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

Start one
Contents

Academy · Research Data Analytics & AI · Fixed-effect assumptions

Fixed-Effect Meta-Analysis: Assumptions and the Conditions Under Which It Is Defensible

The fixed-effect model does not assume that heterogeneity is small. It assumes that there is one true effect and that every study estimates it. This is a claim about the included studies, not a result that can be read from a heterogeneity test, and most reviews cannot support it.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time11 minutes Sources36 peer-reviewed

Peer reviewers seldom question whether a fixed-effect analysis is incorrect; they ask why it was chosen, and most methods sections do not answer that question. The usual justification, that the heterogeneity test was not significant, is inadequate, because that test has low power precisely when reviews are small. This article sets out what the model asserts, which of those assertions a given dataset can support, and how the choice should be reported so that a reviewer can verify it rather than object to it.

Learning objectives

  • What the fixed-effect model asserts, stated as four separate assumptions that can each be examined
  • The distinction between the common-effect (fixed effect, singular) and fixed effects (plural) readings of the same arithmetic
  • Why a non-significant Cochran Q is not evidence of homogeneity, and what it is evidence of
  • Why a between-study variance estimate of exactly zero is a recognised estimation artefact rather than a justification for the model
  • How to write the two methods-section sentences that justify the model rather than merely name it

Common misconceptions

  • “Low heterogeneity justifies a fixed-effect model.” The model’s assumption concerns the included studies, namely the same population, intervention and outcome definition, not a test statistic computed after the fact.
  • “A non-significant Q test shows the studies are homogeneous.” Cochran’s Q, introduced in Biometrics in 1954, follows a chi-squared distribution on k−1 degrees of freedom and has very low power when studies are few. Failure to reject the null hypothesis does not constitute acceptance of it.
  • “If the between-study variance comes out at zero, there is no heterogeneity.” The DerSimonian–Laird estimator is truncated at zero, so it returns zero whenever the data are sparse. In the original paper, this occurred in two of the four worked reviews.
  • “Fixed effect and random effects are two alternatives between which analysts simply choose.” Rice, Higgins and Lumley, in the Journal of the Royal Statistical Society, show that the same weighted average can be defended under a plural fixed-effects interpretation that assumes no common effect at all. The computation is shared; the inferential claim is not.

What the model asserts, as distinct from what it computes

The fixed-effect calculation is straightforward: each study is weighted by the inverse of its within-study variance, and the weighted estimates are averaged. The weights contain no term for variation between studies, and this omission is the defining feature of the model. Deeks, Higgins and Altman set this out in the Cochrane Handbook chapter on analysing data: the random-effects alternative weights each study by the reciprocal of the sum of its squared standard error and the between-study variance, and the fixed-effect weight is that formula with the second term removed 1.

Removing that term constitutes an assertion. It implies that the only reason two studies disagree is sampling error, and that if both had been infinitely large they would have returned the same value. Borenstein, Hedges, Higgins and Rothstein, in Introduction to Meta-Analysis, make the consequence explicit: under the fixed-effect model the summary estimate answers a question about the studies in hand, and its confidence interval describes uncertainty about one parameter shared by all of them 13.

This is a strong claim about clinical and methodological similarity. It is also a claim that can be examined directly, without any statistical analysis, by inspecting the review’s inclusion table. If the trials differ in dose, duration, baseline risk, outcome instrument or length of follow-up, the claim is already doubtful, and no test result can restore it.

Fixed effect (singular) and fixed effects (plural)

Two distinct models share the name, and the resulting confusion is understandable. Rice, Higgins and Lumley re-examined this distinction in the Journal of the Royal Statistical Society: Series A and separated them. The singular fixed-effect or common-effect model assumes one underlying effect. The plural fixed-effects model assumes nothing of the kind: it treats each study as having its own effect and reports an inverse-variance weighted average of those distinct effects, which is a legitimate summary of this particular collection of studies even when they differ substantively 4.

The arithmetic is identical; the interpretation, and therefore what may legitimately be stated in the discussion, is not. Under the plural reading the result cannot be generalised to a new setting, because no shared parameter was ever asserted. Stevens revisited the applicability of the common-effect, fixed-effects and random-effects approaches to systematic reviews and reached a similar conclusion: the choice concerns the target of inference, which should be stated before the model is fitted 2,25.

