Contents
On this page
- Purpose and consequences of standardisation
- The three estimators and the choice of denominator
- Magnitude of the small-sample bias
- Variance of the estimate and its importance
- Influence of the denominator on heterogeneity statistics
- Change scores, opposite-direction scales and multiple outcomes
- Reporting the estimator in the methods section
- Pooling and presentation of a pooled SMD
- Key points
- Application to your study
- Frequently asked questions
Academy · Research Data Analytics & AI · SMD estimation
Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting
The standardised mean difference (SMD) is not a single estimator but a family of estimators, distinguished by the denominator and by whether a small-sample correction is applied. Cohen’s d can be upwardly biased in small samples. Hedges’ g applies a small-sample correction. The specific estimator should therefore be named in the methods section.
Standardising a mean difference by a standard deviation permits the pooling of trials that used different instruments. It also introduces a second source of variation, the denominator, which authors commonly treat as fixed. Three named estimators are in common use; they employ different standard deviations, one is biased upwards in small samples by a known and substantial amount, and most published methods sections state only “standardised mean difference”. This article examines the computational mechanics, the small-sample correction and the information that the methods section should contain.
Learning objectives
- The denominator used for Cohen’s d, Hedges’ g and Glass’s delta, and the reasons this choice is substantive
- The small-sample bias of Cohen’s d, which is upward, and the correction factor that Hedges’ g applies to remove it
- Which estimator Cochrane reviews implement, and the implications for reporting
- How the choice of denominator affects heterogeneity, such that an SMD meta-analysis may appear inconsistent for reasons unrelated to the intervention
- How to handle multiple outcomes, change scores and opposite-direction scales without devising a non-standard estimator
- What the methods section should state so that another analyst can reproduce the results
Common misconceptions
- “Cohen’s d and Hedges’ g are the same thing.” They differ by a small-sample correction factor. Hedges showed in the Journal of Educational Statistics that the factor is 0.79788 at four degrees of freedom, 0.92275 at ten and 0.98491 at fifty — negligible in large trials, not in small ones.
- “The correction is too small to bother with.” Cohen’s d can be upwardly biased in small samples. Hedges’ g applies a small-sample correction; a review of small trials that pools uncorrected estimates accumulates that bias across studies.
- “Standardising makes different instruments comparable.” It makes them poolable. Jung and Aloe, writing in Campbell Systematic Reviews under the title “not so standard after all”, describe the extent to which the resulting values depend on choices that are not visible to the reader.
- “A larger SMD means a bigger clinical effect.” This does not hold across studies with different population variability. A narrower control-group standard deviation inflates the SMD without any change in the clinical effect on the patient.
- “Reporting ‘SMD’ identifies the method.” It identifies the family of estimators. The estimator, the denominator and whether a correction was applied must all be specified.
Purpose and consequences of standardisation
When trials measure the same construct on different instruments — three depression scales, four pain scores, two quality-of-life instruments — their mean differences cannot be averaged directly. The standard solution, introduced by Glass in Educational Researcher and applied at scale by Smith and Glass in American Psychologist, divides each study’s mean difference by a standard deviation, producing a dimensionless quantity measured in standard-deviation units 31,32.
The cost is that the result now depends on two quantities rather than one. The numerator carries the treatment effect; the denominator carries the variability of the population enrolled. Two trials with identical treatment effects will report different standardised mean differences if one recruited a narrow population and the other a broad one, and the trial with the narrower population will appear more effective 2. This is not measurement error; it is inherent in the definition of the metric.
Guyatt and colleagues, in the GRADE guidance on continuous outcomes in the Journal of Clinical Epidemiology, state the problem directly: dividing each study’s mean difference by its standard deviation is the most established approach when instruments differ, but it is sensitive to differences in heterogeneity across populations and can be hard to interpret 6. They list alternatives — dichotomising at a threshold, reporting in the units of the most interpretable instrument, the ratio of means, or minimal important difference units — which are taken up in making sense of effect size. The present article concerns the correct computation of the standardised estimate itself.
The three estimators and the choice of denominator
All three divide a mean difference by a standard deviation. They differ in which standard deviation, and in whether the result is corrected for small-sample bias.
