What Is Effect Size?

Effect size is the number that says how big a difference or a relationship actually is β€” in standard-deviation units, so results from different studies can be compared on the same scale. A p-value can only tell you whether an effect is there; an effect size tells you whether it is worth caring about.

This guide explains effect size in plain language, works through six complete examples β€” a small effect, a huge effect, a one-sample effect, an effect straight from a t-statistic, a paired effect and the no-effect case β€” on the site’s own Cohen’s d Calculator, and shows why you should always report an effect size and its confidence interval alongside your p-value.

Short answer: Effect size is a standardized measure of how big an effect is β€” how large a difference, or how strong a relationship β€” independent of sample size and of the original units. Cohen’s d is the most common one: the difference between two means divided by the pooled standard deviation. A p-value says whether one exists; an effect size says how much.

What Is Effect Size?

A new training programme lifts test scores by 5 points on average. Is that a lot? It depends on the spread: 5 points is a modest difference when scores typically vary by 15 points, and a large one when they vary by only 2. An effect size answers that question by measuring the difference against the spread of the data, so the answer no longer depends on the units the scores were measured in β€” or on how many people were tested.

Effect size is a standardised measure of how big an effect is. It covers two kinds of question: how large the difference between two groups is, and how strong the relationship between two variables is. The most common effect size is Cohen’s d: the difference between two means divided by the pooled standard deviation β€” the gap between the groups measured in standard deviations. Jacob Cohen, who introduced d in the 1960s, put its importance bluntly: ‘The primary product of a research inquiry is one or more measures of effect size, not p values.’

Cohen also warned that the interpretation of effect sizes is a judgment process: the same d can matter a great deal in one field and very little in another. That is why this guide gives you the number and the judgement β€” the benchmarks, the confidence interval, and the plain-language reading of U3 and overlap β€” for every example, with each value computed on the site’s own effect-size calculator.

Effect Size vs P-Value

The two numbers answer different questions. A p-value answers ‘is there something?’ β€” how surprising the data would be if there were no effect at all. An effect size answers ‘how much?’ β€” how large the difference is. A tiny, clinically meaningless difference can reach p < 0.001 in a large enough study, and a real, large effect can fail to reach significance in a small one. One number without the other is half the story: always report the effect size and its confidence interval alongside the p-value.

The three rows below tell that story on the site’s own calculators. Each row is a one-sample t-test of a sample mean against ΞΌ0 = 100, with s = 15 β€” the first two rows with n = 25, the third with n = 200. The middle row changes the sample mean; the bottom row changes only the sample size.

SituationSample meanntp-valueCohen’s dVerdict
(a) 105, n = 25105251.6670.10860.333Fail to reject; small effect
(b) 108.8, n = 25108.8252.9330.00730.587Reject; medium effect
(c) 105, n = 2001052004.714< 0.00010.333Reject; same small effect

Rows (a) and (c) are the telling pair. The sample mean, the spread and the effect size are identical β€” d = 0.333 in both. What changed is only the sample size, and with it the p-value collapsed from 0.1086 (‘not significant’) to below 0.0001. The p-value moved because it mixes the size of the effect with the size of the sample; d did not move, because it measures only the size of the effect. That is what ‘independent of sample size’ means in practice. Try rows (a) and (c) yourself on the one-sample t-test calculator and watch the p-value fall while d stays put. The same pair appears in the worked examples in hypothesis testing explained, where the same one-sample test fails to reject at a mean of 105 and rejects at 108.8.

For the full background on what a p-value is β€” and what it is not β€” see What Is a P-Value?, and for any p-value from a z-score, use the P-Value Calculator from Z-Score.

