Effect Size Calculator
Reviewed by Hamza Gujjar, Algorithm Architect & Software Developer
Last updated: August 17, 2026Understanding Effect Size: Magnitude Beyond Statistical Significance
In empirical scientific research, a p-value indicates whether an observed experimental difference is likely due to random sampling chance, but provides zero information about the magnitude or real-world importance of that difference. With a massive sample size (e.g., n = 1,000,000), even an infinitesimal, meaningless difference achieves high statistical significance (p < 0.0001). Conversely, in small samples, large and vital effects may fail to reach significance.
Effect Size quantifies the standardized strength or magnitude of a relationship between variables, allowing researchers to compare results across different experimental scales, measurement units, and independent clinical trials.
Primary Families of Effect Size Metrics
• Cohen's d (Equal Variances):
d = (μ1 − μ2) / spooled
where spooled = √[ ((n1 − 1)s12 + (n2 − 1)s22) / (n1 + n2 − 2) ]
• Hedges' g (Small-Sample Unbiased Correction):
g = d × [ 1 − 3 / (4(n1 + n2) − 9) ]
Cohen's d is positively biased in small samples (n < 20); Hedges' g removes this estimation bias.
• Glass's Δ (Delta - Control Group Standard Deviation):
Δ = (μtreatment − μcontrol) / scontrol
Used when the experimental treatment alters population variance, making pooled standard deviation invalid.
2. Variance Explained (The r-Family and η2-Family):
• Pearson Correlation Coefficient (r): Measures linear association between continuous variables (−1.0 to +1.0).
• Eta-Squared (η2 in ANOVA): η2 = SSeffect / SStotal
• Partial Eta-Squared (ηp2): ηp2 = SSeffect / (SSeffect + SSerror)
• Omega-Squared (ω2): Unbiased population estimate of variance explained in ANOVA.
3. Categorical Association (Odds Ratio and Risk Ratio):
• Odds Ratio (OR): OR = (a × d) / (b × c) in a 2 × 2 contingency table.
• Risk Ratio (Relative Risk - RR): RR = [ a / (a + b) ] / [ c / (c + d) ].
Effect Size Benchmarks Across Statistical Families
| Effect Size Metric | Small Effect | Medium Effect | Large Effect | Primary Application Area |
|---|---|---|---|---|
| Cohen's d / Hedges' g | 0.20 | 0.50 | 0.80 (> 1.20 = Very Large) | Two-sample t-tests, clinical trials, meta-analyses. |
| Pearson's r | 0.10 | 0.30 | 0.50 | Bivariate correlation, survey research. |
| Eta-Squared (η2 / ηp2) | 0.01 (1%) | 0.06 (6%) | 0.14 (14%) | One-way and factorial ANOVA experimental designs. |
| Omega-Squared (ω2) | 0.01 (1%) | 0.06 (6%) | 0.14 (14%) | Unbiased ANOVA variance accounting. |
| Odds Ratio (OR) | 1.50 | 2.50 | 4.30 | Epidemiology, case-control medical studies. |
| Cramér's V (df = 1) | 0.10 | 0.30 | 0.50 | Chi-Square test of independence for cross-tabulations. |
Step-by-Step Practical Calculation: Treatment Group Cohen's d and Hedges' g
An educational psychology trial tests a cognitive training app on reading comprehension scores:
- Treatment Group (n1 = 15): Mean μ1 = 84.0, s1 = 6.0
- Control Group (n2 = 15): Mean μ2 = 78.0, s2 = 8.0
- Step 1: Calculate Pooled Standard Deviation:
spooled2 = [ (14 × 36) + (14 × 64) ] / (15 + 15 − 2) = [ 504 + 896 ] / 28 = 1,400 / 28 = 50.0.
spooled = √50.0 = 7.071. - Step 2: Calculate Cohen's d:
d = (84.0 − 78.0) / 7.071 = 6.0 / 7.071 = 0.8486 (Large Effect). - Step 3: Apply Hedges' g Small-Sample Bias Correction:
Correction Factor J = 1 − [ 3 / (4(30) − 9) ] = 1 − [ 3 / 111 ] = 1 − 0.0270 = 0.9730.
g = 0.8486 × 0.9730 = 0.8257.
Frequently Asked Questions About Effect Size
Why did Jacob Cohen establish 0.2, 0.5, and 0.8 benchmarks?
