What the number means
A Cohen’s d of 0.50 says the two means differ by one-half of the standard deviation used in the calculation. The sign gives direction under the stated subtraction order:
- d = 0: no standardized mean difference;
- d > 0: the first named group has the higher mean;
- d < 0: the first named group has the lower mean.
Reversing the group order reverses the sign without changing the magnitude. That makes group labels and the standardizer part of the result, not optional details.
The standardized scale helps compare outcomes measured in different units. It does not make every difference equally meaningful. A half-standard-deviation shift can have different practical consequences across outcomes, populations, and study designs.
Cohen’s conventions and their actual status
Cohen explicitly treated such conventional definitions as arbitrary and recommended using relevant substantive knowledge when available.1 The labels are useful as a shared reference when no better benchmark exists. They are not evidence that 0.49 is meaningfully different from 0.50 or that 0.80 is important in every field.
His primary text is unusually direct about the limitation: the definitions are arbitrary, should be replaced when a better basis exists, and were offered with qualifications. Cohen also anchored the values to examples from the behavioral literature he knew; “arbitrary” does not mean meaningless or random. It means an attributed fallback rather than a natural boundary.1
This is why the right sentence is “large by Cohen’s convention,” not simply “large.” Where statistical thresholds come from explains the difference between a formal definition and an attributed convention.
Hedges’ correction and the naming caveat
Modern references commonly use Hedges’ g for a standardized mean difference adjusted for small-sample bias.4 The historical notation is more precise than the modern label suggests. In Hedges’ 1981 paper, g is the uncorrected estimator and the corrected estimator is written gᵁ. Hedges derives the correction as gᵁ = g × c(m), where m is the degrees of freedom used to estimate the standard deviation.3
The correction factor is below 1 and can be approximated by 1 − 3/(4m − 1). Hedges shows that the uncorrected estimator overestimates the population standardized effect on average; the bias approaches zero as the relevant degrees of freedom increase but can be substantial when they are small. This is a continuous correction, not a reason to invent a sample-size cutoff. The paper also shows that multiplying by the factor reduces the estimator’s variance under its stated model.3
This does not establish that Hedges’ g is universally “better than Cohen’s d.” It establishes a narrower point: the correction addresses a defined small-sample bias under specified assumptions. The modern name belongs to later convention; Hedges himself called the corrected quantity gᵁ, not “Hedges’ g.”
Read a reported d in five steps
- Identify the comparison. Which mean was subtracted from which?
- Identify the standardizer. Was d based on a pooled standard deviation, a control-group standard deviation, a repeated-measures standardizer, or another denominator?
- Read the estimate and interval. The point estimate gives magnitude and direction; the confidence interval shows uncertainty.
- Translate back to the outcome. Ask what the difference means in the original units and setting.
- Compare with relevant evidence. Use domain benchmarks, prior distributions, or decision consequences when they exist.
P-value interpretation answers a different question: incompatibility with a null model under stated assumptions. A small p-value does not make an effect large, and a large p-value does not prove the effect is negligible. Eta squared and partial eta squared provide related effect-size measures for ANOVA output.
Why field context can reverse the label
Kraft proposed a tightly scoped education schema based on 1,942 effect sizes from 747 randomized studies of pre-K–12 interventions with standardized achievement outcomes: less than .05 small, .05 to less than .20 medium, and .20 or greater large.5 Under Cohen’s generic convention, d = .20 is small; under Kraft’s stated context, it is large. The contrast does not prove that one author is right everywhere. It shows why a label must carry its source and scope.
Other evidence has a different status. Schäfer and Schwarz described effect-size distributions across psychology and did not propose a replacement cutoff table.6 Lovakov and Agadullina derived social-psychology guidelines from published meta-analyses; the publisher abstract available for verification supports that limited characterization.7 Funder and Ozer offered a reasoned interpretation guideline rather than an empirical percentile distribution.2 These categories should not be merged.
Field-distribution benchmarking is itself disputed: Panzarella and colleagues argue that observed effect-size distributions lack a sufficient theoretical basis for defining magnitude and recommend study-specific context instead.8 That objection does not restore Cohen’s values as universal. It reinforces the practical conclusion that no table can replace the outcome, design, uncertainty, and consequences.
Worked interpretation
Suppose a report gives d = 0.62, 95% CI [0.25, 0.99], with the first group named as the higher-scoring group.
A careful reading is:
- the estimated mean difference is 0.62 standard deviations in the stated direction;
- the point estimate lies between Cohen’s medium and large reference values;
- the interval remains compatible with effects from near the small-to-medium region through approximately the large region;
- practical importance still depends on the outcome scale and study context.
Do not compress that into “the effect is medium.” The interval, denominator, and original scale carry information the label discards.
Effect size is not practical importance
A standardized effect depends partly on the variability of the sample. A numerically identical raw difference can produce different d values in more or less variable populations. Practical importance depends on consequences, costs, duration, reversibility, and the outcome’s real-world units—features d does not encode.
Methodological discussion has therefore emphasized interpreting effect sizes in context and considering cumulative or consequential effects rather than treating a three-bin scale as the answer.2 For alternative standardized metrics, see correlation and standardized association measures.
Common reading errors
- Treating 0.20, 0.50, and 0.80 as natural boundaries.
- Ignoring the confidence interval around d.
- Omitting the sign and group order.
- Comparing d values built with different standardizers as if they were identical.
- Equating statistical significance with magnitude.
- Calling an effect important because it received a “large” label.
A defensible reporting sentence
“The estimated standardized mean difference was d = 0.62 (95% CI [0.25, 0.99]) in favor of the first group. This is between Cohen’s medium and large behavioral-science conventions, but its practical meaning depends on the original outcome scale and context.”
That sentence keeps the convention visible without allowing it to replace interpretation.