In practice, this offers a defensible option that many authors do not recognise. Where the precision of inverse-variance weighting is desired but a common effect cannot be asserted, authors should state this explicitly and interpret the result as a weighted summary of the included studies. What is not permissible is to combine the interpretive latitude of the plural model with the conclusions of the singular model in the same paper. A decision framework for the choice itself is the subject of fixed-effect or random-effects.

Separate examination of the four assumptions

Treating the model as a single binary decision obscures where it usually fails. It is more informative to separate it into its component assumptions, because their failures have different remedies. The first assumption is a common underlying effect. The second is that each study’s within-study variance is known rather than estimated, an approximation that deteriorates as studies become smaller, which is why small trials destabilise fixed-effect intervals more than their weight suggests. The third is independence between the estimates being pooled. The fourth is that the effect measure is one on which a common value is plausible; an odds ratio can be constant across baseline risks where a risk difference cannot 1.

Independence is the assumption most often violated without acknowledgement. Konstantopoulos, in Research Synthesis Methods, worked through three-level meta-analysis in which effect sizes are nested within studies, and the ordinary fixed-effect variance is incorrect in that structure 8. Multiple outcomes from one trial, multiple arms compared with a shared control, or several reports of one cohort all create the same problem. Pooling them as if independent understates the standard error, which is the direction that produces false confidence rather than caution.

The fixed-effect model’s four assumptions, and what breaks each

AssumptionImplication for the datasetWhat breaks itWhat a reviewer looks at
A single common effectEvery included study estimates the same underlying parameter; disagreement is sampling error onlyDifferent populations, doses, comparators, durations or outcome definitionsThe inclusion table and PICO criteria, before any statistical test
Within-study variances knownEach study’s standard error is treated as fixed rather than estimatedVery small studies, where the standard error is itself uncertainSample sizes of the smallest contributing studies
Independent estimatesOne estimate per study, with no shared data between estimatesMultiple outcomes or arms from one trial; overlapping cohorts; nested effect sizesWhether any study appears more than once in the forest plot
A scale on which a common value is plausibleThe effect measure can reasonably be constant across the included settingsRisk differences across widely differing baseline risks; raw scores from non-interchangeable instrumentsWhy this effect measure, rather than another, was chosen

Assembled from the model definitions in Deeks, Higgins and Altman (Cochrane Handbook), Borenstein and colleagues (Introduction to Meta-Analysis), Rice, Higgins and Lumley (JRSS Series A) and Konstantopoulos (Research Synthesis Methods). The right-hand column is editorial guidance, not a quotation from those sources.

Capabilities and limitations of the homogeneity test

Cochran’s Q compares the observed dispersion of study estimates with what sampling error alone would produce, and under the null hypothesis of no differences in effect it follows a chi-squared distribution on k−1 degrees of freedom, where k is the number of studies 1,35. Its weakness is structural rather than fixable: with few studies there is little power to detect real heterogeneity, and with many large studies it becomes sensitive to differences too small to matter 1.

DerSimonian and Laird, in Controlled Clinical Trials, present a worked example in which the statistic performs as intended. In Baum’s review of antibiotic prophylaxis in colon surgery the within-study homogeneity statistic Q was 40.4 on 25 degrees of freedom, indicating dispersion well beyond sampling error and a clear signal against a common effect 30. The more instructive cases are those in which Q is unremarkable and authors treat this as justification for the model.

Higgins and Thompson addressed the interpretive problem directly in Statistics in Medicine, proposing three statistics — H, I-squared and R — to describe heterogeneity independently of the effect metric and the number of studies, illustrated across five datasets. These describe the extent of inconsistency rather than testing a hypothesis about it, which is the quantity that should be reported 17. None of them, however, translates into a threshold at which a fixed-effect model becomes appropriate. No such threshold exists, and asserting one commonly prompts reviewer criticism.

