A well-run meta-analysis can turn dozens of scattered trials into a single number precise enough to change how a drug gets prescribed worldwide. That happened in 2020, when Cochrane researchers pooled 27 trials of antenatal corticosteroids and confirmed that a single course cuts the risk of neonatal death by 22%, a conclusion built from 11,272 women across 20 countries.1 Reaching that kind of number, and trusting it, means working through the same sequence every time. A reviewer chooses a statistical model, checks how much the included trials disagree with each other, searches for studies that never made it into print, and reports each step clearly enough for another researcher to redo the analysis.

This guide walks through that sequence using three real, published meta-analyses rather than invented figures. Each example carries its actual risk ratios, confidence intervals, and heterogeneity statistics, pulled from the original papers and cross-checked against PubMed. Coverage includes choosing between fixed-effect and random-effects models, reading I², Cochran’s Q, and tau² without treating any single number as a pass-fail test, seeing how a forest plot’s diamond gets its width, and using a funnel plot to catch the kind of selective publication that once made a class of antidepressants look considerably more effective than the underlying FDA trial data actually supported.2

In Simple Words: A meta-analysis takes results from several studies that asked a similar question and combines them into one pooled number, weighted so that larger, more precise studies count for more. Reading one comes down to three checks. Does the pooled estimate cross the line of no effect on the forest plot? How much do the individual studies disagree with each other, a property statisticians call heterogeneity? Does the funnel plot look lopsided in a way that suggests missing negative studies, the classic signature of publication bias? A reader comfortable with those three checks can get through most published meta-analyses in a few minutes, and can write one that holds up under peer review.

What a Meta-Analysis Adds That a Narrative Review Cannot

Before working through the mechanics below, it is worth confirming that your studies should be pooled at all. Our guide to the minimum number of studies for a meta-analysis covers the four checks that decide it, including which statistics stop being interpretable below five studies.

A narrative review describes what a set of studies found. A meta-analysis calculates a pooled estimate from them, with a confidence interval that reflects how much total evidence sits behind the number. A forty-patient pilot trial and a five-hundred-patient multicenter study asking the identical question rarely agree exactly, and a narrative summary has no formal way to reconcile the difference beyond an author’s judgment call. Meta-analysis weights each study by its precision, usually the inverse of its variance, so a trial with 2,000 participants pulls the pooled estimate toward itself far more than a trial with 40.

Getting to a trustworthy pooled number starts upstream of any statistics. A defined research question built on the PICO framework, population, intervention, comparison, outcome, sets the boundaries for which studies even qualify. Trials that answer a slightly different question, a different dose, a different population, a different outcome definition, do not belong in the pool regardless of how well they were conducted. Skipping this step is why some published meta-analyses combine studies that were never actually comparable, producing a pooled estimate that describes no real population.

Fixed-Effect and Random-Effects Models

Two statistical models dominate meta-analysis, and picking the wrong one changes both the point estimate and, more importantly, the width of its confidence interval. A fixed-effect model assumes every included trial is estimating the exact same true effect, with any difference between them attributed purely to sampling error. That assumption fits a narrow set of circumstances, essentially identical protocols, populations, and dosing, run across multiple sites of the same trial.

A random-effects model assumes something closer to reality for most clinical questions, on the premise that the true effect varies from study to study because populations, dosing regimens, and follow-up periods genuinely differ, with sampling error sitting on top of that variation. Rebecca DerSimonian and Nan Laird formalized the standard approach to this model in 1986, introducing a between-study variance term (tau²) that gets added to each study’s own sampling variance before the weights are calculated.3 Because random-effects weighting flattens out the influence of any single large trial, its confidence interval is typically wider than a fixed-effect analysis run on the same data, and its point estimate can shift toward the simple average of study results rather than being dominated by the largest study. Most published meta-analyses in clinical medicine now default to random-effects for exactly this reason, reserving the fixed-effect model for genuinely homogeneous data such as planned multi-site replications of one protocol.

Effect measure selection happens alongside the model choice. Continuous outcomes typically use a mean difference or, when studies measured the same construct with different scales, a standardized mean difference such as Hedges’ g or Cohen’s d. Binary outcomes use a risk ratio, an odds ratio, or a risk difference, and the three are not interchangeable, an odds ratio drifts further from the risk ratio as the baseline event rate climbs above roughly 20%. Editverse’s free effect size converter handles the arithmetic between these formats when a manuscript needs to compare results reported in different metrics across included studies.

