What is the main difference between fixed-effect and random-effects models?

A fixed-effect model assumes all studies estimate one shared true effect, so differences are due to chance alone. A random-effects model assumes the true effect varies across studies and estimates the average of that distribution, producing a wider, more conservative confidence interval.

Should I switch to a random-effects model if heterogeneity is high?

No, you should choose the model in the protocol based on your question, not switch after seeing the I-squared. High heterogeneity tells you how to interpret and explore the result; it is not a trigger to change models retrospectively.

Why does a random-effects model give a wider confidence interval?

It adds a second source of variation, the between-study variance, on top of within-study sampling error. That extra uncertainty widens the pooled confidence interval and shifts the study weights so smaller studies count for relatively more.

Which model should I use by default?

For most reviews combining independent studies run by different teams, a random-effects model is the safer default because assuming a single shared effect is hard to justify. A fixed-effect model fits best when studies are near-replicates under identical conditions.

How does each model weight the studies?

Both use inverse-variance weighting. A fixed-effect model weights each study by one over its within-study variance, so precise studies dominate. A random-effects model adds the between-study variance, tau-squared, to each within-study variance before inverting, which flattens the weights so smaller studies count for relatively more.

Does the random-effects model always give a wider interval?

Usually, but not always. Adding between-study variance widens the interval in most cases, but because random-effects flattens the study weights, it can also shift the pooled point estimate when a large study sits at the extreme of the spread. The model can therefore change the estimate itself, not just its precision.

Is fixed-effect the same as fixed effects in regression?

No. The fixed-effect (singular) meta-analysis model assumes one common true effect across studies. Fixed effects in regression is an unrelated technique for controlling unobserved group-level differences. Use the singular term carefully in your methods to avoid confusing reviewers.

Fixed-Effect vs Random-Effects Models

The choice of fixed-effect vs random-effects models is a decision about what you assume is true across your studies. A fixed-effect model assumes every study estimates one single, shared true effect, and any differences between them are sampling error alone. A random-effects model assumes the true effect itself varies from study to study, and estimates the average of that distribution. The two can give noticeably different pooled estimates, confidence intervals, and study weights, so the choice is not cosmetic.

Why the assumption changes the answer

Under a fixed-effect model, the only reason two studies disagree is chance, so the pooled estimate is a precision-weighted average and large studies dominate. Under a random-effects model, you add a second source of variation, the genuine between-study variance, often written as tau-squared. That extra variance does two things: it widens the confidence interval of the pooled estimate, and it flattens the weights so small studies count for relatively more than they would under a fixed-effect analysis. This is why a random-effects result is usually more conservative.

The weighting maths in words

Both models use inverse-variance weighting, but they define the variance differently. Under a fixed-effect model a study’s weight is one divided by its within-study variance, so a precise study with a tiny standard error carries an enormous weight. Under a random-effects model the weight becomes one divided by the sum of the within-study variance and tau-squared. Adding the same tau-squared to every study shrinks the gap between the largest and smallest weights, because the constant addition matters proportionally more to the precise studies. A worked illustration makes this vivid: suppose one trial has a within-study variance of 0.01 and another of 0.25. Under a fixed-effect model their weights are 100 and 4, a 25-fold difference. Add a tau-squared of 0.10 and the weights become roughly 9.1 and 2.9, barely a three-fold difference. The small study has gone from almost ignored to a meaningful voice, which is the whole point of acknowledging that its true effect may genuinely differ.

A subtle naming point matters for reporting. The single-effect model is properly called the fixed-effect (singular) model, because it assumes one fixed common effect; the plural “fixed effects” means something different in regression. The alternative that pools no information across studies, the common-effect versus random-effects distinction, is the axis that actually changes your interval, so use the terms precisely when you write up the methods.

What each model is really estimating

The fixed-effect target

A fixed-effect analysis answers a narrow question: what is the common effect in this exact set of conditions? It is appropriate when the studies are near-replicates, for example several sites running the identical protocol on the same population. In that situation the assumption of one shared effect is plausible, and the model gives the tightest valid interval. The trouble is that genuine clinical and methodological diversity is the norm in most reviews, which is exactly what we cover under heterogeneity in meta-analysis.

