Cohen’s <em>d</em> is an <strong>upward biased</strong> estimator. At eight degrees of freedom it overestimates by 10 %, and neither a better analysis nor a bigger sample fixes the one you already have: it is fixed by multiplying by a factor. That factor is Hedges’ <em>g</em>.
g = d · J, with J = 1 − 3 / (4·df − 1)
The bias does not come from the mean difference, which is unbiased: it comes from the denominator. The sample pooled standard deviation slightly underestimates the population one, and dividing by a number that is too small makes the ratio too large.
The bias depends only on the degrees of freedom, so it can be corrected without touching the data. Hedges gave the exact factor using gamma functions; the approximation J = 1 − 3/(4·df − 1) used here is the one every meta-analysis uses and it agrees with the exact one to the fourth decimal.
Since the bias is always upward, the correction always lowers the number. There is no case where g comes out larger than d.
Below twenty per group, always. There the correction exceeds 2 % and can no longer be called rounding.
In any meta-analysis, always. Small studies contribute the most bias and also tend to be the ones publishing the largest effects; uncorrected, that bias pools and the meta-analytic estimate rises.
And whenever the journal asks, which is increasingly often. Many psychology and biomedical journals now name g explicitly in their guidelines.
With large samples the correction is irrelevant, but it does no harm either: always reporting g is a defensible default, and not having to decide each time saves effort.
It does not fix a lack of power. A corrected g with a confidence interval of [0.05, 1.55] is still a study that cannot tell a small effect from an enormous one.
Nor publication bias. That belongs to the literature, not the estimator, and is examined with other tools: funnel plots, trim-and-fill, p-curve.
Nor a design that does not support the comparison. If the groups were not equivalent to begin with, correcting the estimator does not make the design equivalent.
If what you are missing is the starting d, it comes from the t in from t to Cohen’s d, which already returns the corrected g along the way.
All thirteen effect size conversions, grouped by what each one is for, are in the index.
If you use it in your thesis or in a paper, cite it so that readers of your Methods section know how the calculation was done.
Ciudad-Fernández, V. (2026). Cohen's d to Hedges' g converter [Web application]. Ciudad Science. Retrieved October 14, 2026, from https://analisisdedatospsicologia.com/en/convert/cohens-d-to-hedges-g
For your Methods section: Cohen's d was corrected for small-sample bias to obtain Hedges' g (Hedges, 1981), using the Ciudad Science cohen's d to Hedges' g converter (Ciudad-Fernández, 2026).
@misc{ciudadfernandez2026convdagdehedges,
author = {Ciudad-Fern{\'a}ndez, V{\'i}ctor},
title = {{Cohen's d to Hedges' g converter}},
year = {2026},
publisher = {Ciudad Science},
howpublished = {Web application},
url = {https://analisisdedatospsicologia.com/en/convert/cohens-d-to-hedges-g},
urldate = {2026-10-14},
note = {Accessed 2026-10-14},
}The ones from the test: n₁ + n₂ − 2 for independent samples, and N − 1 for paired samples, where N is the number of pairs. Using the total N in a pre-post design undercorrects.
No, and it is an old confusion. Cohen used the letter g for his uncorrected estimator; today "Hedges’ g" always means the bias-corrected version, and that is what any reviewer will assume.
Yes: both limits are multiplied by the same J. Correcting the point estimate and leaving the interval alone produces an interval that is no longer centered on its estimate.
The correction multiplies, so zero stays zero. J only changes the magnitude, never the sign or the significance.