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

Start one
Contents

Academy · Research Data Analytics & AI · Bayesian random effects

The Bayesian Random-Effects Meta-Analysis Model for Normal Data: The Decisive Role of the Heterogeneity Prior

The Bayesian version of random-effects meta-analysis is the same normal–normal hierarchical model, fitted without assuming that the between-study variance is known. This is both its principal advantage and its principal hazard: with few studies, the result is substantially determined by the prior placed on heterogeneity, so the prior must be stated, justified and tested.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time13 minutes Sources44 peer-reviewed

Frequentist random-effects meta-analysis substitutes a point estimate of the between-study variance into the weights and then constructs an interval as if that estimate were the true value. With four or five studies this assumption does not hold, and the consequences are measurable. The Bayesian model retains the uncertainty about heterogeneity in the result instead of discarding it. The cost is that the analyst must state what was believed about heterogeneity before the data were seen, and a manuscript reporting a Bayesian meta-analysis without stating its priors omits the only component that a reviewer can actually verify.

Learning objectives

  • The structure of the normal–normal hierarchical model, and which quantity the prior is placed on
  • The reasons the Bayesian posterior does not collapse to a between-study variance of zero, and the defect this corrects
  • Defensible selection of a prior for heterogeneity: weakly informative, empirical or derived from the Cochrane database
  • The difference between a credible interval, a confidence interval and a predictive distribution, and which one a clinician needs
  • The form of a prior sensitivity analysis and the reasons it is not optional with fewer than approximately ten studies
  • The reporting required for another analyst to rerun the model and reproduce its results

Common misconceptions

  • “Bayesian methods avoid the small-study problem.” Bayesian methods make the problem visible rather than removing it. With three studies, the posterior for heterogeneity is dominated by the prior, which is a more transparent display of the same shortage of information.
  • “A non-informative prior is the neutral, assumption-free choice.” Röver and colleagues examined weakly informative priors for the heterogeneity parameter in Research Synthesis Methods and show that conventional “uninformative” choices can be strongly informative about the quantity that matters.
  • “A credible interval is the Bayesian name for a confidence interval.” They answer different questions, and neither answers the question a clinician usually has, which needs the predictive distribution for a new study.
  • “Reporting the software is enough detail.” The model is defined by its priors. Naming bayesmeta or brms without the prior specification is analogous to naming a software package without describing the method.
  • “Between-study variance of zero means no heterogeneity.” This is a frequentist artefact of a truncated estimator: DerSimonian and Laird’s own worked examples returned exactly zero for two of four reviews. A posterior distribution does not exhibit this behaviour.

Formal specification of the model

The normal–normal hierarchical model has two levels and three unknowns. At the lower level, each study’s observed effect estimate is treated as a draw from a normal distribution centred on that study’s own true effect, with a variance taken as known from the study’s reported standard error. At the upper level, the study-specific true effects are themselves draws from a normal distribution centred on the overall mean effect, with a between-study variance describing how much they scatter. The unknowns are the overall mean, the between-study variance, and the study-specific effects.

This is exactly the structure of the frequentist random-effects model that DerSimonian and Laird set out in Controlled Clinical Trials, and that Deeks, Higgins and Altman describe in the Cochrane Handbook, where each study is weighted by one over the sum of its squared standard error and the between-study variance 41. The difference is what happens to the between-study variance. The frequentist approach estimates it, substitutes the estimate into the weights, and proceeds as if it were known. The Bayesian approach puts a prior distribution on it and integrates over the resulting uncertainty, so the width of the posterior for the overall mean already includes the fact that heterogeneity is poorly determined 10.

Rubin’s 1981 paper in the Journal of Educational Statistics on estimation in parallel randomised experiments is the origin of the applied form, and it also supplies the assumption that is easiest to forget: exchangeability 42. Treating the study effects as draws from a common distribution amounts to asserting that, before the results were seen, there was no reason to order the studies. If one trial was conducted in a population expected to respond differently, it is exchangeability, rather than the prior, that should be questioned.

Defects of the frequentist approach addressed by the Bayesian model

The Bayesian approach repairs two specific failures of the standard frequentist approach. Precision about which failures are addressed is important, because the general claim that Bayesian methods are better is of no use to a reviewer.

