Introduction
You’ve pulled five trials on a new blood pressure drug. The results don’t match. One reports a big drop, another reports almost nothing, and the other three land somewhere in between.
So what do you do with that?
Before you pool anything, you have to answer one question. Were all five studies chasing the same true effect and just missing it by a little? Or does the drug honestly perform differently depending on who takes it and where?
Your answer picks your model for you.
Meta-analysis offers two. A fixed-effect model says one true effect exists behind every study. A random-effects model says the true effect shifts from study to study.
In this guide, you’ll learn the difference between fixed-effect and random-effects meta-analysis, understand heterogeneity, and know exactly when to use each model.
What is a Fixed-Effect Meta-Analysis?
The fixed-effect model starts from a bold assumption: one true effect size exists, and every study in your review was aiming at it.
Then why do the numbers disagree? Sampling error, says the model. Nothing more. Each trial enrolled a limited number of patients, and that limit alone produces scatter around the real value.
Five thermometers, one patient. The first reads 98.4°F, the second 98.8°F, the third somewhere between. The patient has one temperature. The spread lives in the thermometers.
Since every difference gets blamed on noise, precision decides who gets heard. The model applies inverse variance weighting, variance and confidence intervals all sit underneath this, medical statistics for beginners covers the groundwork if any of it feels shaky.
That has a real consequence. Large trials steer the pooled effect estimate. Put a 4,000-patient study next to a 40-patient study and the small one barely registers.
What is a Random-Effects Meta-Analysis?
The random-effects model refuses that assumption. Different studies, it says, may well be measuring different true effects.
Anyone who has read a set of trials closely will find this familiar. Recruitment differs. Doses differ. Follow-up runs six weeks in one trial and two years in another. A drug tested on healthy 30-year-olds and a drug tested on 75-year-olds with stage 4 CKD are not really being tested under the same conditions.
So the model splits the variation in two.
Within-study variation is the ordinary sampling error inside each trial, the same thing the fixed-effect model already handles. Between-study variation is the new piece. It captures how much the true effects genuinely differ from one study to the next, and it goes by the name between-study variance, or tau squared (τ²).
Five hospitals run the same program. Different neighborhoods, different patients, different staffing. The program helps in all five, but nowhere near equally.
There’s no single number hiding underneath. The pooled effect estimate here is an average across a range of true effects, not one fixed value you’re closing in on.
Fixed Effect vs Random Effects Meta-Analysis Example
Numbers make this obvious faster than definitions do.
Example 1: Consistent results
| Study | Risk Ratio |
| Study A | 0.80 |
| Study B | 0.82 |
| Study C | 0.79 |
| Study D | 0.81 |
Four risk ratios, all sitting within 0.03 of each other. That’s about as much wobble as chance would give you anyway.
Now go back and read the methods. Comparable patients, same dose, same follow-up window. Nothing in there hints that the drug behaves one way in Study A and another way in Study D.
Fixed-effect works fine.
Example 2: Scattered results
| Study | Risk Ratio |
| Study A | 0.60 |
| Study B | 0.82 |
| Study C | 1.10 |
| Study D | 0.70 |
Different picture entirely. Study A found a 40% reduction in risk. Study C found nothing.
And the methods explain why. Four countries. Age ranges that barely overlap. Two dosing schedules. Baseline risk that isn’t remotely comparable across the four populations.
When the spread has reasons behind it, the true effect probably does vary. Random-effects is the honest call.
That’s the whole logic of any fixed effects vs random effects example. Read the forest plot, then read the trials that produced it.
Meta-Analysis
Reading a forest plot is one skill. Producing one is another.
Effect measures, weighting, model choice, heterogeneity, publication bias, twelve weeks of working through real datasets until the output is something you can defend in front of a reviewer.
Understanding Heterogeneity in Meta-Analysis
Heterogeneity is what you call it when studies disagree by more than chance can account for.
A little scatter is normal and expected. The real question is whether the scatter has outgrown sampling error. Two tools help you decide.
Cochran’s Q
Q is a statistical test. It measures each study against the pooled estimate and asks whether the gaps are bigger than they should be. Get a small p-value and you have evidence of heterogeneity.