Jacob Cohen suggested these operational conventions in 1969 based on typical behavioral science distributions. A small effect (d = 0.2) is barely noticeable to casual observers; a medium effect (d = 0.5) is visible to the naked eye (like the height difference between 14-year-old and 15-year-old girls); a large effect (d = 0.8) is unmistakable (like the height difference between 13-year-old and 18-year-old girls).
Can Cohen's d exceed 1.0 or 2.0?
Yes. Cohen's d has no theoretical upper bound. In rigorous engineering interventions or highly potent pharmacology (e.g., comparing epinephrine to placebo for severe anaphylaxis), effect sizes can exceed d = 3.0 or 5.0.
How do you convert between Cohen's d and Pearson's correlation (r)?
Under equal group sample sizes: r = d / √(d2 + 4) and d = 2r / √(1 − r2). For example, d = 0.50 converts to r = 0.50 / √(0.25 + 4) = 0.50 / 2.0615 = 0.2425.
What is the Common Language Effect Size (CLES / Probability of Superiority)?
CLES is the probability that a randomly selected individual from the treatment group has a higher score than a randomly selected individual from the control group. For d = 0.80, CLES = 71.4% (a randomly selected treatment patient outperforms a control patient 71.4% of the time).
Why do meta-analysts prefer Hedges' g over Cohen's d?
Hedges' g removes the systematic positive estimation bias present in Cohen's d when sample sizes are small (n < 20), preventing aggregated meta-analytic effect sizes from becoming artificially inflated.
Non-Parametric Effect Sizes: Cliff's Delta and Vargha-Delaney A
When continuous sample data is heavily skewed, ordinal, or violates Gaussian normality assumptions, parametric Cohen's d is invalid. Statisticians deploy non-parametric rank-based effect size metrics:
δ = [ #(x1 > x2) − #(x1 < x2) ] / (n1 × n2)
Measures the probability that a randomly selected observation from Group 1 is greater than Group 2, minus the reverse probability (−1.0 to +1.0).
• Small Effect: |δ| ≥ 0.147 | Medium: |δ| ≥ 0.330 | Large: |δ| ≥ 0.474.
2. Vargha-Delaney A Measure (A12 / Stochastic Superiority):
A12 = [ R1 / n1 − (n1 + 1) / 2 ] / n2
where R1 is the rank sum of Group 1. An A12 of 0.50 indicates complete stochastic equality; A12 > 0.71 indicates a large positive effect.
3. Wilcoxon Rank-Sum r-Metric:
r = Z / √N
Converts the standard normal Z-statistic from a Mann-Whitney U test into a correlation-scale effect size.
Meta-Analysis Inverse-Variance Weighting and Standard Errors
In medical meta-analyses combining Cohen's d across independent randomized clinical trials, each study is weighted inversely by its squared standard error (estimation variance):
SEd = √[ (n1 + n2) / (n1n2) + d2 / (2(n1 + n2)) ]
Study Weight wi = 1 / (SEd,i)2
Pooled Meta-Analytic Effect Size (Fixed-Effect Model):
d̄pooled = ∑ (wi × di) / ∑ wi
Cohen's U-Measures of Distributional Non-Overlap
| Cohen's d Magnitude | Cohen's U1 (Non-Overlap) | Cohen's U3 (Percentile Standing) | Common Language Effect (CLES) | Percentage of Overlapping Area (OL) |
|---|---|---|---|---|
| d = 0.00 (Zero Effect) | 0.0% | 50.0% | 50.0% | 100.0% |
| d = 0.20 (Small Effect) | 14.8% | 57.9% | 55.6% | 92.3% |
| d = 0.50 (Medium Effect) | 33.0% | 69.1% | 63.8% | 80.3% |
| d = 0.80 (Large Effect) | 47.4% | 78.8% | 71.4% | 68.9% |
| d = 1.20 (Very Large) | 62.2% | 88.5% | 80.2% | 54.9% |
| d = 2.00 (Huge Effect) | 81.1% | 97.7% | 92.1% | 31.7% |
Converting Between Binary Odds Ratios and Continuous Cohen's d
In systematic meta-analyses spanning both epidemiological medical trials (which report binary Odds Ratios) and psychological interventions (which report continuous mean differences), researchers convert across metrics using the Logistic Distribution Transformation (Chinn, 2000):
d = ln(OR) × √3 / π ≈ ln(OR) / 1.814
Cohen's d to Odds Ratio Conversion:
OR = exp[ d × π / √3 ] ≈ exp[ 1.814 × d ]
Example: A medical therapy produces an Odds Ratio of OR = 2.50 (patients are 2.5 times more likely to recover):
d = ln(2.50) / 1.814 = 0.9163 / 1.814 = 0.505 (Medium Continuous Effect Size).