The first is the zero-variance artefact. The DerSimonian–Laird moment estimator is truncated at zero because a negative variance is meaningless, and it frequently returns zero in small reviews — in the original paper, estimated between-study variance was 0.0137 for the Winship review and 0.0208 for the Conn review, and zero for two of the four reviews examined 41. A model that reports zero heterogeneity then produces an interval of fixed-effect width solely because an estimator has reached a boundary. Chung, Rabe-Hesketh and Choi addressed this problem in Statistics in Medicine, and the Bayesian posterior does not have the problem at all: it has mass near zero when the data suggest low heterogeneity, but it never concentrates there 19.

The second is interval coverage. In the review by Veroniki and colleagues of methods for calculating uncertainty in the summary effect, fifteen confidence interval methods were identified and grouped into seven categories, and coverage of the Wald-type normal interval, the common default, fell as low as 0.65 against a nominal 0.95 when heterogeneity was 90% and only two or three studies were pooled 44. The frequentist repair is to use a better interval method; that review recommends the Hartung–Knapp–Sidik–Jonkman approach as standard for meta-analyses of five or more studies, with sensitivity analysis across two or three methods when fewer than ten studies are pooled, and Röver, Knapp and Friede examined the approach and its modification for few studies in BMC Medical Research Methodology 17,44. The Bayesian approach reaches the same end by a different route, carrying the heterogeneity uncertainty through the calculation rather than correcting the interval afterwards. The choice between the two approaches is a modelling preference; adopting neither is not defensible. The prior decision about whether a random-effects target is appropriate at all is covered in fixed-effect or random-effects.

15frequentist confidence interval methods identified, in 7 categories
0.65coverage of the Wald-type normal interval against a nominal 0.95, at 90% heterogeneity with 2–3 studies
0between-study variance returned for two of four worked reviews by the truncated moment estimator

The two frequentist failure modes a Bayesian treatment addresses. Interval methods and coverage from Veroniki AA, Jackson D, Bender R, Kuss O, Langan D, Higgins JPT, Knapp G, Salanti G, Research Synthesis Methods; zero-variance estimates from DerSimonian R, Laird N, Controlled Clinical Trials.

Selection of the prior on heterogeneity as the central modelling decision

With many large studies, the prior on the between-study variance has little influence. With the four, five or six studies that most clinical reviews contain, it has more influence than any other analytical choice, because the data contain very little information about how much the true effects vary. This is not a flaw in the method; it is an accurate reflection of the limitations of the evidence base.

Röver, Bender, Dias, Schmid, Schmidli, Sturtz, Weber and Friede set out the practical options in Research Synthesis Methods, examining weakly informative prior distributions for the heterogeneity parameter and showing that conventional choices intended to be uninformative are often not 9. The alternative to arbitrary selection is the use of external evidence. Turner, Davey, Clarke, Thompson and Higgins analysed empirical data from the Cochrane Database of Systematic Reviews in the International Journal of Epidemiology to predict the extent of heterogeneity by outcome type and comparison type, and Turner, Jackson, Wei, Thompson and Higgins then derived predictive distributions for between-study heterogeneity with simple methods for applying them in Bayesian meta-analysis, in Statistics in Medicine 18,20. Rhodes and colleagues gave an implementation route using meta-regression and pseudo data in the same journal, and Pullenayegum derived an informed reference prior for binary outcomes 16,21. Lilienthal and colleagues extended the empirical-prior approach specifically to health technology assessment with very few studies, in Research Synthesis Methods 2.

In practice, an empirically grounded prior is available for most common outcome types, and using one is considerably easier to defend than either a flat prior or a prior chosen because it was the software default. Whichever prior is chosen, the sensitivity analysis is what makes the paper credible: the model should be refitted under at least two alternative priors, and any change in the conclusion should be reported.

Options for the prior on between-study heterogeneity, and when each is defensible

Prior family or sourceWhat it encodesDefensible whenWhat a reviewer will ask
Flat or very diffuse priorApparent neutrality about the scale of heterogeneityMany large studies, where the data dominateWhy a diffuse prior is not in fact informative about the quantity of interest
Half-normal or half-Cauchy on the heterogeneity standard deviationMass near zero with a declared scale allowing substantial heterogeneityThe scale parameter can be justified from the outcome’s plausible rangeWhere the scale came from, and what happens if it is halved or doubled
Empirical prior from the Cochrane databaseThe distribution of heterogeneity actually observed in comparable reviewsThe outcome and comparison type match a category in the published predictive distributionsThe category matched, and the justification for that match
Informed reference prior for binary outcomesHeterogeneity behaviour specific to event-rate dataThe outcome is binary and the normal approximation is doubtfulWhether the analysis scale matches the one the prior was derived on
Elicited or HTA-specific priorDomain knowledge where very few studies existAn assessment setting with a documented elicitation processWho was elicited, how, and whether the result was pre-registered

