Contents
On this page
- Three pooling methods applied to the 2×2 table
- Zero cells and the reporting of continuity corrections
- Assumptions of Peto’s method
- Fixed-effect weighting in the intravenous magnesium literature
- Testing homogeneity across 2×2 tables
- Reporting the pooling method in the methods section
- Selection of a method in practice
- Unresolved methodological questions
- Key points
- Application to your study
- Frequently asked questions
Academy · Research Data Analytics & AI · Fixed-effect, binary data
Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis
The statement “a fixed-effect model was used to pool odds ratios” does not identify a method. For dichotomous data there are at least three fixed-effect methods; they behave differently when events are rare, and the method most software applies by default is the first to fail. This article compares the methods and the reporting each requires.
For continuous data, fixed-effect pooling involves a single calculation. For 2×2 tables it comprises a family of methods: inverse variance on a log scale, Mantel–Haenszel weighting and Peto’s observed-minus-expected method. These methods agree when events are common and groups are balanced, but diverge in the conditions typical of clinical reviews: rare outcomes, small trials and tables containing zero cells. This article describes what each method computes, where each fails and what a methods section should report to satisfy a statistical reviewer.
Learning objectives
- What the Mantel–Haenszel, Peto and inverse-variance methods each compute, and which effect measures each supports
- Why zero cells invalidate inverse-variance weighting, how a continuity correction alters the estimate, and why adding 0.5 is a methodological choice rather than a convention
- The two conditions under which Peto’s method performs well, and the consequences when they are not met
- How fixed-effect weighting can allow a single trial to dominate a review, illustrated by a case in which one trial held about 90% of the weight
- What the methods section should state for the analysis to be reproducible: method, effect measure, zero-cell handling and rationale
Common misconceptions
- “Fixed effect is synonymous with inverse variance.” For binary data, Mantel–Haenszel weighting, set out by Mantel and Haenszel in the Journal of the National Cancer Institute in 1959, is usually the better-behaved default, and Peto’s method is a distinct third approach.
- “Adding 0.5 to zero cells is standard practice and requires no comment.” Sweeting, Sutton and Lambert examined continuity corrections for rare events in Statistics in Medicine and found that the choice of correction can change the direction of the conclusion.
- “Peto’s method is the appropriate method for rare events in all circumstances.” It assumes balanced group sizes and small effects. Greenland and Salvan documented bias in the one-step approach in Statistics in Medicine, and it supports odds ratios only.
- “Studies with no events in either arm contribute no information.” Under most implementations, such studies are excluded entirely, which constitutes a decision about the evidence base that should be reported to the reader.
- “A narrow pooled confidence interval indicates consistent trials.” In the intravenous magnesium literature, heterogeneity was substantial, at an I-squared of 68%, while one trial carried about 90% of the fixed-effect weight.
Three pooling methods applied to the 2×2 table
Each included trial with a binary outcome yields a 2×2 table: events and non-events in each arm. From this table, an odds ratio, a risk ratio or a risk difference can be computed, and each of the three pooling methods handles these measures differently.
The inverse-variance method takes each study’s effect estimate on the log scale, computes its standard error from the cell counts, and weights it by the inverse of the squared standard error. It is the general-purpose approach and the one carried over from continuous data. Its weakness is that the standard error is computed from the counts, so a sparse table yields an unstable weight and a table with a zero cell yields an undefined one.
The Mantel–Haenszel method, introduced by Mantel and Haenszel in the Journal of the National Cancer Institute, avoids this problem by using weights derived directly from the cell counts rather than from an estimated variance, with a different weight formula for each effect measure 37. Deeks, Higgins and Altman describe it in the Cochrane Handbook as using measure-specific weights on the log odds ratio, log risk ratio or risk difference 1. Robins, Greenland and Breslow derived a general variance estimator for the Mantel–Haenszel odds ratio in the American Journal of Epidemiology, and Greenland and Robins, in Biometrics, showed that the estimator remains consistent with sparse follow-up data, the property that makes it the appropriate default for rare outcomes 31,34. Breslow and Day set out the underlying theory for combining a series of 2×2 tables in the IARC monograph on case–control analysis 36.
