Quick Answer: What Are Cohen's d Thresholds?
Cohen's d measures the standardized difference between two group means. Jacob Cohen proposed three conventional benchmarks in 1988 for behavioral research:
Important: These are conventional rules of thumb for behavioral science, not universal laws. Whether a given d is meaningful depends on the research field, outcome measure, study design, measurement precision, and cost of the intervention.
Cohen's d Interpreter
Cohen's d Effect Size Table (Conventional Thresholds)
The table below shows the three conventional Cohen's d benchmarks and how to describe negative values. These come from Cohen (1988) and are widely cited in psychology, education, and behavioral science — though fields differ in what thresholds carry practical weight.
| Cohen's d | Conventional Label | Notes |
|---|---|---|
| 0 | None | No standardized mean difference between groups |
| ≈ 0.2 | Small | Subtle, may require large samples to detect reliably |
| ≈ 0.5 | Medium | Visible in everyday observation (Cohen, 1988) |
| ≈ 0.8 | Large | Obvious difference between group distributions |
| Negative d | Same magnitude | Sign reflects direction, not quality. |d| = 0.5 means medium regardless of sign |
Conventional thresholds from Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). These are guidelines, not universal standards. See the extended table below for a fuller descriptive reference.
Extended Cohen's d Interpretation Table
The extended table below provides a more granular descriptive guide for interpreting d values across a wider range. This is a descriptive reference, not an official universal classification — field conventions, study design, and context remain the primary guide.
| |Cohen's d| | Conventional Description | General Interpretation |
|---|---|---|
| 0 | None | No standardized mean difference |
| 0.01 – 0.19 | Very Small | Minimal standardized difference; may need very large n to detect |
| 0.20 – 0.49 | Small | Small standardized difference; Cohen's conventional small threshold |
| 0.50 – 0.79 | Medium | Moderate standardized difference; Cohen's conventional medium threshold |
| 0.80 – 1.19 | Large | Large standardized difference; Cohen's conventional large threshold |
| ≥ 1.20 | Very Large | Very large standardized difference; groups have minimal overlap |
What Is Effect Size?
Effect size describes the magnitude of a difference, association, or relationship between variables. It answers a fundamentally different question than a p-value.
Statistical Significance
Answers: Is this difference larger than chance alone predicts? Depends heavily on sample size. A tiny difference can be statistically significant with n = 10,000.
Effect Magnitude
Answers: How large is the difference, in standardized terms? Cohen's d expresses this as the number of standard deviations separating the two group means.
Practical Significance
Answers: Does this difference matter in the real world? Requires domain knowledge, context, and judgment — no formula provides this automatically.
Reporting effect size alongside a p-value gives a more complete picture. The p-value tells you whether a finding is statistically unlikely under the null hypothesis; the effect size tells you how large the observed difference actually is. Neither alone is sufficient for a full interpretation. The p-value guide at Statistics Fundamentals covers this distinction in depth.
What Is Cohen's d?
Cohen's d is a standardized measure of the difference between two group means. "Standardized" means the raw mean difference is divided by a measure of variability — the pooled standard deviation — so the result is unit-free and comparable across studies that use different scales.
d = (M₁ − M₂) / SDpooled
A d = 1.0 means the two group means are separated by exactly one pooled standard deviation. A d = 0.5 means they are separated by half a standard deviation. The exact formula for the standardizer varies by study design — the pooled SD formula shown here applies to two independent groups with roughly equal variances.
Cohen's d Formula: Pooled Standard Deviation
For two independent groups, the pooled standard deviation weights each group's variance by its degrees of freedom:
SDpooled = √[ ((n₁ − 1)×SD₁² + (n₂ − 1)×SD₂²) / (n₁ + n₂ − 2) ]
Then Cohen's d = (M₁ − M₂) / SDpooled. This formula assumes the two groups have similar (homogeneous) variances. For paired designs, repeated measures, or one-sample comparisons, a different standardizer applies — see the section on study designs below.
Worked Examples: Calculating Cohen's d
Scenario: A hypothetical study compares reading comprehension scores between two teaching methods. Group A (standard) and Group B (modified).
| Group | n | Mean (M) | SD |
|---|---|---|---|
| Group A (standard) | 40 | 72.0 | 10.0 |
| Group B (modified) | 40 | 74.0 | 10.0 |
Calculation:
SDpooled = √[((40−1)×10² + (40−1)×10²) / (40+40−2)]
= √[(39×100 + 39×100) / 78]
= √[7800 / 78] = √100 = 10.0
d = (74.0 − 72.0) / 10.0 = 2.0 / 10.0 = 0.20
The modified method produced a mean difference of 2 points, which equals 0.20 pooled standard deviations. By convention this is a small standardized effect. Whether this 2-point difference in reading scores matters practically depends on the scale, context, and costs involved.