The random-effects target

A random-effects analysis answers a broader question: what is the average effect across a population of related but not identical studies? Because it acknowledges that the true effect may differ by setting, dose, or population, its pooled estimate generalises better to new contexts. The cost is a wider interval and a heavier reliance on a tau-squared estimate that is itself unstable when you have only a handful of studies. The widely used DerSimonian-Laird method is the classic estimator, though more robust alternatives exist.

Estimating tau-squared, and why it is fragile

The whole random-effects result hinges on a value you cannot observe directly and must estimate, the between-study variance. With only a few studies that estimate is noisy, and different estimators disagree. The DerSimonian-Laird method, the historical default, tends to understate the variance and yields intervals that are too narrow when studies are scarce. Restricted maximum likelihood, known as REML, is now the preferred default in much guidance because it accounts for the uncertainty in estimating the mean effect, while Paule-Mandel is a strong choice for binary outcomes. Layering the Hartung-Knapp adjustment on top widens the interval to reflect that tau-squared was itself estimated, which is especially important when you have fewer than about ten studies. Reporting which estimator you used is not a footnote; with a small pool it can be the difference between a significant and a non-significant pooled effect.

How to decide in practice

The decision should be made in the protocol, before you see the data, and it should follow from the question rather than from the I-squared statistic. A common mistake is to run a fixed-effect model, notice high heterogeneity, then switch to random-effects after the fact. Heterogeneity informs how you interpret and explore the result, not which model you were always going to use. For most reviews of independent studies conducted by different teams, the safer default is random-effects, because assuming a single shared effect is hard to justify. The reasoning sits inside the wider method described in how to do a meta-analysis.

A decision checklist

Settle the model against these questions, all answerable from the protocol rather than the data:

Are the studies near-replicates of one protocol in one population? If so, a fixed-effect model is defensible. If they are independent studies by different teams, lean random-effects.
Do you want the common effect in these exact conditions, or the average effect across a range of settings you hope to generalise to? The target of inference, not the heterogeneity, decides this.
How many studies do you have? With very few, the tau-squared estimate is unstable, so pre-specify a robust estimator and the Hartung-Knapp adjustment rather than trusting a narrow default interval.
Have you genuinely pre-specified the choice? Switching models after seeing the I-squared is the error reviewers most often catch.

Common mistakes with model choice

The most frequent error is the post-hoc switch: running fixed-effect, seeing high heterogeneity, then quietly moving to random-effects so the result reads better. The reverse is just as bad, flipping to fixed-effect because the random-effects interval crossed the line of no effect. A second mistake is assuming random-effects is always more conservative; when a large study sits at the extreme of the spread, the flatter weights can actually shift the pooled estimate and occasionally narrow it relative to expectation, so the model changes the point estimate, not just its width. A third is reporting a random-effects pooled estimate without its prediction interval, which hides how widely the true effect may vary. Pre-specification and honest reporting close all three.

What it looks like on the plot

On a forest plot, switching from fixed-effect to random-effects typically widens the pooled diamond and shifts the study weights toward the smaller trials. If a result is statistically significant under fixed-effect but not under random-effects, that fragility is worth reporting, not hiding. You can compare both models directly with our meta-analysis calculator.

Reporting the choice honestly

Whichever model you use, state it, justify it from the question, and report the between-study variance alongside the pooled estimate. If heterogeneity is high, a single pooled number may not be the headline at all; a subgroup analysis or meta-regression that explains the variation is often more informative. And before you trust either model, confirm the result survives a sensitivity analysis and that the underlying studies passed a risk of bias assessment. If the modelling and its justification are more than your team wants to carry, our meta-analysis service specifies and runs the synthesis so the model choice is defensible and clearly reported.

Fixed-Effect vs Random-Effects Models Explained