How to Calculate Effect Size in Medical Research (Cohen's d Explained)

Q: What is effect size in medical research?

Effect size is a quantitative measure of the magnitude of a difference or relationship, independent of sample size. While a p-value tells you whether an observed difference is likely to be real (not due to chance), effect size tells you how large that difference actually is in practical terms. Cohen's d is the most common effect size for comparing two means. It expresses the difference between group means in units of pooled standard deviations. A Cohen's d of 0.5 means the group means differ by half a standard deviation — regardless of whether the study had 20 or 2,000 participants.

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

Jacob Cohen's (1988) benchmarks classify effect sizes as: small (d = 0.2), medium (d = 0.5), and large (d = 0.8). In medical research, these thresholds are interpreted contextually: a small effect (d = 0.2) may be clinically meaningful for a preventive intervention with very low cost and risk (e.g., aspirin for cardiovascular prevention), while a large effect (d = 0.8) may still be clinically insufficient if it represents only a modest absolute reduction in disease burden. Always report the effect size alongside its 95% confidence interval and interpret it in the clinical context.

Q: Why is a p-value alone insufficient in medical research?

A p-value only tells you whether an effect is distinguishable from zero given the sample size — it does not tell you how large the effect is. With a sufficiently large sample (thousands of participants), even a clinically trivial difference (e.g., 0.1 mmHg blood pressure reduction) will produce a statistically significant p-value (p < 0.001). Conversely, a small study may fail to reach statistical significance (p = 0.12) even when the true effect size is large and clinically meaningful (d = 0.7). The p-value is determined by both the effect size AND the sample size; it conflates these two distinct pieces of information. Effect size separates them.

Q: How do you calculate Cohen's d?

Cohen's d = (M₁ − M₂) / SD_pooled, where M₁ and M₂ are the group means and SD_pooled is the pooled standard deviation. For equal group sizes: SD_pooled = √[(SD₁² + SD₂²) / 2]. For unequal group sizes: SD_pooled = √[((n₁−1)SD₁² + (n₂−1)SD₂²) / (n₁+n₂−2)]. Example: if a drug group has mean = 120 mmHg (SD = 15) and a placebo group has mean = 130 mmHg (SD = 18), then SD_pooled = √[(15² + 18²)/2] = √[(225+324)/2] = √274.5 = 16.57. Cohen's d = (120−130)/16.57 = −10/16.57 = −0.60 (medium effect, drug lowers blood pressure by 0.60 pooled SDs).

Q: What is the difference between Cohen's d and Hedges' g?

Cohen's d can overestimate the true population effect size when sample sizes are small, because it uses the sample standard deviation as an estimate of the population standard deviation. Hedges' g applies a correction factor (J) to adjust for this small-sample bias: g = d × J, where J ≈ 1 − (3/(4(n₁+n₂)−9)). For large samples (n > 20 per group), Cohen's d and Hedges' g are nearly identical. For small samples (n < 20 per group), Hedges' g is preferred as it is less biased. Most meta-analyses use Hedges' g rather than Cohen's d for this reason.

Q: What is eta squared and when do you use it?

Eta squared (η²) is the effect size measure for ANOVA, representing the proportion of total variance in the outcome explained by the group variable. η² = SS_between / SS_total. Cohen's benchmarks for η²: small = 0.01, medium = 0.06, large = 0.14. Partial eta squared (ηp²) is preferred in factorial ANOVA because it excludes the variance of other factors: ηp² = SS_effect / (SS_effect + SS_error). SPSS reports ηp² by default. Omega squared (ω²) is a less biased alternative to η² for smaller samples. Both η² and ηp² are available in SPSS under the Options menu in the GLM/ANOVA procedure.

Q: Can an effect be statistically significant but not clinically meaningful?

Yes — this is one of the most important concepts in medical research. With large sample sizes, even trivially small differences can produce highly significant p-values. For example, a trial of 10,000 patients comparing two antihypertensive drugs might find a statistically significant difference of 0.8 mmHg in systolic blood pressure (p = 0.003, d = 0.04). While the p-value confirms the difference is real, a 0.8 mmHg reduction has no measurable cardiovascular benefit and is within measurement error for most clinical sphygmomanometers. The reverse is also possible: a small study with d = 0.70 (a large, clinically meaningful effect) might fail to reach p < 0.05 simply because it was underpowered. Effect size and clinical significance must always be interpreted alongside p-values.

Q: How do I report effect size in a medical research paper?

For independent samples t-test: 'The intervention group showed significantly lower systolic blood pressure (M = 128.4 mmHg, SD = 14.2) compared to controls (M = 134.6 mmHg, SD = 15.8); t(118) = 2.31, p = 0.023, Cohen's d = 0.41 (95% CI 0.05–0.77), indicating a small-to-medium effect size.' For ANOVA: 'There was a significant main effect of treatment group on HbA1c, F(2,147) = 8.42, p < 0.001, ηp² = 0.103 (medium-to-large effect).' Always include the effect size measure, its numeric value, a 95% confidence interval if available, and a verbal descriptor (small/medium/large). CONSORT 2010 requires effect size reporting for RCTs; APA 7th edition requires it for any inferential statistical test.

Q: What effect size measure should I use for non-parametric tests?