Scenario: A hypothetical clinical trial compares anxiety symptom scores (lower = better) between a treatment group and a control group after 8 weeks.
| Group | n | Mean (M) | SD |
|---|---|---|---|
| Control | 30 | 52.0 | 12.0 |
| Treatment | 30 | 46.0 | 12.0 |
Calculation:
SDpooled = √[((30−1)×144 + (30−1)×144) / (30+30−2)]
= √[(29×144 + 29×144) / 58]
= √[8352 / 58] = √144 = 12.0
d = (46.0 − 52.0) / 12.0 = −6.0 / 12.0 = −0.50
|d| = 0.50
The treatment group's mean is 6 points lower than the control. The d is negative because the treatment mean was subtracted from the control — this reflects direction only. The magnitude, 0.50, is a conventionally medium standardized effect.
Scenario: A hypothetical experiment compares reaction times (ms) between two groups with different training levels.
| Group | n | Mean (ms) | SD |
|---|---|---|---|
| Group 1 (less trained) | 25 | 420 | 50 |
| Group 2 (more trained) | 25 | 380 | 50 |
Calculation:
SDpooled = √[((25−1)×2500 + (25−1)×2500) / (25+25−2)]
= √[(24×2500 + 24×2500) / 48]
= √[120000 / 48] = √2500 = 50.0
d = (420 − 380) / 50.0 = 40 / 50 = 0.80
The more trained group responded 40 ms faster. This 40 ms raw difference equals 0.80 pooled standard deviations — a conventionally large effect. Whether a 40 ms reaction time difference is practically meaningful depends on the application (e.g., aviation safety vs. a casual game).
Scenario: The same hypothetical anxiety study from Example 2, but this time the subtraction is done in the opposite order (Treatment − Control instead of Control − Treatment).
d = (Mtreatment − Mcontrol) / SDpooled
= (46.0 − 52.0) / 12.0 = −0.50
Key point: d = −0.50 has exactly the same magnitude as d = 0.50. The negative sign reflects only which group was placed first in the formula. Whether d is negative depends on the order of subtraction, not on whether the effect is "bad." Always report |d| for the magnitude and explain the direction separately.
Common mistake: Treating a negative d as automatically indicating a harmful or undesirable outcome. The sign is purely a labeling convention that depends on which group is M₁ and which is M₂.
Scenario: A hypothetical online marketing study with n = 500 per group measures purchase conversion rates. The difference is tiny but statistically significant.
| Group | n | Mean Score | SD |
|---|---|---|---|
| Control (original page) | 500 | 10.00 | 5.00 |
| Treatment (redesigned page) | 500 | 11.00 | 5.00 |
SDpooled = 5.00 (equal SDs, equal n — simplifies directly)
d = (11.00 − 10.00) / 5.00 = 1.00 / 5.00 = 0.20
t ≈ d × √(n/2) = 0.20 × √250 = 0.20 × 15.81 ≈ 3.16 → p < 0.01
Lesson: Large samples detect small effects reliably. Always report the effect size alongside the p-value. A significant p-value with d = 0.20 warrants very different conclusions than a significant p-value with d = 0.80.
Scenario: A hypothetical pilot study with only n = 6 per group finds a large estimated effect but fails to reach statistical significance.
| Group | n | Mean | SD |
|---|---|---|---|
| Control | 6 | 50 | 12 |
| Intervention | 6 | 62 | 14 |
SDpooled = √[((6−1)×144 + (6−1)×196) / (6+6−2)]
= √[(720 + 980) / 10] = √170 ≈ 13.04
d = (62 − 50) / 13.04 = 12 / 13.04 ≈ 0.92
t ≈ d × √(n₁n₂/(n₁+n₂)) = 0.92 × √3 ≈ 1.59 → p ≈ 0.14 (not significant)
A large estimated effect that does not reach statistical significance. With n = 6 per group, the standard error is large and power is low. The wide confidence interval around d makes the estimate uncertain. This does not mean no effect exists — it means the sample is too small to detect one reliably.