The Peto method operates differently again. As summarised in the Cochrane Handbook, it sums observed-minus-expected statistics with hypergeometric variances, giving a summary log odds ratio of the sum of the observed-minus-expected terms divided by the sum of the variances, with a standard error of one over the square root of that same sum of variances 1. Yusuf, Peto and colleagues popularised the method in their overview of beta blockade in myocardial infarction in Progress in Cardiovascular Diseases, where the trials were large and balanced and the method’s assumptions held 33.
Three fixed-effect methods for dichotomous data, compared
| Inverse variance | Mantel–Haenszel | Peto | |
|---|---|---|---|
| What it weights by | One over the squared standard error of the log effect estimate | Measure-specific weights built from the cell counts | Hypergeometric variances of the observed-minus-expected statistic |
| Effect measures supported | Odds ratio, risk ratio, risk difference | Odds ratio, risk ratio, risk difference | Odds ratio only |
| Behaviour with rare events | Weights become unstable; estimate unreliable | Remains consistent with sparse data | Handles sparse data without a correction |
| Single zero cell | Undefined; needs a continuity correction | No correction needed for the odds ratio | No correction needed |
| No events in either arm | Dropped from ratio analyses | Dropped from ratio analyses | Dropped |
| Key assumption to declare | Within-study standard errors adequately estimated | None beyond the common-effect assumption itself | Balanced arm sizes and a small underlying effect |
| Use it when | You have only published estimates and standard errors, not cell counts | Default for binary outcomes, and the safer choice when events are rare | Arms balanced, events rare, expected effect small, odds ratio is the question |
Method definitions from Deeks, Higgins and Altman (Cochrane Handbook) and Mantel and Haenszel (Journal of the National Cancer Institute); sparse-data behaviour from Greenland and Robins (Biometrics) and Robins, Greenland and Breslow (American Journal of Epidemiology); Peto’s conditions from Yusuf, Peto and colleagues (Progress in Cardiovascular Diseases) and Greenland and Salvan (Statistics in Medicine). Continuity-correction behaviour from Sweeting, Sutton and Lambert (Statistics in Medicine). The bottom row is editorial guidance drawn from those sources rather than a quotation.
Zero cells and the reporting of continuity corrections
A trial with no events in one arm has an odds ratio of zero or infinity and a standard error that cannot be computed. Software typically handles this without notification, usually by adding a small constant (conventionally 0.5) to every cell of that table and then proceeding as normal. The added constant is not neutral: it shrinks the study’s estimate towards no effect and inflates its apparent precision, and the magnitude of this distortion depends on the degree of imbalance between the arms.
Sweeting, Sutton and Lambert examined this issue directly in Statistics in Medicine, under the title of what to add to nothing, and compared corrections and alternatives for rare-event meta-analysis. Their conclusion is significant for any analyst who accepts a default: the choice of correction can change the magnitude, and sometimes the direction, of the pooled result 16. Rücker, Schwarzer and Carpenter, also in Statistics in Medicine, investigated the problem further, and Günhan, Röver and Friede, in Research Synthesis Methods, set out random-effects approaches for the specific case of few studies with rare events 2,9.
Two practical consequences follow. First, Mantel–Haenszel pooling of odds ratios does not require a correction for single zero cells, which is one of the strongest arguments for using it as the default on rare outcomes 1,16. Second, trials with no events in either arm contribute no information about a ratio measure and are excluded by most implementations; consequently, the number of studies in the forest plot may be smaller than the number in the review, and this discrepancy should be stated. If the review’s interest lies in whether an event occurs at all, a risk difference retains those trials in the analysis, at the cost of an effect measure that is less likely to be constant across baseline risks.