Prior families and their behaviour from Röver, Bender, Dias, Schmid, Schmidli, Sturtz, Weber and Friede (Research Synthesis Methods); empirical distributions from Turner, Davey, Clarke, Thompson and Higgins (International Journal of Epidemiology) and Turner, Jackson, Wei, Thompson and Higgins (Statistics in Medicine); binary-outcome reference prior from Pullenayegum (Statistics in Medicine); HTA application from Lilienthal and colleagues (Research Synthesis Methods). The right-hand column is editorial guidance, not a quotation.

Credible intervals and the interval relevant to clinical practice

A 95% credible interval for the overall mean is the interval containing 95% of the posterior probability for that parameter. It is not equivalent to a confidence interval, and the distinction should be stated correctly in a manuscript rather than approximately. Both, however, describe the mean of the distribution of effects, and in a model whose whole premise is that true effects vary, the mean is not what a clinician treating the next patient needs.

The clinician requires the predictive distribution for the effect in a new, comparable study. Riley, Higgins and Deeks argued the frequentist case for prediction intervals in BMJ; the Bayesian model produces the predictive distribution directly, as a by-product of the fit, and it is typically much wider than the interval for the mean 22,25. Reporting it is the clearest single advantage of the Bayesian presentation and also the element most easily omitted.

The model also gives shrunken estimates for each individual study: the posterior for each study is pulled towards the overall mean by an amount that depends on its own precision and on the estimated heterogeneity. This was Rubin’s original point, and it is useful in reviews in which one small outlying trial dominates discussion: the shrunken estimate shows the result of that trial once the rest of the evidence is taken into account 42. The translation of these quantities into information on which a clinician can act is discussed in making sense of effect size.

Reporting a Bayesian analysis to permit verification

A Bayesian meta-analysis is reproducible only if the priors, the computation and the convergence checks are all reported. In practice, the priors are most often omitted, which renders the rest of the methods section unverifiable, since the same data analysed with a different prior on heterogeneity constitute a different analysis. The specimen below contrasts the two versions.

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

Unverifiable

A Bayesian random-effects meta-analysis was conducted. Non-informative priors were used throughout. The pooled mean difference was 0.42 (95% credible interval 0.15 to 0.69), indicating a probable benefit of the intervention.

Which priors? “Non-informative” names no distribution, and conventional diffuse priors on heterogeneity are often strongly informative.

What was the posterior for heterogeneity? Omitted, so the reader cannot determine whether the data informed it at all.

Does the conclusion survive a different prior? No sensitivity analysis is reported, although it is the essential check with six studies.

What should a clinician expect in a new setting? The credible interval for the mean is not that quantity.

Was the computation sound? No software, method or diagnostics are reported, so there is nothing to reproduce.

Reproducible

We fitted a normal–normal hierarchical random-effects model to the six trials. The prior for the overall mean was normal with mean zero and standard deviation 2, which is diffuse relative to the plausible range of this outcome. The prior for the between-study heterogeneity standard deviation was a half-normal with scale 0.5, taken from the predictive distribution published for subjective outcomes in pharmacological comparisons rather than chosen by us; the posterior median was 0.21. The posterior mean difference was 0.42 (95% credible interval 0.15 to 0.69), and the 95% posterior predictive interval for a new comparable trial was −0.09 to 0.93, which includes no effect. Refitting with a half-Cauchy prior and with the scale doubled moved the posterior mean by less than 0.03 and did not change the sign of the predictive interval’s lower bound. Fitted with bayesmeta, which integrates the posterior numerically, so no sampling diagnostics apply; code and data are in the supplement.

Both priors are stated as distributions with their parameters, and the heterogeneity prior is attributed to an external source.

The posterior for heterogeneity is reported, so the reader can see what the data added.

Two prior sensitivity analyses are reported with their effect on the conclusion.

A posterior predictive interval is given, and its crossing of no effect is stated rather than buried.

The computational method is named and the absence of sampling diagnostics is explained, not left ambiguous.