For Mann-Whitney U test (non-parametric alternative to independent t-test), the appropriate effect size is r = Z / √N, where Z is the standardised test statistic and N is total sample size. Cohen's benchmarks for r: small = 0.1, medium = 0.3, large = 0.5. For Wilcoxon signed-rank test (paired), use the same r = Z / √N formula. For Kruskal-Wallis test (non-parametric ANOVA alternative), use epsilon squared (ε²) or eta squared computed from the H statistic. For chi-square tests of association, use Cramér's V: small = 0.1, medium = 0.3, large = 0.5 (for 2×2 tables; thresholds adjust for larger tables). Always match the effect size measure to the statistical test used.

Q: What is a minimum clinically important difference and how does it relate to effect size?

The Minimum Clinically Important Difference (MCID) is the smallest change in a patient-reported or clinical outcome that patients or clinicians consider meaningful in practice. It is determined by patient surveys, expert consensus, or anchor-based methods, and exists independently of statistical significance. Effect size translates the statistical result into standardised units, but MCID translates it into clinically relevant units. For example, for the SF-36 Quality of Life scale, the MCID is approximately 5 points. An intervention producing a 3-point improvement with d = 0.30 and p = 0.02 is statistically significant with a small effect — but falls below the MCID, meaning patients would not notice the improvement. When an MCID is established for your outcome, report both the effect size and whether the observed difference exceeds the MCID.

A trial of 12,000 patients finds that a new antihypertensive drug significantly reduces systolic blood pressure compared to placebo (p < 0.001). The headline looks impressive. But when you read the data, the actual reduction is 1.2 mmHg — a difference far smaller than measurement error and with no proven cardiovascular benefit. The p-value was tiny because the sample was enormous, not because the treatment effect was meaningful. This scenario, repeated across thousands of published trials, is why effect size has become one of the most important concepts in modern medical research. Effect size tells you not just whether a difference exists, but how large it is in practical terms. Cohen's d is the most widely used effect size measure for comparing two group means, and understanding it is essential for every researcher who conducts, reads, or reports clinical research.

What Is Effect Size?

Effect size is a standardised, sample-size-independent quantification of the magnitude of a result. It answers the question that a p-value cannot: "How big is this difference, really?"

Where the p-value depends on both the true effect size and the sample size, an effect size measure isolates the magnitude of the effect and expresses it in a way that is interpretable regardless of the number of participants. A Cohen's d of 0.50 means the group means are separated by half a pooled standard deviation — whether the study enrolled 30 participants or 3,000.

Effect size serves several distinct purposes in medical research:

Clinical interpretation: Translates statistical results into a language that answers whether the difference matters to patients or clinicians
Sample size planning: Required input for power calculations; you must estimate the expected effect size to determine how many participants you need
Meta-analysis: The foundational unit of systematic review and meta-analysis; pooling results across studies requires a common effect size metric
Journal reporting: Required by APA 7th edition, CONSORT 2010, and most leading clinical journals for any inferential test
Cross-study comparison: Allows meaningful comparison of interventions across different studies using different measurement scales

Key Definition Effect size = a standardised measure of the magnitude of a difference or relationship. It tells you HOW BIG an effect is, independently of sample size. P-values tell you whether an effect is statistically detectable in your sample. Effect size tells you whether it is clinically or practically important.

Why the P-Value Alone Is Insufficient

The p-value is not a measure of effect size. It is a probability: specifically, the probability of observing a result as extreme as the one obtained (or more extreme) if the null hypothesis were true. Three facts about p-values explain why they are insufficient on their own:

1. P-values are inflated by sample size

The p-value is mathematically driven by both the true effect size and the sample size. As sample size increases, even a trivially small effect will produce a smaller and smaller p-value. Given enough participants, the statistical test will detect any non-zero difference — no matter how clinically meaningless. Conversely, a large and clinically important effect may fail to reach p < 0.05 in a small study, leading to a false conclusion that the treatment does not work.

p = 0.0001

Cohen's d = 0.04

n = 50,000 per group

Trivial effect, huge sample.
Statistically "significant" but
clinically meaningless.

Misleading p-value

p = 0.12

Cohen's d = 0.70

n = 20 per group

Large clinically important effect,
underpowered study fails to
reach significance threshold.

Underpowered study

p = 0.038

Cohen's d = 0.52

n = 60 per group

Medium effect, adequate power.
p-value and effect size both
support the conclusion.

Complete reporting

p = 0.049

Cohen's d = 0.18

n = 500 per group

Just significant, but effect
is trivial. Journal-publishable
but clinically uninformative.

Trivial effect, significant p

2. P-values cannot be directly compared across studies

Two studies showing p = 0.03 may reflect completely different magnitudes of effect if they had different sample sizes. Effect sizes can be directly compared across studies — they exist on a common standardised scale. This is why systematic reviews pool effect sizes, not p-values.

3. P-values are binary thresholds applied to continuous evidence

The conventional p < 0.05 threshold treats results as either "significant" or "not significant" — a binary classification of a continuous probability. Cohen's d exists on a continuous scale that captures gradations of clinical importance (trivial, small, medium, large) that p < 0.05 vs p = 0.06 does not. A result with p = 0.051 and d = 0.68 is more clinically informative than p = 0.049 and d = 0.19, but the p-value threshold would call the latter "significant" and the former "not significant."