Meta-analysis planning steps for choosing an effect measure and statistical model

Reading a Forest Plot Step by Step

A forest plot is simply the visual form of a weighted average, and every element on it carries information. Each horizontal line represents one included trial, with a square marking its point estimate and the line’s length showing its 95% confidence interval. Square size scales with the study’s weight in the pooled analysis, so a glance at the plot shows which trials are actually driving the result before reading a single number. A diamond at the bottom represents the pooled estimate, its width equal to the pooled confidence interval, and its horizontal position relative to the vertical line of no effect (1.0 for a risk ratio, 0 for a mean difference) is the first thing worth checking.

The 2020 Cochrane review on antenatal corticosteroids offers real numbers to walk through this mechanically rather than abstractly. Emma McGoldrick and colleagues combined 26 trials covering 11,183 infants for the respiratory distress syndrome outcome and found a risk ratio of 0.71, with a 95% confidence interval running from 0.65 to 0.78.1 On a forest plot, every one of those 26 trial lines would sit at least partly to the left of 1.0, and the pooled diamond sits entirely left of it too, since the upper bound of 0.78 never reaches the no-effect line. A diamond that does not touch the vertical line is what statistical significance looks like on a forest plot. GRADE rated this particular finding as high-certainty evidence, meaning further research is unlikely to change the estimate meaningfully.

The same review’s neonatal death outcome pooled 22 trials and 10,609 infants into a risk ratio of 0.78 (95% CI, 0.70 to 0.87), translating to roughly 2.6 fewer deaths per 100 treated infants (95% CI, 1.5 to 3.6 fewer).1 Reporting the absolute risk difference alongside the relative risk ratio matters clinically. A 22% relative reduction sounds identical whether the baseline risk is 1% or 20%, but the number of infants who need treatment to prevent one death differs by a factor of twenty between those two scenarios. Editverse’s free meta-analysis calculator generates both the pooled estimate and a forest plot from a set of study-level inputs, which is a faster way to sanity-check a hand-built plot before it goes into a manuscript.

Measuring Heterogeneity with I², Cochran’s Q, and Tau²

Heterogeneity is the technical term for how much included studies disagree beyond what sampling error alone would predict, and three statistics quantify it from different angles. Cochran’s Q tests whether the disagreement is statistically significant at all, but its power depends heavily on how many trials are pooled, so a meta-analysis with only four or five studies can carry real heterogeneity that Q fails to flag, while a meta-analysis with sixty studies can flag trivial heterogeneity as significant. Julian Higgins and colleagues addressed that weakness by developing I², which expresses heterogeneity as a percentage of total variation across studies rather than a raw significance test, making it comparable across meta-analyses of different sizes. A widely cited 2003 paper translated the statistic into practical guidance for a clinical audience, suggesting that values near 25% represent low heterogeneity, near 50% moderate, and near 75% substantial, while cautioning against applying those cutoffs mechanically since the clinical importance of heterogeneity depends on the question being asked.4

A 2018 JAMA Psychiatry meta-analysis of resistance exercise training and depressive symptoms shows what a high I² looks like in a real dataset. Brett Gordon and colleagues pooled 54 effect sizes from 33 randomized trials involving 1,877 participants and found a moderate antidepressant benefit, Hedges’ d of 0.66 (95% CI, 0.48 to 0.83), alongside substantial heterogeneity, a Cochran’s Q of 216.92 on 53 degrees of freedom (P less than .001) and an I² of 76.0% (95% CI, 72.7% to 79.0%).5 Rather than treating that heterogeneity as a reason to distrust the pooled estimate, the authors ran a meta-regression across four pre-specified moderators, total prescribed exercise volume, participants’ baseline health status, whether allocation and outcome assessment were blinded, and whether the intervention produced measurable strength gains. Blinding status turned out to explain part of the variation, trials with blinded assessment reported smaller effects, while the other three moderators did not reach significance. Investigating its likely sources with subgroup analysis or meta-regression, rather than either ignoring it or abandoning the pooled estimate altogether, is the proper response to high heterogeneity.

Tau² measures the same underlying concept on the scale of the effect size itself rather than as a percentage, and it feeds directly into how wide a random-effects confidence interval ends up. Gerta Rücker and colleagues cautioned that I² alone can be misleading because its value depends partly on the precision of the individual studies rather than purely on true effect variation, which is one reason a careful methods section reports tau² and Cochran’s Q alongside I² rather than I² in isolation. Editverse’s free heterogeneity calculator computes all three statistics from study-level data, and the site’s dedicated guide on managing and assessing heterogeneity in systematic reviews covers subgroup and meta-regression strategies in more depth than fits here.