The difference lies not in rigour of thought but in disclosure. With six studies, the heterogeneity prior substantially determines the width of the result, so an unstated prior makes the analysis unreviewable; stating the prior, attributing it and showing the result under alternatives converts the main weakness of the model into its main claim to transparency.

Software and the reporting of computational details

Röver’s bayesmeta package, documented in the Journal of Statistical Software, is purpose-built for the normal–normal model and computes the posterior without Markov chain Monte Carlo, which removes convergence diagnostics from the list of things that can go wrong; Röver and Friede extended it to meta-regression in Computer Methods and Programs in Biomedicine 5,10. Bürkner’s brms, also in the Journal of Statistical Software, fits the model as a Bayesian multilevel model in Stan and is the appropriate choice when the structure is more complex than two levels 15. Lunn, Thomas, Best and Spiegelhalter’s WinBUGS, in Statistics and Computing, is the historical reference implementation and remains the citation behind a great deal of published work 38. Warn, Thompson and Spiegelhalter, in Statistics in Medicine, worked through the binary-outcome case on the absolute risk difference and relative risk scales, which matters because the normal approximation that this page assumes is weakest exactly there — Günhan, Röver and Friede address few studies with rare events in Research Synthesis Methods, and the fixed-effect alternatives for binary data are covered in fixed-effect methods for dichotomous data 11,31.

Where sampling is used, convergence must be reported rather than assumed: the number of chains, iterations and warm-up, and the diagnostics inspected. Where sampling is not used, this should also be stated, because a reader cannot infer it from the package name alone.

Model averaging, publication bias and moderators

Two extensions have become sufficiently common that reviewers ask about them. The first is model-averaged Bayesian meta-analysis, which, rather than choosing between a fixed-effect and a random-effects model, weights both by their posterior plausibility. Gronau, Heck, Berkhout, Haaf and Wagenmakers give a primer in Advances in Methods and Practices in Psychological Science, Bartos and colleagues apply it in medicine in Statistics in Medicine, and Berkhout and colleagues provide a tutorial implementation in Behavior Research Methods 1,7,8. This is a genuine answer to the model-choice problem rather than an evasion of it, though it does not remove the need to state a target of inference.

The second is publication-bias adjustment. Maier, Bartos and Wagenmakers set out robust Bayesian meta-analysis in Psychological Methods, and Bartos and colleagues extended it by averaging across complementary adjustment methods in Research Synthesis Methods; the earlier selection-model tradition runs through Silliman in the Journal of the American Statistical Association and Smith, Givens and Tweedie in Biostatistics 3,6,40. These methods adjust for small-study effects of the kind Sterne and colleagues catalogue in BMJ, and they should be interpreted as sensitivity analyses rather than corrections: they indicate how fragile the result is to an unobservable process, not what the result would have been 23.

For moderators, van Lissa, van Erp and Clapper describe Bayesian regularized meta-regression in Research Synthesis Methods, which is useful when there are more candidate moderators than studies, a common situation in which conventional meta-regression cannot be fitted 4. Our biostatistics team treats the choice between these extensions as a protocol decision rather than an exploratory one.

Unresolved methodological questions

There is no consensus default prior for the between-study variance, and the papers cited above do not pretend otherwise: Röver and colleagues set out a menu of weakly informative options, the Turner papers supply empirical distributions derived from the Cochrane database, and Pullenayegum and Lilienthal and colleagues offer alternatives for particular settings. Which to prefer depends on the outcome type and on how much external evidence exists, and the literature’s own recommendation is to show the sensitivity rather than to settle the question 9.

Nor is there agreement on whether Bayesian or improved frequentist intervals should be preferred in small reviews. The Veroniki review recommends the Hartung–Knapp–Sidik–Jonkman method within the frequentist framework; the Bayesian literature addresses the same problem at the level of the model. Both are defensible, neither is established as superior, and a manuscript that chooses one and reports the other as a sensitivity analysis is in a stronger position than one that asserts a winner.

Key points

Key points

The Bayesian random-effects model has the same two-level normal–normal structure as the frequentist model but is fitted without assuming that the between-study variance is known. This removes two specific defects: the spurious zeros of the truncated moment estimator, which appeared in two of the four reviews in DerSimonian and Laird’s original paper, and intervals whose coverage degrades substantially in small, heterogeneous reviews. The cost is that the prior on heterogeneity substantially determines the result when studies are few. The prior should therefore be stated as a distribution with its parameters, an empirically derived prior should be preferred over a default, the posterior for heterogeneity and a posterior predictive interval should be reported, and the model should be refitted under at least two alternative priors.

