Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Meta-analysis · Definition and examples

Meta-Analysis: Definition, Method and a Worked Example

A meta-analysis is a weighted average in which the weights reflect precision rather than sample size. The first decision is whether to pool at all, and the most common error is selecting the model on the basis of a heterogeneity test.

Pubrica AcademyWritten by a research writing specialist Reviewed bya research methodology reviewer Published15 October 2026 Last checked15 October 2026 Reading time14 minutes Sources8 primary

Definition of meta-analysis

A meta-analysis is the statistical combination of results from more than one study addressing the same question, producing a pooled estimate with a confidence interval. The term was introduced by Glass in 1976, in what began as his presidential address to the American Educational Research Association1, and his own three-word description of it — “analysis of analyses”2 — remains the clearest concise description available.

Two sourcing notes. Quoted as Glass (1976, p. 3) in Kulik and Kulik's review, not read in the original. And The volume and issue could not be confirmed from the publisher or from any primary record. The James Lind Library's record prints the issue number in the volume position, so a citation copied from there will be wrong. Check the journal directly before printing a volume.

The method that made random-effects meta-analysis routine in clinical research is DerSimonian and Laird's, published in Controlled Clinical Trials in 19863; their estimator is still the default in most software.

A meta-analysis is not a systematic review. The review is the process — the question, the eligibility criteria, the search, the screening, the appraisal. The meta-analysis is an optional statistical step at the end of it. A review can be systematic with no meta-analysis in it, and a pooled estimate calculated without a systematic search is not a meta-analysis in any useful sense, because nobody knows what was left out. Our page on structuring review tables covers what the review around it has to report.

Circumstances in which pooling is inappropriate

This section comes second rather than last because the decision comes before the arithmetic. The Cochrane Handbook states it directly: “A systematic review need not contain any meta-analyses.”5 And on the condition: “Meta-analysis should only be considered when a group of studies is sufficiently homogeneous”5

Where studies are too clinically or methodologically diverse, “it may be misleading to quote an average value for the intervention effect”5 — an average of incommensurable things is a number with no referent. The failure mode is not a wide confidence interval; it is a tidy pooled estimate of a quantity that does not exist.

Two related limits worth knowing before you plan subgroup work: “It is very unlikely that an investigation of heterogeneity will produce useful findings unless there is a substantial number of studies”5, and meta-regression “should generally not be considered when there are fewer than ten studies”5

Weighting of studies

A meta-analysis is a weighted average, and the weights are the part most often misunderstood. The weight given to each study is “the weight given to each study is chosen to be the inverse of the variance of the effect estimate.”5 — the inverse-variance approach, chosen because “This choice of weights minimizes the imprecision (uncertainty) of the pooled effect estimate.”5

It follows that “larger studies, which have smaller standard errors, are given more weight than smaller studies.”5 — but note the mechanism. Weight comes from the standard error. Sample size affects it only through that, and the Handbook shows standard deviations entering alongside sample size in the calculation5. So:

Table 1. A worked meta-analysis. The figures in this table are illustrative and invented for the purpose of showing how the columns relate to one another. They are not from any study and must not be cited.

StudynMean differenceSEWeight
Study A120−0.420.2118.6
Study B96−0.310.1825.3
Study C210−0.550.2414.2
Study D180−0.380.1536.4
Study E64−0.190.315.5

Pooled estimate: −0.38 (95% CI −0.50 to −0.26). I² = 12%.

Weights are inverse-variance, so they are determined by the standard error rather than directly by the sample size. Larger studies usually have smaller standard errors and therefore larger weights, but not always: Study C has the largest sample and the third-smallest weight because its standard error is the second largest, while Study D carries the most weight on the second-largest sample. A worked example should report the standard errors so that the weights can be checked.

The table this replaces

Worked examples should be checked for internal consistency. A table that gives a study with 120 participants a weight of 25% and a study with 180 participants a weight of 20%, without the standard errors that would explain the difference, cannot illustrate inverse-variance weighting. An example should also report the pooled estimate if it claims that the combined effect is more reliable than any single study.

One further consequence, which matters for reading forest plots: under a random-effects model “a random-effects meta-analysis will award relatively more weight to smaller studies.”5 So the same set of studies will carry different weights under the two models, and the pooled estimate can move.