P-Value What it tells you

Probability of observing this result if H₀ is true
Whether the difference is distinguishable from zero
Affected by sample size — increases with larger n
Does not describe the size of the effect
Cannot be compared between studies directly
Required for hypothesis testing decision

Effect Size What it tells you

Standardised magnitude of the difference or relationship
How large the effect is in practical terms
Independent of sample size — stable across studies
Describes practical or clinical importance
Directly comparable and poolable across studies
Required for sample size calculation and meta-analysis

Cohen's d: Formula and Derivation

Cohen's d is the most widely used effect size measure for comparing the means of two independent groups. It was formalised by Jacob Cohen in his landmark 1988 textbook Statistical Power Analysis for the Behavioral Sciences, and it standardises the mean difference by the pooled standard deviation of the two groups, making it interpretable on a common scale regardless of the original unit of measurement.

Cohen's d — Independent Samples

d = (M₁ − M₂) / SD_pooled

M₁ = mean of group 1 (intervention/treatment)

M₂ = mean of group 2 (control/comparator)

SD_pooled = pooled standard deviation. For equal group sizes: SD_pooled = √[(SD₁² + SD₂²) / 2]. For unequal group sizes: SD_pooled = √[((n₁−1)SD₁² + (n₂−1)SD₂²) / (n₁+n₂−2)]

Step-by-Step Calculation

The calculation breaks into three steps that you can execute with any calculator or spreadsheet:

Calculate the mean difference: Subtract one group mean from the other. The sign of Cohen's d indicates direction (positive = group 1 higher; negative = group 1 lower) but the magnitude is what matters for interpretation.
Calculate the pooled standard deviation: For equal sample sizes, average the two variances (SD²) and take the square root. For unequal sizes, use the degrees-of-freedom weighted formula that gives more weight to the larger group.
Divide: Mean difference ÷ pooled SD = Cohen's d. The result tells you how many pooled standard deviations apart the two group means are.

Paired Samples (Pre-Post or Matched) Cohen's d

d = M_diff / SD_diff

M_diff = mean of the within-subject differences (post − pre for each participant)

SD_diff = standard deviation of those within-subject differences

For paired designs, use the standard deviation of the difference scores, not the pooled SD of the two time points. This accounts for the correlation between measurements and gives a different (usually larger) effect size than the independent-samples formula applied to the same data.

Hedges' g: The Small-Sample Correction

Cohen's d uses sample standard deviations to estimate population standard deviations. In small samples (n < 20 per group), this produces a slight upward bias in the effect size estimate. Hedges' g applies a correction factor J to remove this bias:

Hedges' g — Bias-Corrected Effect Size

g = d × J where J ≈ 1 − 3 / (4(n₁+n₂) − 9)

When to use Hedges' g: When n per group is small (<20), or in meta-analyses where you are pooling effect sizes from studies with variable sample sizes. For n ≥ 30 per group, g and d differ by less than 2% and either is acceptable. Most meta-analysis software reports Hedges' g rather than Cohen's d by default.

Cohen's d Interpretation Thresholds

Jacob Cohen proposed his benchmarks as tentative, context-dependent guidelines — not absolute rules. He explicitly stated that "what is small in one context may be large in another." Nevertheless, his thresholds have become the universal reference point in medical research:

Trivial

< 0.2

The two groups overlap almost entirely. The difference is unlikely to be noticed by patients or clinicians even if statistically significant in a large trial. Example: a 0.3 kg weight difference after 12 weeks of dietary intervention.

Small

0.2 – 0.5

Visible to a careful observer but not obvious in everyday clinical practice. May be meaningful for population-level interventions with very low cost or risk. Example: aspirin's effect on cardiovascular events in primary prevention (d ≈ 0.20–0.25).

Medium

0.5 – 0.8

Noticeable to a clinician and likely to be visible in aggregate patient outcomes. Cohen considered this the "default" expected effect for a moderately effective intervention. Example: a second-line antihypertensive reducing SBP by 8–12 mmHg vs placebo.

Large

0.8 – 1.2

Immediately obvious in clinical practice. The group distributions overlap modestly. Most clinicians would observe the difference in individual patients. Example: an effective antibiotic vs placebo for an acute bacterial infection (fever resolution time).

Very Large

> 1.2

Striking, unmistakeable differences rarely seen in pharmacological trials but common in comparisons of two very different conditions (e.g., disease vs healthy control groups) or highly effective surgical versus medical interventions.

Context Modifies These Thresholds Cohen's benchmarks were derived from psychological research and should not be applied mechanically in all medical contexts. A drug reducing 30-day mortality by 1.5% absolute risk may have Cohen's d = 0.08 (trivial by the scale above) but represents a clinically enormous benefit across millions of patients. Always contextualise the effect size against: (1) the baseline risk of the outcome, (2) the cost and risk of the intervention, (3) the MCID (minimum clinically important difference) for the specific outcome measure, and (4) what effect sizes are typical for effective treatments in your therapeutic area.

Clinical Trial Examples

Antihypertensive RCT: The Statistical Significance Trap

Double-blind RCT • New ARB vs placebo • Primary outcome: systolic blood pressure at 12 weeks • n = 2,400 per group

A large industry-funded RCT evaluates a new angiotensin receptor blocker against placebo in 4,800 mildly hypertensive adults. The primary endpoint is change in systolic blood pressure (SBP) at 12 weeks. The sample is large by design to guarantee statistical significance.

