Meta-analysis in psychology is a statistical method that combines the effect sizes from many independent studies on the same question into a single pooled effect. Instead of treating any one experiment as the final word, it weights each study by its precision, so larger and more reliable studies carry more influence. The result is an estimate that is more accurate, and far harder to dismiss, than the scattered findings it draws from.
For a field built on small samples and surprising headlines, this matters enormously. A dozen studies on the same hypothesis can point in different directions purely by chance. Meta-analysis in psychology turns that noisy pile of results into one defensible number, complete with a confidence interval and an honest account of how much the studies disagree.
Psychology has a structural problem that meta-analysis is unusually well suited to solve. Most individual studies recruit modest samples, which gives them low statistical power and wide margins of error. Two well-run experiments can easily produce opposite conclusions, not because the underlying truth changed, but because each caught a different slice of random variation.
This is the engine behind the replication crisis. When a single striking study fails to repeat, it is tempting to conclude the effect is fake. A better move is to ask what happens when you combine every credible study on the question. By pooling results and weighting them by precision, meta-analysis reveals whether an effect is real but small, genuinely absent, or strongly dependent on context. That shift, from celebrating one decisive study to weighing the whole body of evidence, is exactly why meta-analysis now sits at the top of the evidence hierarchy in psychological science.
You cannot average raw results from studies that used different scales, tasks, and sample sizes. What you can combine is the effect size, a standardized number that expresses the strength of a finding independently of the original measurement units. Two metrics dominate psychology. The first is Cohen's d, a standardized mean difference that reports how far two group means sit apart in standard-deviation units. The second is the correlation coefficient r, used when the question is about association rather than group difference.
The art of preparing a meta-analysis is converting every primary study into the same metric, then attaching a standard error to each one. Studies with tighter standard errors get more weight in the final weighted average. If you are still building intuition for these numbers, our guide to calculating standardized effect sizes walks through the conversions, and the effect size calculator handles the arithmetic for a single comparison.
The method's founding example is also its clearest. In the mid-1970s Gene Glass coined the term meta-analysis, and together with Mary Lee Smith he synthesized hundreds of psychotherapy outcome studies. Their pooled result indicated that the average treated client improved more than roughly 75 percent of comparable untreated people, a conclusion no single trial could have established with the same authority. It reframed a contentious debate about whether therapy worked into a quantitative question about how much.
Other examples show the method's range. Cross-cultural syntheses of conformity experiments pooled decades of replications of the classic Asch line-judgement task to ask whether conformity has weakened over time and whether it differs across cultures. More recently, large coordinated efforts on ego depletion, the idea that self-control draws on a limited resource, found a much smaller effect than earlier optimistic summaries, precisely because they accounted for unpublished and underpowered studies. Each case demonstrates the same lesson: pooling many studies produces a more honest picture than any headline finding.
Fixed-effect versus random-effects models
Once your effect sizes are ready, you choose how to pool them. A fixed-effect model assumes every study estimates one identical true effect, and any differences between them are pure sampling error. That assumption is rarely plausible in psychology, where samples, age groups, measures, and procedures vary from lab to lab.
The random-effects model is the realistic default. It assumes the true effect itself varies across studies around an average, so the analysis estimates both that average and the spread around it. In practice this gives smaller studies a fairer share of influence and produces wider, more honest confidence intervals. Choosing the model is not a cosmetic decision: it changes the pooled estimate, its precision, and how you interpret the result. The companion guide on how to do a meta-analysis details the between-study variance estimators (REML, Paule-Mandel) and the Hartung-Knapp adjustment that make a random-effects result defensible.
Psychology is unusual in running two distinct meta-analytic frameworks, and a doctoral review should state which one it uses and why. The Hedges-Olkin tradition, described above and dominant in clinical and experimental psychology, pools observed effect sizes under fixed-effect or random-effects models and treats each study's effect as it was measured.