Detecting Publication Bias with Funnel Plots and Egger’s Test

Publication bias occurs when studies with positive or statistically significant results reach print more often than studies with null or negative results, skewing any meta-analysis built only from the published literature. Erick Turner and colleagues demonstrated the scale of the problem directly in 2008 by comparing FDA regulatory filings against the published literature for 74 antidepressant trials covering 12,564 patients across 12 drugs.2 Thirty-one percent of those trials, representing 3,449 participants, were never published at all. Among trials the FDA classified as positive, 37 of 38 reached print. Among trials the FDA classified as negative or questionable, most either disappeared entirely or were written up in a way that read as positive. The published literature alone suggested 94% of antidepressant trials were positive; the complete FDA dataset showed the true figure was 51%. Pooling only the published trials inflated the apparent drug effect by 32% on average, ranging from 11% to 69% depending on the specific drug.

A funnel plot is the standard visual check for this kind of distortion. Each study is plotted with its effect size on the horizontal axis and a measure of its precision, often standard error, on the vertical axis, with more precise (larger) studies toward the top. In the absence of bias, smaller and less precise studies scatter symmetrically around the pooled estimate, producing a shape resembling an inverted funnel. Missing small studies on one side, typically the side showing null or unfavorable results, produce visible asymmetry. Matthias Egger and colleagues formalized this visual check into a statistical test in 1997, regressing each study’s standardized effect against its precision and testing whether the resulting intercept differs from zero, which happens when small imprecise studies systematically diverge from what larger studies show.6

When Egger’s test or a visual funnel plot inspection flags likely asymmetry, Sue Duval and Richard Tweedie’s trim-and-fill method offers one way to estimate its impact. The method removes the most extreme small studies driving the asymmetry, uses the remaining symmetric core to estimate a corrected pooled effect, then adds back mirrored placeholder studies to restore symmetry and recalculate a confidence interval that reflects the likely presence of unpublished data.7 Trim-and-fill is best treated as a sensitivity check rather than a definitive correction; it estimates what the answer might look like if the missing studies existed and matched the pattern of the ones that are visible, which is a real assumption rather than a certainty. Editverse’s free publication bias calculator runs Egger’s test and generates a funnel plot from study-level inputs, and the full mechanics of trim-and-fill and related detection methods are covered in Editverse’s guide to advanced methods for detecting publication bias in systematic reviews.

Risk of bias assessment feeding into a meta-analysis of clinical trials

PRISMA 2020 and Transparent Meta-Analysis Reporting

A statistically sound meta-analysis still fails peer review if a reader cannot reconstruct what was actually done, which is what the PRISMA 2020 statement exists to prevent. Matthew Page’s working group published the current 27-item checklist in 2021, and several items map directly onto the steps covered above.8 Item 12 asks authors to name the effect measure chosen for each outcome. Items 13a through 13f require a description of how studies were grouped, which synthesis method and software combined them, how heterogeneity was explored, and what sensitivity analyses were run. Item 14 covers the methods used to assess bias from missing results, funnel plots and Egger’s test among them, and items 20a through 20d require the actual synthesis results, heterogeneity statistics, and sensitivity findings to appear in the paper rather than being left implicit.

Editverse’s full walkthrough of the PRISMA 2020 guidelines and complete checklist covers every item in detail, including the flow diagram that tracks records from initial search through final inclusion. Building that diagram from raw screening numbers is easier with Editverse’s free PRISMA 2020 flow diagram generator, which produces a publication-ready figure directly from the counts at each screening stage.

Subgroup Analysis, Sensitivity Analysis, and Meta-Regression

Once a pooled estimate and its heterogeneity are in hand, three techniques probe the result further. Subgroup analysis splits included studies by a pre-specified characteristic, dosage tier, patient age band, study region, and compares pooled effects across the groups, which is exactly what explained part of the heterogeneity in the resistance-exercise example above. Sensitivity analysis reruns the pooled estimate after excluding studies one at a time, or after excluding an entire category such as trials at high risk of bias, to check whether the overall conclusion depends heavily on any single study or design choice. Meta-regression extends subgroup analysis to continuous study-level variables, testing whether something like mean participant age or year of publication predicts the size of the effect across studies, though it needs a reasonably large number of included studies, generally ten or more, to produce a trustworthy result.

All three techniques should be specified in a protocol before the data are pooled, not chosen afterward based on which subgroup produces the most favorable result. A subgroup comparison run only after seeing that it might strengthen the headline finding is a form of the same selective reporting that publication bias detection tries to catch at the study level, just moved inside a single analysis.