ARB Group

M = 138.4 mmHg

SD = 14.2 mmHg

n = 2,400

Placebo Group

M = 139.8 mmHg

SD = 14.7 mmHg

n = 2,400

Step 1: Mean Difference M₁ − M₂ = 138.4 − 139.8 = −1.4 mmHg
Step 2: Pooled SD (equal n) SD_pooled = √[(14.2² + 14.7²) / 2] = √[(201.64 + 216.09) / 2] = √[208.87] = 14.45 mmHg
Step 3: Cohen's d d = −1.4 / 14.45 = −0.097 = Trivial
Reported P-Value t(4798) = −3.36, p = 0.0008 — highly significant

Interpretation: The p-value of 0.0008 appears to endorse this drug powerfully. The effect size tells the truth: d = −0.097 is a trivial effect. The drug reduces SBP by 1.4 mmHg on average — a difference that falls below measurement reproducibility for most clinical sphygmomanometers (typical device-to-device variation: ±2–4 mmHg) and far below the minimum clinically meaningful SBP reduction established in cardiovascular outcomes trials (approximately 5 mmHg for measurable event reduction). The huge sample size of 4,800 made the trivial difference statistically detectable. This drug would correctly be classified as clinically ineffective despite its highly significant p-value.

Weight Loss Intervention: A Genuine Medium Effect

RCT • 16-week structured dietary + behavioural intervention vs standard care • Outcome: body weight (kg) • n = 45 per group

A randomised trial of a structured dietary-behavioural intervention versus standard dietary advice in overweight adults (BMI 27–35) measures weight loss at 16 weeks. This is a typical adequately-powered pilot RCT.

Intervention Group

M = 81.4 kg

SD = 11.2 kg

n = 45

Standard Care

M = 87.8 kg

SD = 12.6 kg

n = 45

Step 1: Mean Difference M₁ − M₂ = 81.4 − 87.8 = −6.4 kg
Step 2: Pooled SD (equal n) SD_pooled = √[(11.2² + 12.6²) / 2] = √[(125.44 + 158.76) / 2] = √[142.10] = 11.92 kg
Step 3: Cohen's d d = −6.4 / 11.92 = −0.537 = Medium
Hedges' g (small-sample correction) J = 1 − 3/(4×90 − 9) = 1 − 3/351 = 0.9915
g = 0.537 × 0.9915 = 0.532 (negligible correction at n=45)
Reported Result t(88) = −2.54, p = 0.013, Cohen's d = −0.54, 95% CI −0.96 to −0.11

Interpretation: A medium effect size (d = 0.54) with statistical significance (p = 0.013) and a 95% CI that does not cross zero (−0.96 to −0.11). The −6.4 kg mean weight loss in the intervention group versus −0.0 kg in standard care (implied by these means) represents a clinically meaningful difference — exceeding the commonly cited MCID of 5% body weight for metabolic benefits in overweight adults. This is an example of statistical significance and clinical significance aligning correctly. Note how including the effect size and its 95% CI gives readers far more information than the p-value alone.

Cognitive Rehabilitation After Stroke: A Large Effect in a Small Pilot

Pilot RCT • 8-week structured cognitive rehabilitation vs usual care • Outcome: MoCA score (0–30) • n = 18 per group

A small pilot RCT of cognitive rehabilitation following ischaemic stroke measures Montreal Cognitive Assessment (MoCA) score at 8 weeks. The sample is too small to expect statistical significance, but the effect size will inform the power calculation for the definitive trial.

Rehabilitation Group

M = 22.8

SD = 3.4 MoCA points

n = 18

Usual Care

M = 19.3

SD = 3.9 MoCA points

n = 18

Step 1: Mean Difference M₁ − M₂ = 22.8 − 19.3 = 3.5 MoCA points
Step 2: Pooled SD (equal n) SD_pooled = √[(3.4² + 3.9²) / 2] = √[(11.56 + 15.21) / 2] = √[13.39] = 3.66 points
Step 3: Cohen's d d = 3.5 / 3.66 = 0.957 = Large
Hedges' g (small n correction) J = 1 − 3/(4×36 − 9) = 1 − 3/135 = 0.9778
g = 0.957 × 0.9778 = 0.936
Reported Result t(34) = 2.02, p = 0.051, Hedges' g = 0.94, 95% CI −0.01 to 1.87
Interpretation: large effect, underpowered study, wide CI

Interpretation: The study just misses the conventional p < 0.05 threshold (p = 0.051) — but the effect size is large (g = 0.94) and the 3.5-point MoCA difference exceeds the established MCID of 2 points for this scale. The wide 95% CI (−0.01 to 1.87) reflects the small sample size and substantial uncertainty, but the point estimate strongly justifies proceeding to a definitive trial. Using this effect size to power the main trial: with d = 0.94, α = 0.05, power = 80%, the required sample size is approximately n = 20 per group; for 90% power, n = 26 per group. Without the effect size estimate from this pilot, the main trial cannot be planned.

Antibiotic Duration RCT: Paired Pre-Post Design

Crossover pilot • 5-day vs 7-day amoxicillin for community-acquired pneumonia • Outcome: CAP symptom score • n = 52 participants (paired)

A within-patient crossover study measures symptom score at day 7 and day 10 in patients with CAP. The paired design uses each patient as their own control, increasing precision. Cohen's d for paired data uses the standard deviation of the within-patient difference scores.