Application to your study

Application to your study

  1. Both priors should be specified in the protocol as distributions with parameters. One prior applies to the overall mean and one to the heterogeneity standard deviation. “Non-informative” does not constitute a specification.
  2. A published empirical prior for the outcome type should be sought before a new prior is constructed. Predictive distributions derived from the Cochrane database exist for many outcome and comparison categories and are considerably easier to defend than a default.
  3. Exchangeability should be assessed explicitly. If the studies can be ordered by expected effect before the results are seen, it is the hierarchical assumption, rather than the prior, that requires justification.
  4. The posterior for heterogeneity should be reported, not only the mean effect. Its median and interval indicate the extent to which the data informed the model.
  5. A posterior predictive interval for a new comparable study should be reported. This is the quantity relevant to clinical decisions; it is a by-product of the fit and is usually considerably wider than the interval for the mean.
  6. The model should be refitted under at least two alternative heterogeneity priors. Both the distributional family and the scale should be varied, and the extent to which the posterior changes should be reported. With fewer than approximately ten studies, this step should be treated as mandatory rather than optional.
  7. The method used to compute the posterior should be stated. The method may be numerical integration or sampling; for sampling, the chains, iterations, warm-up and diagnostics inspected should be reported.
  8. The frequentist analysis should be reported alongside the Bayesian analysis. A random-effects estimate with a Hartung–Knapp–Sidik–Jonkman interval, reported in a single line, pre-empts the objection that the Bayesian result is an artefact of the prior.

Frequently asked questions

It is a two-level hierarchical model. Each study’s observed effect is treated as a normal draw around that study’s own true effect, with variance taken as known from its standard error; the study-specific true effects are in turn treated as normal draws around an overall mean, with a between-study variance describing their scatter. Priors are placed on the overall mean and on the heterogeneity parameter, and the result is a joint posterior distribution rather than point estimates. The structure is identical to the frequentist random-effects model described by DerSimonian and Laird and by Deeks, Higgins and Altman in the Cochrane Handbook; what differs is that the uncertainty about heterogeneity is carried through rather than substituted away.

A prior derived from external evidence is preferable to a software default. Turner, Davey, Clarke, Thompson and Higgins used empirical data from the Cochrane Database of Systematic Reviews to characterise the extent of heterogeneity by outcome and comparison type, and Turner, Jackson, Wei, Thompson and Higgins converted that analysis into predictive distributions with simple methods for applying them in Bayesian meta-analysis. Pullenayegum derived an informed reference prior for binary outcomes, and Lilienthal and colleagues extended the approach to health technology assessment with very few studies. If no category matches, Röver and colleagues set out weakly informative options in Research Synthesis Methods, and the sensitivity analysis then carries the argument.

No, and this is the most common misunderstanding in applied Bayesian meta-analysis. Röver, Bender, Dias and colleagues examined weakly informative priors for the heterogeneity parameter in Research Synthesis Methods and showed that distributions chosen to appear neutral can be strongly informative about the between-study standard deviation, particularly in its upper tail, which is the region that determines the width of the resulting interval. A flat prior on a variance is not flat on a standard deviation, and neither is flat in any sense that a clinician would recognise. The defensible options are an empirically derived prior or a weakly informative prior with a stated, justified scale, accompanied in either case by a sensitivity analysis.

Fewer than are required for a reliable frequentist random-effects interval; however, with two or three studies no method can recover information that is absent. The Bayesian model makes the shortage visible: the posterior for heterogeneity will closely resemble the prior, and the predictive interval will be wide. This is a more useful output than a narrow frequentist interval built on a truncated variance estimate; Veroniki and colleagues found Wald-type coverage falling as low as 0.65 against a nominal 0.95 when heterogeneity was 90% and only two or three studies were pooled. Below approximately five studies, the prior sensitivity analysis should be reported prominently, and authors should consider whether a structured synthesis is the more appropriate output.