Assumptions of Peto’s method
Peto’s method is attractive because it handles sparse data without a continuity correction and yields narrow intervals. Both properties derive from assumptions that are easily violated. It assumes that the treatment and control groups in each trial are of similar size, and that the underlying effect is small, with the odds ratio close to one 1,33. Greenland and Salvan, in Statistics in Medicine, documented bias in the one-step approach, which increases as group sizes become unbalanced and as the true effect departs from the null 30.
It is also restricted to odds ratios. If the clinical question of the review concerns absolute risk (the number of additional events per thousand treated), Peto’s method cannot answer it, and converting a Peto odds ratio to an absolute scale reintroduces the assumption that the analyst sought to avoid. The presentation of results in a form on which clinicians can act is discussed in making sense of effect size.
None of this renders the method incorrect; it renders it conditional, and the conditions should be stated in the methods section: balanced arms, rare events, a small expected effect, and the odds ratio as the measure of interest. Where these conditions hold, Peto’s method is defensible and often preferable. Where they do not, Mantel–Haenszel is the safer choice.
Fixed-effect weighting in the intravenous magnesium literature
The clearest demonstration of the problem is drawn from the published literature. Sterne and colleagues, in their BMJ recommendations on examining and interpreting funnel plot asymmetry, use intravenous magnesium after myocardial infarction as a worked example 6,29. In that analysis, heterogeneity was substantial, at an I-squared of 68%, and the single ISIS-4 trial received about 90% of the fixed-effect weight 6. A review in which one trial holds nine tenths of the weight is not a synthesis; it is effectively that single trial, with the remaining trials contributing little.
The subsequent MAGIC trial reported an odds ratio of 1.0 with a 95% confidence interval of 0.8 to 1.1, contradicting the earlier pooled benefit that smaller trials had suggested 6. The episode is cited as a caution about small-study effects and publication bias, and it is also a caution about fixed-effect weighting specifically: because the weights are proportional to precision without an upper limit, a single very large trial can determine the result while a heterogeneity statistic of 68% appears in the same table without comment 6.
The implication for drafting is specific and practical. Per-study weights should be reported, not only the pooled estimate. If one study exceeds approximately half of the total weight, this should be stated in the text, and the appropriateness of a pooled estimate should be reconsidered. Whether a fixed-effect model was defensible in a dataset with such dispersion is discussed in fixed-effect assumptions, and the choice between models is set out in fixed-effect or random-effects.
Substantial heterogeneity and near-total weight concentration in the same fixed-effect analysis, and the trial that followed. Source: Sterne JAC, Sutton AJ, Ioannidis JPA, Terrin N, et al. Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. BMJ.
Testing homogeneity across 2×2 tables
Assessment of heterogeneity for binary data has its own complications. Cochran’s Q, introduced in Biometrics, remains the standard test, and DerSimonian and Laird give a worked instance in Controlled Clinical Trials: in Baum’s review of antibiotic prophylaxis in colon surgery, the within-study homogeneity statistic Q was 40.4 on 25 degrees of freedom 32. The three statistics of Higgins and Thompson (H, I-squared and R), proposed in Statistics in Medicine and illustrated on five datasets — describe the extent of inconsistency independently of the effect metric and the number of studies 18.
With sparse tables, the asymptotic approximations underlying these statistics become less reliable. Reis, Hirji and Afifi set out exact and asymptotic tests for homogeneity in several 2×2 tables in Statistics in Medicine, and the exact approaches exist precisely because the usual chi-squared reference distribution is unreliable when expected cell counts are small 23. Turner, Omar, Yang, Goldstein and Thompson proposed a multilevel model framework for meta-analysis of trials with binary outcomes in the same journal, which handles the binomial nature of the data directly rather than through a normal approximation on a log scale 21.
The between-study variance is similarly unstable in this setting. The worked examples of DerSimonian and Laird returned estimates of 0.0137 for the Winship review and 0.0208 for the Conn review, and zero for two of the four reviews examined; an estimate of zero in sparse data usually reflects the estimator reaching its truncation boundary rather than a finding about the trials 14,32.
