Cohen's d Explained: Effect Size, Formula, and Interpretation

What Is Cohen's d? (Definition)

The statistic was introduced by psychologist Jacob Cohen in his 1988 textbook Statistical Power Analysis for the Behavioral Sciences, where he argued that researchers needed a way to communicate effect magnitude separately from statistical significance. It became the most widely reported effect size in the behavioral sciences and is now standard across medicine, education, and data science.

Understanding Cohen's d requires grasping one key concept first: raw mean differences are hard to compare across studies because they depend on the measurement scale. A mean difference of 10 points means something very different on a 100-point IQ scale versus a 20-point pain scale. Dividing by the standard deviation removes the units, making the number interpretable on its own.

Why Effect Size Matters Beyond the P-Value

A p-value answers one narrow question: "If the null hypothesis were true, how often would data this extreme arise by chance?" It says nothing about the size of the difference. With enough participants, any trivial difference eventually becomes statistically significant.

Consider two studies on the same intervention. Study A tests 40 participants and finds d = 0.80, p = 0.04. Study B tests 10,000 participants and finds d = 0.03, p = 0.001. Study B is more statistically significant — but the effect is so small it almost certainly has no practical value. Cohen's d surfaces that distinction immediately.

The Core Idea: Standardized Difference

Think of Cohen's d as converting two overlapping distributions into a single number that captures how far apart their centers are, relative to how spread out they are. When d = 0, the distributions are perfectly centered on the same value. When d = 1.0, the means are separated by one full standard deviation — roughly 84% of one group falls below the mean of the other.

The standardization step is what makes comparison across studies possible. A meta-analysis can pool a pain study measured in centimeters on a visual analog scale with a mood study measured on a 7-point Likert scale, because both produce Cohen's d values on the same dimensionless scale.

Cohen's d Formula Reference

Independent Samples Formula

Use this when the two groups are separate — for example, a treatment group and a control group with different participants in each.

Pooled Standard Deviation

The pooled standard deviation combines the variability from both groups into one estimate. It gives more weight to whichever group has more observations.

Paired Samples Formula

Use this when the same participants are measured twice — before and after an intervention, or under two different conditions.

Paired Cohen's d tends to be larger than the independent-samples version for the same data, because within-subject designs reduce the variance by controlling for individual differences. Be explicit in any report about which formula you used.

Variable Definitions Table

How to Calculate Cohen's d (Step-by-Step)

Calculating Cohen's d by hand takes five steps. The process is straightforward once you have the group means and standard deviations.

Step-by-Step Walkthrough

Common Calculation Mistakes

Worked Examples

Example 1 — Psychology: Cognitive Training Study

Example 2 — Medicine: Blood Pressure Drug Trial

Example 3 — Education: Teaching Method Comparison

Example 4 — A/B Testing: Website Checkout Conversion

Cohen's d Interpretation Scale

Understanding Overlap Between Distributions

Another way to interpret Cohen's d is through the overlap of the two group distributions. When d = 0, the groups are perfectly aligned. As d grows, the distributions pull apart and the overlap shrinks. The table below shows the percentage of the combined distribution that the two groups share at each benchmark level.

Context-Dependent Interpretation

Cohen himself acknowledged that his 0.2/0.5/0.8 benchmarks were arbitrary and should serve only as rough guides. Different disciplines have accumulated their own baseline expectations:

Interactive Cohen's d Calculator

Cohen's d Compared to Related Metrics

Cohen's d vs. Hedges' g

Hedges' g applies a correction factor J to reduce the positive bias that Cohen's d has in small samples. The correction is: g = d × J, where J = 1 − (3 / (4·df − 1)) and df = n₁ + n₂ − 2. For samples of 20 or more per group, the difference between d and g is below 5% and generally negligible. For samples under 10 per group, Hedges' g is the better choice — and is the standard in meta-analyses.

Cohen's d vs. Glass's Delta (Δ)

Glass's delta uses only the control group's standard deviation as the denominator, rather than a pooled estimate: Δ = (M_treatment − M_control) / SD_control. This is appropriate when you have strong reason to believe the treatment changed the variance — for example, if an intervention increases variability in outcomes alongside the mean. When group variances are roughly equal, Glass's delta and Cohen's d produce similar values.

Cohen's d vs. the P-Value

Cohen's d vs. Pearson's r

Both are standardized effect sizes, but they suit different study designs. Cohen's d compares two group means on a continuous outcome; Pearson's r measures the linear association between two continuous variables. The two can be converted: r = d / √(d² + 4). A Cohen's d of 0.5 corresponds to r ≈ 0.24; d = 0.8 corresponds to r ≈ 0.37. Learn more about correlation and scatter plots here.