Why a between-study variance of zero does not justify the model

Statistical software frequently returns a between-study variance of exactly zero, and authors frequently interpret this as confirmation of homogeneity. It is more accurately interpreted as a boundary artefact. The DerSimonian–Laird moment estimator is truncated at zero because a negative variance is meaningless, so it returns zero whenever the method-of-moments calculation yields a negative value, which occurs routinely with few or small studies 30.

The original paper demonstrates the pattern on its own examples. Estimated between-study variance using the weighted non-iterative estimator was 0.0137 for the Winship review and 0.0208 for the Conn review, and zero for two of the four reviews examined 30. For the Winship data the resulting overall effect was 0.389 with a standard error of 0.058 30. That two of the four reviews reached the boundary in the paper that introduced the estimator gives a reasonable indication of how often this result should be expected.

Subsequent work has examined the consequences of this behaviour. Veroniki and colleagues reviewed methods to estimate the between-study variance and its uncertainty in Research Synthesis Methods, and Viechtbauer examined the bias and efficiency of the competing estimators in the Journal of Educational and Behavioral Statistics; both make clear that a point estimate of the between-study variance from a handful of studies carries very little information 7,15. Langan and colleagues’ simulation comparison of heterogeneity variance estimators reaches a consistent conclusion. In a review of six trials, an estimate of zero is informative about the estimator rather than about the data.

0.0137between-study variance, Winship review
0.0208between-study variance, Conn review
0two of the four reviews examined

Between-study variance estimated with the weighted non-iterative estimator across four worked reviews, two of which returned the truncation boundary of zero. Source: DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials.

Consequences of violating the assumption

The principal failure is less an incorrect point estimate than a confidence interval that is too narrow and a weighting that effectively assigns the conclusion to a single trial. Because fixed-effect weights are proportional to precision without an upper limit, a single large study can dominate completely, and its particular features then become the review’s finding. Sterne and colleagues, in BMJ, document this pattern in the intravenous magnesium literature, which is treated in detail in fixed-effect methods for dichotomous data 10,29.

Interval coverage deteriorates in the same direction. In the review by Veroniki and colleagues of methods for calculating uncertainty in the summary effect, coverage of the Wald-type normal confidence interval fell as low as 0.65 against a nominal 0.95 when heterogeneity was 90% and only two or three studies were pooled 36. That work concerns the random-effects interval, so it does not directly measure fixed-effect behaviour; it does, however, demonstrate the general principle that intervals built on an assumption the data do not support can fall well short of their nominal coverage, and the fixed-effect interval makes the stronger assumption of the two 36.

Poole and Greenland, in the American Journal of Epidemiology, add a caution in the opposite direction: random-effects analyses are not always more conservative 23. Changing the model is therefore not a safeguard; it changes the question being addressed.

Reporting the model in a verifiable manner

Most objections to a fixed-effect analysis are, in substance, objections to an unexplained one. A weak results paragraph reports a pooled estimate and leaves the model, the weighting and the dispersion for the reader to reconstruct from a forest plot. A strong results paragraph states the target of inference, names the model and the estimator, and reports the heterogeneity that the model assumes away, and it is scarcely longer.

The same pooled result, reported two ways (a constructed specimen — the figures in both versions are illustrative placeholders, not findings from any study)

Invites the objection

Pooling the eight trials gave a significant benefit of treatment (SMD 0.39, 95% CI 0.28 to 0.50, p<0.001). Heterogeneity was not significant, so a fixed-effect model was used.

Why a fixed-effect model? Justified by a test result, which is the one justification the test cannot provide.

What is the estimate of? The target of inference is never stated, so the discussion’s generalisation is unsupported.

How much heterogeneity was there? “Not significant” is a hypothesis-test outcome, not a quantity.

What would the alternative have shown? Unknown, so the reader cannot tell whether the conclusion depends on the model.

Answers it in advance

Because the eight trials used the same intervention protocol, outcome instrument and follow-up window, we specified a common-effect target of inference in the protocol and fitted an inverse-variance fixed-effect model. The pooled standardised mean difference was 0.39 (95% CI 0.28 to 0.50). Cochran Q was non-significant on seven degrees of freedom, which we report as weak evidence rather than as evidence of homogeneity, and I-squared was 12%. A DerSimonian–Laird random-effects analysis returned a materially unchanged estimate with a wider interval; both are given in the supplement.

