The standardized mean difference (SMD) is a family of effect-size measures that express the difference between two means relative to a chosen standard deviation. Unlike a raw mean difference, it has no physical measurement units. It is widely used in statistics to describe group differences, combine research findings measured on different scales, and assess covariate balance between comparison groups. Cohen’s , Hedges’ , and Glass’s are common forms, distinguished principally by their standardizing denominator and treatment of small-sample bias.
Definition and interpretation
For two populations with means and , a population standardized mean difference can be written as
where is a specified reference standard deviation. The population means are expected values; the reference standard deviation describes variability in one or both populations.
The choice of is part of the definition, not merely a computational detail. If both populations have a common standard deviation , the natural definition is
When their variances differ, several reasonable standardizers exist, and they need not produce the same effect size.
An SMD of means that the first mean exceeds the second by half of the selected reference standard deviation. It does not mean that the first mean is 50% larger, or that every individual in the first group scores higher. Zero indicates equal means, not identical probability distributions.
The sign depends on the order of subtraction and the direction of the measurement scale. For example, a positive difference can indicate either better or worse performance depending on whether high scores represent success or difficulty.
Main estimators
Cohen’s
For two independent samples, let be the sample means, their sample standard deviations, and their sample sizes. A common definition of Cohen’s is
where
Here, is a pooled estimate of a common population variance, with
This denominator weights each sample variance by its degrees of freedom. It is not the standard deviation of all observations combined: the latter also reflects separation between the group means.
The label “Cohen’s ” is sometimes used for other standardization conventions. Consequently, the formula or denominator should be specified when comparing reported values.
Hedges’
Under independent sampling from normal populations with equal variances, has a small-sample bias away from zero. Hedges’ applies a correction:
where
Here, is the gamma function. A widely used approximation is
For the pooled-standard-deviation estimator under the normal, equal-variance model, the exact correction makes unbiased for , where the relevant expectations exist. The correction is most consequential in small samples and approaches one as sample sizes increase.
Hedges’ is especially common in meta-analysis, where systematic small-sample bias can accumulate across studies.
Glass’s
Glass’s standardizes by the standard deviation of a designated reference or control group:
This definition is useful when the intervention may change variability as well as the mean, making a pooled denominator difficult to interpret. Its population target is
rather than a difference standardized by a common or average standard deviation.
Glass’s can also have finite-sample bias; using a control-group denominator does not itself provide a bias correction.
Example
Suppose two independent groups each contain 50 observations. Their sample means are 75 and 70, and both sample standard deviations are 10. Then
With ,
The estimated difference is therefore approximately half a standard deviation. The correction is small because the combined sample is reasonably large.
If both groups’ scores are transformed using , with , both the mean difference and standard deviations are multiplied by , so the SMD is unchanged. A negative reverses its sign. Nonlinear transformations generally change its magnitude.
Sampling uncertainty and hypothesis tests
An SMD is an estimate with sampling uncertainty. Its precision depends on sample size, group allocation, the underlying distributions, and the denominator used.
For independent groups under the equal-variance model, one common large-sample approximation is
An approximate variance for the corresponding is obtained by multiplying this expression by . More refined estimators use different finite-sample adjustments.
The standard error is the square root of the estimated sampling variance. A normal-approximation confidence interval takes the form
where is the appropriate standard-normal quantile.
For the usual pooled two-sample -test,
Under independent normal observations with equal variances, this statistic has a noncentral Student’s -distribution when the population difference is nonzero. Its noncentrality parameter is
Inverting that distribution provides confidence intervals for that need not be symmetric around the estimated effect.
The SMD and statistical significance answer different questions. A small effect can be estimated precisely in a large sample, while a larger estimated effect can remain uncertain in a small sample. A p-value is therefore not a substitute for reporting the effect size and its uncertainty.
Paired and repeated-measures data
For paired observations, such as measurements before and after an intervention, the denominator requires particular care.
Let . One standardized change measure is
where is the standard deviation of individual differences. Because
the correlation between measurements affects the magnitude of .
Other measures standardize the mean change by a baseline standard deviation or an average of the pre- and post-measurement variances. These estimate different quantities. A paired-sample SMD standardized by change-score variability is not directly interchangeable with an independent-groups SMD standardized by between-person variability.
The standardizer and the study design must therefore be identified when comparing or combining standardized effects. Within-group change also does not, by itself, establish an intervention effect.
Use in meta-analysis
SMDs allow studies to be combined when they measure a sufficiently similar construct using different instruments—for example, different tests intended to assess the same skill. Before pooling, score directions must be aligned so that positive and negative values have consistent meanings.
A common inverse-variance pooled estimate is
In a common-effect model, , where is the estimated sampling variance. In a random-effects model, weights commonly take the form
where represents between-study effect heterogeneity.
Standardization removes measurement units, but it does not guarantee substantive comparability. Different instruments may capture different constructs or differ in measurement reliability. Populations may also differ in variability: the same raw mean difference yields a larger SMD in a more homogeneous population.
Effects from studies sharing participants, comparison groups, or outcomes are statistically dependent. Their synthesis requires accounting for that dependence rather than treating all estimates as independent.
Use in assessing covariate balance
In causal inference, standardized differences are used to compare covariates between groups before and after matching or weighting. For a continuous covariate, a common definition is
Unlike the degrees-of-freedom-weighted denominator used in the usual Cohen’s , this denominator gives equal weight to the two group variances. Other conventions use a baseline or reference-group standard deviation.
For a binary covariate, one convention replaces the means by proportions and the variances by :
Absolute standardized differences are commonly used as descriptive balance diagnostics rather than significance tests. A threshold such as is a convention, not proof that confounding has been eliminated. Equal means can coexist with different variances, distributional shapes, or relationships among covariates, and balance on measured variables does not establish balance on unmeasured variables.
Interpretation and limitations
Conventional benchmarks associated with Cohen describe absolute values around , , and as small, medium, and large. These are rough conventions rather than universal standards. Practical importance depends on the outcome, population, costs, and substantive context.
Several limitations are especially important:
- Dependence on variability. The denominator reflects population heterogeneity and measurement properties, not just the size of the mean difference.
- Unequal variances. When population variances differ, a common-standard-deviation interpretation may be inappropriate. The pooled denominator can also depend on the relative group sizes.
- Sensitivity to extreme observations. Both means and standard deviations can be strongly affected by outliers. Robust alternatives generally define different effect measures.
- Limited distributional information. An SMD describes standardized mean separation, not all differences between distributions or the probability that one observation exceeds another without additional assumptions.
- No automatic causal meaning. Standardizing a difference does not remove selection effects, confounding, or other sources of bias.
- Undefined or unstable denominators. A zero reference standard deviation makes the measure undefined; a poorly estimated denominator can make the estimate unstable.
Reporting the raw means, standard deviations, sample sizes, direction of subtraction, standardization convention, and uncertainty interval makes an SMD interpretable and reproducible.