Real-World Applications

Psychology Research

Cohen's d is nearly universal in experimental psychology. Studies on cognitive training, therapy outcomes, group differences in test performance, and treatment comparisons all routinely report it. Meta-analyses in clinical psychology — such as those examining the effectiveness of cognitive-behavioral therapy versus waitlist controls — aggregate Cohen's d values from dozens of primary studies to estimate the overall effect. The APA Publication Manual explicitly recommends reporting effect sizes for all primary analyses.

Medical and Clinical Trials

Regulatory agencies and clinical journals increasingly require effect size reporting alongside significance levels. In drug trials comparing a treatment to a placebo on continuous outcomes (blood pressure, biomarkers, quality-of-life scores), Cohen's d gives clinicians an intuitive sense of whether the group difference translates to meaningful patient benefit — something a p-value cannot communicate on its own.

Education Research

Researcher John Hattie's landmark meta-analysis Visible Learning synthesized over 800 meta-analyses covering millions of students and expressed all effects in Cohen's d units. He proposed d = 0.40 as the "hinge point" — the threshold at which an educational intervention is worth implementing. This work made Cohen's d the standard reporting metric across educational psychology worldwide. Explore the foundations of hypothesis testing on Statistics Fundamentals.

A/B Testing in Tech and Product

Product teams at technology companies use Cohen's d to understand whether changes to interfaces, algorithms, or features produce meaningful user behavior differences. With very large sample sizes typical of web experiments, even a d of 0.05 can be statistically significant — making effect size reporting essential for separating business-relevant changes from statistical noise.

Meta-Analysis

Meta-analysis aggregates Cohen's d values across multiple independent studies to estimate the overall effect of an intervention or phenomenon. Because d is dimensionless, studies measuring outcomes on different scales can be meaningfully combined. The resulting pooled d and its confidence interval provide a much more reliable estimate of truth than any single study. See the confidence intervals guide for how uncertainty is quantified around effect size estimates.

Cohen's d Formula Cheat Sheet

Entity & Terminology Glossary

Common Misconceptions

Related Topics on Statistics Fundamentals

Cohen's d fits into a broader framework of inferential statistics. These guides on Statistics Fundamentals connect directly to what you've learned here:

Academic Sources

• Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. — The original text establishing the d formula and 0.2/0.5/0.8 benchmarks.
• Hedges, L. V., & Olkin, I. (1985). Statistical Methods for Meta-Analysis. Academic Press. — Introduces the bias-corrected g statistic.
• Hattie, J. (2009). Visible Learning: A Synthesis of Over 800 Meta-Analyses Relating to Achievement. Routledge. — Establishes d = 0.40 as the education benchmark.
• Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science. Frontiers in Psychology, 4, 863. — A practical guide to effect size reporting, freely available via doi.org.
• Fritz, C. O., Morris, P. E., & Richler, J. J. (2012). Effect size estimates: Current use, calculations, and interpretation. Journal of Experimental Psychology: General, 141(1), 2–18.

Frequently Asked Questions

What is Cohen's d used for?

Cohen's d measures the practical magnitude of a difference between two group means. It tells you how large the effect is in standard deviation units — independent of sample size. Researchers use it in psychology, medicine, education, and A/B testing whenever a p-value alone is not enough to judge whether a result matters in practice.

How do you interpret Cohen's d?

Jacob Cohen's benchmarks treat d = 0.2 as a small effect, d = 0.5 as medium, and d = 0.8 as large. These thresholds are domain-dependent: a d of 0.2 can be clinically meaningful in medicine, while a d of 0.8 might be unremarkable in some laboratory settings. Always consider the specific field's norms and the practical stakes of the decision.

Can Cohen's d be negative?

Yes. A negative value simply means the second group mean is larger than the first. The sign reflects which group was placed in the numerator, not whether the result is good or bad. When reporting effect size magnitude, take the absolute value; keep the sign only when the direction of the difference carries theoretical meaning.

Is Cohen's d better than the p-value?

They answer different questions and work best together. The p-value indicates whether an effect is likely real given sampling variability; Cohen's d tells you how large that effect is. A study with a large sample can produce a statistically significant p-value for a Cohen's d of 0.03 — real but practically trivial. Reporting both is the current standard in most peer-reviewed journals.

What is a good effect size for Cohen's d?

"Good" depends entirely on the context. In education, Hattie's research suggests d ≥ 0.4 is needed to justify an intervention. In medicine, d = 0.2 can represent a clinically important difference if the outcome is mortality or serious morbidity. In technology A/B testing, a d as small as 0.1 can be worth implementing at sufficient scale. There is no single threshold for "good."

When should I use Hedges' g instead of Cohen's d?

Use Hedges' g when your sample size is below about 20 per group, or when conducting a meta-analysis that pools studies of varying sizes. For most individual studies with adequate samples, the two values differ by less than 5% and the distinction is rarely consequential in practice.