The justification is clinical and methodological, decided before the data.

The target of inference is named, so the discussion’s scope is bounded.

Heterogeneity is reported as a quantity, with the test’s limitation acknowledged.

The sensitivity analysis shows the conclusion does not rest on the model choice.

Neither version is better written. The second removes four distinct grounds for criticism and requires approximately sixty additional words. It also places the decision in the protocol, which is the only stage at which a fixed-effect model can legitimately be chosen.

Settings in which the model is rarely defensible

Some situations can be excluded before any computation. A review that pools across clearly different populations or intervention intensities has no common effect to estimate, whatever the test indicates. A review assembled from observational studies with varying confounder adjustment is pooling estimates of different quantities. A review in which the outcome is measured on instruments that are not interchangeable faces a measurement problem before a modelling problem; see standardised mean difference estimation for what the standardisation does and does not fix.

Two further cases warrant mention. Where the included trials deliberately span a range of settings in order to establish generalisability, the fixed-effect model contradicts the review’s own purpose. Where the review is small (two or three studies), neither model is well supported, because the heterogeneity can neither be estimated nor assumed away 6,7. In that situation the more transparent output is a structured narrative presenting the individual estimates, rather than a pooled estimate with a spuriously narrow interval. Pubrica’s systematic review and biostatistics teams treat this as a routine finding rather than a failure.

Unresolved questions

Three questions remain unresolved. First, there is no agreed number of studies below which pooling should not be attempted; recommendations exist, but they are conventions rather than findings. Second, the methodological literature has not converged on whether the plural fixed-effects interpretation should be offered routinely as a third option in applied reviews, despite the argument of Rice, Higgins and Lumley that it is often the most accurate description of what was computed 4. Third, the choice of between-study variance estimator remains contested — the Veroniki and Langan reviews list many candidates with different biases and no uniform winner 7.

The sources do agree on procedure: the target of inference should be stated first, the model chosen to match it, heterogeneity reported whichever model is chosen, and the result under the alternative model presented. That sequence is defensible even where the statistics are ambiguous.

Key points

Key points

The fixed-effect model asserts one true effect shared by every included study. This is a claim about the inclusion criteria rather than a conclusion from a heterogeneity test, and Cochran Q cannot establish it, because failure to reject a null hypothesis does not constitute acceptance of it. A between-study variance of zero usually reflects the DerSimonian–Laird estimator reaching its truncation boundary, as occurred in two of the four reviews in the original paper. Where a common effect cannot be defended clinically, authors should either state a plural fixed-effects target of inference explicitly or use a random-effects model, and should report heterogeneity in either case.

Application to your study

Application to your study

  1. The target of inference should be stated before any model is fitted. A single sentence should state whether the analysis estimates a single shared effect, summarises the particular studies included, or estimates the mean of a distribution of effects. The choice of model follows from this.
  2. The inclusion table should be audited against the common-effect claim. Population, intervention intensity, comparator, outcome instrument and follow-up should be listed for every study. Any column that varies materially argues against the singular model.
  3. The Q test should not be treated as permission to use the model. It should be reported with its degrees of freedom and alongside I-squared, and the text should state that a non-significant Q provides weak evidence rather than evidence of homogeneity.
  4. Authors should check whether the between-study variance estimate is exactly zero. If it is, this should be stated and the estimator named, because truncation at zero is a property of the method rather than a finding about the data.
  5. Study weights should be examined, not only the pooled estimate. If one study holds most of the fixed-effect weight, the review’s conclusion is effectively that study’s conclusion; weight percentages should be reported so that readers can see this.
  6. Non-independence should be checked explicitly. Any trial contributing two outcomes, two arms against a shared control, or a second publication of the same cohort breaks the independence assumption and needs a multilevel or multivariate approach.
  7. The alternative model should be run and reported. A brief sensitivity analysis presenting the estimate under the alternative model anticipates the most common reviewer request for any pooled analysis.
  8. With two or three studies, authors should consider not pooling. Neither model is supported at that size; a structured narrative presenting the individual estimates is more transparent than a narrow interval.