A 95% credible interval is the range containing 95% of the posterior probability for the parameter, given the data and the priors. A 95% confidence interval is a procedure with the property that intervals constructed in this way contain the true value 95% of the time across repeated samples. They are different objects and should be named correctly in a manuscript. More importantly, both describe the mean of the distribution of effects, and in a random-effects model this is not what a clinician requires for the next patient; that requires the posterior predictive distribution for a new study, the case for which Riley, Higgins and Deeks argued in BMJ.

Yes, and this is one of its clearest practical advantages. The DerSimonian–Laird moment estimator is truncated at zero, so it returns exactly zero whenever the method-of-moments calculation becomes negative, which occurs frequently with few or small studies. In the original paper, the estimate was 0.0137 for the Winship review and 0.0208 for the Conn review, and zero for two of the four reviews examined. A posterior distribution for heterogeneity can place substantial mass near zero but does not collapse onto it, so the resulting interval for the mean effect does not silently become a fixed-effect interval. Chung, Rabe-Hesketh and Choi address the frequentist aspect of this problem in Statistics in Medicine.

Five elements are required, and the first is the one most often missing. First, both priors, written as named distributions with their parameters and with a source or justification for the heterogeneity prior. Second, the posterior for the heterogeneity parameter, not only for the mean. Third, a posterior predictive interval for a new comparable study. Fourth, a prior sensitivity analysis varying both the family and the scale, with its effect on the conclusion. Fifth, the computational detail: the package, whether the posterior was integrated numerically or sampled, and, for sampling, the chains, iterations, warm-up and convergence diagnostics. Provision of code and data in a supplement makes all of this verifiable rather than merely asserted.

The sources do not establish either approach as superior, and acknowledging this is more defensible than choosing one and claiming superiority. Within the frequentist framework, Veroniki and colleagues recommend the Hartung–Knapp–Sidik–Jonkman method as standard for meta-analyses of five or more studies, with sensitivity analysis across two or three methods when fewer than ten studies are pooled. The Bayesian approach addresses the same small-sample problem at the level of the model rather than by correcting the interval. A practical position is to choose one as the pre-specified primary analysis for stated reasons, report the other in a single line, and note where they disagree.