Cohen’s d: The Formula

For two independent groups, Cohen’s d is the difference between the two means divided by the pooled standard deviation β€” the spread of the two groups combined:

d = (M1 βˆ’ M2) / SDpooled

The pooled standard deviation, SDpooled, is the weighted average of the two groups’ standard deviations, weighted by their degrees of freedom:

SDpooled = √[((n1 βˆ’ 1)s12 + (n2 βˆ’ 1)s22) / (n1 + n2 βˆ’ 2)]

Pooling matters because it uses all the data to estimate one common spread, instead of betting everything on a single group’s standard deviation. The formula is the same one the NIST/SEMATECH e-Handbook of Statistical Methods gives for the two-sample t-test for equal means β€” the t-test and d share the same pooled spread, which is why a t-test result and an effect size are so easy to convert between (see Worked Example 4 below).

Notice what the formula does not contain: the sample size. Doubling n leaves d exactly where it was β€” the confidence interval gets narrower, but the point estimate does not move. That is the whole point of a standardised measure: a d of 0.5 means a difference of half a standard deviation whether the study had 20 people or 2,000, and whether the scores were measured in points, pounds or milliseconds. The number is, in the words of a New Zealand schools’ booklet on effect sizes, ‘independent of the original units of measurement’.

Worked Example 1: A Small Effect

A tutoring programme’s students score 105 on average, against 100 for students without it. Both groups have a standard deviation of 15, and each group has 25 students. Run the numbers on the Cohen’s d Calculator:

QuantityValue
Cohen’s d0.333
Hedges’ g0.328
95% confidence interval[βˆ’0.225, 0.892]
Pooled standard deviation15.000
Cohen’s U363.1%
Distribution overlap86.8%
SizeSmall

The difference is 5 points β€” a third of a standard deviation, so a small effect. But notice the confidence interval: [βˆ’0.225, 0.892]. It includes 0, so these data are still consistent with no effect at all, and also with an effect as large as 0.89. That is what a small sample does to a small effect: it leaves a wide interval. U3 of 63.1% means about 63% of the tutored students score above the average of the untutored group, and the two distributions overlap by 86.8% β€” the groups look more alike than different.

Worked Example 2: A Huge Effect

Now make the difference bigger and the spread smaller: a treatment group scores 120 against a control group’s 100, with a standard deviation of 10 and 30 students in each group. Try the same numbers on the Cohen’s d Calculator:

QuantityValue
Cohen’s d2.000
Hedges’ g1.974
95% confidence interval[1.380, 2.620]
Pooled standard deviation10.000
Cohen’s U397.7%
Distribution overlap31.7%
SizeHuge

A d of 2.000 means the groups are two full standard deviations apart. It is the largest band on the calculator’s scale β€” huge β€” and the interval [1.380, 2.620] sits far from zero, so there is no doubt about the direction or the size. U3 of 97.7% means 97.7% of the treatment group scores above the control group’s average, and the distributions overlap by only 31.7%. When you see numbers like these, the practical question flips: the effect is so large that you should check the data for mistakes before you celebrate it.

Worked Example 3: One Sample Against a Known Value

Sometimes there is no second group β€” just one sample measured against a known value. A sample of 36 bottles has a mean fill of 108 ml and a standard deviation of 12 ml, against a target of 100 ml. Here d is the distance from the target, in standard deviations: d = (108 βˆ’ 100) / 12 = 0.667. Check the one-sample numbers on the Cohen’s d Calculator:

QuantityValue
Cohen’s d0.667
Hedges’ g0.652
95% confidence interval[0.306, 1.028]
Pooled standard deviation12.000
Cohen’s U374.8%
Distribution overlap73.9%
SizeMedium

A medium effect, and this time the interval [0.306, 1.028] stays above zero β€” the effect is not only medium, it is clearly there. In U3 terms, about 74.8% of the sample distribution sits above the target of 100.