Lesson: Effect size and statistical significance are distinct. Small samples can produce large estimated effects that remain non-significant due to insufficient statistical power. A pilot study finding d ≈ 0.92 is a reason to conduct a properly powered follow-up study — not to conclude the effect is proven.
Cohen's d vs Statistical Significance
The four measures below each answer a different question. A complete analysis typically requires reporting all of them rather than relying on any single number.
| Measure | What it tells you |
|---|---|
| Cohen's d | Magnitude of standardized mean difference — independent of sample size |
| p-value | Evidence against the null hypothesis; highly sensitive to sample size |
| Confidence interval | Uncertainty range around the effect estimate; indicates precision |
| Sample size (n) | Directly influences precision, power, and the p-value — not the effect size |
p < 0.05 → the result is statistically unlikely under H₀ (not necessarily large or important)
d = 0.8 → the groups differ by 0.8 standard deviations (not necessarily statistically significant in small samples)
A complete report includes both.
Cohen's d vs Hedges' g
Both measures quantify the standardized mean difference between two groups. The difference lies in a small-sample bias correction.
Cohen's d
Divides by the pooled SD. Slightly overestimates the population effect size in small samples. Widely reported and understood across fields.
Hedges' g
Applies a correction factor J to reduce upward bias. Preferred for small samples (n < 20 per group) and meta-analyses where combining effect sizes from multiple studies requires precision.
In practice the difference between d and g is small when sample sizes exceed 20 per group. For meta-analyses and small-sample research, Hedges' g is the safer choice. The two are not identical and should not be treated interchangeably when precision matters.
Other Effect Size Measures
Cohen's d applies to comparisons between two means. Different analyses require different effect size measures. The table below covers common options.
| Analysis | Common effect size measure |
|---|---|
| Two independent means | Cohen's d / Hedges' g |
| Paired / repeated means | Standardized mean difference (design-appropriate) |
| Correlation (bivariate) | Pearson's r, r² |
| t-test | Cohen's d |
| ANOVA | η² / partial η² / ω² |
| Chi-square association | Cramér's V |
| Logistic / 2×2 table | Odds ratio / Risk ratio |
| Regression (proportion of variance) | R², f² |
The choice of effect size measure must match the statistical analysis. Using Cohen's d for a chi-square test, or r for an ANOVA, produces results that cannot be meaningfully compared to the relevant benchmark tables. See the hypothesis testing section for guidance on selecting the right test and measure.
Cohen's d for Different Study Designs
The pooled SD formula shown above applies to two independent groups. Other designs require a different standardizer — using the wrong one produces a d value that is not comparable to the conventional thresholds.
Independent Groups
The standard case. Uses SDpooled as the denominator. Both groups are measured once, independently. This is the formula demonstrated in the examples above.
Paired Samples and Repeated Measures
When the same participants are measured twice (pre/post or two conditions), a common approach divides the mean difference score by the standard deviation of the difference scores. This produces a larger d than the independent-groups formula, because within-subject correlation is not part of the denominator. The two cannot be directly compared.
Pretest-Posttest Designs
Some researchers use the pooled SD of both time points; others use only the pretest SD as the standardizer (Glass's Δ). The choice matters because each produces a different value and a different interpretation. Reporting which standardizer was used is essential.
One-Sample Designs
Compares a sample mean against a known population value. The standardizer is the sample standard deviation. The resulting d indicates how many standard deviations the sample mean sits above or below the reference value.
How to Interpret Cohen's d in Practice
The conventional benchmarks are a starting point, not a final answer. When interpreting a d value, work through these questions in addition to looking up its size category.
A d of 0.50 from n = 10 has a very wide confidence interval. The same d from n = 200 is far more precise. Always consider the confidence interval.
In some fields, d = 0.2 is considered large relative to typical findings. In others, d = 0.8 is routine. Context within the literature matters more than the benchmark.
A d = 0.1 in a drug trial for a life-threatening condition may justify action. The same d in a low-stakes educational intervention may not. Practical significance is not synonymous with standardized magnitude.
A 0.2 SD difference in blood pressure may translate to measurable health outcomes. A 0.2 SD difference on a poorly validated questionnaire scale may mean very little.
Confidence Intervals for Effect Size
A single effect size estimate carries sampling uncertainty. A confidence interval quantifies that uncertainty: a 95% CI for d means that if the study were repeated many times, 95% of the resulting intervals would contain the true population effect size.
Wide confidence intervals indicate the estimate is imprecise, typically due to small sample sizes. Narrow intervals — from large, well-powered studies — support more confident conclusions about the magnitude of the effect.
