The choice between a random effects vs fixed effect meta-analysis is one of the most consequential statistical decisions you will make when synthesizing evidence. The model you select determines how study weights are calculated, how wide your confidence intervals become, how your pooled effect size is interpreted, and ultimately whether your conclusions generalize beyond the specific studies in your review. Despite its importance, the decision is frequently made by default rather than by design, researchers select whichever model their software defaults to, or whichever their supervisor used last, without interrogating the assumptions underneath.
This guide explains the fundamental difference between the two models, walks through how each computes a pooled estimate, compares the three most common tau-squared estimators (DerSimonian-Laird, REML, and Paule-Mandel), and provides a practical decision framework for choosing the right model for your meta-analysis. In our meta-analyses, we default to random-effects because clinical diversity across studies is the norm, identical populations, interventions, and outcome measurement are the exception, not the rule.
What Is the Difference Between Fixed-Effect and Random-Effects Meta-Analysis?
The difference is rooted in what you believe about the studies you are combining. A fixed-effect model assumes that every study in your meta-analysis estimates the same single true effect. Differences between observed study results are attributed entirely to sampling error, the random variation that occurs because each study enrolls a finite number of participants. Under this assumption, there is one true effect size in the population, and every study is trying to estimate that same number.
A random-effects model assumes that the true effect size varies from study to study. Each study estimates its own true effect, and those true effects are drawn from a distribution of possible true effects. The variability between study-level true effects is called between-study heterogeneity, quantified by the variance parameter tau-squared. Under this assumption, your pooled estimate represents the mean of the distribution of true effects rather than a single common effect.
This distinction is not merely philosophical. It changes the mathematics of pooling, the interpretation of the result, and the width of the confidence interval around the pooled estimate.
Under a fixed-effect model, you are asking: "What is the best estimate of the single true effect, given these data?" Under a random-effects model, you are asking: "What is the mean of the distribution of true effects across the population of studies that could have been conducted?" The Cochrane Handbook for Systematic Reviews of Interventions (Higgins et al., 2023) frames the distinction precisely this way and notes that the random-effects question is usually the more clinically relevant one.
How the Fixed-Effect Model Works
In a fixed-effect meta-analysis, each study receives a weight that is the inverse of its within-study variance. Larger studies, those with smaller standard errors, receive proportionally greater weight because they provide more precise estimates of the single true effect. The pooled effect size is a weighted average of the individual study effects, where the weights are determined solely by each study's precision.
The formula for the weight of study i under a fixed-effect model is:
w_i = 1 / v_i
where v_i is the within-study variance (the square of the standard error) of study i. The pooled estimate is the sum of (w_i multiplied by the effect size of study i) divided by the sum of all w_i.
Because all variability between studies is attributed to sampling error, the confidence interval around the pooled estimate reflects only within-study uncertainty. If studies are truly homogeneous, estimating the same true effect, this approach is statistically efficient because it maximizes the precision of the pooled estimate.
The fixed-effect model works well in narrow contexts. When you are combining results from multiple centers within a single multi-center trial, or pooling nearly identical replications of the same experiment in the same population, the assumption of a common true effect may be justified. In these cases, the fixed-effect model provides the tightest possible confidence interval and the most precise pooled estimate.
However, in most systematic reviews, studies differ in patient populations, intervention dosing and delivery, comparator conditions, outcome measurement instruments, follow-up durations, and risk of bias. The Cochrane Handbook acknowledges that some degree of clinical and methodological heterogeneity is almost always present, which makes the fixed-effect assumption difficult to justify in practice.
How the Random-Effects Model Works
A random-effects meta-analysis adds a second source of variability: between-study variance, denoted tau-squared. Each study's weight now incorporates both within-study variance and between-study variance:
w_i = 1 / (v_i + tau-squared)
This adjustment has a profound effect on how studies are weighted. In a fixed-effect model, a very large study might receive 60-70% of the total weight. In a random-effects model, adding tau-squared to the denominator compresses the weights, large studies still receive more weight, but the gap between large and small studies narrows. Small studies gain relative influence because the between-study variance component is the same for all studies regardless of size.
The pooled estimate under random-effects represents the estimated mean of the distribution of true effects. The confidence interval reflects uncertainty about this mean, incorporating both within-study sampling error and between-study heterogeneity. As a result, random-effects confidence intervals are wider than fixed-effect confidence intervals whenever tau-squared is greater than zero.
DerSimonian-Laird (DL) Estimator
The DerSimonian-Laird method (DerSimonian and Laird, 1986) is the most widely used approach for estimating tau-squared. It uses a method-of-moments estimator derived from the Cochran Q statistic. DL is computationally simple, it requires only one pass through the data, and is the default in many software packages including RevMan, Comprehensive Meta-Analysis, and the meta package in R.
The DL estimator computes tau-squared as:
tau-squared = max(0, (Q - df) / C)
where Q is the Cochran heterogeneity statistic, df is the number of studies minus one, and C is a constant derived from the fixed-effect weights. If Q is less than or equal to df, tau-squared is set to zero.