Common Errors That Undermine a Meta-Analysis Manuscript

A handful of mistakes account for a disproportionate share of the revision requests and rejections that meta-analysis submissions receive. Combining outcomes measured on genuinely different scales without converting to a common metric first is one. Reporting I² without also reporting the confidence interval around it, which can be wide enough to span “low” and “substantial” heterogeneity in a small meta-analysis, is another. Choosing a random-effects model after seeing that it narrows an unwelcome confidence interval, rather than because the underlying studies are genuinely heterogeneous, is a third pattern reviewers catch quickly, since the model choice should follow from the clinical and methodological diversity of the included studies rather than from which model produces a more publishable result. Editverse’s dedicated guide on common mistakes in meta-analysis and how to avoid them works through a longer list with fixes for each.

How Editverse Supports Meta-Analysis and Systematic Review Manuscripts

Running a defensible meta-analysis and writing it up so a reviewer can follow every decision are two different skill sets, and manuscripts often stall at the second one even when the underlying statistics are sound. Editverse’s systematic review and meta-analysis team works across both, from protocol design and model selection through PRISMA-compliant reporting and pre-submission checks against the specific journal’s requirements.

  • Statistical model selection and heterogeneity assessment for the actual dataset in hand
  • Forest plot and funnel plot construction, publication-ready for journal submission
  • Publication bias evaluation using Egger’s test and trim-and-fill where appropriate
  • Full PRISMA 2020 checklist compliance review before submission

Working on a meta-analysis or systematic review manuscript?

Editverse helps researchers choose the right model, interpret heterogeneity correctly, and report results in a way that survives peer review.

Get Meta-Analysis Support

Frequently Asked Questions About Meta-Analysis

What is the difference between a systematic review and a meta-analysis?

A systematic review is the structured process of searching, screening, and appraising all available studies on a question. A meta-analysis is the statistical step within that process, and only applies when included studies are similar enough to combine into one pooled estimate. A systematic review can exist without a meta-analysis, when studies are too different to pool, but a meta-analysis presumes a systematic review happened first.

When should a fixed-effect model replace a random-effects model?

A fixed-effect model fits when the included studies are close to clinical and methodological replicates of one another, such as pre-planned multi-site arms of a single trial protocol. For most meta-analyses combining independently conducted trials with differing populations or dosing, random-effects is the safer default, since it does not assume away real between-study variation.

How high does I² need to be before it is a problem?

There is no fixed cutoff that applies universally. Values near 75% or higher, as in the resistance-exercise example above, warrant investigating likely sources through subgroup analysis or meta-regression rather than abandoning the pooled estimate. The clinical direction and consistency of individual study effects matter alongside the raw I² percentage.

Does a symmetric funnel plot rule out publication bias?

No. A funnel plot with fewer than roughly ten included studies rarely has enough data points to judge asymmetry reliably in either direction, and some forms of bias, such as selective outcome reporting within an otherwise published trial, do not produce funnel plot asymmetry at all. Egger’s test and a check of trial registries against the published literature, the approach Turner and colleagues used for antidepressants, catch some bias that visual inspection alone misses.

Can a meta-analysis use both randomized and observational studies together?

Generally not in the same pooled estimate. Randomized and observational designs carry different baseline risks of confounding, and mixing them without a clear methodological justification, and separate sensitivity analysis, undermines the interpretability of the pooled result. Most journals expect these study types to be analyzed and reported separately even within one review.

What sample size does a meta-regression need to be reliable?

Ten included studies per covariate is a commonly cited minimum, and meta-regression results from reviews with fewer studies than that should be described as exploratory rather than confirmatory. A regression run across five studies and three covariates has too little statistical power to distinguish a real moderator from noise.

References

  1. McGoldrick E, Stewart F, Parker R, Dalziel SR. Antenatal corticosteroids for accelerating fetal lung maturation for women at risk of preterm birth. Cochrane Database Syst Rev. 2020;12(12):CD004454. PMID: 33368142
  2. Turner EH, Matthews AM, Linardatos E, Tell RA, Rosenthal R. Selective publication of antidepressant trials and its influence on apparent efficacy. N Engl J Med. 2008;358(3):252-260. PMID: 18199864
  3. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials. 1986;7(3):177-188. PMID: 3802833
  4. Higgins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557-560. PMID: 12958120
  5. Gordon BR, McDowell CP, Hallgren M, Meyer JD, Lyons M, Herring MP. Association of efficacy of resistance exercise training with depressive symptoms: meta-analysis and meta-regression analysis of randomized clinical trials. JAMA Psychiatry. 2018;75(6):566-576. PMID: 29800984
  6. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-analysis detected by a simple, graphical test. BMJ. 1997;315(7109):629-634. PMID: 9310563
  7. Duval S, Tweedie R. Trim and fill: a simple funnel-plot-based method of testing and adjusting for publication bias in meta-analysis. Biometrics. 2000;56(2):455-463. PMID: 10877304
  8. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. PMID: 33782057

Editverse Research Team | Updated August 2026