Multiple Regression Effect Sizes: Cohen's f-Squared
In multiple linear regression, the standardized effect size for the overall model or an added subset of predictors is quantified by Cohen's f2:
f2 = R2 / (1 − R2)
2. Partial / Incremental Effect Size for Added Predictor (B):
f2partial = [ R2AB − R2A ] / [ 1 − R2AB ]
• Small Effect: f2 = 0.02 | Medium: f2 = 0.15 | Large: f2 = 0.35.
The 10-Point Evidence-Based Effect Size Reporting Protocol
- Report Effect Sizes for All Primary Endpoints: Mandate standardized effect sizes alongside every reported p-value.
- Present 95% Confidence Intervals for Effect Sizes: Quantify estimation precision using non-central t distribution confidence intervals for d.
- Apply Hedges' g for Samples Below 20: Eliminate small-sample positive estimation bias by applying Hedges' correction factor.
- Use Glass's Delta When Variances Differ: If treatment expands standard deviation, standardize exclusively against the control group s.
- Interpret Against Empirical Domain Benchmarks: In educational software, an intervention effect of d = 0.25 is considered massive, whereas in pharmacology d = 0.50 is modest.
- Convert to Common Language Metrics: Translate Cohen's d to Probability of Superiority (CLES) for non-technical stakeholders.
- Report Omega-Squared for ANOVA: Use ω2 rather than η2 to report unbiased population variance accounted for.
- Calculate Standardized Mean Differences in Meta-Analysis: Weight studies by inverse variance (1/SE2) to generate pooled summary effects.
- Avoid Dichotomous "Significance" Clifftops: Treat effect sizes as continuous indicators of clinical importance regardless of whether p = 0.049 or p = 0.051.
- Verify Pre-Registration Alignment: Ensure reported effect sizes correspond to primary pre-registered outcome variables.
Detailed Effect Size FAQs
What is the difference between standardized and unstandardized effect sizes?
An unstandardized effect size is measured in raw physical units (e.g., blood pressure reduced by 8 mmHg, salary increased by $4,500/year). A standardized effect size (Cohen's d = 0.65) divides raw difference by standard deviation, allowing direct comparison across different tests and measurement scales.
Why is partial eta-squared often larger than standard eta-squared?
In multi-factor ANOVA, partial eta-squared (ηp2) removes variance accounted for by other factors from the denominator, inflating the reported percentage of variance explained.
Can an effect size be statistically insignificant?
Yes. A study with n = 8 patients might observe a large effect (d = 0.90), but due to low statistical power, the 95% confidence interval will span [−0.15, +1.95], resulting in p > 0.05.
What is Mahalanobis Distance as a multivariate effect size?
Mahalanobis Distance (^2$) generalizes Cohen's d to multivariate outcomes (MANOVA), standardizing group separation using the pooled covariance matrix.
How does range restriction attenuate effect sizes?
If a study samples only top-performing university students (restricted range), observed correlations (r) and standardized differences (d) are artificially compressed compared to the general population.
Why do some researchers prefer unstandardized effect sizes?
When physical measurement units are universally understood (e.g., kilograms lost, survival months extended), unstandardized effect sizes are more transparent and directly interpretable for medical clinicians and policy makers.
Constructing Exact Confidence Intervals for Cohen's d
Because the sampling distribution of Cohen's d follows a Non-Central t-Distribution (T ~ t(df, δ)) rather than a standard Gaussian normal distribution, computing exact 95% Confidence Intervals requires inverting the non-central cumulative distribution function:
Observed test statistic tobs = d × √(n1n2 / (n1 + n2))
Find lower non-centrality parameter δlower where Ft(tobs, df, δlower) = 0.975.
Find upper non-centrality parameter δupper where Ft(tobs, df, δupper) = 0.025.
Convert back to Cohen's d scale: CI95% = [ δlower / √(n1n2 / (n1 + n2)), δupper / √(n1n2 / (n1 + n2)) ].