References

  1. Berkhout SW, Haaf JM, Gronau QF, Heck DW, Wagenmakers EJ. A tutorial on Bayesian model-averaged meta-analysis in JASP. Behavior Research Methods. 2023. read 9 October 2026. arxiv.org
  2. Lilienthal J, Sturtz S, Schurmann C, Maiworm M, Rover C, Friede T, Bender R. Bayesian random-effects meta-analysis with empirical heterogeneity priors for application in health technology assessment with very few studies. Research Synthesis Methods. 2023. read 9 October 2026. arxiv.org
  3. Maier M, Bartos F, Wagenmakers EJ. Robust Bayesian meta-analysis: addressing publication bias with model-averaging. Psychological Methods. 2023;28:107-122. read 9 October 2026. arxiv.org
  4. van Lissa CJ, van Erp S, Clapper EB. Selecting relevant moderators with Bayesian regularized meta-regression. Research Synthesis Methods. 2023;14:301-322. read 9 October 2026. arxiv.org
  5. Rover C, Friede T. Using the bayesmeta R package for Bayesian random-effects meta-regression. Computer Methods and Programs in Biomedicine. 2023;229:107303. read 9 October 2026. arxiv.org
  6. Bartos F, Maier M, Wagenmakers EJ, Doucouliagos H, Stanley TD. Robust Bayesian meta-analysis: model-averaging across complementary publication bias adjustment methods. Research Synthesis Methods. 2022;14:99-116. read 9 October 2026. arxiv.org
  7. Gronau QF, Heck DW, Berkhout SW, Haaf JM, Wagenmakers EJ. A primer on Bayesian model-averaged meta-analysis. Advances in Methods and Practices in Psychological Science. 2021;4:1-19. read 9 October 2026. arxiv.org
  8. Bartos F, Gronau QF, Timmers B, Otte WM, Ly A, Wagenmakers EJ. Bayesian model-averaged meta-analysis in medicine. Statistics in Medicine. 2021;40:6743-6761. read 9 October 2026. arxiv.org
  9. Rover C, Bender R, Dias S, Schmid CH, Schmidli H, Sturtz S, Weber S, Friede T. On weakly informative prior distributions for the heterogeneity parameter in Bayesian random-effects meta-analysis. Research Synthesis Methods. 2021;12(4):448-474. doi:10.1002/jrsm.1475 read 9 October 2026. doi.org/10.1002/jrsm.1475
  10. Rover C. Bayesian random-effects meta-analysis using the bayesmeta R package. Journal of Statistical Software. 2020;93(6):1-51. doi:10.18637/jss.v093.i06 read 9 October 2026. doi.org/10.18637/jss.v093.i06
  11. Gunhan BK, Rover C, Friede T. Random-effects meta-analysis of few studies involving rare events. Research Synthesis Methods. 2020;11:74-90. read 9 October 2026. arxiv.org
  12. Amrhein V, Greenland S, McShane B. Retire statistical significance. Nature. 2019;567(7748):305-307. doi:10.1038/d41586-019-00857-9 read 9 October 2026. doi.org/10.1038/d41586-019-00857-9
  13. 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
  14. Ahlbom A. Significance testing: why does it prevail? European Journal of Epidemiology. 2017;32(1):1-2. doi:10.1007/s10654-016-0220-0 read 9 October 2026. doi.org/10.1007/s10654-016-0220-0
  15. Burkner PC. brms: An R package for Bayesian multilevel models using Stan. Journal of Statistical Software. 2017;80:1-28. read 9 October 2026. arxiv.org
  16. Rhodes KM, Turner RM, White IR, Jackson D, Spiegelhalter DJ, Higgins JPT. Implementing informative priors for heterogeneity in meta-analysis using meta-regression and pseudo data. Statistics in Medicine. 2016;35:5495-5511. read 9 October 2026. arxiv.org
  17. 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
  18. Turner RM, Jackson D, Wei Y, Thompson SG, Higgins JPT. Predictive distributions for between-study heterogeneity and simple methods for their application in Bayesian meta-analysis. Statistics in Medicine. 2015;34(6):984-998. doi:10.1002/sim.6381 read 9 October 2026. doi.org/10.1002/sim.6381
  19. Chung Y, Rabe-Hesketh S, Choi IH. Avoiding zero between-study variance estimates in random-effects meta-analysis. Statistics in Medicine. 2013;32:4071-4089. read 9 October 2026. arxiv.org
  20. Turner RM, Davey J, Clarke MJ, Thompson SG, Higgins JPT. Predicting the extent of heterogeneity in meta-analysis, using empirical data from the Cochrane Database of Systematic Reviews. International Journal of Epidemiology. 2012;41:818-827. read 9 October 2026. arxiv.org
  21. Pullenayegum EM. An informed reference prior for between-study heterogeneity in meta-analyses of binary outcomes. Statistics in Medicine. 2011;30:3082-3094. read 9 October 2026. arxiv.org
  22. 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
  23. 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
  24. 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
  25. 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
  26. Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. 2009. read 9 October 2026. medcalc.org
  27. 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
  28. Bax L, Yu LM, Ikeda N, Tsuruta H, Moons KG. Development and validation of MIX: comprehensive free software for meta-analysis of causal research data. BMC Medical Research Methodology. 2006;6(1):50. doi:10.1186/1471-2288-6-50 read 9 October 2026. doi.org/10.1186/1471-2288-6-50
  29. 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
  30. Leandro G. Meta-analysis in medical research. Blackwell Publishing / BMJ Books (book). 2005. read 9 October 2026. link.springer.com
  31. Warn DE, Thompson SG, Spiegelhalter DJ. Bayesian random effects meta-analysis of trials with binary outcomes: methods for absolute risk difference and relative risk scales. Statistics in Medicine. 2002;21:1601-1623. read 9 October 2026. ageconsearch.umn.edu
  32. 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
  33. Smith DD, Givens GH, Tweedie RL. Adjustment for publication bias and quality bias in Bayesian meta-analysis. Biostatistics. 2001;4:277-304. read 9 October 2026. arxiv.org
  34. 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
  35. 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
  36. Sutton AJ, Lambert PC, Hellmich M, Abrams KR, Jones DR. Meta-analysis in practice: a critical review of available software. Meta-Analysis in Medicine and Health Policy (Marcel Dekker, book chapter). 2000. read 9 October 2026. link.springer.com
  37. 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
  38. Lunn DJ, Thomas A, Best N, Spiegelhalter D. WinBUGS - a Bayesian modelling framework: concepts, structure, and extensibility. Statistics and Computing. 2000;10:325-337. read 9 October 2026. ageconsearch.umn.edu
  39. Hedges LV, Vevea JL. Fixed- and random-effects models in meta-analysis. Psychological Methods. 1998;3. read 9 October 2026. arxiv.org
  40. Silliman NP. Hierarchical selection models with applications in meta-analysis. Journal of the American Statistical Association. 1997;92:926-936. read 9 October 2026. arxiv.org
  41. 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
  42. Rubin DB. Estimation in parallel randomized experiments. Journal of Educational Statistics. 1981;6(4):377-401. doi:10.3102/10769986006004377 read 9 October 2026. doi.org/10.3102/10769986006004377
  43. 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
  44. 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). The Bayesian Random-Effects Meta-Analysis Model for Normal Data: The Decisive Role of the Heterogeneity Prior. Pubrica. https://pubrica.com/academy/meta-analysis/what-is-the-bayesian-random-effects-meta-analysis-model-for-normal-data/

