Effect size in psychology is a standardized number that expresses how large a relationship or group difference is, independent of sample size. Where a p-value only says whether an effect could plausibly be zero, an effect size says how big, how consistent, and how practically meaningful that effect actually is. It is the difference between knowing a program works and knowing whether it works well enough to justify its cost.

What an effect size actually measures

Every effect size answers one question: how many standard deviations separate two conditions? Standardizing by variability is what makes results comparable across studies that use different scales. A 5-point gain means little on its own. A 5-point gain relative to a standard deviation of 10 is half a standard deviation, and that number travels.

The two dominant families work differently. The d family standardizes a mean difference by pooled variability, so d = 0.50 means the groups sit half a standard deviation apart. The r family expresses the strength of association on a −1 to +1 scale. They are convertible: d of 0.2 equals r of about 0.10, d of 0.5 equals r of about 0.24, and d of 0.8 equals r of about 0.37.

The mechanism behind the conversion is shared variance. An effect size is ultimately a statement about overlap between distributions, not distance between averages. Two groups separated by d = 0.5 still overlap substantially, which is why a "medium" effect rarely means every treated person beats every untreated one.

The benchmarks and their numbers

Jacob Cohen's 1988 conventions remain the reference point. He proposed d = 0.20 as small, 0.50 as medium, and 0.80 as large. For correlations, the equivalents are roughly r = 0.10, 0.30, and 0.50. These are conventions, not laws, and Cohen said so explicitly.

Context recalibrates them sharply. An r of 0.30 sounds modest until you recognize it matches the correlation between ibuprofen and reduced headache pain in trial data. Meanwhile, the average psychological intervention in meta-analytic reviews lands around d = 0.30 to 0.50, and many replicated social-psychology effects fall below d = 0.20.

The counterintuitive detail: in large samples, trivial effects become statistically significant. With 10,000 participants per group, a d of 0.05 produces a p-value below 0.001. Significance tracks sample size; magnitude tracks reality. That asymmetry is why effect size reporting became mandatory rather than optional.

Reading the metrics side by side

Different effect sizes suit different designs, and confusing them inflates or deflates findings. The table below maps the common ones to their use case and interpretation.

Metric Best for Scale Rule of thumb Watch out for
Cohen's d Two-group mean difference Unbounded, usually 0–2 0.2 small, 0.5 medium, 0.8 large Inflated in small samples
Hedges' g Same, small samples Unbounded Slightly smaller than d Correction shrinks d by ~1–4% at n < 20
Pearson's r Continuous association −1 to +1 0.1, 0.3, 0.5 Curvilinear data yield r near zero
Odds ratio Binary outcomes 0 to ∞ 1.5 modest, 2.5 moderate Not comparable to d directly
Eta squared ANOVA variance explained 0 to 1 0.01, 0.06, 0.14 Sums to 1 across all factors
Cliff's delta Ordinal or non-normal data −1 to +1 0.15, 0.33, 0.47 Less familiar to reviewers

Hedges' g exists because d is biased upward in small samples. The correction is small at typical study sizes but matters when pooling many tiny trials. Choosing d where g belongs is a quiet source of inflated meta-analytic estimates.

Why small effects can still matter

A d of 0.10 looks negligible against Cohen's labels. Whether it matters depends entirely on what it predicts. Small effects compound across many people and many decisions, and they can shift population-level outcomes substantially when the exposure is universal.

Consider a screening or policy decision applied to millions. Moving a cutoff by an effect of d = 0.10 changes how many people cross a threshold, and that count can be enormous even if any individual's odds barely move. The same logic explains why public-health interventions with tiny per-person effects can outperform dramatic clinical ones.

The professional caveat: effect size does not equal practical significance. A large effect on a trivial outcome is worthless. A small effect on mortality, employment, or school completion can be transformative. Judge magnitude against the outcome's stakes, not against Cohen's table alone.

Edge cases that distort the number

Several conditions reliably warp effect-size estimates, and pooling them without correction produces confident but wrong conclusions.