The main limitation of DerSimonian-Laird is that it can underestimate tau-squared, particularly when the number of studies is small (Veroniki et al., 2016). Because it is a moment-based estimator, it does not produce confidence intervals for tau-squared itself, and it can yield a point estimate of zero even when true heterogeneity exists. This underestimation means the random-effects confidence interval may be too narrow, partially defeating the purpose of using a random-effects model.
Restricted Maximum Likelihood (REML) Estimator
REML is an iterative likelihood-based method that provides less biased estimates of tau-squared compared to DerSimonian-Laird, especially with fewer than 15-20 studies. REML accounts for the fact that estimating the mean effect size "uses up" one degree of freedom, which the unrestricted maximum likelihood (ML) estimator does not. This correction reduces downward bias in the variance estimate.
REML is the default estimator in the metafor package in R and is increasingly recommended by methodological guidelines. Its main disadvantage is computational, it requires iterative optimization, but in practice this is negligible with modern software. Veroniki et al. (2016) conducted a comprehensive simulation study comparing 16 tau-squared estimators and found that REML performed well across a range of scenarios, particularly when the number of studies was small.
Paule-Mandel (PM) Estimator
The Paule-Mandel estimator is an iterative generalized method-of-moments estimator that has gained attention as a robust alternative to both DL and REML. It solves for the value of tau-squared such that the generalized Q statistic equals its expected value under the random-effects model. Simulation studies have shown that PM produces confidence intervals with close to nominal coverage, even with as few as five studies.
PM is less widely implemented in software than DL or REML, but it is available in the metafor package and is referenced as a recommended estimator in recent Cochrane methods guidance. For researchers who want a balance between statistical rigor and robustness, Paule-Mandel is worth considering.
Comparison Table: Fixed-Effect vs Random-Effects Meta-Analysis
| Feature | Fixed-Effect Model | Random-Effects Model |
|---|---|---|
| Core assumption | One true effect across all studies | True effect varies across studies |
| Sources of variance | Within-study only | Within-study + between-study (tau-squared) |
| Weighting | Inverse of within-study variance | Inverse of (within-study variance + tau-squared) |
| Weight distribution | Large studies dominate | Weights more balanced across studies |
| Confidence interval | Narrower | Wider (unless tau-squared = 0) |
| Interpretation of pooled estimate | Estimate of the single common effect | Estimate of the mean of the distribution of true effects |
| Generalizability | Limited to studies included | Generalizes to population of similar studies |
| When I-squared = 0% | Standard result | Collapses to fixed-effect result |
| Common estimator | Inverse variance (Mantel-Haenszel for binary) | DerSimonian-Laird, REML, Paule-Mandel |
| Cochrane recommendation | Use when studies are truly identical | Default for most clinical meta-analyses |
| Tau-Squared Estimator | Method | Strengths | Weaknesses |
|---|---|---|---|
| DerSimonian-Laird | Method of moments | Simple, fast, widely available | Underestimates tau-squared with few studies |
| REML | Iterative likelihood | Less biased, good small-sample properties | Requires iteration, less intuitive |
| Paule-Mandel | Iterative generalized moments | Robust, good CI coverage | Less widely implemented |
How Model Choice Affects Results: A Worked Example
Consider a meta-analysis of seven randomized controlled trials examining the effect of an intervention on systolic blood pressure. Suppose the observed mean differences (mmHg) and standard errors are as follows:
| Study | Mean Difference | Standard Error | Fixed-Effect Weight (%) | Random-Effects Weight (%) |
|---|---|---|---|---|
| Study A | -8.2 | 1.1 | 35.7 | 19.8 |
| Study B | -5.1 | 2.0 | 10.8 | 16.2 |
| Study C | -6.9 | 1.5 | 19.2 | 18.3 |
| Study D | -3.4 | 2.5 | 6.9 | 14.1 |
| Study E | -7.5 | 1.3 | 25.6 | 19.1 |
| Study F | -4.0 | 3.0 | 4.8 | 12.0 |
| Study G | -2.1 | 4.5 | 2.1 | 7.8 |
Under a fixed-effect model, Study A receives 35.7% of the weight because it has the smallest standard error. The pooled mean difference is approximately -6.8 mmHg (95% CI: -7.9 to -5.7). Notice that the three largest studies (A, C, E) together account for over 80% of the weight.
Under a random-effects model with DerSimonian-Laird estimation and a tau-squared of approximately 2.1, the weights become more evenly distributed. Study A still receives the most weight but only 19.8% instead of 35.7%. The pooled mean difference shifts to approximately -5.6 mmHg (95% CI: -7.4 to -3.8). The point estimate has moved toward the results of the smaller, less precise studies, and the confidence interval has widened from a width of 2.2 mmHg to 3.6 mmHg.
In this example, both models yield statistically significant results (neither confidence interval crosses zero). But the random-effects estimate is more conservative, smaller in magnitude and less precise. In some meta-analyses, this difference is enough to change a statistically significant result to a non-significant one, which is precisely why model choice matters.