Heterogeneity and the interpretation of I²

Heterogeneity is variation in the true effect between studies, as distinct from the variation you would expect from sampling alone. Three quantities describe it and they are not interchangeable.

Table 2. Three measures of heterogeneity

What it isWhat it is good for
Cochran’s QA test statistic. “The usual test statistic (Cochran's Q) is computed by summing the squared deviations of each study's estimate” from the pooled estimate.Testing whether there is more variation than chance alone would produce. Weak at it: “The test is known to be poor at detecting true heterogeneity among studies as significant.” and “the test arguably has excessive power when there are many studies, especially when those studies are large.”
τ²“An alternative quantification of heterogeneity in a meta-analysis is the among-study variance” — the between-study variance, in the units of the effect measure.“This is more useful for comparisons of heterogeneity among subgroups, but values depend on the treatment effect scale.”
I²“I 2 describes the percentage of total variation across studies that is due to heterogeneity rather than chance.”A relative, scale-free summary of the proportion of observed variation that is real. It does not indicate the magnitude of the heterogeneity in clinical terms; τ² does.

A common misinterpretation of I²

I-squared is frequently described as the percentage of studies that disagree. It is not a count of studies at all: it is the share of the observed variation attributable to real heterogeneity rather than to sampling error. Its formula makes that explicit: “I 2 = 100%×(Q − df)/Q, where Q is Cochran's heterogeneity statistic and df the degrees of freedom.” And “Negative values of I 2 are put equal to zero so that I 2 lies between 0% and 100%.”

On thresholds. The 25%, 50% and 75% figures are quoted constantly, and they come with a caveat in the same paper. Higgins and colleagues wrote that “we would tentatively assign adjectives of low, moderate, and high to I 2 values of 25%, 50%, and 75%.”4 — note “tentatively” — and said in terms that “A naive categorisation of values for I 2 would not be appropriate for all circumstances”4 They also warn that “conclusions of homogeneity in meta-analyses of small numbers of studies are often unjustified.”4, so a low I² across five studies is not reassurance.

Fixed-effect and random-effects models

The two models make different assumptions about what the studies are estimating, and the choice is an assumption rather than a finding.

Table 3. The two models

Fixed-effectRandom-effects
What it assumes“all effect estimates are estimating the same underlying intervention effect”.Studies estimate “different, yet related, intervention effects”, and “we assume that they follow (usually) a normal distribution”
What the pooled estimate isAn estimate of that one common effect.An estimate of the mean of the distribution of effects.
WeightingInverse variance within studies only.Inverse variance plus the between-study variance, so “a random-effects meta-analysis will award relatively more weight to smaller studies.”
Treatment of heterogeneityIt does not account for it — fixed-effect analyses “ignore heterogeneity”.It incorporates it into the interval, which is usually wider.

An indefensible basis for model selection

The Cochrane Handbook states that the choice “should never be made on the basis of a statistical test for heterogeneity”. Conducting a Q test and switching to a random-effects model because p < 0.10 is the most common way in which this decision is made, and it is not a defensible basis. Authors should decide, and state, what they are prepared to assume about the underlying effect, and should pre-specify it in the protocol before the heterogeneity has been observed.

Publication bias and funnel plots

A funnel plot plots each study's effect against its precision. If smaller, less precise studies systematically show larger effects, the plot looks asymmetric. The usual inference is publication bias — and it is only one of several explanations. The Handbook is explicit: “Small-study effects may be due to reasons other than non-reporting bias.”5

Table 4. Five causes of small-study effects

CauseWhat it is
Non-reporting biasStudies with unfavourable results are less likely to be published, are published later, or are published in less visible outlets.
Bias in study resultsSmaller studies are on average more prone to methodological flaws, and flawed trials tend to show larger effects.
True heterogeneityThe effect genuinely differs between studies — for example, by baseline risk or by the intensity with which the intervention was delivered.
ArtefactualOdds ratios and standardised mean differences are correlated with their own standard errors, so asymmetry can appear with no bias and no heterogeneity at all.
ChanceWith few studies, asymmetry can readily arise by chance.

The second of those has an evidential basis worth quoting: “Trials with flaws or limitations tend to show larger intervention effects”5 The fourth is the one almost always omitted, and it is the one that should make you cautious, because it means asymmetry can appear with no bias and no heterogeneity at all — purely from the choice of effect measure.