Glass’s delta uses the control group’s standard deviation alone. The logic is that the intervention may itself change the spread of outcomes, so the control group gives an uncontaminated estimate of the population’s variability 29. Glass and Smith’s class-size analysis in Educational Evaluation and Policy Analysis is the canonical application. The disadvantage is reduced precision: a single group’s standard deviation is estimated less well than a pooled one, so the estimator is noisier.
Cohen’s d uses the standard deviation pooled across both arms, which is more precise when the intervention does not change the variance. It is the usual default and the usual meaning of an uncorrected standardised mean difference.
Hedges’ g is Cohen’s d multiplied by a correction factor that removes the upward bias present in small samples 12,28. Hedges and Olkin develop the distribution theory in Statistical Methods for Meta-Analysis, building on Hedges’ earlier work and on the non-central t-distribution results of Johnson and Welch in Biometrika; Walker and Saw, in the Journal of the American Statistical Association, supply the underlying distributional result for linear combinations of t-variables, and Lehmann and Scheffé’s completeness results in Sankhya sit behind the unbiasedness argument 27,30,35.
Deeks, Higgins and Altman record the practical consequence in the Cochrane Handbook: the standardised mean difference implemented for Cochrane reviews is Hedges’ adjusted g, which is similar to Cohen’s d but includes a correction for small-sample bias 1. An analysis conducted in the Cochrane manner in standard software will very probably have computed this estimator; this should nevertheless be verified rather than assumed, because the output may be labelled only “SMD”.
The three named standardised mean difference estimators
| Estimator | Denominator | Small-sample bias | Use it when |
|---|---|---|---|
| Glass’s delta Glass; Glass and Smith | Control group standard deviation only | Biased upwards; expectation is the true effect divided by the correction factor | The intervention is expected to change the spread of outcomes, not only the mean |
| Cohen’s d | Standard deviation pooled across both arms | Biased upwards in small samples, by the same mechanism | A more precise denominator is required and variances are comparable across arms |
| Hedges’ g (adjusted) Hedges; Hedges and Olkin | Pooled standard deviation, multiplied by a correction factor | Corrected; factor is 0.79788 at 4 df, 0.92275 at 10 df, 0.98491 at 50 df | Default for systematic reviews; the estimator implemented for Cochrane reviews |
Denominators and bias behaviour from Glass (Educational Researcher), Glass and Smith (Educational Evaluation and Policy Analysis), Hedges (Journal of Educational Statistics) and Hedges and Olkin (Statistical Methods for Meta-Analysis). The Cochrane implementation is as stated by Deeks, Higgins and Altman in the Cochrane Handbook. The right-hand column is editorial guidance drawn from those sources.
Magnitude of the small-sample bias
The bias is of practical, not merely theoretical, importance. Hedges, in the Journal of Educational Statistics, derived the distribution of Glass’s estimator and showed that its expectation is the true effect size divided by a correction factor, so the uncorrected estimator overestimates 28. Cohen’s d can be upwardly biased in small samples. Hedges’ g applies a small-sample correction.
The correction factor approaches one as sample size increases, which is why the bias is negligible in large trials but material in small ones. Hedges gives the factor as 0.79788 at four degrees of freedom, 0.92275 at ten, and 0.98491 at fifty 28. Multiplying by the factor removes the bias; dividing the uncorrected estimate by its reciprocal is the same operation seen from the other side.
Two practical implications follow. First, a meta-analysis of small trials that pooled uncorrected estimates has an upward bias in its summary, study by study, and the direction is always the same: towards a larger apparent effect. Second, because the factor depends on each study’s own degrees of freedom, the bias is not uniform across the forest plot; it is largest in precisely the small studies to which random-effects weighting gives relatively greater influence. Where the smallest included trials have groups of fewer than ten participants, the limitations should state whether the correction was applied.
The small-sample bias in the uncorrected estimator, and the correction factor that removes it, at three sample sizes. Source: Hedges LV. Distribution theory for Glass’s estimator of effect size and related estimators. Journal of Educational Statistics.
Variance of the estimate and its importance
Pooling needs a standard error for every study, and for a standardised mean difference the standard error depends on both the sample sizes and the effect size itself 3,12. That dependency has a consequence that authors frequently do not anticipate: the weight a study receives is not a function of its size alone, so two trials of equal size can receive different weights because their observed effects differ.