Day 7 Score (End of 5-day tx)

M = 8.4

SD = 3.8

n = 52

Day 10 Score (End of 7-day tx)

M = 6.1

SD = 3.2

n = 52

Within-Patient Differences (Day 10 − Day 7) M_diff = 6.1 − 8.4 = −2.3 points (score decreases = improvement)
SD_diff = 2.9 (SD of individual difference scores, not SDs of each time point)
Paired Cohen's d d = M_diff / SD_diff = −2.3 / 2.9 = −0.793 = Medium–Large
Reported Result t(51) = −5.72, p < 0.001, Cohen's d_paired = −0.79, 95% CI −1.13 to −0.45
Interpretation: medium-large meaningful clinical improvement from day 7 to day 10

Interpretation: The paired design detects a medium-large effect (d = 0.79) of extending treatment from 5 to 7 days, with p < 0.001 and a narrow 95% CI (−1.13 to −0.45) indicating high precision. Note that the paired d is computed using SD_diff (the SD of the change scores), not the pooled SD of the two time points — a critical distinction. If the independent-samples formula had been incorrectly applied to these paired data, d would have been falsely smaller (≈ 0.65) because the correlation between paired measurements would not have been accounted for. This example also demonstrates how a medium-large effect size, combined with the clinical MCID for the CAP symptom score, informs the clinical recommendation to prefer 7-day over 5-day treatment in more severe presentations.

Effect Size in ANOVA: Eta Squared and Omega Squared

When comparing more than two groups — the domain of ANOVA — Cohen's d is not applicable because it is defined for exactly two groups. Three effect size measures are used for ANOVA:

η² Eta Squared

Proportion of total variance explained by the group variable. Simple to calculate: SS_between / SS_total. Tends to slightly overestimate population effect size, especially in smaller samples.

0.01Small

0.06Medium

0.14Large

ηp² Partial Eta Squared

Proportion of variance explained by a specific factor controlling for other factors in a factorial ANOVA. SPSS reports ηp² by default in GLM output. Always larger than η² in factorial designs.

0.01Small

0.06Medium

0.14Large

ω² Omega Squared

Less biased estimate of the population effect size than η². Preferred for smaller samples. Formula: (SS_between − (k−1)MS_within) / (SS_total + MS_within). Not automatically reported by SPSS — must be calculated manually.

0.01Small

0.06Medium

0.14Large

Example: One-Way ANOVA Comparing Three Analgesic Regimens

A clinical trial randomises 180 post-operative patients to three analgesic regimens (NSAID, opioid, combined) and measures pain score (0–10) at 4 hours:

Source	SS	df	MS	F	p	Effect Size
Between Groups (Regimen)	84.6	2	42.3	9.24	< 0.001	η² = 0.094
Within Groups (Error)	813.4	177	4.60	—	—	—
Total	898.0	179	—	—	—	—

Eta Squared η² = SS_between / SS_total = 84.6 / 898.0 = 0.094 — Medium
Omega Squared (Less Biased) ω² = (SS_between − (k−1)MS_within) / (SS_total + MS_within)
ω² = (84.6 − 2 × 4.60) / (898.0 + 4.60) = (84.6 − 9.2) / 902.6 = 75.4 / 902.6 = 0.084 — Medium

Both η² = 0.094 and ω² = 0.084 exceed Cohen's medium threshold of 0.06, indicating that analgesic regimen explains approximately 8–9% of the variance in post-operative pain score — a practically meaningful effect. The ANOVA effect size tells you that the choice of regimen matters; post-hoc tests (Tukey, Bonferroni) then identify which specific pairs of regimens differ significantly.

Effect Size for Other Statistical Tests

Cohen's d and eta squared are the most common effect size measures, but every type of statistical test has its appropriate counterpart:

Statistical Test	Effect Size Measure	Formula	Benchmarks (small / medium / large)
Independent t-test	Cohen's d	(M₁−M₂) / SD_pooled	0.2 / 0.5 / 0.8
Paired t-test	Cohen's d (paired)	M_diff / SD_diff	0.2 / 0.5 / 0.8
One-way ANOVA	Eta squared (η²)	SS_between / SS_total	0.01 / 0.06 / 0.14
Factorial ANOVA	Partial eta squared (ηp²)	SS_effect / (SS_effect+SS_error)	0.01 / 0.06 / 0.14
Chi-square (2×2)	Phi (φ)	√(χ²/N)	0.1 / 0.3 / 0.5
Chi-square (r×c)	Cramér's V	√(χ²/(N×df_min))	0.1 / 0.3 / 0.5 (2×2)
Pearson correlation	r	Correlation coefficient directly	0.1 / 0.3 / 0.5
Mann-Whitney U	r = Z/√N	Standardised Z from output / √N	0.1 / 0.3 / 0.5
Wilcoxon signed-rank	r = Z/√N	Standardised Z from output / √N	0.1 / 0.3 / 0.5
Logistic regression	Odds ratio / Nagelkerke R²	From regression output	OR: 1.5 / 2.5 / 4.3 (approx.)

Statistical Significance vs Practical Significance

The distinction between statistical significance and practical (clinical) significance is one of the most important conceptual shifts in modern evidence-based medicine. They are related but not synonymous, and conflating them is responsible for a significant proportion of misinterpreted clinical research.

The Core Distinction Statistical significance (p < 0.05) means the effect is probably real — not due to chance given your sample. Practical/clinical significance means the effect is large enough to matter in real-world patient care. An effect can be one without the other: real but trivially small, or large and clinically important but not yet confirmed due to an underpowered study.