The forest plot under each model tells a visual story as well. A fixed-effect forest plot shows the diamond (pooled estimate) pulled toward the large-sample studies. A random-effects forest plot shows a wider diamond positioned closer to the center of all the individual study estimates. For guidance on reading and creating forest plots, see our step-by-step meta-analysis guide.
When to Use Each Model
The decision between fixed-effect and random-effects should be pre-specified in your protocol and grounded in your understanding of the studies you are combining. Here is a decision framework.
Use a fixed-effect model when all studies share the same population, intervention, comparator, outcome, and design, and you believe the only reason for differences between study results is sampling variability. This is most common in multi-center trials where each site follows the same protocol, or in highly controlled laboratory replications.
Use a random-effects model when you expect clinical or methodological differences between studies. This includes differences in patient demographics, intervention dosing, comparator types, outcome measurement instruments, follow-up periods, or risk of bias. Because nearly all systematic reviews of independent studies involve some degree of heterogeneity, random-effects is the safer default.
Use a random-effects model when your goal is to generalize beyond the specific studies in your review to a broader population of similar studies that could have been conducted. If you want your conclusion to apply to future studies in different settings, random-effects is the appropriate model because it explicitly accounts for variation across settings.
Consider a fixed-effect model as a sensitivity analysis even when random-effects is your primary model. If both models yield the same conclusion, your result is robust. If they disagree, investigate why, the discrepancy usually reveals something important about heterogeneity in your evidence base.
Avoid choosing the model based on the observed I-squared value. Selecting random-effects when I-squared is high and fixed-effect when I-squared is low constitutes data-driven model selection, which inflates type I error and undermines the pre-specified analysis plan. Your model should be chosen based on your a priori expectations about between-study variability, not on the data you observe.
Common Mistakes in Model Selection
Several recurring errors undermine the credibility of meta-analyses, and many of them relate directly to the fixed-effect versus random-effects decision.
Selecting the model post hoc based on results. Some analysts run both models and report whichever gives a more favorable (usually more significant) result. This is a form of selective reporting. Pre-specify your model in the protocol and stick with it. Report the alternative model as a sensitivity analysis, but base your primary conclusions on the pre-specified model.
Equating I-squared of 0% with no heterogeneity. An I-squared of 0% means that the observed variability between study results is no greater than what you would expect from sampling error alone. It does not prove that the true effects are identical. With fewer than ten studies, the power to detect heterogeneity is low. Cochran's Q test is similarly underpowered in small meta-analyses. The absence of statistical evidence for heterogeneity is not evidence of absence.
Ignoring tau-squared and reporting only I-squared. I-squared is a relative measure, it tells you what percentage of the observed variability is attributable to heterogeneity rather than chance. But it does not tell you the magnitude of the heterogeneity on the scale of the outcome. Tau-squared does. A meta-analysis of blood pressure reductions might have I-squared of 60%, but tau-squared of only 1.5 mmHg-squared, meaning the standard deviation of true effects across studies is about 1.2 mmHg, which may or may not be clinically important. Always report both statistics.
Using DerSimonian-Laird without acknowledging its limitations. DerSimonian-Laird is the default in most software and most published meta-analyses. But if your meta-analysis includes fewer than 15 studies, consider using REML or Paule-Mandel as your primary estimator, or at minimum as a sensitivity analysis. The Veroniki et al. (2016) simulation study provides evidence that DL's underestimation of tau-squared can lead to confidence intervals that are too narrow, a problem that is more severe with fewer studies.
Confusing the fixed-effect model with the common-effect assumption in all contexts. In network meta-analysis and multivariate meta-analysis, the term "fixed effects" can refer to study-level fixed effects (treating each study as a separate parameter) rather than a common-effect assumption. Make sure you understand which "fixed" you are discussing.
Failing to consider prediction intervals. Even under a random-effects model, the confidence interval around the pooled mean tells you only about the uncertainty in the mean of the effect distribution. A prediction interval tells you where you would expect the true effect of a new study to fall. Prediction intervals are almost always wider than confidence intervals and provide a more honest assessment of the range of plausible treatment effects. If you want to know whether the intervention will be effective in the next clinical setting, the prediction interval is more informative than the confidence interval.
Not reporting a sensitivity analysis with the alternative model. Reviewers and editors increasingly expect to see both models reported. If your pre-specified model is random-effects, include a fixed-effect sensitivity analysis in your supplementary materials. This costs nothing and substantially strengthens the manuscript's credibility.
Understanding the distinction between random-effects and fixed-effect meta-analysis is fundamental to producing trustworthy evidence synthesis. The model you choose shapes your weights, your confidence intervals, your point estimate, and the scope of your conclusions. By grounding the decision in a priori assumptions about between-study variability, and by reporting sensitivity analyses with the alternative model, you produce meta-analyses that withstand methodological scrutiny and provide the evidence base that clinical decision-makers need. For a comprehensive walkthrough of the entire meta-analysis process, including model selection within the broader workflow, see our step-by-step meta-analysis guide. And for deeper coverage of heterogeneity assessment, including subgroup analysis and meta-regression, see our I-squared and heterogeneity guide.