Lin and Aloe evaluated the available estimators for the standardised mean difference in Statistics in Medicine and compared their behaviour, including their variance estimators, across realistic conditions. Jung and Aloe, in Campbell Systematic Reviews, extend the argument that the family is less standardised than its name suggests — different defensible choices about the denominator and the variance produce measurably different pooled results from identical data 2,3. Rosenthal’s chapter on parametric effect-size measures in The Handbook of Research Synthesis remains a useful survey of the wider family 25.
None of these papers nominates a single best estimator for all circumstances, and the practical response is that adopted in the literature itself: authors should name the estimator used and demonstrate that the conclusion does not depend on it. Viechtbauer’s metafor paper in the Journal of Statistical Software documents which estimator each function computes and should be consulted before the methods sentence is written 10.
Influence of the denominator on heterogeneity statistics
This is the consequence most frequently overlooked. Because each study’s standardised mean difference is scaled by that study’s own standard deviation, variation between studies in population variability appears as variation in the effect estimate, and therefore as heterogeneity 2. A set of trials with identical treatment effects, conducted in populations of differing breadth, will produce a dispersed forest plot and a non-trivial I-squared.
Higgins and Thompson proposed three statistics in Statistics in Medicine — H, I-squared and R — to describe heterogeneity independently of the effect metric and the number of studies, illustrated on five datasets 17. These statistics describe dispersion but do not identify its source. In an SMD analysis, an important part of the explanation may lie in the denominators rather than the interventions, which constitutes a different limitation from the one authors usually report.
Two inexpensive checks are recommended. First, the control-group standard deviation should be tabulated alongside each study’s standardised estimate; if the standard deviations vary by a factor of two or more across studies, this should be noted. Second, the analysis may be repeated as a raw mean difference on the subset of studies sharing one instrument, as a sensitivity analysis: if the heterogeneity largely disappears, the standardisation was contributing to it. The model-level consequences of heterogeneity are covered in fixed-effect or random-effects, and whether a common-effect model survives any of this is dealt with in fixed-effect assumptions.
Change scores, opposite-direction scales and multiple outcomes
Three situations arise in almost every review of continuous outcomes, and in each of them analytical errors can occur without being apparent.
Change scores versus final values. A standardised mean difference computed from change-from-baseline scores uses the standard deviation of the change, which is usually smaller than the standard deviation of the final value; the two are therefore not on the same scale and should not be pooled together without careful consideration. The usual advice is to standardise using the standard deviation of final values even where the mean difference comes from change scores, and whichever approach is adopted should be stated 1.
Scales that run in opposite directions. Where one instrument scores higher as the patient improves and another scores lower, one set of signs must be reversed. Although this is a trivial operation, it is performed incorrectly often enough to justify a statement in the methods and a visual check of the direction labels on the forest plot.
Multiple outcomes per study. López-López, Page, Lipsey and Higgins set out the options for dealing with effect-size multiplicity in systematic reviews in Research Synthesis Methods: select one outcome per study on a pre-specified rule, combine outcomes within a study, or model the dependence multivariately 4. Selecting the outcome for each study after the results are known is not acceptable, because it converts a statistical problem into a reporting-bias problem. The same paper explains why the selection rule belongs in the protocol rather than in the analysis script.
Reporting the estimator in the methods section
Most reviewer objections about standardised mean differences are requests for information that was not withheld deliberately; the author was simply unaware that it had to be stated. Four items address almost all of them: which estimator, which standard deviation, whether a small-sample correction was applied, and how multiple outcomes were handled.
The same continuous-outcome analysis, reported two ways (a constructed specimen — the figures in both versions are illustrative placeholders, not findings from any review)
Invites the objection
Because the included trials used different instruments, results were expressed as standardised mean differences. The pooled SMD was 0.41 (95% CI 0.19 to 0.63), favouring the intervention, representing a moderate effect.
Which estimator? “Standardised mean difference” names a family of at least three.
Which standard deviation, and was a correction applied? Unstated, and the uncorrected estimator is biased upwards by roughly a quarter at four degrees of freedom.
Change scores or final values? They are not on the same scale and should not be mixed silently.
Were opposite-direction scales handled? Nothing confirms the signs were reconciled.
How were multiple outcomes chosen? Unspecified, so selective outcome selection cannot be excluded.
What does “moderate” mean? A conventional label applied to a standardised unit, with no clinical anchor.