Frequently asked questions

Not on that basis alone. I-squared describes the proportion of observed dispersion not attributable to sampling error; it does not test, and cannot establish, the common-effect assumption. Higgins and Thompson introduced it in Statistics in Medicine, alongside H and R, precisely to describe the extent of inconsistency independently of the effect metric and the number of studies — not to provide a decision threshold. With few studies I-squared is itself imprecisely estimated, so a low value may simply reflect low power to detect dispersion. The defensible justification for a fixed-effect model is clinical and methodological similarity among the included studies, specified before the analysis.

Usually, it indicates that the DerSimonian–Laird moment estimator produced a negative value and was truncated at zero, because a negative variance has no meaning. This is common, not rare, with few or small studies. In the paper that introduced the estimator, DerSimonian and Laird reported between-study variance of 0.0137 for the Winship review and 0.0208 for the Conn review, and zero for two of the four reviews they examined. A zero estimate should be treated as uninformative rather than as evidence of homogeneity, the estimator used should be named, and reporting an interval or a sensitivity analysis across estimators should be considered; the Veroniki and Langan reviews in Research Synthesis Methods both show the estimators disagree.

The singular form assumes one common underlying effect shared by all studies, so the summary estimates that parameter. The plural form, re-examined by Rice, Higgins and Lumley in the Journal of the Royal Statistical Society: Series A, assumes each study has its own effect and reports an inverse-variance weighted average of those distinct effects. The computation is the same; the claim and the permissible interpretation are not. Under the plural reading the included studies may be described, but the result may not be generalised to a new setting, because no shared parameter was ever asserted. Stevens reaches a compatible conclusion: the model should be chosen according to the target of inference, which should be stated before fitting.

When the common-effect assumption holds, yes: the model is more efficient and its narrower interval is justified. However, the assumption rarely holds in clinical reviews, and switching models is not a safeguard either. Poole and Greenland showed in the American Journal of Epidemiology that random-effects analyses are not always more conservative, so the choice cannot be made by selecting the apparently cautious option. The decision should depend on the quantity to be estimated: an effect in these studies, under these conditions, or an average effect across a distribution of settings. The decision itself is set out on the Academy’s fixed-effect versus random-effects page.

Q compares the observed dispersion of the study estimates with the dispersion sampling error alone would produce. Under the null hypothesis of no differences in effect it follows a chi-squared distribution on k−1 degrees of freedom, where k is the number of studies. The criticism concerns power: with few studies Q rarely reaches significance even when real heterogeneity is substantial, and with many large studies it detects differences too small to matter. DerSimonian and Laird report a clear positive case (Q of 40.4 on 25 degrees of freedom in Baum’s review of antibiotic prophylaxis in colon surgery), but the ambiguous cases are the common ones, and they are routinely over-interpreted.

Not as independent estimates. The fixed-effect variance formula assumes the pooled estimates are independent, and two outcomes from one trial share participants, so the standard error will be understated, which is the direction that produces unwarranted confidence. Konstantopoulos set out the three-level structure in Research Synthesis Methods, where effect sizes are nested within studies, and van Houwelingen, Arends and Stijnen in Statistics in Medicine cover the multivariate alternatives. Options are to select one outcome per study in advance, to pool outcomes within a study first, or to fit a multilevel or multivariate model. Choosing per study after seeing the results is a reporting-bias problem as well as a statistical one.

There is no agreed threshold; any proposed threshold reflects a convention. What the sources do establish is that with two or three studies the between-study variance cannot be estimated usefully (Veroniki and colleagues found Wald-type interval coverage falling as low as 0.65 against a nominal 0.95 when heterogeneity was 90% and only two or three studies were pooled), and that a fixed-effect model in that situation substitutes an unverifiable assumption for the missing information. In practice, very small reviews should present the individual estimates and reason about them in the text, with any pooled figure presented as secondary.