Case Study: Meta-Analytic Clinical Trial Synthesis
A medical meta-analysis synthesizes 4 independent randomized controlled trials evaluating a novel antidepressant versus active control:
| Study Identifier | Sample Size (n1, n2) | Observed Cohen's d | Standard Error (SEd) | Study Weight (wi = 1/SE2) | Weighted Product (wi × di) |
|---|---|---|---|---|---|
| Trial Alpha (2021) | n1 = 40, n2 = 40 | d1 = 0.450 | 0.2265 | 19.49 | 8.771 |
| Trial Beta (2022) | n1 = 100, n2 = 100 | d2 = 0.520 | 0.1432 | 48.78 | 25.366 |
| Trial Gamma (2023) | n1 = 25, n2 = 25 | d3 = 0.380 | 0.2854 | 12.28 | 4.666 |
| Trial Delta (2024) | n1 = 150, n2 = 150 | d4 = 0.480 | 0.1162 | 74.06 | 35.549 |
| Summary Pooled | N = 630 Patients | d̄ = 0.482 | SE = 0.0804 | ∑w = 154.61 | ∑wd = 74.352 |
Conclusion: The pooled summary effect size is d̄ = 74.352 / 154.61 = 0.481 (Medium Effect Size) with 95% Confidence Interval [0.323, 0.639], establishing highly significant and clinically meaningful therapeutic efficacy across trials.
Additional Effect Size FAQs
How does heteroscedasticity affect Cohen's d?
When group variances are unequal (s12 ≠ s22), standard pooled Cohen's d is biased. Researchers use Glass's Delta (standardizing against the control group) or Shieh's generalized standardized mean difference.
What is the relationship between Cohen's d and the Area Under the Curve (AUC)?
In diagnostic medicine and machine learning classification, Cohen's d maps directly to the Receiver Operating Characteristic Area Under the Curve: AUC = Φ(d / √2). An effect size of d = 0.80 yields an AUC of Φ(0.80 / 1.414) = Φ(0.566) = 0.714 (71.4% discriminative accuracy).
Can effect size be used for statistical power calculations?
Yes. A-priori power analysis requires an expected effect size (Cohen's d or f) to calculate the non-centrality parameter (δ = d × √(n/2)) and determine required sample size.
What is Coe's Table of Effect Size Interpretations?
Robert Coe (2002) published benchmark tables linking Cohen's d to percentile gains. For example, an educational effect of d = 0.50 moves the average student from the 50th percentile to the 69th percentile of achievement.
Historical Foundation: The Effect Size Revolution in Meta-Analysis
The mathematical formalization of standardized effect sizes arose through a series of foundational breakthroughs in behavioral statistics and quantitative synthesis:
- Jacob Cohen (1969): Published Statistical Power Analysis for the Behavioral Sciences, introducing standardized mean differences (d), variance explained metrics (f2, η2), and operational rules of thumb (0.2, 0.5, 0.8) to rescue psychological research from severely underpowered studies.
- Gene V. Glass (1976): Coined the term Meta-Analysis to describe the statistical pooling of standardized treatment outcomes across hundreds of distinct psychotherapy trials, introducing Glass's Δ to handle unequal experimental variances.
- Larry Hedges (1981): Developed the unbiased small-sample estimator Hedges' g alongside formal asymptotic sampling distributions and inverse-variance weighting frameworks that form the bedrock of the Cochrane Collaboration and modern evidence-based medicine.
Effect Sizes in Structural Equation Modeling (SEM) and Path Analysis
In complex multivariable path models, researchers analyze both direct and indirect effect sizes:
Standardized Path Coefficient βstd = βunstd × ( σX / σY )
Total Effect Size: Sum of direct standardized path coefficients plus the product of mediating path coefficients: Total Effect = βdirect + ∑ (βmediator,1 × βmediator,2).
Effect Size Troubleshooting and Diagnostics Matrix
| Observed Diagnostic Anomaly | Underlying Cause | Methodological Risk | Remediation Strategy |
|---|---|---|---|
| Massive Cohen's d (> 2.50) in Small Sample (n = 8) | Publication bias, selective reporting, or severe baseline outlier distortion. | Extreme overestimation of treatment efficacy; replication failure. | Calculate Hedges' g; report median-based non-parametric Cliff's Delta. |
| Partial Eta-Squared Sum Exceeds 1.0 (100%) in Factorial ANOVA | ηp2 removes non-factor error variance from denominator. | Misinterpreting ηp2 as total non-overlapping variance explained. | Report generalized eta-squared (ηG2) or omega-squared (ω2). |
| Discrepancy Between Parametric d and Non-Parametric A12 | Heavy skewness, floor/ceiling effects, or extreme response outliers. | Parametric d is distorted by non-normal error distribution. | Report Vargha-Delaney A or Cliff's Delta alongside median differences. |
| Attenuated Correlation (r = 0.20) in High-Selectivity Cohort | Range restriction on independent predictor variable. | Severe underestimation of true population association strength. | Apply Thorndike's Case II Range Restriction Correction formula. |
Glossary of Standardized Effect Size Terminology
- Standardized Mean Difference (Cohen's d):
- The difference between two group means divided by the pooled standard deviation, expressing gap magnitude in standard deviation units.