The clearest way to understand this distinction is through the concept of the Minimum Clinically Important Difference (MCID): the smallest change in an outcome that patients or clinicians consider meaningful. The MCID is established independently of any study results — through patient surveys, expert panels, or anchor-based methods. When a clinical trial reports its results, you should compare the observed effect to the MCID:

Effect > MCID and p < 0.05: Statistically significant AND clinically meaningful. The ideal result.
Effect > MCID but p ≥ 0.05: Clinically meaningful but not yet confirmed statistically — most likely an underpowered study. Requires a larger confirmatory trial. Report effect size and CI.
Effect < MCID and p < 0.05: Statistically confirmed but clinically trivial. Common in large trials. The treatment does something, but not enough to matter. Often produced by oversized industry trials designed to guarantee significance for any measurable effect.
Effect < MCID and p ≥ 0.05: Neither statistically significant nor clinically important. The treatment genuinely does not work at a meaningful level, or the study is underpowered for even the smallest clinically relevant effect.

Common Interpretation Mistakes

Mistake 1: Confusing small effect size with no clinical importance

Cohen explicitly noted that his benchmarks were tentative and context-dependent. A drug that reduces the absolute risk of stroke by 1.5% over 10 years may have Cohen's d = 0.12 (classified "trivial") but, applied across a population of millions, could prevent hundreds of thousands of strokes. Population-level interventions with very low cost and risk profile can be enormously important even with small standardised effect sizes. The absolute difference, the NNT (number needed to treat), and the population prevalence all modulate the real-world impact of even small effect sizes.

✓ Fix: Always supplement effect size with the absolute difference and NNT when reporting clinical trial results. Classify effect size using Cohen's benchmarks as a guideline, not a verdict. State explicitly: "While the standardised effect size is small (d = 0.18), the absolute reduction of 1.5% in 10-year stroke risk represents a clinically meaningful benefit at the population level."

Mistake 2: Using the independent-samples formula for paired data

For paired or repeated-measures designs (pre-post, crossover, matched), the Cohen's d formula using the pooled SD of the two groups is incorrect. The paired design explicitly exploits the correlation between measurements to reduce error variance. Applying the independent-samples formula ignores this correlation and typically underestimates the true effect size. Conversely, using the SD of only the pre-measurement as the standardiser (another common error) inflates d.

✓ Fix: For paired designs, calculate d as the mean of the difference scores divided by the standard deviation of the difference scores (SD_diff), not the pooled SD of the two time points. Report clearly which formula was used: "Cohen's d for paired samples" vs "Cohen's d for independent samples."

Mistake 3: Reporting eta squared from factorial ANOVA as if it were partial eta squared

In factorial ANOVA (multiple independent variables), SPSS reports partial eta squared (ηp²) by default, but many researchers label it as "eta squared" in papers. Eta squared (classical) and partial eta squared differ substantially in factorial designs: partial η² for each factor excludes the variance of other factors from the denominator, making it systematically larger than classical η². Reporting ηp² as η² overstates the proportion of total variance explained by each factor.

✓ Fix: Always specify whether you are reporting η² or ηp² and use the correct label in your results table. In factorial ANOVA, report ηp² and label it explicitly as partial eta squared. Recognise that the sum of ηp² values across all factors in a factorial ANOVA will typically exceed 1.0 — this is normal and not an error.

Mistake 4: Interpreting effect size confidence intervals that cross zero as "no effect"

A Cohen's d 95% confidence interval of −0.05 to 0.85 crosses zero and would be statistically consistent with a null effect — but the point estimate of d = 0.40 represents a meaningful effect, and the wide CI reflects an underpowered study rather than evidence of no effect. Concluding "there was no effect" from a non-significant result with a wide CI containing large positive values is a type II error masquerading as a valid conclusion.

✓ Fix: When the 95% CI for effect size is wide and crosses zero, report the result as "inconclusive" rather than "no effect." State: "The 95% CI for Cohen's d (−0.05 to 0.85) includes both clinically important positive effects and trivial effects, indicating that the study was insufficiently powered to draw a definitive conclusion." Calculate and report the minimum detectable effect at the achieved power to contextualise the result.

Mistake 5: Failing to report effect size at all

A 2021 audit of 200 randomly selected clinical trials in three major general medical journals found that more than 35% reported no effect size measure beyond the raw mean difference. APA 7th edition (2020) requires effect sizes for all inferential tests; CONSORT 2010 requires them for RCT primary endpoints. Despite these requirements, omission remains common, particularly in surgical and procedural trials where statistical analysis training is historically less rigorous.

✓ Fix: Before submission, use a checklist: (1) every t-test result should include Cohen's d; (2) every ANOVA should include ηp² or ω²; (3) every chi-square should include Cramér's V or φ; (4) all effect sizes should have 95% CIs where feasible. Many online calculators and SPSS (Analyze → Compare Means → tick "Effect size") now provide these automatically.

Mistake 6: Applying Cohen's benchmarks across all contexts without adjustment

Cohen derived his benchmarks from psychological research. Medical research has different standards. In oncology, a hazard ratio of 0.75 (reducing mortality by 25%) is clinically dramatic; in pharmacology, a d of 0.3 for a first-line medication versus placebo may be entirely unacceptable if better alternatives exist. In surgical research, effect sizes are often large (d > 1.0) because surgical interventions produce large measurable changes. Using Cohen's benchmarks mechanically without considering the therapeutic area, comparator, and clinical context leads to misclassification of effects as "small" when they are impactful, or "large" when they are still insufficient.