PRISMA 2020, published by Page and colleagues in PLOS Medicine, requires authors to specify the synthesis methods, including the model, the choice of effect measure and how heterogeneity was assessed, and to report them as pre-specified or as post hoc. It does not adjudicate between models, and it does not provide a threshold. In practice, the model choice must be visible and attributable to a stated rationale, which is why writing the target of inference into the protocol makes the reporting requirement straightforward rather than awkward to satisfy.

References

  1. 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
  2. Stevens RJ. The applicability to systematic reviews of common effect, random effects and fixed effects approaches to meta-analysis. 2026. doi:10.1177/09622802261439260 read 9 October 2026. doi.org/10.1177/09622802261439260
  3. Efthimiou O, Rucker G, Schwarzer G, Higgins J, Egger M, Salanti G. Network meta-analysis of rare events using the Mantel-Haenszel method. Statistics in Medicine. 2019;38(16):2992-3012. doi:10.1002/sim.8158 read 9 October 2026. doi.org/10.1002/sim.8158
  4. Rice K, Higgins JPT, Lumley T. A re-evaluation of fixed effect(s) meta-analysis. Journal of the Royal Statistical Society: Series A. 2018;181:205-227. read 9 October 2026. arxiv.org
  5. McKenzie JE, Beller EM, Forbes AB. Introduction to systematic reviews and meta-analysis. Respirology. 2016;21(4):626-637. doi:10.1111/resp.12783 read 9 October 2026. doi.org/10.1111/resp.12783
  6. 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
  7. Veroniki AA, Jackson D, Viechtbauer W, Bender R, Bowden J, Knapp G, Kuss O, Higgins JPT, Langan D, Salanti G. Methods to estimate the between-study variance and its uncertainty in meta-analysis. Research Synthesis Methods. 2015;7(1):55-79. doi:10.1002/jrsm.1164 read 9 October 2026. doi.org/10.1002/jrsm.1164
  8. Konstantopoulos S. Fixed effects and variance components estimation in three-level meta-analysis. Research Synthesis Methods. 2011;2(1):61-76. doi:10.1002/jrsm.35 read 9 October 2026. doi.org/10.1002/jrsm.35
  9. 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
  10. 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
  11. 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
  12. Higgins JPT, Thompson SG, Spiegelhalter DJ. A re-evaluation of random-effects meta-analysis. Journal of the Royal Statistical Society: Series A. 2009;172(1):137-159. read 9 October 2026. discovery-pp.ucl.ac.uk
  13. Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. 2009. read 9 October 2026. medcalc.org
  14. 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
  15. 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
  16. van Houwelingen HC, Arends LR, Stijnen T. Advanced methods in meta-analysis: multivariate approach and meta-regression. Statistics in Medicine. 2002;21:589-624. read 9 October 2026. arxiv.org
  17. 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
  18. 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
  19. 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
  20. 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
  21. 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
  22. Reis I, Hirji KF, Afifi A. Exact and asymptotic tests for homogeneity in several 2x2 tables. Statistics in Medicine. 1999;18:893-906. read 9 October 2026. ageconsearch.umn.edu
  23. Poole C, Greenland S. Random-effects meta-analyses are not always conservative. American Journal of Epidemiology. 1999;150:469-475. read 9 October 2026. ageconsearch.umn.edu
  24. Follmann DA, Proschan MA. Valid inference in random effects meta-analysis. Biometrics. 1999;55:732-737. read 9 October 2026. arxiv.org
  25. Hedges LV, Vevea JL. Fixed- and random-effects models in meta-analysis. Psychological Methods. 1998;3. read 9 October 2026. arxiv.org
  26. 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
  27. Eysenck HJ. Meta-analysis and its problems. BMJ. 1994;309:789-792. read 9 October 2026. link.springer.com
  28. Pagliaro L, D'Amico G, Sorensen TI, Lebrec D, Burroughs AK, Morabito A, Tine F, Politi F, Traina M. Prevention of first bleeding in cirrhosis: a meta-analysis of randomized trials of nonsurgical treatment. Annals of Internal Medicine. 1992;117(1):59-70. read 9 October 2026. link.springer.com
  29. Teo KK, Yusuf S, Collins R, Held PH, Peto R. Effects of intravenous magnesium in suspected acute myocardial infarction: overview of randomised trials. BMJ. 1991;303(6816):1499-1503. read 9 October 2026. link.springer.com
  30. 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
  31. Greenland S, Robins J. Estimation of a common effect parameter from sparse follow-up data. Biometrics. 1985;41:55-68. read 9 October 2026. cochrane.org
  32. Hedges LV, Olkin I. Statistical methods for meta-analysis. 1985. read 9 October 2026. shiny.lapclin-aids.ini.fiocruz.br
  33. Breslow NE, Day NE. Combination of results from a series of 2x2 tables; control of confounding. In: Statistical Methods in Cancer Research, Volume 1: The analysis of case-control data (IARC Scientific Publications No. 32). 1980. read 9 October 2026. cochrane.org
  34. Mantel N, Haenszel W. Journal of the National Cancer Institute. 1959;22:719-748. read 9 October 2026. medcalc.org
  35. 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
  36. 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). Fixed-Effect Meta-Analysis: Assumptions and the Conditions Under Which It Is Defensible. Pubrica. https://pubrica.com/academy/latest-topics/an-overview-of-fixed-effects-assumptions-for-meta-analysis/