Small samples inflate d, sometimes dramatically; studies with n < 20 per group can report effects twice their true size. Publication bias removes null results from the literature, pushing pooled estimates upward by 10–30% in some fields. Range restriction shrinks correlations when the sample is more homogeneous than the population. Measurement error attenuates observed effects; unreliable scales understate true associations. Selective reporting of outcomes lets authors publish the largest of several measured effects. Extreme groups designs inflate d by selecting participants at distribution tails.

Each distortion has a correction. Hedges' g addresses small-sample bias. Funnel plots and trim-and-fill methods probe publication bias. Reliability correction adjusts for attenuation. None of these fixes are automatic, and many published effect sizes carry no correction at all.

Time course: how effect sizes behave across a study's life

Effect sizes are not fixed properties. They shift as designs, samples, and analyses change, and the pattern is documented across replication efforts.

Stage What happens to the effect size Typical pattern
Original publication Often the largest estimate Selected for significance
First direct replication Shrinks noticeably 30–50% smaller common
Meta-analytic pooling Converges toward a stable value Weighted by precision
Large preregistered trial Usually the most modest Closest to truth
Follow-up assessment Decays over time Booster effects fade
Cross-cultural extension Varies by context Moderators emerge

The shrinkage from original to replication is the single most important pattern for readers to internalize. It is not fraud; it is selection plus regression to the mean plus the many small analytic choices that maximize a first result. The remedy is preregistration and large samples, which produce estimates that hold up.

Misconceptions worth correcting

The most common error treats effect size as a property of an intervention. It is a property of a comparison, in a population, with a measure, at a time. Change any of those and the number changes.

A second error assumes the benchmarks are universal. Cohen's labels came from mid-twentieth-century social science and were explicitly provisional. Applying d = 0.80 as "large" in education, where typical effects run near 0.40, misreads the landscape. In pharmacology, an effect that size would be extraordinary.

A third error confuses effect size with variance explained. A d of 0.50 corresponds to r of about 0.24, which is roughly 6% of variance. That sounds trivial but describes a real, detectable, and often useful group difference. Percentage of variance explained understates practical impact because it squares an already-standardized quantity.

A fourth error converts effect sizes into individual predictions. Knowing that a program produces d = 0.40 does not tell you whether a specific person will benefit. Group averages and individual outcomes are different questions, and the effect size answers only the first.

Practical actions for readers and researchers

Applying effect sizes well takes a few disciplined habits, whether you are consuming research or producing it.

For readers: locate the effect size before the p-value, then check the sample size and design. Compare the reported number against the field's typical range rather than Cohen's generic table. Ask whether the outcome is the one you care about. Treat single-study effects as provisional until replication.

For researchers: report at least one standardized effect size with a confidence interval, not a point estimate alone. Prefer Hedges' g for small samples and report the correction. Preregister the primary outcome to prevent selective reporting. Power analyses should target the smallest effect of interest, not Cohen's defaults. When pooling, model heterogeneity rather than assuming one true effect.

For practitioners: translate d into a concrete outcome before deciding. A d of 0.30 on a test score with a 15-point standard deviation is about 4.5 points. That number, not the standardized statistic, is what stakeholders can weigh. State the conversion explicitly so decision-makers see the units they understand.

Limits of what an effect size can tell you

An effect size cannot establish causation, cannot prove a mechanism, and cannot substitute for a well-designed study. It summarizes a comparison; it does not explain why the comparison turned out as it did.

Estimates are always conditional on the measure's reliability, the sample's composition, and the analysis choices made. Two meta-analyses of the same literature can reach different pooled values through different inclusion rules and weighting schemes. The number is an estimate with uncertainty, and confidence intervals around effect sizes are frequently wide enough to span two benchmark categories.

Finally, the field's benchmark conventions are inherited rather than derived. There is no natural law making 0.50 "medium." Cohen chose it to be useful, and subsequent evidence has both vindicated and challenged that choice. Use the benchmarks as orientation, not verdict. The number that matters is the one tied to a real outcome, in a real population, with its uncertainty stated plainly.