Reporting the pooling method in the methods section
Almost every objection raised by statistical reviewers to a binary-outcome meta-analysis reduces to one of four questions: which method was used, which effect measure, how zero cells were handled, and how many studies entered the pooled analysis. A single sentence naming the method leaves all four unanswered; three sentences can address all four.
The same binary-outcome analysis, reported in two ways (a constructed specimen; the study counts are illustrative placeholders, not findings from any review)
Invites the objection
Odds ratios were pooled across the included trials using a fixed-effect model. Heterogeneity was assessed using the I-squared statistic. Analyses were performed in standard meta-analysis software.
Which fixed-effect method? At least three methods exist for binary data, and they diverge in the conditions that matter most.
What happened to the zero cells? This is unstated, so the reader cannot determine whether a correction was applied or what effect it had.
How many studies entered the pool? Trials with no events may have been excluded without notification.
Which software and which default? “Standard software” is not a method; software packages have different defaults.
Was the choice pre-specified? Nothing excludes the possibility that the choice was made after the results were seen.
Answers it in advance
Odds ratios were pooled using the Mantel–Haenszel fixed-effect method, chosen in the protocol because the primary outcome was expected to be uncommon and the method remains consistent with sparse data. No continuity correction was applied, since Mantel–Haenszel odds-ratio pooling does not require one for tables with a single zero cell; four trials with no events in either arm contributed no information to the ratio analysis and were therefore excluded, leaving 18 of 22 trials in the pooled estimate, and these four are listed in the supplement. We report Cochran Q with its degrees of freedom, I-squared, and per-study weight percentages, and repeated the analysis with the Peto method and with a 0.5 continuity correction as pre-specified sensitivity analyses.
The method is named, with a stated reason tied to the expected event rate.
Zero-cell handling is explicit, including why no correction was needed.
The discrepancy between review size and pooled size is reported, and the excluded trials are identifiable.
Heterogeneity and per-study weights are reported, so weight concentration is visible.
The choice is identified as pre-specified, with two sensitivity analyses that test it.
The second version is approximately one hundred words longer and pre-empts five standard statistical-review objections. The most damaging is usually the third: a reader who discovers that four trials disappeared between the flow diagram and the forest plot is unlikely to trust the remainder of the analysis.
Selection of a method in practice
A workable default sequence is as follows. If events are reasonably common and arms are not grossly unbalanced, the three methods will give similar answers and the choice is of little consequence; Mantel–Haenszel should be used and reported. If events are rare, Mantel–Haenszel on the odds ratio is preferred, since it requires no continuity correction for single zero cells, and if a correction was used, a sensitivity analysis with and without it should be reported. If arms are balanced, events are rare and the expected effect is small, Peto’s method is defensible, and the three conditions should be stated. Inverse variance should be reserved for cases in which only effect estimates and standard errors from publications, rather than cell counts, are available; this situation is common and is itself a reason to report the data on which the analysis was based.
Two further aspects of method selection often go unexamined. The effect measure matters as much as the pooling method: Harbord, Egger and Sterne addressed small-study effects in meta-analyses of controlled trials with binary endpoints in Statistics in Medicine, and the funnel-plot behaviour of odds ratios in sparse data is itself a known artefact 6,12. In addition, the software default is rarely the method an analyst would choose: Bradburn, Deeks and Altman documented the metan implementation in the Stata Technical Bulletin and Harris and colleagues in the Stata Journal, Viechtbauer documented metafor in the Journal of Statistical Software, and they do not all default to the same method or the same correction 11,26.
Where the review is the deliverable rather than an exercise, this work should be undertaken before extraction rather than after. Pubrica’s systematic review and biostatistics teams specify the pooling method, effect measure and zero-cell rule in the protocol, which also means the extraction form collects cell counts rather than only published ratios.
Unresolved methodological questions
Rare-event meta-analysis is an active area of research without a consensus method. Sweeting and colleagues, Rücker and colleagues, and Günhan and colleagues each propose or evaluate different approaches, and they do not converge on a single recommendation. The most defensible position in a manuscript is to name the method used, present the result under at least one alternative, and state that the literature is unsettled; this is more defensible than asserting a best practice that the sources do not support.