A limit on the tests. The Handbook states that “tests for funnel plot asymmetry should be used only when there are at least 10 studies included in the meta-analysis.”5 The minimum of ten studies applies to statistical tests for funnel plot asymmetry. No statement was identified that sets a minimum number of studies before a funnel plot merits inspection, so this site does not make that claim.

Certainty of evidence for the pooled result

A pooled estimate is not the end of the analysis. GRADE provides the structured judgement of how much confidence the body of evidence supports7, and it is what a summary of findings table reports alongside each number.

Table 5. GRADE: five domains that can lower certainty

DomainWhat it is
Risk of biasLimitations in the design or execution of the included studies.
InconsistencyUnexplained heterogeneity of results across studies.
IndirectnessThe evidence does not directly address the question — a different population, intervention, comparator or outcome.
ImprecisionConfidence intervals wide enough to include meaningfully different conclusions.
Publication biasThere is reason to believe that the available studies are a biased selection of those conducted.

Table 6. GRADE: three factors that can raise it

FactorWhat it is
Large magnitude of effectCan raise certainty by one or two levels.
All plausible confounding would reduce the effectWhere existing biases would have operated against the observed effect.
Dose-response gradientA graded relationship between exposure and outcome.

The three upgrading factors apply mainly to observational evidence, which starts lower. The practical point for a meta-analysis is that a narrow confidence interval around a pooled estimate does not by itself mean high certainty: risk of bias in the contributing studies, inconsistency between them and indirectness of the evidence all bear on it independently of the arithmetic.

Reporting a meta-analysis

A meta-analysis is reported inside a systematic review, against PRISMA 20206. What that requires of the synthesis specifically: the summary estimate, its precision, and the measures of statistical heterogeneity, for every synthesis conducted — plus the characteristics and risk of bias of the contributing studies.

The forest plot is the conventional presentation, and its parts have meanings worth stating because they are frequently described wrongly: each study is a square or block at its point estimate, whose area is that study’s weight, with a horizontal line for its confidence interval; the diamond is the pooled estimate, drawn differently precisely so it cannot be mistaken for a study8. Our page on review tables and figures covers this in full.

Practical application

  1. The decision whether to pool should be made, and justified, before any calculation. A review need not contain a meta-analysis.
  2. The model should be pre-specified in the protocol on the basis of assumptions about the underlying effect and never selected from a heterogeneity test afterwards.
  3. I² should be reported with τ² or a prediction interval, because I² is a relative measure and conveys nothing about the magnitude of the heterogeneity in clinical terms.
  4. Higgins’s thresholds should be cited as the tentative descriptors they were, not as categories, and a low I² across a small number of studies should not be interpreted as homogeneity.
  5. Where a funnel plot is drawn, all five causes of asymmetry should be discussed rather than publication bias being concluded, and asymmetry tests should not be conducted on fewer than ten studies.
  6. The certainty of each pooled outcome should be rated with GRADE and reported alongside the estimate, together with the reason for any downgrading.
  7. Worked tables should be checked: weights follow the standard error, so weights that run against the sample sizes should be explained by the reported standard errors.

Key points

A meta-analysis is the weighted statistical combination of studies addressing one question, and the weights are inverse-variance, reflecting precision rather than sample size. The first decision is whether to pool at all, because an average of incommensurable studies yields a number with no referent. I² is the percentage of total variation attributable to heterogeneity rather than chance, not the percentage of studies that disagree, and its 25/50/75 thresholds were offered tentatively by their own authors. The choice between fixed-effect and random-effects models should rest on the assumptions made, never on a heterogeneity test. Funnel plot asymmetry has five possible causes, of which publication bias is one.

The statistical combination of results from more than one study addressing the same question, to produce a single pooled estimate with a confidence interval. The term was introduced by Glass in 1976, who described it as the “analysis of analyses”. It sits within a systematic review: the review identifies and appraises the studies, and the meta-analysis is the optional statistical step at the end.

It is the percentage of the total variation across studies that is due to real heterogeneity rather than to chance — that is Higgins and colleagues’ own definition. It is not the percentage of studies that disagree, although it is often described in that way; it is a ratio of variance components, not a count of studies. Its formula is 100% × (Q − df) / Q, and negative values are set to zero so that it ranges from 0% to 100%.