Answers it in advance
The nine trials used four different instruments for the same construct, so effects were expressed as Hedges’ adjusted g — the mean difference divided by the standard deviation pooled across arms, multiplied by the small-sample correction factor — standardised throughout on the standard deviation of final values rather than of change scores. Scores from the two instruments on which higher values indicate worse status were sign-reversed before pooling. Where a trial reported more than one eligible instrument, we used the one pre-specified in the protocol as primary and report the alternatives in the supplement. The pooled estimate was 0.41 (95% CI 0.19 to 0.63). Control-group standard deviations ranged from 4.2 to 11.8 across trials, so part of the observed heterogeneity (I-squared 48%) reflects differing population variability rather than differing treatment effect; a sensitivity analysis restricted to the four trials using one instrument, pooled as a raw mean difference, is reported alongside.
The estimator is named, including the correction, so it can be reproduced.
The standardising standard deviation is specified, and the change-score question is settled explicitly.
Sign reversal is reported, so the direction of the pooled effect is verifiable.
The outcome selection rule is pre-specified and the alternatives are available.
The range of control-group standard deviations is given, so the reader can see standardisation’s contribution to heterogeneity.
No conventional magnitude label is asserted; a same-instrument sensitivity analysis is offered instead.
The weak version is not poorly written, and its estimate may well be correct; it is, however, unverifiable, because six distinct decisions were made and none was recorded. The strong version adds approximately one hundred words, and the final addition is the most valuable: reporting the spread of the denominators indicates to the reader that part of the inconsistency arises from the metric, a limitation that no reviewer can infer from the forest plot alone.
Pooling and presentation of a pooled SMD
Once every study has a standardised estimate and its standard error, pooling proceeds as for any other continuous effect measure. DerSimonian and Laird’s worked examples in Controlled Clinical Trials give a sense of the scale of quantities involved: for the Winship review the weighted non-iterative overall effect was 0.389 with a standard error of 0.058 26. This is a summary on a standardised scale, and it also illustrates the central interpretive problem: a clinician cannot act on a value of 0.389 without back-translation into a measurable quantity.
Correct estimation is therefore necessary but not sufficient. The estimator determines whether the result is correct; the presentation determines whether it can be used. For the back-translation options (minimal important difference units, the units of the most interpretable instrument and the ratio of means), see making sense of effect size in meta-analysis, and for assistance with either step, Pubrica’s biostatistics and meta-analysis teams work to the same reporting standard.
Key points
Key points
A standardised mean difference is defined by its denominator and by whether a small-sample correction was applied; the abbreviation “SMD” alone therefore does not identify the quantity computed. Glass’s delta uses the control-group standard deviation and Cohen’s d the pooled standard deviation. Cohen’s d can be upwardly biased in small samples. Hedges’ g applies a small-sample correction, with correction factors of 0.79788 at four degrees of freedom, 0.92275 at ten and 0.98491 at fifty. Cochrane reviews implement Hedges’ adjusted g. Authors should name the estimator, state whether standardisation was based on final values or change scores, report how multiple outcomes were selected, and report the range of control-group standard deviations, because part of the observed heterogeneity may reflect the denominator rather than the intervention.
Application to your study
Application to your study
- The methods section should name the specific estimator, not only the family. Hedges’ adjusted g, Cohen’s d or Glass’s delta should be specified, with a clause stating which standard deviation forms the denominator.
- The quantity computed by the software should be confirmed. The function documentation should be consulted to determine whether the small-sample correction is applied by default; an output column labelled “SMD” does not indicate this.
- The correction should be applied if any study has small groups. Cohen’s d can be upwardly biased in small samples, and the bias is concentrated in the small studies that random-effects weighting favours.
- A single decision between change scores and final values should be made and reported. Their standard deviations differ, so combining them without disclosure places two different scales in one forest plot.
- Signs should be reversed for opposite-direction instruments before pooling, and the reversal reported. The direction labels of the forest plot should then be checked against a trial with well-understood results.
- The rule for selecting among multiple outcomes should be specified in the protocol. Options include one pre-specified instrument per study, combination within studies or a multivariate model, selected before any results are examined.
- The control-group standard deviation should be tabulated alongside each study’s estimate. If these values vary by a factor of two or more, this should be reported in the limitations as a source of heterogeneity distinct from the treatment effect.