✓ Fix: Contextualise your effect size against: (a) what effect sizes are typical for effective treatments in this therapeutic area; (b) the MCID for the specific outcome; (c) the cost/risk of the intervention; and (d) the existing treatment alternatives. State: "A Cohen's d of 0.42 is categorised as small-to-medium by Cohen's (1988) benchmarks and exceeds the MCID of 4 points on the [outcome scale], consistent with clinical meaningfulness in this patient population."

Journal Reporting Standards

Major guidelines and journals now require effect size reporting. Here is what to include and how to phrase it:

✓ Independent Samples t-Test — APA / CONSORT

"The intervention group demonstrated significantly lower HbA1c at 24 weeks (M = 7.1%, SD = 0.8%) compared to the control group (M = 7.7%, SD = 0.9%); t(118) = 3.97, p < 0.001, Cohen's d = 0.71 (95% CI 0.34–1.07), indicating a medium-to-large effect. The observed mean difference of 0.6% (95% CI 0.30%–0.90%) exceeds the minimum clinically important difference of 0.5% for HbA1c in this population."

✓ One-Way ANOVA — Factorial Design

"There was a statistically significant effect of analgesic regimen on post-operative pain score at 4 hours, F(2, 177) = 9.24, p < 0.001, ηp² = 0.094 (medium effect). Pairwise comparisons (Tukey HSD) revealed that the combined regimen (M = 3.8, SD = 2.1) produced significantly lower pain scores than NSAID alone (M = 5.9, SD = 2.2; p < 0.001, d = 0.97) and opioid alone (M = 5.2, SD = 2.0; p = 0.003, d = 0.69). The NSAID and opioid groups did not differ significantly (p = 0.28, d = 0.33)."

✓ Paired t-Test / Pre-Post Design

"Symptom severity scores decreased significantly from pre-intervention (M = 24.3, SD = 5.8) to post-intervention (M = 17.6, SD = 5.1); t(51) = 8.94, p < 0.001. The mean within-patient reduction was 6.7 points (SD of differences = 5.4), Cohen's d_paired = 1.24 (95% CI 0.87–1.61), indicating a large effect. This change exceeded the minimum clinically important difference of 4 points established for this scale."

✕ Incomplete Report — Fails APA / CONSORT Requirements

"The intervention group had significantly lower HbA1c at 24 weeks compared to controls (p < 0.001)."

How to Get Effect Size from SPSS Independent t-test: Analyze → Compare Means → Independent Samples T-Test → Options → tick "Effect size." SPSS reports Cohen's d with 95% CI. For paired t-test: Analyze → Compare Means → Paired Samples T-Test → same Option. For ANOVA: Analyze → General Linear Model → Univariate → Options → tick "Estimates of effect size" — SPSS reports partial eta squared (ηp²). For chi-square: partial eta squared is not applicable — compute Cramér's V manually from the chi-square statistic: V = √(χ² / (N × df_min)).

Frequently Asked Questions

What is effect size in medical research? +

Effect size is a standardised, sample-size-independent measure of the magnitude of a difference or relationship. While a p-value tells you whether an effect is statistically detectable, effect size tells you how large it is. Cohen's d is the most common effect size for comparing two group means: it expresses the mean difference in units of the pooled standard deviation. A Cohen's d of 0.5 means the groups are separated by half a pooled standard deviation — whether the study had 20 or 20,000 participants. Effect size is essential for clinical interpretation, sample size planning, and meta-analysis.

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

Jacob Cohen's (1988) benchmarks: small = 0.2, medium = 0.5, large = 0.8. In medical research, these thresholds require contextual interpretation. A small effect (d = 0.2) may be clinically meaningful for a widely applied, very low-risk intervention (e.g., aspirin for primary prevention). A large effect (d = 0.8) may still be clinically insufficient if better alternatives achieve d > 1.5. Always interpret the effect size alongside the absolute difference, the MCID for the outcome, the cost/risk of the intervention, and typical effect sizes for effective treatments in the same therapeutic area.

Why is a p-value alone insufficient in medical research? +

A p-value is mathematically determined by both the true effect size and the sample size. With enough participants, any non-zero difference produces a tiny p-value — even a clinically trivial 0.1 mmHg blood pressure reduction in a trial of 50,000 patients will produce p < 0.0001. Conversely, a clinically important effect (d = 0.7) in an underpowered study (n = 20 per group) may fail to reach p < 0.05. The p-value conflates magnitude and sample size. Effect size separates them: d = 0.04 with p < 0.001 tells you the trial had enormous power but a trivial effect. d = 0.70 with p = 0.12 tells you the trial was underpowered for a meaningful effect.

How do you calculate Cohen's d step by step? +

Step 1: Calculate the mean difference: M₁ − M₂. Step 2: Calculate pooled SD. For equal n: SD_pooled = √[(SD₁² + SD₂²)/2]. For unequal n: SD_pooled = √[((n₁−1)SD₁² + (n−−1)SD₂²)/(n₁+n₂−2)]. Step 3: d = (M₁ − M₂) / SD_pooled. Example: Group 1 M=120, SD=15; Group 2 M=130, SD=18. SD_pooled = √[(225+324)/2] = √274.5 = 16.57. d = (120−130)/16.57 = −10/16.57 = −0.60 (medium effect). The negative sign indicates Group 1 is lower; the magnitude 0.60 is the effect size.