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.

Fixed-effect defensibility check you can copy

This checklist should be completed before a fixed-effect model is fitted, and the completed version retained with the protocol. Any item that cannot be answered in writing is likely to be raised by a reviewer. The final block contains the two sentences intended for the methods section.

FIXED-EFFECT DEFENSIBILITY CHECK
Review: ______________________   Outcome: ______________________   k = ______

1. TARGET OF INFERENCE  (decide before fitting, not after)
   [ ] A single common effect shared by all included studies
   [ ] A weighted summary of these particular studies, no common effect claimed
   [ ] An average effect across a distribution of settings
   Chosen: ______________________   Recorded in protocol? Y / N

2. CLINICAL AND METHODOLOGICAL SIMILARITY
   Does every included study share:
   [ ] the same population definition
   [ ] the same intervention and intensity / dose
   [ ] the same comparator
   [ ] the same outcome and instrument
   [ ] the same follow-up window
   Any box unticked is an argument against the common-effect model. Note which:
   ____________________________________________________________

3. STRUCTURAL ASSUMPTIONS
   [ ] One estimate per study (no shared participants between estimates)
   [ ] No trial contributes two outcomes or two arms against one control
   [ ] No overlapping cohorts or duplicate publications
   [ ] Smallest study is large enough for its standard error to be stable
   If any fails: multilevel or multivariate model required, not fixed effect.

4. HETEROGENEITY, REPORTED AS A QUANTITY
   Cochran Q = ________ on ________ df     p = ________
   I-squared = ________%      H = ________
   Between-study variance estimate = ________   Estimator used: ____________
   [ ] If the estimate is exactly zero, I have said so and named the estimator
   [ ] I have NOT used a non-significant Q as the justification for the model

5. WEIGHT CONCENTRATION
   Largest single study weight = ________%
   [ ] If above roughly half, I have said in the text that the summary largely
       reflects that study, and reported the per-study weights

6. SENSITIVITY ANALYSIS
   Fixed-effect estimate ........ ________ (95% CI ________ to ________)
   Random-effects estimate ...... ________ (95% CI ________ to ________)
   [ ] Both reported; conclusion stated to be robust / not robust to model choice

7. THE TWO METHODS SENTENCES
   "We specified a ______________ target of inference in the protocol because
   ____________________________________________________. We therefore fitted
   a ______________ model using the ______________ estimator, and report
   Cochran Q, I-squared and the between-study variance alongside the summary
   estimate; the alternative model is reported as a sensitivity analysis."

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 reviews the choice of model in a planned meta-analysis against the inclusion criteria rather than the heterogeneity output, and drafts methods text that makes the choice verifiable. Where the included studies will not support a common effect, this is identified while the protocol can still be amended, and the alternatives are set out: plural fixed effects, random effects with a declared variance estimator, or a structured synthesis without pooling.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Pubrica reviews the inclusion table to determine whether a fixed-effect analysis is defensible before it is conducted.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

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.