Two narrower gaps should be noted. There is no agreed threshold for the degree of arm imbalance at which Peto’s method should be abandoned, only the general finding that bias increases with imbalance. In addition, whether studies with no events in either arm should be excluded or retained through a risk-difference analysis is a judgement about the review question rather than a statistical fact. The approach taken, and its rationale, should be stated.
Key points
Key points
For binary outcomes, “fixed effect” denotes a family of methods rather than a single method. Mantel–Haenszel is the appropriate default because it remains consistent when events are rare and requires no continuity correction for single zero cells; inverse variance is appropriate when only published estimates and standard errors are available; Peto is defensible when arms are balanced, events are rare and the expected effect is small, and it is restricted to odds ratios. The method, effect measure, zero-cell rule and number of studies entering the pooled analysis should be reported, together with per-study weights: in the intravenous magnesium analysis, I-squared was 68% while one trial carried about 90% of the fixed-effect weight.
Application to your study
Application to your study
- Cell counts, rather than published ratios, should be collected. The extraction form should capture events and totals in both arms for every trial, because the choice of pooling method depends on access to the 2×2 table.
- The pooling method and its rationale should be specified in the protocol. Mantel–Haenszel is the default for binary outcomes; Peto is appropriate only where it can be stated that arms are balanced, events are rare and the expected effect is small.
- The zero-cell rule should be specified before the data are examined. This includes whether a correction is applied, which correction, and how trials with no events in either arm are handled.
- The flow diagram should be reconciled with the forest plot. If the pooled analysis contains fewer studies than the review, the number should be stated and the excluded trials listed.
- Per-study weight percentages should be reported. If any single trial exceeds roughly half of the total weight, this should be stated in the text, and the appropriateness of a pooled estimate reconsidered.
- The alternative method should be run as a sensitivity analysis. For example, Mantel–Haenszel against Peto, or analyses with and without a continuity correction, each reported in a single line.
- The procedure applied by the statistical software should be verified. The documentation for the specific command should be consulted to confirm the method and default correction rather than assumed.
- The unsettled state of rare-event methodology should be acknowledged. Where a result depends on the choice of approach, that dependence is a finding and should be reported in the limitations.
Frequently asked questions
Mantel–Haenszel, in most cases. Its weights are derived directly from the cell counts rather than from an estimated variance, so it remains well behaved when events are rare: Greenland and Robins showed in Biometrics that it remains consistent with sparse follow-up data, and Robins, Greenland and Breslow derived a general variance estimator for it in the American Journal of Epidemiology. Inverse variance is appropriate when only published effect estimates and standard errors, rather than the 2×2 tables, are available, which occurs sufficiently often to be relevant. With common events and balanced arms, the two methods agree closely, so the choice is of little consequence; the divergence appears precisely in the conditions typical of clinical reviews.
When three conditions hold simultaneously: the treatment and control arms in each trial are of similar size, events are rare, and the underlying effect is expected to be small. Under these conditions, it handles sparse data without a continuity correction and yields narrow intervals, which is why Yusuf, Peto and colleagues used it in their beta-blockade overview in Progress in Cardiovascular Diseases. Outside these conditions, it is biased; Greenland and Salvan documented bias in the one-step approach in Statistics in Medicine, and the bias increases as arms become unbalanced and as the true effect departs from the null. It also supports odds ratios only, so it cannot answer a question posed in terms of absolute risk.
Two cases should be distinguished. A trial with no events in one arm can be pooled by Mantel–Haenszel on the odds ratio without any correction, which is the most straightforward approach. A trial with no events in either arm carries no information about a ratio measure and is excluded by most implementations; the number of such trials should therefore be reported and the trials listed, since otherwise the forest plot will contain fewer studies than the review without explanation. If a continuity correction is applied, the correction should be specified and the result without it presented: Sweeting, Sutton and Lambert found in Statistics in Medicine that the choice of correction can change the magnitude and sometimes the direction of the pooled result.