Worked Example 4: From a t-Statistic, With No Raw Means

Published papers often report only a t-statistic. That is enough. With t = 2.5 and 25 people in each group, d = t Γ— √(1/n1 + 1/n2) = 2.5 Γ— 0.283 = 0.707. Recover d from t on the Cohen’s d Calculator:

QuantityValue
Cohen’s d0.707
Hedges’ g0.696
95% confidence interval[0.136, 1.279]
Degrees of freedom48
Cohen’s U376.0%
Distribution overlap72.4%
SizeMedium

A medium effect recovered from a single t-value and the two group sizes β€” no means, no standard deviations. The interval [0.136, 1.279] excludes zero, so the effect clears the ‘is there something?’ bar as well as the ‘how much?’ one.

Worked Example 5: Paired Data

When the same people are measured twice β€” before and after, say β€” the effect is the average change divided by the standard deviation of the changes, written dz to mark the paired design. Twenty people improve by 5 points on average, with a standard deviation of the changes of 10: dz = 5 / 10 = 0.500. Check the paired example on the Cohen’s d Calculator:

QuantityValue
Cohen’s dz0.500
Hedges’ g0.479
95% confidence interval[0.035, 0.965]
Standard deviation of differences10.000
Cohen’s U369.1%
Distribution overlap80.3%
SizeMedium

A medium effect, but a fragile one: the interval [0.035, 0.965] only just clears zero. With 20 pairs, a genuine medium effect can still be hard to pin down β€” which is why the interval matters as much as the point estimate.

Worked Example 6: No Effect at All

The boundary case: two groups with equal means β€” 100 against 100, with a standard deviation of 10 and 30 in each group. Check the boundary case on the Cohen’s d Calculator:

QuantityValue
Cohen’s d0.000
Hedges’ g0.000
95% confidence interval[βˆ’0.506, 0.506]
Pooled standard deviation10.000
Cohen’s U350.0%
Distribution overlap100.0%
SizeNegligible

Zero difference, zero effect: U3 of 50.0% means exactly half of one group sits above the other’s average, and the distributions overlap completely (100.0%). The interval [βˆ’0.506, 0.506] says the data are consistent with anything from a small negative to a small positive effect β€” ‘no effect detected’ is not the same as ‘no effect exists’.

How to Interpret d: The Benchmarks

Cohen’s rule of thumb, from his 1969 book Statistical Power Analysis for the Behavioral Sciences, is that d = 0.2 is small, d = 0.5 is medium and d = 0.8 is large. They are rough defaults, not universal standards β€” a d of 0.2 in a tightly controlled laboratory study can be disappointing, while the same 0.2 in a noisy field study can be a real achievement. The interpretation of effect sizes is a judgment process, and the field decides what counts as big.

|d|SizeCohen’s U3Distribution overlap
0.0Negligible50.0%100.0%
0.2Small57.9%92.0%
0.5Medium69.1%80.3%
0.8Large78.8%68.9%
1.0Very large84.1%61.7%
2.0Huge97.7%31.7%

Two ways to feel the numbers. Cohen’s U3 is the percentage of one group that scores above the other group’s average: at d = 0.5, about 69% of the treatment group beats the control average. Distribution overlap is how much of the two curves cover each other: at d = 0.5 the curves still overlap by about 80% β€” a medium effect still leaves the groups mostly intertwined β€” while at d = 2.0 only about a third of the curves overlap. For a clinically meaningful difference, an effect size as a standardised mean difference should usually at least approach 0.5.

Hedges’ g and Glass’s Delta

Cohen’s d has a small upward bias in small samples, so for small n report Hedges’ g instead β€” or alongside d. The correction multiplies d by J = 1 βˆ’ 3 / (4N βˆ’ 9), where N is the total sample size. With 25 in each group (N = 50) the factor is 0.984, so d = 0.333 becomes g = 0.328. The correction is small here, but it grows as samples shrink, which is why many journals now ask for g specifically. Glass’s delta (Ξ”) is the third member of the family: it divides the mean difference by the control group’s standard deviation only, instead of the pooled SD. Use delta when the treatment itself changed the spread β€” for example, when the treatment group is much more variable than the control β€” so that the treatment’s extra variability does not shrink the effect.

