages. Entering 50 and 40 instead of 0.50 and 0.40 fails, because the square root of a number above 1 cannot be passed to arcsin.
- Two proportions only. h compares a pair. It is not a summary for a table with several groups; for association in larger tables use a measure such as Cramér’s V.
- It is not a difference in percentage points. Reporting h as if it were a percentage gap, or translating “h = 0.2” into “a 20-point difference”, is wrong; the table above shows how the same gap maps to different values.
- Keep the sign convention explicit. State which group is p₁, or report |h|.
- Do not confuse it with Cohen’s d. The 0.2, 0.5, 0.8 labels are shared, but d is built from means and standard deviations, and h from proportions.
- A benchmark is not a threshold for importance. Do not call a result meaningful only because it passed 0.50.
When to prefer h over a simple difference
Use h when you need one number that supports both a power calculation and a plain-language size label across studies whose baseline rates differ. If every group in your field sits near the same baseline, and the audience is clinical or operational, the raw difference in proportions (and its confidence interval) is often easier to communicate; h is most useful alongside it, not instead of it. The data table above shows the cost of relying on the raw gap alone: a four-point gap at 5% vs 1% gives h = 0.2507, which is larger than the 0.2014 of a ten-point gap at 50% vs 40%.
Using Cohen’s h for sample size and power
Cohen’s h is the effect size input for power analysis of proportions, which is why it appears in the pwr R package next to the functions pwr.p.test, pwr.2p.test and pwr.2p2n.test. As an indicative figure, using a normal approximation for two equal groups, a two-sided α of 0.05 and power of 0.80 (z-values 1.96 and 0.8416), the required size per group is n ≈ 2 × (1.96 + 0.8416)² ÷ h². With h = 0.2014 that is 2 × 7.849 ÷ 0.04055 ≈ 387 per group (calculated for this page). Run the pwr function for the exact design you plan, and see the guide on sensitivity power analysis and the minimum detectable effect for the reverse question of what h a fixed sample can detect.
How to report Cohen’s h (APA style)
Report the effect size alongside the test and its confidence interval. For each primary outcome the APA Publication Manual (7th edition) asks for an effect size with a confidence interval, and CONSORT 2010 asks for effect sizes with their precision. A typical sentence:
The response rate was higher in the intervention group (50%) than in the control group (40%), Cohen’s h = 0.20.
Give h to two decimals, say which group is p₁ if you report a signed value, and name the benchmark source if you use a label (“small by Cohen’s conventions”).
How to calculate Cohen’s h in Excel, R or Python
Excel
=2*(ASIN(SQRT(0.5))-ASIN(SQRT(0.4)))
Returns 0.201358. Put the proportions in cells (for example A2 and B2) and use =2*(ASIN(SQRT(A2))-ASIN(SQRT(B2))). This is the form given by the Real Statistics resource.
R
library(pwr)
h = ES.h(0.5, 0.4) # 0.2013579
pwr.2p.test(h = h, sig.level = 0.05, power = 0.80)
Python
import math
def cohens_h(p1, p2):
phi = lambda p: 2 * math.asin(math.sqrt(p))
return phi(p1) - phi(p2)
print(round(cohens_h(0.5, 0.4), 4)) # 0.2014
R from a 2 × 2 table
If your data are already a 2 × 2 contingency table, the rcompanion package’s cohenH function takes the table directly and derives the two proportions from the rows (the default) or the columns. Its documentation defines the statistic as Φ₂ − Φ₁, with Φ = 2·asin(sqrt(P)), the reverse order of subtraction from the pwr function ES.h. The magnitude is identical and only the sign differs, so when you move between packages, or compare your output with a published value, check which proportion is subtracted from which before interpreting a negative number.
Related effect sizes
When the outcome is a mean rather than a proportion, use Cohen’s d; when you model a continuous outcome with several predictors, see Cohen’s f² for regression. Odds ratios are another common way to express a difference between two proportions; h differs in that it is symmetric around 0.5 and variance-stabilised rather than multiplicative.
Sources
- Champely, S. pwr: Basic Functions for Power Analysis, version 1.3-0 (R package), function ES.h documentation: CRAN pwr package page, read 2026-11-18.
- Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates, as cited by the pwr documentation.
- Cohen’s h, Wikipedia, read 2026-11-18 (definition, directional and nondirectional forms, conventions and Cohen’s caution).
- rcompanion R package, cohenH function documentation: cohenH: Cohen’s h to compare proportions for 2 x 2 contingency tables, read 2026-11-18.
- Cohen’s h Effect Size, Real Statistics, read 2026-11-18 (Excel formula).
Data: Cohen's h for six pairs of proportions, computed from the arcsine square-root formula
Each row gives two proportions, the gap between them in percentage points, their arcsine square-root transforms (φ = 2·arcsin(√p), radians) and Cohen's h = φ₁ − φ₂. Values were computed from the formula documented in the pwr R package (ES.h) and checked against its example for 0.5 vs 0.4.
| Proportion pair (p1 vs p2) | Gap (percentage points) | phi1 (radians) | phi2 (radians) | Cohen's h |
|---|---|---|---|---|
| 5% vs 1% | 4 | 0.451 | 0.2003 | 0.2507 |
| 20% vs 10% | 10 | 0.9273 | 0.6435 | 0.2838 |
| 30% vs 20% | 10 | 1.1593 | 0.9273 | 0.232 |
| 50% vs 40% | 10 | 1.5708 | 1.3694 | 0.2014 |
| 55% vs 35% | 20 | 1.671 | 1.2661 | 0.4049 |
| 90% vs 80% | 10 | 2.4981 | 2.2143 | 0.2838 |
Sources
- pwr R package, ES.h documentation (Cohen 1988 formula) (accessed 2026-11-18)
- Cohen's h (Wikipedia) (accessed 2026-11-18)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