It is common, it is a choice, and it is not neutral. Adding a constant shrinks that study’s estimate towards no effect and inflates its apparent precision, with the distortion depending on the degree of imbalance between the arms. Sweeting, Sutton and Lambert examined corrections and alternatives for rare-event meta-analysis in Statistics in Medicine and found that the choice can change the conclusion; Rücker, Schwarzer and Carpenter investigated the problem further in the same journal. If a correction is unavoidable, the constant used should be stated and a sensitivity analysis without it reported. The preferable first step is usually to choose a method that does not require a correction.
Data should be pooled on the scale on which a common effect is most plausible and presented on the scale on which a clinician can act; these are often different scales. Odds ratios and risk ratios tend to be more stable across differing baseline risks than risk differences, which is why they are the usual pooling scales. A risk difference, however, retains trials with no events in either arm in the analysis and answers the absolute question directly. Peto’s method is restricted to odds ratios. The scale should be decided in the protocol and the reason stated, and if results are converted for presentation, the baseline risk used for the conversion should be reported.
Per-study weight percentages should be reported and examined. Fixed-effect weights are proportional to precision without an upper limit, so one very large trial can determine the result while heterogeneity statistics in the same table suggest that the trials disagree. The worked example in the BMJ recommendations of Sterne and colleagues illustrates this clearly: in the intravenous magnesium analysis, I-squared was 68% and the single ISIS-4 trial received about 90% of the fixed-effect weight. The subsequent MAGIC trial reported an odds ratio of 1.0 with a 95% confidence interval of 0.8 to 1.1, contrary to the earlier pooled benefit. If one study exceeds approximately half of the total weight, this should be stated in the text.
Less reliably than with common events, and this limitation should be stated explicitly. Cochran’s Q and the statistics derived from it rely on asymptotic approximations that become less reliable when expected cell counts are small; Reis, Hirji and Afifi set out exact and asymptotic tests for homogeneity in several 2×2 tables in Statistics in Medicine precisely for this reason. Turner, Omar, Yang, Goldstein and Thompson proposed a multilevel framework in the same journal that models the binomial data directly rather than through a normal approximation. In a sparse review, the heterogeneity statistics should be reported with an explicit caveat about their reliability rather than interpreted at face value.
No, and the sources acknowledge this. Sweeting, Sutton and Lambert; Rücker, Schwarzer and Carpenter; and Günhan, Röver and Friede each evaluate or propose different approaches for rare events and few studies without converging on a single recommendation. The defensible position in a manuscript is to name the method used and the reason for its selection, report the result under at least one alternative, and state explicitly that the methodological literature is unsettled. A reviewer is considerably more likely to accept an acknowledged uncertainty than a confident claim of best practice that the cited literature does not support.