- A same-instrument subset should be analysed as a raw mean difference. If heterogeneity largely disappears, standardisation was contributing to it, and this should be addressed in the discussion.
Frequently asked questions
The denominator is the same — the standard deviation pooled across both arms — but Hedges’ g multiplies the result by a correction factor that removes the upward bias present in small samples. Hedges derived the relevant distribution theory in the Journal of Educational Statistics and gives the factor as 0.79788 at four degrees of freedom, 0.92275 at ten and 0.98491 at fifty. In a trial with hundreds of participants per arm the two estimators are indistinguishable, whereas Cohen’s d can be upwardly biased in small samples. Deeks, Higgins and Altman record in the Cochrane Handbook that the standardised mean difference implemented for Cochrane reviews is Hedges’ adjusted g.
When there is reason to believe that the intervention changes the spread of outcomes as well as the mean. Glass’s delta uses the control group’s standard deviation alone, on the argument that the control arm gives an uncontaminated estimate of the population’s variability, and Glass and Smith’s class-size analysis in Educational Evaluation and Policy Analysis is the canonical use. The disadvantage is reduced precision: one group’s standard deviation is estimated less well than a pooled one, so the estimator is noisier, and Hedges showed that the uncorrected version is biased upwards. In most clinical reviews the pooled denominator is preferred, and the choice should be stated with its reason in either case.
It matters when the smallest included trials are genuinely small, and the direction is always the same: towards a larger apparent effect. Cohen’s d can be upwardly biased in small samples. Hedges’ g applies a small-sample correction, with a correction factor of 0.79788 at four degrees of freedom, 0.92275 at ten and 0.98491 at fifty, so the problem diminishes rapidly as sample size increases. It still deserves attention in meta-analysis because the bias is unevenly distributed across the included studies: it is largest in the small trials, and a random-effects model gives those relatively more weight than a fixed-effect model does.
Not without careful consideration, because the two use different denominators. The standard deviation of a change score is usually smaller than the standard deviation of the final value, so the same treatment effect produces a larger standardised estimate when computed from change scores. The common approach is to standardise on the standard deviation of final values even when the mean difference itself comes from change-from-baseline analyses, which keeps the scale consistent. Whichever approach is adopted should be stated explicitly in the methods and applied uniformly. Combining the two without disclosure is one of the less conspicuous ways in which an SMD meta-analysis becomes uninterpretable.
Possibly because of the denominators. Each study’s standardised estimate is scaled by its own standard deviation, so differences in how broad or narrow the enrolled population was translate directly into differences in the effect estimate. Trials with identical treatment effects conducted in populations of differing variability will produce a dispersed forest plot. Guyatt and colleagues note in the GRADE continuous-outcomes guidance that the standardised approach is sensitive to differences in heterogeneity across populations for precisely this reason. Two inexpensive diagnostic checks are available: tabulating the control-group standard deviations, and repeating the analysis as a raw mean difference on the subset sharing one instrument.
Authors may select one outcome per study according to a rule specified in the protocol, combine outcomes within each study, or model the dependence; the decision should, however, be made before the results are seen. López-López, Page, Lipsey and Higgins set out the options for effect-size multiplicity in Research Synthesis Methods, and the choice must be pre-specified because selecting the most favourable instrument after extraction is a reporting-bias problem rather than a statistical one, and reviewers increasingly check for it. Whichever approach is adopted, the estimates from the instruments not used should be reported in a supplement, so that the reader can see that the conclusion does not depend on the selection.
The sources do not nominate one for all circumstances, and it is more defensible to acknowledge this than to assert that one is superior. Lin and Aloe evaluated the available estimators in Statistics in Medicine and found differences in behaviour across conditions; Jung and Aloe, in Campbell Systematic Reviews, argue under the title “not so standard after all” that defensible variation in these choices produces measurably different pooled results from the same data. What is settled is the convention: systematic reviews default to Hedges’ adjusted g, which is what Cochrane implements. The practical requirement is to name the estimator used and to demonstrate that the conclusion does not depend on it.
The documentation for the specific function, rather than the package overview, should be consulted. Viechtbauer’s metafor paper in the Journal of Statistical Software documents which estimator each measure argument computes, and other packages differ in whether the small-sample correction is applied by default and in which denominator they use. An output column headed “SMD” is not a specification. A reliable check is to compute one study’s estimate by hand from the published means, standard deviations and sample sizes and to confirm that it matches the software’s value for that row; this takes a few minutes and resolves the question definitively.