What is the difference between Cohen's d and Hedges' g? +

Cohen's d slightly overestimates the true population effect size in small samples because sample SDs are biased estimates of population SDs. Hedges' g applies a correction factor J ≈ 1 − 3/(4(n₁+n₂) − 9) to remove this small-sample bias: g = d × J. For n ≥ 30 per group, d and g differ by less than 2% — either is acceptable. For n < 20 per group, use Hedges' g. Meta-analyses almost universally report Hedges' g rather than Cohen's d because studies with varying sample sizes contribute pooled estimates, and the correction improves accuracy in smaller contributing studies.

What is eta squared and when do you use it? +

Eta squared (η²) is the effect size for ANOVA, equal to the proportion of total variance in the outcome explained by the group variable: η² = SS_between / SS_total. Cohen's benchmarks: small = 0.01, medium = 0.06, large = 0.14. Use η² for one-way ANOVA. For factorial (multi-factor) ANOVA, SPSS reports partial eta squared (ηp²) by default: ηp² = SS_effect / (SS_effect + SS_error). This excludes variance from other factors in the denominator, making it larger than η² in factorial designs. Omega squared (ω²) is a less biased alternative for small samples. Always label which measure you are reporting.

Can an effect be statistically significant but not clinically meaningful? +

Yes — and this is extremely common in large clinical trials. A trial of 10,000 patients comparing two antihypertensives might find a statistically significant SBP difference of 0.8 mmHg (p = 0.003, d = 0.04). The p-value confirms the effect is real. The d = 0.04 (trivial) confirms it is clinically meaningless. The reverse also occurs: a pilot trial of 20 patients with d = 0.72 (large, clinically important effect) may produce p = 0.09 simply because the study was underpowered. This is why both p-value and effect size must be reported and interpreted together, alongside the MCID for the specific outcome measure.

How do I report effect size in a medical research paper? +

For t-test: "t(df) = value, p = value, Cohen's d = value (95% CI lower–upper), indicating a [small/medium/large] effect." For ANOVA: "F(df₁, df₂) = value, p = value, ηp² = value ([medium] effect)." Always include the effect size measure name, numeric value, 95% CI where available, and a verbal descriptor. CONSORT 2010 requires effect size for RCT primary endpoints; APA 7th edition requires it for all inferential tests. In SPSS: Analyze → Compare Means → [T-test or ANOVA] → Options → tick "Effect size" to get Cohen's d with 95% CI automatically.

What effect size measure should I use for non-parametric tests? +

For Mann-Whitney U and Wilcoxon signed-rank tests, use r = Z / √N, where Z is the standardised test statistic from the SPSS output and N is the total sample size. Cohen's benchmarks for r: small = 0.1, medium = 0.3, large = 0.5. For Kruskal-Wallis, use epsilon squared (ε²) = H / (N²−1)/(N+1) or eta squared derived from H. For chi-square tests, use Cramér's V = √(χ² / (N × df_min)), where df_min is the smaller of (rows−1) or (columns−1). Cramér's V benchmarks: small = 0.1, medium = 0.3, large = 0.5 (for 2×2 tables; thresholds adjust upward for larger tables).

What is a minimum clinically important difference and how does it relate to effect size? +

The MCID (Minimum Clinically Important Difference) is the smallest change in an outcome that patients or clinicians consider meaningful in practice. It is determined independently of any study — through patient surveys, expert panels, or anchor-based methods — and provides a clinical benchmark against which effect sizes can be evaluated. Effect size (d) standardises the result; MCID contextualises it. Compare the observed mean difference to the MCID: if the observed difference > MCID and d ≥ 0.2, the result is likely both statistically and clinically important. If d = 0.5 but the mean difference falls below the MCID, the effect is statistically medium but clinically insufficient for this particular outcome and patient group. Always report both the standardised effect size and whether the absolute difference exceeds the established MCID.

Need to calculate p-values and effect sizes?

StatClinic's free statistical tools give you t-test results, ANOVA, and effect size measures ready for thesis and journal submission. No SPSS required.

Open StatClinic Now →

How to Calculate Effect Size in Medical Research(Cohen's d Explained)

What Is Effect Size?

Why the P-Value Alone Is Insufficient

1. P-values are inflated by sample size

2. P-values cannot be directly compared across studies

3. P-values are binary thresholds applied to continuous evidence

P-Value What it tells you

Effect Size What it tells you

Cohen's d: Formula and Derivation

Step-by-Step Calculation

Hedges' g: The Small-Sample Correction

Cohen's d Interpretation Thresholds

Clinical Trial Examples

Antihypertensive RCT: The Statistical Significance Trap

Weight Loss Intervention: A Genuine Medium Effect

Cognitive Rehabilitation After Stroke: A Large Effect in a Small Pilot

Antibiotic Duration RCT: Paired Pre-Post Design

Effect Size in ANOVA: Eta Squared and Omega Squared

η² Eta Squared

ηp² Partial Eta Squared

ω² Omega Squared

Example: One-Way ANOVA Comparing Three Analgesic Regimens

Effect Size for Other Statistical Tests

Statistical Significance vs Practical Significance

Common Interpretation Mistakes

Mistake 1: Confusing small effect size with no clinical importance

Mistake 2: Using the independent-samples formula for paired data

Mistake 3: Reporting eta squared from factorial ANOVA as if it were partial eta squared

Mistake 4: Interpreting effect size confidence intervals that cross zero as "no effect"

Mistake 5: Failing to report effect size at all

Mistake 6: Applying Cohen's benchmarks across all contexts without adjustment

Journal Reporting Standards

Frequently Asked Questions

Need to calculate p-values and effect sizes?

How to Calculate Effect Size in Medical Research
(Cohen's d Explained)