What Is an Effect Size

By Kevin Fernando · Sep 22, 2026 · 9 min read

Effect size is a standardized metric that quantifies how large a difference or relationship actually is, separate from whether that difference is statistically significant. While a p-value tells you whether an effect exists, effect size tells you how big it is. A result can be statistically significant but have negligible practical importance, or vice versa. Common effect sizes include Cohen's d for group comparisons and correlation coefficients for relationships.

When you read a research report, you encounter two numbers that sound similar but measure entirely different things: the p-value and the effect size. The p-value tells you whether a difference or relationship is unlikely to have occurred by chance. The effect size tells you how large that difference or relationship actually is. A study can report a p-value of 0.001, indicating an extremely unlikely result, yet describe a difference so small that it has no practical importance. Conversely, a study with a larger p-value might reveal a substantial real-world effect that matters despite greater uncertainty. Understanding effect size is essential for interpreting research honestly, because statistical significance alone can mislead you into treating trivial findings as important discoveries.

What effect size actually measures

Effect size quantifies the magnitude of a difference or relationship in data. It answers the question: how big is the effect? Not whether it exists, but how large it is. If a study compares two groups, effect size measures the distance between them in standardized units. If a study examines a relationship between two variables, effect size measures the strength of that association. Effect size is always independent of sample size. A study with 30 participants and a study with 3,000 participants can report the same effect size if the actual magnitude of the difference is identical. This independence from sample size is what makes effect size so valuable: it tells you about the phenomenon itself, not about how many people were studied.

How effect size differs from statistical significance

A p-value is the probability of observing data as extreme or more extreme than what was collected, assuming the null hypothesis is true. It answers a narrow question: is this result unlikely by chance? Statistical significance means the answer is yes. But statistical significance does not mean the effect is large or practically important. P-values indicate whether an effect exists, but they do not tell how large or important that effect is. This is the critical distinction that confuses many readers of research. A very large study can detect a tiny difference and produce a p-value below 0.05, making the result statistically significant even though the actual difference is negligible. Conversely, a small study might fail to reach statistical significance even when the true effect is substantial, simply because the sample size was too small to detect it reliably. Effect size separates the question of existence from the question of magnitude. You need both numbers to interpret research correctly.

Standardized versus unstandardized effect sizes

An unstandardized effect size preserves the original units of measurement. If a study measures blood pressure in millimeters of mercury, the unstandardized effect size is reported in millimeters of mercury. If a study measures income in dollars, the unstandardized effect size is in dollars. Unstandardized effect sizes communicate practical meaning directly: they tell you the real-world magnitude in the units that matter to the question. A standardized effect size removes the units and expresses the magnitude relative to the variability in the data. Cohen's d, for example, divides the difference between groups by the standard deviation, producing a unitless number. Correlation coefficients like Pearson's r are also standardized, ranging from -1 to +1 regardless of the original measurement scale. Standardized effect sizes enable comparison across studies that use different measures or units, making them invaluable for research synthesis.

When to use standardized effect sizes

Standardized metrics allow researchers to pool results across multiple studies and conduct meta-analyses, where findings from different investigations are combined to estimate an overall effect. When studies measure the same construct using different instruments or different measurement units, standardized effect sizes make direct comparison possible. A meta-analysis of cognitive interventions might combine studies that measured memory using different tests; standardized effect sizes allow those disparate results to be synthesized into a single estimate. Standardized effect sizes also facilitate comparison within a field: if you are reviewing literature on a topic, standardized metrics let you see at a glance which studies found large effects and which found small ones, even when the original measurements differ.

When to use unstandardized effect sizes

Unstandardized effect sizes communicate the practical significance of a finding in terms that stakeholders understand immediately. If a medication reduces blood pressure by 8 millimeters of mercury, that number means something to a clinician and a patient. If a training program increases productivity by 12 percentage points, that is actionable information. Standardized effect sizes like Cohen's d convey the magnitude relative to variability, which is useful for research synthesis but can obscure the real-world meaning. When you need to understand whether an effect matters in practice, unstandardized effect sizes in the original units are often more informative than their standardized counterparts.

Common types of effect size and how to interpret them

Researchers use different effect size metrics depending on the type of data and comparison. For group comparisons, Cohen's d is standard. For relationships between continuous variables, correlation coefficients are used. For binary outcomes like success or failure, odds ratios are common. Each has interpretation thresholds that help you judge whether an effect is small, medium, or large. These thresholds are not universal rules but guidelines based on typical patterns in research. Researchers typically use Cohen's guidelines of Pearson's r =.10,.30, and.50, and Cohen's d = 0.20, 0.50, and 0.80 to interpret observed effect sizes as small, medium, or large, respectively [F11, F12]. However, these guidelines were not based on quantitative estimates and are only recommended if field-specific estimates are unknown. Different fields have developed their own thresholds. For example, in gerontology research, recommended Cohen's d thresholds are 0.15, 0.40, and 0.75 for small, medium, and large effects respectively, reflecting the reality that smaller effects are often meaningful in that context.

Cohen's d for comparing group means

Cohen's d measures the standardized difference between two group means. It divides the difference by the pooled standard deviation, producing a unitless number that allows comparison across studies. A Cohen's d of 0.20 represents a small effect, meaning the groups differ by one-fifth of a standard deviation. A d of 0.50 represents a medium effect, a difference of half a standard deviation. A d of 0.80 represents a large effect, a difference of four-fifths of a standard deviation. These thresholds help you interpret what you are seeing, but context matters. In education, a small effect on test scores might represent meaningful learning gains. In pharmaceutical research, a small effect on a surrogate biomarker might be clinically trivial. The threshold tells you the statistical magnitude; you must judge the practical importance yourself.