References
- Deeks JJ, Higgins JPT, Altman DG. Analysing data and undertaking meta-analyses. Cochrane Handbook for Systematic Reviews of Interventions. 2019:241-284. doi:10.1002/9781119536604.ch10 read 9 October 2026. doi.org/10.1002/9781119536604.ch10
- Jung J, Aloe AM. Standardized mean differences: not so standard after all. Campbell Systematic Reviews. 2025;21(3):e70056. doi:10.1002/cl2.70056 read 9 October 2026. doi.org/10.1002/cl2.70056
- Lin L, Aloe AM. Evaluation of various estimators for standardized mean difference in meta-analysis. Statistics in Medicine. 2021;40(2):403-426. doi:10.1002/sim.8781 read 9 October 2026. doi.org/10.1002/sim.8781
- Lopez-Lopez JA, Page MJ, Lipsey MW, Higgins JPT. Dealing with effect size multiplicity in systematic reviews and meta-analyses. Research Synthesis Methods. 2018;9(3):336-351. doi:10.1002/jrsm.1310 read 9 October 2026. doi.org/10.1002/jrsm.1310
- Rover C, Knapp G, Friede T. Hartung-Knapp-Sidik-Jonkman approach and its modification for random-effects meta-analysis with few studies. BMC Medical Research Methodology. 2015;15:99. doi:10.1186/s12874-015-0091-1 read 9 October 2026. doi.org/10.1186/s12874-015-0091-1
- Guyatt GH, Thorlund K, Oxman AD, Walter SD, Patrick D, Furukawa TA, Johnston BC, Karanicolas P, Akl EA, Vist G, Kunz R, Brozek J, Kupper LL, Martin SL, Meerpohl JJ, Alonso-Coello P, Christensen R, Schunemann HJ. GRADE guidelines: 13. Preparing summary of findings tables and evidence profiles - continuous outcomes. Journal of Clinical Epidemiology. 2013;66(2):173-183. doi:10.1016/j.jclinepi.2012.08.001 read 9 October 2026. doi.org/10.1016/j.jclinepi.2012.08.001
- Johnston BC, Thorlund K, da Costa BR, Furukawa TA, Guyatt GH. New methods can extend the use of minimal important difference units in meta-analyses of continuous outcome measures. Journal of Clinical Epidemiology. 2012;65(8):817-826. doi:10.1016/j.jclinepi.2012.02.008 read 9 October 2026. doi.org/10.1016/j.jclinepi.2012.02.008
- Riley RD, Higgins JPT, Deeks JJ. Interpretation of random effects meta-analyses. BMJ. 2011;342:d549. doi:10.1136/bmj.d549 read 9 October 2026. doi.org/10.1136/bmj.d549
- Sterne JAC, Sutton AJ, Ioannidis JPA, Terrin N, Jones DR, Lau J, Carpenter J, Rucker G, Harbord RM, Schmid CH, Tetzlaff J, Deeks JJ, Peters J, Macaskill P, Schwarzer G, Duval S, Altman DG, Moher D, Higgins JPT. Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. BMJ. 2011;343:d4002. doi:10.1136/bmj.d4002 read 9 October 2026. doi.org/10.1136/bmj.d4002
- Viechtbauer W. Conducting meta-analyses in R with the metafor package. Journal of Statistical Software. 2010;36(3):1-48. doi:10.18637/jss.v036.i03 read 9 October 2026. doi.org/10.18637/jss.v036.i03
- Johnston BC, Thorlund K, Schunemann HJ, Xie F, Murad MH, Montori VM, Guyatt GH. Improving the interpretation of quality of life evidence in meta-analyses: the application of minimal important difference units. Health and Quality of Life Outcomes. 2010;8:116. doi:10.1186/1477-7525-8-116 read 9 October 2026. doi.org/10.1186/1477-7525-8-116
- Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. 2009. read 9 October 2026. medcalc.org
- Bax L, Yu LM, Ikeda N, Tsuruta N, Moons KGM. A systematic comparison of software dedicated to meta-analysis of causal studies. BMC Medical Research Methodology. 2007;7:40. doi:10.1186/1471-2288-7-40 read 9 October 2026. doi.org/10.1186/1471-2288-7-40
- Viechtbauer W. Bias and efficiency of meta-analytic variance estimators in the random-effects model. Journal of Educational and Behavioral Statistics. 2005;30(3):261-293. doi:10.3102/10769986030003261 read 9 October 2026. doi.org/10.3102/10769986030003261
- Borenstein M, Hedges L, Higgins J, Rothstein H. Comprehensive Meta-Analysis version 2. Biostat (software). 2005. read 9 October 2026. link.springer.com
- Leandro G. Meta-analysis in medical research. Blackwell Publishing / BMJ Books (book). 2005. read 9 October 2026. link.springer.com
