Fixed vs Random Effects Meta-Analysis: Which Model to Report

Most meta-analyses in medical journals report a pooled estimate, a forest plot and a single sentence in the methods saying which model was used. The sentence is often the least examined part of the paper, and it is one of the places where reviewers and statistical editors most often send authors back. The question behind it is simple. Do you believe that every included study is estimating one common effect, or do you believe the studies are estimating different effects that happen to cluster around an average?

Those two beliefs lead to different models, different confidence intervals and sometimes different conclusions. The choice is not a cosmetic one. It is also not a choice that a p-value for heterogeneity can make for you, despite how often authors reach for one. This guide walks through what each model assumes, why software defaults deserve a second look, how the Hartung-Knapp adjustment and prediction intervals fit in, and what to write so that the choice reads as a deliberate decision rather than a habit.

What the Two Models Actually Assume

The fixed-effect model assumes that there is one true effect underlying every included study. Differences between study estimates are attributed to sampling error alone. If a trial in one hospital found a hazard ratio of 0.82 and a trial in another found 0.64, the fixed-effect model treats the gap as noise around a single number. Each study is weighted mainly by its precision, so the largest trials dominate the pooled estimate, and the confidence interval reflects only within-study sampling variation.

The random-effects model assumes instead that the true effects themselves vary from study to study. Those effects are treated as draws from a distribution, usually a normal distribution with some between-study variance, often written tau squared. The pooled estimate then describes the average of that distribution. Smaller studies receive relatively more weight than they would under a fixed-effect model, and the confidence interval is wider because it accounts for two sources of uncertainty: sampling error within studies and variation between them.

The two models answer different questions. A fixed-effect pooled estimate says what the common effect most likely is, assuming the studies are functionally the same experiment. A random-effects pooled estimate says what the average effect is across the population of settings the studies represent. For a clinician deciding whether a drug helps patients in general, the second question is usually closer to what matters, because the included trials rarely used identical populations, doses, comparators or outcome definitions.

Why the Default Is Not a Neutral Choice

Many authors inherit their model from a previous paper, a supervising statistician or the default setting in a statistical package. That is understandable, but it is worth knowing where the defaults came from. For many years the most widely used random-effects estimator in systematic reviews was the DerSimonian and Laird method, which is simple and fast and was the standard in older versions of common review software. It remains in wide use. Its known weakness is that it can understate uncertainty when there are few studies.

Review software has been catching up. A Cochrane training presentation from January 2025 describes new random-effects options in RevMan, including restricted maximum likelihood (REML) estimation and the Hartung-Knapp-Sidik-Jonkman adjustment, which had previously been available mainly in dedicated statistical code. Authors using RevMan may now see these options listed where they once saw only the older methods. The practical lesson is to check the options in whatever tool you use, state the estimator you chose, and avoid letting the default make the decision silently.

The Cochrane Handbook, in its chapter on meta-analysis, treats the choice as a judgment that should follow from the review question and the diversity of the included studies. Secondary guides and training materials make the same point in slightly different words. The choice is a methodological decision you make and defend, not a box to tick.

A Heterogeneity Test Will Not Decide It for You

A common shortcut goes like this: run the heterogeneity test, and if the p-value is below 0.10 or 0.05, use random effects; otherwise use fixed effect. This approach has been criticized for good reason. Cochran's Q has low power when there are few studies, so a small, genuinely heterogeneous review can produce a non-significant test and get the fixed-effect model it should not have. With many large studies, the same test can flag trivial differences as significant. Neither outcome tells you whether the studies are conceptually similar enough to share a common effect.

The I squared statistic, which describes the proportion of variability attributable to between-study differences rather than chance, is more informative than a p-value, but it still does not settle the model question. It is a descriptive summary, and its uncertainty is large when there are few studies. A reviewer will accept an I squared value in your results. A reviewer will be less satisfied if the I squared value is the only justification in your methods.

The more defensible route is to decide in advance, on substantive grounds. Are the trials run in different populations, care settings or eras? Did they use different doses, comparators, follow-up periods or outcome definitions? Would a clinician expect the effect to differ between them? If the answer to these questions is yes, a random-effects model is usually the honest choice, even if the heterogeneity statistics look modest. If the included studies are close replications of one protocol, the fixed-effect model may be defensible, and the review should say why it expects them to be effectively the same.