The Three Families of Effect Size

Cohen’s d is not the only effect size β€” the literature knows fifty to a hundred different measures. A useful way to sort them, used in a federal review of the evidence, is three families: (a) the standardised mean difference family, including the d index; (b) the standardised regression coefficient family, including the correlation coefficient; and (c) the odds ratio family, including the risk difference and the risk ratio.

The d family covers differences between means, and it is what every example above uses. The r family covers relationships: a correlation is an effect size too β€” r = 0.3 and r = 0.7 are a moderate and a strong relationship, not just a ‘significant’ one β€” and the correlation coefficient calculator reports r with its own interval. The odds-ratio family covers binary outcomes, such as whether a treatment raises the odds of recovery, and is the standard language of medicine and epidemiology. The three families answer the same question in different units: how much, not merely whether.

Always Report the Effect Size β€” With Its Confidence Interval

A p-value on its own is an incomplete report: it says whether an effect is there, while saying nothing about its size or its precision. The complete report is three numbers together β€” the effect size, its confidence interval and the p-value. The interval does the extra work: a narrow interval around a medium d is a confident claim, while a wide interval around the same d says the data cannot tell a small effect from a huge one. Read the interval the same way in every example above: does it cross zero, and does the whole interval sit on the side of a size that matters? A treatment effect can be statistically significant but so small as to be clinically meaningless β€” and only the effect size and its interval can tell you that.

Common Mistakes

  • Reading 0.2, 0.5 and 0.8 as laws of nature. They are Cohen’s rough defaults from the behavioural sciences. A small effect in physics can be a triumph; a large one in education can be suspicious. Judge the number against the field, and against what the effect would mean in the real world.
  • Reporting the p-value without the effect size. Significance without size is a headline without a story: with enough data, any difference becomes ‘significant’. The effect size is what turns the verdict into information.
  • Confusing statistical significance with practical importance. ‘Significant’ means ‘unlikely to be chance’; it says nothing about whether the difference matters. A d of 0.1 that reaches p < 0.001 is still a d of 0.1.
  • Using d on badly skewed data without saying so. Cohen’s d divides by a standard deviation, and a standard deviation is a poor summary when the distribution is badly skewed or has heavy outliers. Say so, or use a robust alternative β€” do not present the number as if the data were normal.
  • Forgetting the interval. A point estimate of 0.5 with an interval of [βˆ’0.1, 1.1] is a guess, not a finding. In every example above the interval, not the point estimate, is what settles whether the effect is even there.

What is a good effect size?

There is no universal ‘good’ value: a good effect size is one big enough to matter in its field. As a starting point, Cohen’s bands are 0.2 small, 0.5 medium and 0.8 large β€” but a d of 0.2 can be a breakthrough in a field where effects are hard to move, and a d of 0.8 can be unimpressive where they are easy. For a standardised mean difference to be clinically meaningful, it should usually at least approach 0.5. The honest answer to ‘is this good?’ is always another question: good for what?

What is the difference between effect size and p-value?

The effect size says how big an effect is; the p-value says how surprising the data would be if there were no effect. The effect size depends only on the size of the difference and the spread of the data; the p-value also depends on the sample size. So the same d = 0.333 fails to reach significance with 25 people per group and reaches p < 0.0001 with 200 β€” the effect did not change, only the amount of evidence for it. Report both: the p-value for ‘is there something?’, the effect size (with its interval) for ‘how much?’.

How do you interpret Cohen’s d?

Read it as a distance in standard deviations: d = 0.5 means the two means are half a standard deviation apart. Then translate it into plain words with U3 and overlap β€” at d = 0.5, about 69% of the treatment group scores above the control average, and the two distributions overlap by about 80%. Then judge it against the field: 0.2 is small, 0.5 medium and 0.8 large by Cohen’s rough defaults, but what counts as big depends on the context and on what the difference would mean in practice. And never read d without its confidence interval: if the interval includes 0, the data are still consistent with no effect at all.