- Higgins JPT, Thompson SG. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine. 2002;21(11):1539-1558. doi:10.1002/sim.1186 read 9 October 2026. doi.org/10.1002/sim.1186
- Hartung J, Knapp G. On tests of the overall treatment effect in meta-analysis with normally distributed responses. Statistics in Medicine. 2001;20:1771-1782. read 9 October 2026. arxiv.org
- Glasziou P, Irwig L, Bain C, Colditz G. Systematic reviews in health care: a practical guide. Cambridge University Press (book). 2001. read 9 October 2026. link.springer.com
- Egger M, Davey Smith G, Altman DG. Systematic reviews in health care: meta-analysis in context. BMJ Books (book). 2001. read 9 October 2026. link.springer.com
- Rosenberg MS, Adams DC, Gurevitch J. MetaWin: statistical software for meta-analysis version 2. Sinauer Associates (software). 2000. read 9 October 2026. link.springer.com
- Sutton AJ, Abrams KR, Jones DR, Sheldon TA, Song F. Methods for meta-analysis in medical research. Wiley (book). 2000. read 9 October 2026. link.springer.com
- Normand SLT. Meta-analysis software: a comparative review - DSTAT version 1.10. The American Statistician. 1995;49:298-309. read 9 October 2026. link.springer.com
- Arthur W, Bennett W, Huffcutt A. Choice of software and programs in meta-analysis research: does it make a difference? Educational and Psychological Measurement. 1994;54:776-787. read 9 October 2026. link.springer.com
- Rosenthal R. Parametric measures of effect size. In: Cooper H, Hedges LV, eds. The Handbook of Research Synthesis. 1994. read 9 October 2026. cochrane.org
- DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188. doi:10.1016/0197-2456(86)90046-2 read 9 October 2026. doi.org/10.1016/0197-2456(86)90046-2
- Hedges LV, Olkin I. Statistical methods for meta-analysis. 1985. read 9 October 2026. shiny.lapclin-aids.ini.fiocruz.br
- Hedges LV. Distribution theory for Glass's estimator of effect size and related estimators. Journal of Educational Statistics. 1981;6(2):107-128. read 9 October 2026. faculty.ucmerced.edu
- Glass GV, Smith ML. Meta-analysis of research on the relationship of class-size and achievement. Educational Evaluation and Policy Analysis. 1979;1:2-16. read 9 October 2026. faculty.ucmerced.edu
- Walker GA, Saw JG. The distribution of linear combinations of t-variables. Journal of the American Statistical Association. 1978;73:876-878. read 9 October 2026. faculty.ucmerced.edu
- Smith ML, Glass GV. Meta-analysis of psychotherapy outcome studies. American Psychologist. 1977;32:752-760. read 9 October 2026. faculty.ucmerced.edu
- Glass GV. Primary, secondary, and meta-analysis of research. Educational Researcher. 1976:3-8. read 9 October 2026. faculty.ucmerced.edu
- Cochran WG. The combination of estimates from different experiments. Biometrics. 1954;10(1):101-129. doi:10.2307/3001666 read 9 October 2026. doi.org/10.2307/3001666
- Lehmann EL, Scheffe H. Completeness, similar regions and unbiased estimation. Sankhya. 1950:305-340. read 9 October 2026. faculty.ucmerced.edu
- Johnson NL, Welch BL. Applications of the non-central t-distribution. Biometrika. 1939;31:362-389. read 9 October 2026. faculty.ucmerced.edu
- Veroniki AA, Jackson D, Bender R, Kuss O, Langan D, Higgins JPT, Knapp G, Salanti G. Methods to calculate uncertainty in the estimated overall effect size from a random-effects meta-analysis. doi:10.1002/jrsm.1319 read 9 October 2026. doi.org/10.1002/jrsm.1319
Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
Pubrica Academy. (2026, October 9). Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting. Pubrica. https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/
Pubrica Academy. “Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting.” Pubrica, 9 Oct. 2026, pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/.
