A fixed-effect meta-analysis (also called the common-effect model) assumes every included study estimates one single, shared true effect, and that the only reason the observed results differ is within-study sampling error. A random-effects meta-analysis instead assumes the true effect varies from study to study and that the studies sample from a distribution of true effects, so the model estimates both the average of that distribution and how widely it spreads. That single assumption, one fixed effect versus a distribution of effects, drives everything that follows: how studies are weighted, how large the pooled estimate is, and how wide its confidence interval ends up.
The assumption behind each model
Under the fixed-effect model, the question being answered is narrow and precise: what is the one true effect that all of these studies share? Differences between study results are treated as noise around that constant. Because every study is presumed to target the same underlying value, the model is willing to let large, precise studies dominate the answer.
Under the random-effects model, the question is broader: what is the average effect across a population of related-but-not-identical studies, and how much does the true effect itself move between them? Here, variation between studies is not just noise. Part of it reflects genuine differences in populations, doses, follow-up windows, and measurement, captured by the between-study variance. This is why the two models can give meaningfully different conclusions even on the exact same data set, a point worth understanding before you learn how pooling actually works.
How each model weights studies
Both models begin from inverse-variance weighting: each study is weighted by the reciprocal of its variance, so more precise studies count for more. The difference lies in what counts as the relevant variance.
Fixed-effect weights
In the fixed-effect model the weight assigned to a study is simply one divided by its within-study variance (the square of its standard error). Small studies, which have large standard errors, receive very little weight, and a single very large trial can carry most of the pooled answer. The pooled estimate is essentially a precision-weighted average that leans hard toward the biggest, tightest studies.
Random-effects weights
The random-effects model adds a second component to every study's variance: the tau-squared term, which is the estimated between-study variance. Each weight becomes one divided by the sum of the within-study variance and tau-squared. Because tau-squared is added to every denominator, it shrinks the gap between large and small studies. The practical consequence is that random effects gives more weight to smaller studies relative to the fixed-effect approach. When between-study variance is large, the weights flatten toward equality, and modest studies that the fixed-effect model would have ignored start to influence the result.
The role of the DerSimonian-Laird estimator
To use random effects you have to estimate tau-squared, and the classic method is the DerSimonian-Laird estimator. It derives tau-squared from Cochran's Q heterogeneity statistic, comparing how much variation the studies actually show against how much would be expected from sampling error alone. If observed variation exceeds the expected amount, the excess is attributed to tau-squared; if it does not, tau-squared is set to zero and the random-effects result collapses back onto the fixed-effect result. DerSimonian-Laird is fast, transparent, and the historical default in most software, though it can underestimate uncertainty when the number of studies is small or heterogeneity is high. Restricted maximum likelihood is a common modern alternative, and naming the estimator you used is part of a clean methods section.
How the choice changes your results
Switching from fixed to random effects changes two visible things at once: the pooled point estimate and the width of its confidence interval.
- The pooled estimate can shift because reweighting toward smaller studies changes the average. If the smaller studies trend in a different direction than the large ones, the random-effects estimate moves toward them.
- The confidence interval gets wider under random effects, because tau-squared adds extra uncertainty to every weight. A result that looks statistically clear under a fixed-effect model can become non-significant once between-study variance is honestly carried through.
- Relative study influence flattens, which you can read directly off the weight column when you interpret a forest plot line by line.
Neither wider nor narrower is automatically "better." A narrow fixed-effect interval is honest only when the one-shared-effect assumption is genuinely plausible. When it is not, that narrow interval is overconfident, and the broader random-effects confidence interval is the more truthful summary of what the evidence supports.
How heterogeneity informs the choice
The decision is not arbitrary; it is anchored in measured heterogeneity. Two quantities matter. Tau-squared measures the absolute between-study variance on the scale of the effect, and I-squared expresses the percentage of total variation that is due to real between-study differences rather than chance. A related index, H-squared, captures the ratio of total to within-study variation. When these statistics indicate that studies disagree more than sampling error can explain, the one-true-effect premise of the fixed-effect model is hard to defend.
As a working rule, with high heterogeneity random effects is usually more defensible, because it explicitly models the spread that the data are showing you. With near-zero heterogeneity the two models converge and the choice barely matters numerically, though many methodologists still report random effects so the conclusion does not depend on an assumption that could fail in an update. To build intuition for these statistics, work through what I-squared and tau-squared really tell you.
Why the prediction interval is a random-effects concept
A confidence interval describes uncertainty about the average effect. It does not tell a reader what effect to expect in a new, future setting. That forward-looking question is answered by a prediction interval, which incorporates tau-squared directly and therefore exists only inside the random-effects framework; a fixed-effect model, assuming no between-study variation, has nothing to predict a distribution from. When between-study variance is substantial, the prediction interval is often dramatically wider than the confidence interval and can even cross the line of no effect, a sobering and honest disclosure. If your synthesis carries real heterogeneity, plan to report a prediction interval alongside the pooled estimate.
Practical guidance and how to justify it to reviewers
Decide the model from the clinical and methodological reasoning first, not from whichever gives a prettier p-value. Ask whether it is biologically and contextually plausible that every study estimates the identical true effect. If the studies differ in population, intervention intensity, outcome definition, or follow-up, a single shared effect is rarely credible, and random effects is the principled default for most reviews of real-world literature.
To satisfy reviewers, do the following in your methods and results:
- State the chosen model and the reason, tied to expected clinical and methodological diversity, not to the result it produced.
- Name the tau-squared estimator, for example DerSimonian-Laird or restricted maximum likelihood, so the analysis is reproducible.
- Report I-squared, tau-squared, and the Cochran Q test so readers can judge the heterogeneity for themselves.
- Pre-specify the model in the protocol, which removes any suspicion that it was chosen after seeing the data.
- Consider presenting both models as a sensitivity analysis when the conclusion is close to a threshold, and show the prediction interval when heterogeneity is high.
Once the model is chosen and justified, the mechanics are quick. You can pool your data and see the weights, pooled estimate, and confidence interval update live in our forest plot builder, then compare model assumptions against your heterogeneity output before you commit the result to the manuscript.