References
- Deeks JJ, Higgins JPT, Altman DG. Analysing data and undertaking meta-analyses. Cochrane Handbook for Systematic Reviews of Interventions. 2019:241-284. doi:10.1002/9781119536604.ch10 read 9 October 2026. doi.org/10.1002/9781119536604.ch10
- 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
- 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
- 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
- Riley RD, Higgins JPT, Deeks JJ. Interpretation of random effects meta-analyses. BMJ. 2011;342:d549. doi:10.1136/bmj.d549 read 9 October 2026. doi.org/10.1136/bmj.d549
- Sterne JAC, Sutton AJ, Ioannidis JPA, Terrin N, Jones DR, Lau J, Carpenter J, Rucker G, Harbord RM, Schmid CH, Tetzlaff J, Deeks JJ, Peters J, Macaskill P, Schwarzer G, Duval S, Altman DG, Moher D, Higgins JPT. Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. BMJ. 2011;343:d4002. doi:10.1136/bmj.d4002 read 9 October 2026. doi.org/10.1136/bmj.d4002
- Viechtbauer W. Conducting meta-analyses in R with the metafor package. Journal of Statistical Software. 2010;36(3):1-48. doi:10.18637/jss.v036.i03 read 9 October 2026. doi.org/10.18637/jss.v036.i03
- Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. 2009. read 9 October 2026. medcalc.org
- Rucker G, Schwarzer G, Carpenter J. Statistics in Medicine. 2008;27:746-763. read 9 October 2026. lshtm-eprints-repo-prod.lshtm-eprints.cdl.cosector.com
- Harris RJ, Bradburn MJ, Deeks JJ, Harbord RM, Altman DG, Sterne JAC. metan: fixed- and random-effects meta-analysis. The Stata Journal. 2008;8(1):3-28. doi:10.1177/1536867x0800800102 read 9 October 2026. doi.org/10.1177/1536867x0800800102
- 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
- Harbord RM, Egger M, Sterne JA. Statistics in Medicine. 2006;25:3443-3457. read 9 October 2026. lshtm-eprints-repo-prod.lshtm-eprints.cdl.cosector.com
- 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
- 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
- Leandro G. Meta-analysis in medical research. Blackwell Publishing / BMJ Books (book). 2005. read 9 October 2026. link.springer.com
- Sweeting MJ, Sutton AJ, Lambert PC. What to add to nothing? Use and avoidance of continuity corrections in meta-analysis of rare events. Statistics in Medicine. 2004;23:1351-1375. read 9 October 2026. ageconsearch.umn.edu
- Bradburn MJ, Deeks JJ, Altman DG. Metan: an alternative meta-analysis command. Stata Technical Bulletin. 2003. read 9 October 2026. link.springer.com
- Higgins JPT, Thompson SG. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine. 2002;21(11):1539-1558. doi:10.1002/sim.1186 read 9 October 2026. doi.org/10.1002/sim.1186
- Hartung J, Knapp G. On tests of the overall treatment effect in meta-analysis with normally distributed responses. Statistics in Medicine. 2001;20:1771-1782. read 9 October 2026. arxiv.org
- 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
- Turner RM, Omar RZ, Yang M, Goldstein H, Thompson SG. A multilevel model framework for meta-analysis of clinical trials with binary outcomes. Statistics in Medicine. 2000;19:3417-3432. read 9 October 2026. arxiv.org
- 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
- 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
- Hedges LV, Vevea JL. Fixed- and random-effects models in meta-analysis. Psychological Methods. 1998;3. read 9 October 2026. arxiv.org
- Bradburn MJ, Deeks JJ, Altman DG. sbe24: metan - an alternative meta-analysis command. Stata Technical Bulletin. 1998;44:4-15. read 9 October 2026. ageconsearch.umn.edu
- 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
- Colditz GA, Brewer TF, Berkey CS, Wilson ME, Burdick E, Fineberg HV, Mosteller F. Efficacy of BCG vaccine in the prevention of tuberculosis: meta-analysis of the published literature. JAMA. 1994;271:698-702. read 9 October 2026. ageconsearch.umn.edu
- 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
- 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
- Greenland S, Salvan A. Bias in the one-step method for pooling study results. Statistics in Medicine. 1990;9:247-252. read 9 October 2026. cochrane.org
- Robins J, Greenland S, Breslow NE. A general estimator for the variance of the Mantel-Haenszel odds ratio. American Journal of Epidemiology. 1986;124:719-723. read 9 October 2026. cochrane.org
- DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188. doi:10.1016/0197-2456(86)90046-2 read 9 October 2026. doi.org/10.1016/0197-2456(86)90046-2
- Yusuf S, Peto R, Lewis J, Collins R, Sleight P. Beta blockade during and after myocardial infarction: an overview of the randomized trials. Progress in Cardiovascular Diseases. 1985;27:335-371. read 9 October 2026. cochrane.org
- 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
- Hedges LV, Olkin I. Statistical methods for meta-analysis. 1985. read 9 October 2026. shiny.lapclin-aids.ini.fiocruz.br
- 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
- Mantel N, Haenszel W. Journal of the National Cancer Institute. 1959;22:719-748. read 9 October 2026. medcalc.org
- 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
- 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). Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis. Pubrica. https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/
Pubrica Academy. “Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis.” Pubrica, 9 Oct. 2026, pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/.