Pubrica Academy (2026) Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting. Pubrica. Available at: https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/ (Accessed: 15 October 2026).
Pubrica Academy. “Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting.” Pubrica. October 9, 2026. https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/
Pubrica Academy, ‘Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting’ (Pubrica, 9 October 2026) <https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/> accessed 15 October 2026.
Pubrica Academy. Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting. [Internet]. Pubrica; 2026 Oct 9 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/
Pubrica Academy, “Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting,” Pubrica, Oct. 9, 2026. [Online]. Available: https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/
@misc{standardised_mean_difference_smd_estim_2026,
author = {{Pubrica Academy}},
title = {Standardised Mean Difference: Cohen’s d, Hedges’ g and Glass’s Delta in Meta-Analysis Reporting},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/concepts-definitions/standardised-mean-difference-smd-estimation/},
note = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
SMD specification and reporting checklist you can copy
Sections 1 to 4 should be completed at the protocol stage, as they determine the data that the extraction form must collect. Section 6 is the methods paragraph, assembled from the preceding answers. Section 5 is the manual check confirming that the software performed the intended computation.
STANDARDISED MEAN DIFFERENCE SPECIFICATION
Review: ______________________ Construct: ______________________
Instruments encountered: ______________________________________________
1. ESTIMATOR
[ ] Hedges' adjusted g (pooled SD, small-sample correction applied)
[ ] Cohen's d (pooled SD, no correction)
[ ] Glass's delta (control-group SD only)
Reason for this choice:
____________________________________________________________
Smallest group size across included trials: n = ______
If n is small, correction applied? Y / N (uncorrected is biased upwards)
2. DENOMINATOR DETAIL
Standardising SD taken from: [ ] final values [ ] change scores
Applied uniformly to all studies? Y / N
Where a trial reported only change-score SDs, the approach taken was:
____________________________________________________________
3. DIRECTION OF SCALES
Instruments on which a HIGHER score is WORSE:
____________________________________________________________
[ ] signs reversed before pooling
[ ] forest plot direction labels checked against one known trial
4. MULTIPLE ELIGIBLE OUTCOMES PER STUDY
[ ] one pre-specified instrument per study; priority order:
1. ____________ 2. ____________ 3. ____________
[ ] combined within study (method: ______________)
[ ] multivariate model of the dependence
[ ] estimates from non-selected instruments reported in supplement
5. HAND-CHECK (do this once, it takes minutes)
Pick one study. n1 ____ mean1 ____ SD1 ____ n2 ____ mean2 ____ SD2 ____
Hand-computed estimate: ________ Software value for that row: ________
[ ] they match; the estimator is confirmed
[ ] they do not match; resolve before proceeding
6. REPORTING
Control-group SD range across studies: ______ to ______
[ ] range reported in the text as a source of heterogeneity
[ ] I-squared, H and Cochran Q with df reported
[ ] same-instrument subset re-analysed as a raw mean difference
[ ] no conventional "small / moderate / large" label asserted without a
clinical anchor
7. METHODS PARAGRAPH TO DROP IN
"Because the included trials used ______ different instruments for the same
construct, effects were expressed as ______________ (the mean difference
divided by ______________, with / without a small-sample correction).
Standardisation used the standard deviation of ______________. Scores from
instruments on which higher values indicate worse status were sign-reversed
before pooling. Where more than one eligible instrument was reported, the
instrument pre-specified as primary in the protocol was used and the
alternatives are given in the supplement. Control-group standard deviations
ranged from ______ to ______ across trials, so part of the observed
heterogeneity reflects differing population variability rather than
differing treatment effect." Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.
How Pubrica can help
Pubrica’s biostatistics team specifies the estimator, the standardising standard deviation and the outcome-selection rule before extraction, verifies one study manually against the software output so that the estimator is confirmed rather than assumed, and drafts the methods paragraph so that another analyst can reproduce the results. Where standardisation contributes materially to heterogeneity, the team quantifies this contribution and sets out the options for back-translation in presentation.