Reporting a 95% CI around d is recommended by the APA Publication Manual and various statistical reporting guidelines. A CI that crosses zero is consistent with the possibility of no effect. A CI that does not include small benchmark values (e.g., 0.2) supports a conclusion about minimum effect size.
Is a Large Effect Size Always Important?
Standardized magnitude does not automatically equal practical importance. The same d can mean different things in different contexts:
Medicine
A d = 0.3 reduction in a validated pain scale may be clinically meaningful and justify prescribing a medication. A d = 0.8 improvement on an intermediate biomarker may matter less if it does not translate to patient outcomes.
Education
A d = 0.2 improvement in standardized test scores across a national sample represents millions of students. The same d from a single classroom experiment carries far more uncertainty about generalizability.
Business and Marketing
A d = 0.1 difference in conversion rates across high-volume traffic can produce substantial revenue differences. Whether the cost of implementing a change is justified by that difference is a business decision, not a statistical one.
Psychology and Social Science
Cohen's original benchmarks were calibrated for behavioral science. They may not transfer directly to other domains. Researchers increasingly recommend comparing effect sizes to prior findings in the same domain rather than to the universal benchmarks alone.
Common Mistakes When Interpreting Cohen's d
How to Report Cohen's d
Reporting should include enough information for a reader to understand and evaluate the finding without additional lookup. The examples below illustrate good practice.
"The difference between the groups was small, Cohen's d = 0.23, 95% CI [0.04, 0.42]."
"The groups differed by a medium standardized effect, Cohen's d = 0.54, t(58) = 3.12, p = .003."
"The observed standardized difference was large, Cohen's d = 0.87, n1 = 45, n2 = 47."
A complete report should include:
The effect size estimate (d) · Confidence interval where available · Sample sizes (n₁ and n₂) · The statistical test and test statistic · The p-value · A brief statement of the conventional size category · Context for practical significance when relevant
Frequently Asked Questions About Effect Size
What are Cohen's d thresholds?
Cohen's d = 0.2 is the conventional small threshold, d = 0.5 is medium, and d = 0.8 is large. These were introduced by Jacob Cohen in 1988 for behavioral science. They are widely used as a first reference point but should be interpreted in the context of the specific research field and outcome measure.
What is a small effect size?
A small effect size corresponds to Cohen's d ≈ 0.2. The two group means differ by about 0.2 standard deviations. Small effects can still be practically important, especially in large-population interventions where even a small average improvement affects many people.
What is a medium effect size?
A medium effect size corresponds to Cohen's d ≈ 0.5. Cohen described this as an effect that would be visible to a careful observer in everyday experience. Medium effects often represent meaningful differences in behavioral and educational research.
What is a large effect size?
A large effect size corresponds to Cohen's d ≈ 0.8. The group means differ by 0.8 pooled standard deviations, indicating a substantial separation between group distributions. A d of 0.8 means the distributions of two groups have about 47% non-overlap.
Is Cohen's d = 0.2 always small?
By convention, yes. But conventional labels are not evaluative judgments. A d of 0.2 applied to a policy that affects 10 million people may have enormous practical consequences. A d of 0.2 in a controlled lab task with no real-world stakes may be entirely trivial.
What does Cohen's d = 1.0 mean?
A d of 1.0 indicates the two group means are separated by exactly one pooled standard deviation. This is well above the large threshold of 0.8. With normally distributed data, a d of 1.0 corresponds to roughly 55% non-overlap between the two distributions — meaning about 84% of one group scores above the average of the other.
What does a negative Cohen's d mean?
A negative d means the second group mean (M₂) is larger than the first (M₁) in the formula d = (M₁ − M₂) / SDpooled. The sign indicates direction only, not quality. A d of −0.8 has the same large magnitude as d = 0.8. The interpretation of which direction is "better" depends entirely on the context of the outcome variable.
What is the difference between effect size and p-value?
The p-value measures how consistent the data are with the null hypothesis — it is heavily influenced by sample size. Cohen's d measures the magnitude of the standardized difference, which is largely independent of sample size. A p-value below 0.05 does not tell you the effect is large; a large d does not guarantee a significant p-value in a small sample. Both are needed.
How do I calculate Cohen's d step by step?
Step 1: Calculate M₁ and M₂ (group means). Step 2: Calculate SD₁ and SD₂ (group standard deviations). Step 3: Calculate SDpooled = √[((n₁−1)SD₁² + (n₂−1)SD₂²)/(n₁+n₂−2)]. Step 4: Divide the mean difference by the pooled SD: d = (M₁ − M₂) / SDpooled. Use the effect size calculator to automate these steps.