Pubrica Academy (2026) Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis. Pubrica. Available at: https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/ (Accessed: 15 October 2026).
Pubrica Academy. “Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis.” Pubrica. October 9, 2026. https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/
Pubrica Academy, ‘Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis’ (Pubrica, 9 October 2026) <https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/> accessed 15 October 2026.
Pubrica Academy. Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis. [Internet]. Pubrica; 2026 Oct 9 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/
Pubrica Academy, “Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis,” Pubrica, Oct. 9, 2026. [Online]. Available: https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/
@misc{fixed_effect_methods_used_for_dichotom_2026,
author = {{Pubrica Academy}},
title = {Mantel–Haenszel, Peto or Inverse Variance: Fixed-Effect Methods for Binary Outcomes in Meta-Analysis},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/concepts-definitions/fixed-effect-methods-used-for-dichotomous-data/},
note = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Binary-outcome pooling specification you can copy
This specification should be completed at the protocol stage, before extraction, because it determines what the extraction form must collect. The final block converts directly into the methods paragraph.
BINARY-OUTCOME POOLING SPECIFICATION
Review: ______________________ Outcome: ______________________
Expected event rate in controls: ______% Arms balanced across trials? Y / N
1. EFFECT MEASURE
[ ] Odds ratio [ ] Risk ratio [ ] Risk difference
Pooling scale: ______________ Presentation scale: ______________
If converting for presentation, baseline risk used: ______%
Reason for this measure:
____________________________________________________________
2. POOLING METHOD
[ ] Mantel-Haenszel (default for binary outcomes; consistent with sparse data)
[ ] Inverse variance (only estimates and standard errors available, no cell counts)
[ ] Peto (requires ALL THREE boxes below)
[ ] arms of similar size in each trial
[ ] events rare
[ ] expected effect small, odds ratio is the question of interest
Reason recorded in protocol: Y / N
3. ZERO-CELL RULE (decide before seeing the data)
Single zero cell:
[ ] no correction (Mantel-Haenszel odds ratio)
[ ] continuity correction of ________ applied
No events in either arm:
[ ] excluded from ratio analysis, count and list reported
[ ] retained via risk-difference analysis
Sensitivity analysis with / without correction planned? Y / N
4. EXTRACTION FORM MUST CAPTURE
[ ] events, intervention arm [ ] total, intervention arm
[ ] events, control arm [ ] total, control arm
[ ] published effect estimate and CI (as a cross-check, not the input)
[ ] whether the trial reported zero events in either arm
5. WHAT WILL BE REPORTED
[ ] Method named, with reason
[ ] Zero-cell handling stated
[ ] Trials in review ______ vs trials in pooled analysis ______ (reconciled)
[ ] Cochran Q with df, and I-squared
[ ] Per-study weight percentages; largest single weight ______%
[ ] Sensitivity analysis: alternative method ______________
[ ] Statement that rare-event methodology is unsettled, if applicable
6. METHODS PARAGRAPH TO DROP IN
"Odds ratios / risk ratios / risk differences were pooled using the
______________ fixed-effect method, pre-specified in the protocol because
____________________________________________________. ______________ was
applied to tables containing a zero cell. ______ trials reported no events
in either arm and contributed no information to the ratio analysis; these
are listed in the supplement, leaving ______ of ______ trials in the pooled
estimate. We report Cochran Q with degrees of freedom, I-squared and
per-study weights, and repeated the analysis using ______________ as a
pre-specified 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 specifies the pooling method, effect measure and zero-cell rule before extraction begins, so that the extraction form collects cell counts rather than only published ratios, and conducts the sensitivity analyses that statistical reviewers typically request. Where a single trial dominates the weighting or the event rate is too low for any method to be reliable, this is reported in writing while the analysis plan can still be modified.