The Hartung-Knapp Adjustment and Its Limits

The standard random-effects confidence interval can be too narrow when the number of studies is small, which is common in medical meta-analyses of rare outcomes or specialist interventions. The Hartung-Knapp adjustment, sometimes written Hartung-Knapp-Sidik-Jonkman, changes how the variance of the pooled estimate is calculated and generally produces wider and more realistic intervals in those settings. In a 2017 methodological paper, Jackson and colleagues concluded that the modified method may be a suitable replacement for the conventional random-effects approach, and suggested running a standard random-effects analysis alongside it as a sensitivity check.

Editors sometimes ask whether an adjusted interval changes the conclusion. If it does, that is important information and belongs in the discussion, not buried in a supplement. A result whose interval crosses the null under one estimator and not under another is a finding about the evidence, and a careful reader will want to see it.

Prediction Intervals: The Number Clinicians Actually Need

A random-effects confidence interval describes uncertainty about the average effect. It does not describe where the effect in a new setting is likely to fall. A prediction interval does that. It estimates the range within which the true effect in a future study, or a different population, would plausibly lie, given the between-study variation observed in your data.

This distinction matters more than it first appears. A pooled estimate can be statistically significant while the prediction interval spans no effect at all, which means that a new study in a similar setting could plausibly show benefit, harm or nothing. A 2022 reanalysis of published meta-analyses by Siemens and colleagues found that applying Hartung-Knapp adjustments and 95 percent prediction intervals altered the interpretation of a notable share of significant results. The exact proportion depends on the sample, and the study should be read on its own terms, but the lesson for authors is straightforward. When heterogeneity is present, a prediction interval is worth reporting, and it is worth interpreting honestly.

Prediction intervals are not always estimable. They need enough studies to estimate between-study variance with any stability, and many reviews do not meet that threshold. When you cannot compute one, say so, and do not imply that the confidence interval is the prediction interval.

What to Write in the Methods Section

PRISMA 2020 item 13 asks authors to describe the methods used to synthesize results, and to give a rationale for the choice of synthesis method. Item 13 is split into lettered sub-items covering eligibility, data preparation, the synthesis method and model, heterogeneity exploration and sensitivity analyses. Reviewers read these sub-items closely, and most of the questions they raise about model choice trace back to a sentence that is either missing or vague.

Every element of that paragraph answers a question a reviewer is likely to ask. It states the model and the reason for it, it says when the decision was made, it names the estimator and the adjustment, it reports the sensitivity analysis, and it names the software and version so that another team can reproduce the result. Replace the example details with your own, and do not copy the sentence if your design differs.

If you ran both fixed-effect and random-effects analyses, present the random-effects estimate as primary when the studies are clinically diverse and state that the fixed-effect estimate was a comparison. Presenting two sets of numbers without saying which one carries the conclusion invites a reviewer to choose the more convenient one for you.

Common Reporting Mistakes

The first mistake is switching models after seeing the results. If a fixed-effect analysis gives a significant pooled estimate and the random-effects version does not, the temptation to keep the first one is strong, and reviewers have become alert to it. The safer course is to report both, explain the difference, and make the conclusion explicitly depend on the interval you can justify in advance.

The second mistake is describing a random-effects model without naming the variance estimator or adjustment. "Random-effects model" covers several distinct computations, and two papers that both say it may not agree on their intervals. The third is interpreting a non-significant heterogeneity test as proof that the studies are homogeneous. A non-significant result in a small review means the test could not detect heterogeneity, which is a different statement.

The fourth mistake is leaving the model choice out of the abstract and the discussion. A reader who sees only the pooled estimate will assume a model that may not match the paper's claims. If the conclusion depends on the random-effects interval or on a prediction interval that crosses the null, the abstract should say so.

Where This Leaves Authors

The model you report is a statement about what your studies are estimating, and that statement has to survive scrutiny. Authors who choose it on substantive grounds, record the choice before extracting data, name the estimator and adjustment, and report a prediction interval when heterogeneity is present rarely get stuck in review over the model. Authors who choose it after looking at the p-values usually do.

Some uncertainty remains, and it is worth saying so. Guidance on estimators, adjustments and prediction intervals has been moving, and software options are changing faster than most journal author instructions. Before you finalize the methods, check the current version of the reporting guideline your target journal names, and check whether your review software offers the estimator you intend to describe. If the software does not offer it, say what you used instead and why.

The practical implication is simple. Write the model decision into your protocol or registration before you extract data, justify it in a single sentence that refers to the diversity of your studies, and report the interval you can defend alongside the one you think is conventional. A reader who can see your reasoning will trust your pooled estimate more than a reader who has to guess which model you used and why.

Further Reading