When should I use Hedges' g instead of Cohen's d?
Prefer Hedges' g when sample sizes are small (below 20 per group) or when conducting a meta-analysis. Cohen's d slightly overestimates the true population effect in small samples; Hedges' g applies a correction factor to reduce that bias. For large samples the difference is negligible.
Can Cohen's d exceed 1?
Yes. Cohen's d has no upper bound. Values above 1.0 indicate the group means are more than one standard deviation apart. Values of 2.0 or higher occur in comparisons between populations that differ dramatically, such as height differences between adult populations from very different demographic groups.
How does sample size affect Cohen's d?
Cohen's d is designed to be relatively independent of sample size — it does not grow as n increases. What does change with larger n is the precision of the estimate (narrower confidence intervals) and the statistical power to detect the effect. A d of 0.5 computed from n = 20 and n = 200 should estimate the same population effect size, but the latter is far more precise. Use the sample size calculator to plan adequately powered studies.
What effect size is needed for a t-test to be significant?
There is no fixed answer — it depends on sample size and alpha level. A d of 0.5 may be significant with n = 30 per group but not with n = 8 per group. The minimum detectable effect size for a given n and alpha can be found using a power analysis. See the t-test calculator and the power guide on Statistics Fundamentals.
Is Cohen's d the same as the standardized mean difference in meta-analysis?
Closely related, but not always identical. Meta-analyses frequently use Hedges' g (the bias-corrected version) or Glass's Δ (which uses only the control group SD). When combining effect sizes across studies, it is important that all studies use the same standardizer so the values are comparable.
How do I convert between Cohen's d and Pearson's r?
The conversion formula for equal sample sizes is: r = d / √(d² + 4). For d = 0.5, r ≈ 0.24. For d = 0.8, r ≈ 0.37. The reverse is: d = 2r / √(1 − r²). These conversions assume equal group sizes; unequal sizes require a modified formula.
What is a good effect size to aim for in study design?
The target effect size for study design should come from prior research, theory, or the minimum effect that would be practically meaningful in the field. Using Cohen's conventional values (d = 0.2, 0.5, 0.8) as a substitute for domain knowledge in sample size calculations often leads to either underpowered or unnecessarily large studies.
Sources & Further Reading
Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. — The original source for the 0.2/0.5/0.8 thresholds. Cohen explicitly cautioned against treating them as universal laws.
Hedges, L. V., & Olkin, I. (1985). Statistical Methods for Meta-Analysis. Academic Press. — Introduced the bias-correction factor for Hedges' g. Standard reference for meta-analytic effect size methods.
NIST/SEMATECH e-Handbook of Statistical Methods. National Institute of Standards and Technology. itl.nist.gov — U.S. government statistical reference covering effect size, hypothesis testing, and related methodology.
Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, 863. doi:10.3389/fpsyg.2013.00863 — Open-access practical guide to computing and reporting Cohen's d and related measures for t-tests and ANOVA.
American Psychological Association (2020). Publication Manual of the American Psychological Association (7th ed.). APA. — Requires reporting effect sizes alongside significance tests in submitted manuscripts. Section 6.24 covers effect size reporting conventions.
Penn State STAT 415: Introduction to Mathematical Statistics. Pennsylvania State University. online.stat.psu.edu — Open-access course covering hypothesis testing and effect size in the context of formal statistical inference.
Related Pages on Statistics Fundamentals
Key Takeaways
The Three Thresholds Are Conventions
Cohen proposed 0.2, 0.5, and 0.8 as rules of thumb for behavioral research in the absence of other information. He did not intend them as universal standards applicable to every field and every outcome variable.
Effect Size ≠ Statistical Significance
A p-value below 0.05 does not mean the effect is large. A Cohen's d above 0.8 does not guarantee a significant p-value in small samples. Both pieces of information contribute to a complete interpretation. See the p-value guide for further detail.
Study Design Determines the Formula
The pooled SD formula applies to independent groups. Paired, repeated-measures, and one-sample designs use different standardizers. Applying the wrong formula produces a d that cannot be interpreted against the conventional benchmarks.
Negative d Is Not Automatically Bad
The sign of d depends on the order of group subtraction. A d of −0.5 has the same medium magnitude as d = 0.5. Interpret the direction in the context of the outcome variable, not as an evaluation of quality.