- Hedges' g:
- An unbiased small-sample variation of Cohen's d multiplying d by a degree-of-freedom correction factor (J = 1 − 3/(4df − 1)).
- Glass's Delta (Δ):
- Standardized mean difference computed using exclusively the control group standard deviation when treatment alters population variance.
- Eta-Squared (η2):
- Proportion of total experimental variance accounted for by a specific treatment factor in ANOVA (SSeffect / SStotal).
- Omega-Squared (ω2):
- An unbiased population estimate of variance explained in ANOVA adjusting for sample degrees of freedom.
- Common Language Effect Size (CLES):
- The probability that a randomly sampled individual from the experimental group scores higher than a randomly sampled control individual.
- Cliff's Delta:
- Non-parametric ordinal effect size measuring the dominance probability of Group 1 over Group 2 without distributional assumptions.
- Cohen's f2:
- Standardized effect size metric in multiple regression quantifying explained variance relative to unexplained residual variance (R2 / (1 − R2)).
Reporting Standards: APA 7th Edition and CONSORT Guidelines
Leading scientific organizations and regulatory bodies mandate reporting standardized effect sizes for all empirical investigations:
- American Psychological Association (APA 7th Edition): Requires reporting of effect sizes for all primary and secondary statistical outcomes, explicitly noting that p-values alone are insufficient for scientific communication.
- CONSORT 2010 Statement for Clinical Trials: Mandates reporting both unstandardized effect sizes (with confidence intervals in original physical units) and standardized metrics to allow cross-study synthesis in systematic Cochrane reviews.
- Open Science Badges & Pre-Registration: Encourages specifying target effect sizes a-priori to justify sample size recruitment targets before collecting observational data.
Executive Summary: Best Practices for Effect Size Reporting
To ensure reproducible and transparent scientific reporting:
- Always Accompany P-Values with Effect Sizes: Never report a significance test without an accompanying standardized or unstandardized effect size estimate.
- Provide 95% Confidence Intervals: Quantify the uncertainty of the effect size estimate using non-central distribution confidence intervals.
- Use Hedges' g for Small Samples: Apply small-sample degree-of-freedom corrections whenever group sample sizes are under 20.
- Interpret Against Practical Domain Benchmarks: Evaluate effect sizes relative to historical literature in the specific field rather than blindly relying on generic Cohen thresholds.
Step-by-Step Selection: Choosing the Optimal Effect Size Metric
Follow this systematic decision tree to identify the correct effect size index for any empirical study design:
- Two Independent Group Means (Continuous): If sample sizes are large (n > 20 per group) and variances are equal, report Cohen's d; if sample sizes are small (n < 20), report Hedges' g; if treatment alters group variance, report Glass's Δ.
- Two Groups with Skewed / Ordinal Data: Deploy non-parametric effect sizes such as Cliff's Delta (δ), Vargha-Delaney A measure (A12), or the Wilcoxon r rank-biserial correlation.
- Multi-Group Continuous Comparisons (ANOVA): In one-way or factorial ANOVA, report Omega-Squared (ω2) or Epsilon-Squared (ε2) for unbiased population variance accounting.
- Multiple Linear Regression: Report Cohen's f2 for overall and partial explained variance increments (f2 = R2 / (1 − R2)).
- 2 × 2 Categorical Proportions: Report Odds Ratios (OR) for case-control epidemiological studies and Relative Risk (RR) for prospective cohort clinical trials.
Interpreting Practical Clinical Importance vs Statistical Abundance
In modern scientific methodology, researchers emphasize that statistical significance measures evidence against chance, while standardized effect size communicates substantive real-world impact. When analyzing large-scale clinical registries, education initiatives, and genomic association studies, reporting exact effect sizes alongside 95% confidence intervals ensures that findings are interpreted with scientific integrity, clinical relevance, and statistical transparency.