The choice should rest on what the analyst is willing to assume about the underlying effect, not on a heterogeneity test. A fixed-effect model assumes that every study estimates the same underlying effect; a random-effects model assumes that they estimate different but related effects drawn from a distribution. The Cochrane Handbook states explicitly that the choice “should never be made on the basis of a statistical test for heterogeneity”.

The inverse of the variance of its effect estimate, that is, its precision. A larger study usually carries more weight because a larger sample tends to produce a smaller standard error, but the weight is computed from the standard error, not from the sample size. Two studies of the same size can carry quite different weights, and under a random-effects model weight shifts relatively towards the smaller studies.

Not on its own. The Cochrane Handbook lists five causes of small-study effects, of which non-reporting bias is one: the others are bias in the results of smaller studies, genuine heterogeneity, chance, and an artefactual cause — odds ratios and standardised mean differences are correlated with their own standard errors, so asymmetry can appear with no bias and no heterogeneity at all. In addition, tests for asymmetry should not be used with fewer than ten studies.

No. The Handbook states that a systematic review need not contain any meta-analyses, and that pooling should be considered only where the studies are sufficiently homogeneous. Where they are not, an average can be actively misleading, and a structured narrative synthesis is the more appropriate approach.

References

  1. Glass GV. Primary, secondary, and meta-analysis of research. Educational Researcher. 1976. Originally delivered as the author's AERA presidential address. read 15 October 2026. View on James Lind LibraryThe paper that introduced the term meta-analysis. No copy could be opened and the publisher record was not reachable, so the volume and issue are not printed here: the James Lind Library's own record prints them in a form that omits the volume, and no source we could open settles it.
  2. Kulik JA, Kulik C-LC. Meta-analysis in Education. ERIC document ED297015. read 15 October 2026. View on ERICCited only as the source through which Glass's own phrasing is quoted, since the 1976 paper itself could not be opened.
  3. 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 15 October 2026. View on Controlled Clinical TrialsThe paper that gave the random-effects method its standard estimator. Verified from the article itself plus a database record; every field agrees.
  4. Higgins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557–60. doi:10.1136/bmj.327.7414.557 read 15 October 2026. View on The BMJThe paper that introduced I-squared. The source for its definition, its formula, and the authors' own warning that their suggested thresholds are tentative and not appropriate in all circumstances.
  5. Higgins JPT, Thomas J, Chandler J, Cumpston M, Li T, Page MJ, Welch VA (editors). Cochrane Handbook for Systematic Reviews of Interventions. Version 6.5, updated August 2024. Cochrane, 2024. read 15 October 2026. View on CochraneThe current edition. Chapter III covers reporting and tables, chapter 9 covers summarising study characteristics, chapter 10 covers meta-analysis and forest plots, chapter 14 covers summary of findings tables, and chapters 7 and 25 cover risk of bias.
  6. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71 read 15 October 2026. View on The BMJThe current reporting guideline for systematic reviews. Item 5 requires the eligibility criteria to be specified, which is where a review's PICO ends up.
  7. Guyatt G, Oxman AD, Akl EA, Kunz R, Vist G, Brozek J, et al., Schünemann HJ. GRADE guidelines: 1. Introduction — GRADE evidence profiles and summary of findings tables. Journal of Clinical Epidemiology. 2011;64:383–394. doi:10.1016/j.jclinepi.2010.04.026 read 15 October 2026. View on Journal of Clinical EpidemiologyThe foundational paper of the GRADE series. The issue number is not printed on the copy we opened and is not given here.
  8. Cochrane. Considerations and recommendations for figures in Cochrane reviews: graphs of statistical data. In: Cochrane Handbook for Systematic Reviews of Interventions, version 6.5, chapter III.S1. Cochrane, 2024. read 15 October 2026. View on CochraneThe technical supplement that specifies how a forest plot must be drawn, including the marker for each study, what its area means, and how a pooled estimate is to be distinguished from a study. Last updated August 2024.

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 systematic review and biostatistics teams conduct meta-analyses from start to finish: the decision on whether pooling is appropriate, the model and its justification, heterogeneity and sensitivity analyses, funnel plots where the number of studies supports them, GRADE ratings and PRISMA-ready reporting. Where a completed extraction sheet is available, the synthesis can be commissioned as a discrete piece of work.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.