Frequently Asked Questions

Tap a question to see the answer.

Effect size is a plain number for ‘how big?’ β€” how large a difference between two groups is, or how strong a relationship between two variables is. Unlike a p-value, it is not a yes-or-no verdict, and it does not move when you collect more data. The most common one, Cohen’s d, is the difference between two means divided by the pooled standard deviation: d = 0.5 means the groups are half a standard deviation apart.

For two independent groups, take the difference between the two means and divide it by the pooled standard deviation: d = (M1 βˆ’ M2) / SD_pooled. For a one-sample design, divide the sample mean minus the target value by the sample SD. If you have only a t-statistic, use d = t Γ— √(1/n1 + 1/n2). The Cohen’s d Calculator does all four versions and reports Hedges’ g and the 95% confidence interval alongside.

An effect size of 0 means there is no difference at all between the groups (or no relationship between the variables): the two means are equal, U3 is 50.0% β€” exactly half of one group sits above the other’s average β€” and the distributions overlap completely, by 100.0%. Note that a confidence interval containing 0 means ‘no effect detected’, which is not the same as ‘no effect exists’.

Yes. The sign shows the direction of the difference: a negative d means the first mean is lower than the second (or the sample is below the target). The size is judged by the absolute value β€” a d of βˆ’0.5 is a medium effect in the opposite direction β€” so say which group was first when you report it.

Hedges’ g is Cohen’s d with a small-sample correction: g = d Γ— J, where J = 1 βˆ’ 3 / (4N βˆ’ 9) and N is the total sample size. The correction removes the slight upward bias that d has in small samples. With 25 in each group, J is 0.984, so d = 0.333 becomes g = 0.328. Report g when samples are small; many journals now ask for it.

By Cohen’s rough defaults, 0.5 is a medium effect β€” and in many fields a medium effect is a worthwhile one: at d = 0.5, about 69% of the treatment group scores above the control average. But ‘good’ depends on the context: for a standardised mean difference to be clinically meaningful it should usually at least approach 0.5, while in tightly controlled laboratory work a 0.5 can be disappointing. Judge it against the field, not the number alone.

Because the p-value only answers ‘is there something?’ and it moves with sample size, while the effect size answers ‘how much?’ and does not. The same d = 0.333 fails to reach significance at n = 25 (p = 0.1086) and reaches p < 0.0001 at n = 200. Report the effect size and its confidence interval alongside the p-value, or the reader cannot tell a large effect from a trivial one backed by a huge sample.

Glass’s delta is the effect size to use when the treatment itself has changed the spread of the data. Instead of the pooled standard deviation, it divides the mean difference by the control group’s standard deviation only. That way, extra variability created by the treatment cannot shrink the effect. Use it when the treatment and control groups have clearly different spreads.

Conclusion

Effect size is what turns a statistical verdict into useful information. A p-value says whether an effect is there; an effect size says how much of it there is, in units that survive any change of scale and any change of sample size. Learn the three habits: report d or g with its confidence interval, never the p-value alone; translate the number into U3 and overlap before deciding what it means; and judge ‘big’ against the field, not against 0.2, 0.5 and 0.8 as if they were laws. Work the examples above on the Cohen’s d Calculator until the numbers feel like judgements, not just digits.

Check Tools

Cohen’s d Calculator

The paired calculator for this guide: Cohen’s d and Hedges’ g with the 95% confidence interval, U3 and overlap for two groups, paired data, one sample or a t-statistic. Try the Cohen’s d Calculator

One-Sample T-Test Calculator

The calculator behind the p-value contrast: the t-statistic, the p-value and Cohen’s d for a sample mean against a claimed value. Run the one-sample t-test

Correlation Coefficient Calculator

Effect sizes are not only differences between means: a correlation r is an effect size too. Compute r with its own interval and see the r family in action. Calculate a correlation coefficient

Explore All Statistical Tools

Browse the full set of calculators and visualisers on the site. Browse all the statistical tools