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.

Bayesian meta-analysis specification template

This specification should be completed at the protocol stage. Sections 1 to 4 define the model; section 6 is the methods paragraph, which can be assembled directly from the preceding sections. Recording these decisions in advance is essential, because a prior chosen after the forest plot has been seen is not a prior.

BAYESIAN RANDOM-EFFECTS META-ANALYSIS SPECIFICATION
Review: ______________________   Outcome: ______________________   k = ______
Effect measure and scale of analysis: ______________________

1. MODEL STRUCTURE
   [ ] Normal-normal hierarchical (two level), normal approximation within studies
   [ ] Binomial likelihood at study level  (required if events are rare)
   [ ] Additional levels / moderators: ______________________
   Exchangeability justified?  Y / N
   If any study is expected a priori to differ systematically, say which and why:
   ____________________________________________________________

2. PRIOR FOR THE OVERALL MEAN EFFECT
   Distribution: ______________   Parameters: ______________
   Justification (relative to the plausible range of this outcome):
   ____________________________________________________________

3. PRIOR FOR BETWEEN-STUDY HETEROGENEITY  (the decisive choice)
   Parameterised on:  [ ] standard deviation   [ ] variance
   Distribution: ______________   Scale / parameters: ______________
   Source:
     [ ] empirical predictive distribution, category matched: ______________
     [ ] informed reference prior for binary outcomes
     [ ] weakly informative, scale justified as: ______________
     [ ] elicited; elicitation process documented at: ______________
   NOT acceptable as a source: "software default", "non-informative"

4. PRIOR SENSITIVITY ANALYSES  (plan at least two)
   A. Alternative family: ______________       B. Scale x2 and x0.5: ______________
   Decision rule: the conclusion is robust if ____________________________

5. COMPUTATION
   Package and version: ______________________
   [ ] posterior integrated numerically (no sampling diagnostics apply)
   [ ] sampled:  chains ______  iterations ______  warm-up ______
                 diagnostics inspected: ______________________
   [ ] code and data to be deposited at: ______________________

6. QUANTITIES TO REPORT
   [ ] posterior mean / median for the overall effect, with 95% credible interval
   [ ] posterior median and interval for the heterogeneity parameter
   [ ] 95% posterior predictive interval for a new comparable study
   [ ] shrunken posterior estimate for each study, if any single study is contested
   [ ] prior sensitivity results, with the movement in the posterior stated
   [ ] frequentist random-effects estimate as a one-line comparison, with the
       interval method named

7. METHODS PARAGRAPH TO DROP IN
   "We fitted a normal-normal hierarchical random-effects model to the ______
   included studies. The prior for the overall effect was ______________ with
   ______________. The prior for the between-study heterogeneity standard
   deviation was ______________ with scale ______, taken from
   ______________________ rather than chosen by the authors. We report the
   posterior for the heterogeneity parameter and a 95% posterior predictive
   interval for a new comparable study, and refitted the model under
   ______________ and with the prior scale doubled and halved as pre-specified
   sensitivity analyses. Analyses used ______________; code and data are
   available at ______________."

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 model and both priors before data extraction, sources an empirically derived heterogeneity prior where one exists for the outcome type, conducts the prior sensitivity analyses, and drafts the methods paragraph so that another analyst could reproduce the results. Where the number of studies is too small for any prior to be uncontroversial, the team states this and reports the result as a range across priors rather than as a single figure.

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.

On the basis of the outcome type and the number of available studies, Pubrica can source a defensible heterogeneity prior before any model is fitted.

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.