Correlation coefficients for relationships

Pearson's correlation coefficient r measures the strength of a linear relationship between two continuous variables. It ranges from -1 to +1, where -1 is a perfect negative relationship, +1 is a perfect positive relationship, and 0 is no relationship. Interpretation thresholds help you classify the strength. Values below +0.29 are considered a weak correlation. Values between ±0.30 and ±0.49 indicate a moderate correlation. Values between ±0.50 and ±1 suggest a strong correlation. These thresholds are guidelines, not absolutes. A correlation of 0.35 is stronger than 0.20 but weaker than 0.60. The sign of r indicates direction: a negative r means that as one variable increases, the other tends to decrease. The magnitude indicates strength: the farther r is from zero, the tighter the relationship. Remember that correlation does not imply causation; a strong correlation between two variables means they move together, not that one causes the other.

Odds ratios for binary outcomes

An odds ratio is an effect size measure that quantifies the magnitude of a phenomenon when outcomes are binary, such as disease present or absent, success or failure. An odds ratio of 1 means there is no difference in odds between groups. An odds ratio of 2 indicates a difference in odds between groups. An odds ratio of 0.5 indicates a difference in odds between groups. Odds ratios are an effect size measure that quantifies the magnitude of a phenomenon. Like other effect sizes, odds ratios are independent of sample size. A small study and a large study can report the same odds ratio if the true relationship is identical. Odds ratios are an effect size measure commonly used in epidemiology and clinical research.

A worked example: comparing two groups

Suppose a researcher compares two teaching methods on student test scores. Method A is used with 30 students who score a mean of 85 with a standard deviation of 10. Method B is used with 28 students who score a mean of 78 with a standard deviation of 12. To calculate Cohen's d, first compute the pooled standard deviation, which weights the standard deviations by sample size: pooled SD = √[(29 × 10² + 27 × 12²) / 56] ≈ 11.01. Then divide the difference in means by the pooled standard deviation: d = (85 − 78) / 11.01 ≈ 0.64. This d of approximately 0.64 falls between the medium threshold of 0.50 and the large threshold of 0.80, suggesting a medium-to-large effect. Method A students scored roughly 0.64 standard deviations higher than Method B students. Now suppose the study included 3,000 students per group instead of 30 and 28, with identical means and standard deviations. The Cohen's d would still be 0.64. The larger sample size would produce a smaller p-value, making the result more statistically significant, but the effect size itself would not change. This illustrates why effect size and statistical significance are independent: the magnitude of the difference is the same regardless of how many people were studied.

Why effect size matters for research synthesis

Effect size enables researchers to combine results across multiple studies through meta-analysis. When a field has conducted many studies on the same question, meta-analysts pool the effect sizes to estimate an overall effect and identify patterns. Without standardized effect sizes, this synthesis would be impossible: studies using different measures could not be compared. Effect size also allows readers to judge practical significance independent of sample size. A p-value of 0.001 sounds impressive, but if the effect size is trivial, the finding may not matter in practice. Conversely, a p-value of 0.10 might seem weak, but if the effect size is large, the result could represent a meaningful phenomenon that a larger sample would confirm. Effect size prevents the common mistake of treating statistical significance as equivalent to practical importance. Large studies can detect tiny effects and produce highly significant p-values for findings that have no real-world consequence. Small studies can fail to detect large effects and produce non-significant p-values for phenomena that matter substantially. Effect size gives you the information you need to distinguish signal from noise.

Limitations and common misreadings

Effect size thresholds like those of Cohen are context-dependent. A small effect in one field may represent a meaningful real-world impact in another. In clinical medicine, a small reduction in mortality is often considered important even if the effect size is modest. In social psychology, a small effect on attitudes might be trivial. The thresholds are starting points for interpretation, not universal rules. Field-specific thresholds, when available, should guide your judgment. Effect size also does not establish causation. A strong correlation between two variables means they are associated, but it does not mean one causes the other. A large difference between groups in an observational study might reflect confounding rather than a true causal effect. Effect size quantifies the magnitude of what was observed, but you must evaluate the study design and context to determine whether that magnitude reflects a real causal relationship. Additionally, effect size can be inflated by publication bias: studies with large effects are more likely to be published than studies with small effects, so the published literature may overestimate typical effect sizes in a field. When reviewing research, consider the full body of evidence, not just the largest or most impressive individual studies.

Do this automatically

Calculating effect sizes manually requires researchers to compute standardized metrics like Cohen's d or correlation coefficients for each analysis, then repeat these calculations across multiple studies when conducting meta-analyses, a process that multiplies with every new dataset or research question examined. Software can automate these computations, apply the appropriate standardized or unstandardized formulas based on study design, and aggregate results across studies without manual recalculation. Academia research runs this kind of analysis end to end and cites every figure back to its original source, across as many subjects as you need, with Mimir attaching a citation to every figure. It will not write your paper for you. It gathers and cites the evidence; the argument stays yours, and deliberately so. Every plan includes it, including the free one.

Start your first report and run it on something you already know well.

This article is general information and research commentary. It is not professional advice, and it does not account for your particular circumstances.