The Hunter-Schmidt tradition, also called psychometric meta-analysis or validity generalization, takes a different view that is influential in industrial-organizational, educational, and individual-differences research. It argues that an observed correlation is systematically biased downward by study artifacts, chiefly unreliable measurement and range restriction in the sample. Before pooling, it corrects each correlation for attenuation due to measurement error, dividing by the square root of the product of the two variables' reliabilities, and corrects for range restriction using the study's selection ratio. When reliabilities are reported inconsistently, it applies an artifact distribution to correct the pooled estimate as a whole rather than study by study. The payoff is an estimate of the true-score relationship between constructs, not merely the relationship between fallible measures, which is why personnel-selection research uses it to argue that a validity coefficient generalizes across settings once artifacts are removed.
The choice between the traditions is substantive. Psychometric corrections raise the pooled effect and widen its interval, and reviewers will expect you to justify the reliabilities and selection ratios you used, so commit to one framework in the protocol rather than switching after seeing the data.
Heterogeneity and publication bias: the two threats you must report
A pooled number is only trustworthy if you are candid about two problems. The first is heterogeneity, the genuine variation in the true effect from study to study. It is commonly summarized with the I-squared statistic, which expresses the percentage of variation that reflects real differences rather than chance. High heterogeneity is not a failure; it is a signal to investigate moderators such as age, design, or outcome measure. Our guide to measuring heterogeneity with I-squared explains how to interpret and act on it.
The second threat is publication bias, often called the file drawer problem. Studies with significant, exciting results get published; null results quietly disappear into desk drawers. If your meta-analysis only sees the published tip of the iceberg, it will overstate the effect. Funnel plots and formal tests help detect this asymmetry so you can flag it honestly.
Modern psychology goes well beyond the funnel plot, and a doctoral review is expected to use several of these tools. Egger's regression tests funnel asymmetry formally, trim-and-fill imputes the studies that asymmetry implies are missing, and PET-PEESE uses the relationship between effect size and standard error to estimate what an infinitely precise study would have found. p-curve and p-uniform analyze the distribution of statistically significant p-values to judge whether a literature carries genuine evidential value or merely reflects selective reporting and p-hacking, while three-parameter selection models estimate the effect under an explicit model of how significance shapes what gets published. The structural answer to the file-drawer problem is pre-registration and large multi-site Registered Replication Reports, such as the Many Labs projects, which publish results regardless of outcome and have repeatedly pulled once-celebrated effects toward zero.
The standard way to display all of this at once is the forest plot, which shows each study's effect size, its confidence interval, and the pooled diamond at the bottom. Learning to read one is a core skill, covered in our guide to reading a forest plot.
The workflow is disciplined and reproducible. You define a precise question and pre-register the protocol. You search multiple databases and screen studies against explicit inclusion criteria. You extract the data needed to compute an effect size and its standard error from every included study. You choose a pooling model, run the synthesis, and produce a forest plot. Finally you probe heterogeneity, test for publication bias, and run sensitivity checks to see whether any single study is driving the result.
If you want the full procedure laid out in detail, with the formulas and software steps, see our companion guide on how to do a meta-analysis. When it is time to assemble your own figures, the forest plot generator turns extracted study data into a publication-ready plot.
A meta-analysis is the quantitative heart of a larger systematic review. Our guide to systematic reviews in psychology and behavioral science covers that wider review process. The review is the transparent process of finding and appraising every relevant study; the meta-analysis is the statistical step that pools their numbers. Done well, the pair produces the most credible evidence psychology can offer on a question, which is why journals, funders, and thesis committees prize them.
That rigor is also why meta-analysis is demanding to execute. Extracting effect sizes consistently, choosing the right model, and defending every methodological decision takes time and statistical fluency. Whether you are a doctoral student synthesizing the literature for a thesis chapter or a research team preparing a publishable review, the difference between a defensible pooled estimate and a misleading one lives in those details.