Worth knowing, though: Q struggles when you only have four or five studies. It simply lacks the power to detect much. A non-significant Q is not proof that your trials agree.
The I² statistic
I² asks something else. Of all the variation you can see, what share reflects real differences between studies rather than chance?
| I² | Interpretation |
| 0–25% | Low |
| 25–50% | Moderate |
| 50–75% | Substantial |
| >75% | High |
Climbing I² values generally push you toward a random-effects model.
What Is Tau Squared (τ²)?
Tau squared (τ²) estimates how far apart the true effect sizes actually are. It’s the between-study variance, reported on whatever scale you’re measuring your effect on.
Students mix up τ² and I² constantly, so here’s the split:
I² answers “is there meaningful heterogeneity here?” and gives you a percentage.
τ² answers “how much?” and gives you an actual quantity.
When τ² comes out at zero, your studies look like they share one true effect. As it climbs, those true effects sit further and further apart, and sticking with a fixed-effect model gets harder to defend.
Biostatistics
Q, I² and tau squared are three answers to three different questions.
Confusing them is the most common reason a methods section gets queried. Variance, weighting, confidence intervals and effect measures, the foundation everything above is built on, taught from the ground up.
How Does the DerSimonian-Laird Method Work?
DerSimonian-Laird is the workhorse of random-effects meta-analysis. Open RevMan, tick the random-effects box, and this is usually what runs behind the scenes. Many people use it without ever noticing the name.
Two steps, essentially.
It looks at how much the study results actually spread out, and from that spread it estimates the between-study variance, tau squared (τ²).
Then it takes that estimate back into the weighting. Every study’s weight now reflects two things: its own precision, plus the variability across the whole set.
The formula isn’t the point. Knowing what’s being adjusted for is.
How Are Studies Weighted?
Weighting is just the model deciding whose results matter most.
Fixed-effect
Inverse variance weighting, and nothing else. Precision rules. One large, tightly estimated trial can end up carrying most of the pooled result while three small ones sit in the margins contributing next to nothing.
Random-effects
Now τ² gets added into every study’s weight. Because that addition is identical across studies, it compresses the gaps between them.
Weights even out. Small studies pick up influence they didn’t have before, and the big trials give some of theirs back.
Same data, two models, two different pooled estimates. This is where that difference comes from.
How Do These Models Affect the Forest Plot?
If forest plots are new, the diamond, the whiskers and the weight column all carry specific meaning. How to write a systematic review and meta-analysis covers reading one before you build one.
You can spot all of this on the forest plot without doing any math.
Run fixed-effect and the diamond at the bottom comes out narrow. Tight confidence interval. Glance at the weight column and a couple of large trials will be holding most of it.
Switch to random-effects and the diamond widens. The confidence interval stretches because the model is now paying for between-study variation on top of within-study error. Weights look more even, and the pooled estimate can drift toward the smaller trials.
Wider isn’t worse. It’s more truthful.
Fixed Effect vs Random Effects: Which Should You Use?
Start here.
| Use Fixed Effect | Use Random Effects |
| Similar populations | Different populations |
| Same protocol | Different protocols |
| Minimal heterogeneity | Moderate or high heterogeneity |
| Low I² | Higher I² |
| One common effect | Multiple true effects |
Now a warning that people ignore too often.
Don’t let I² alone make this decision for you. With a handful of studies it bounces around unreliably, and with very large studies it can look alarming when the actual differences are trivial.
Read the trials first. Were the patients comparable? Same intervention, same outcome definition, same follow-up? If the answer is no, your true effects almost certainly vary, and no statistic is going to talk you out of that.
One more practical note. Reviewers expect your model choice to appear in the protocol, written down before you’ve seen a single result.
That’s not a soft expectation, PROSPERO asks for your synthesis method directly. Registering a systematic review on PROSPERO covers what goes in that field and when.
It also has to appear in your write-up. Item 13 of the PRISMA 2020 checklist asks how you pooled results, or why you chose not to.
Conclusion
Two models, two beliefs. Fixed-effect says a single true effect is out there. Random-effects says the true effects vary and you’re averaging across them.
Heterogeneity is the dividing line. Cochran’s Q, the I² statistic, and tau squared (τ²) each describe some part of how much your studies disagree, and DerSimonian-Laird is the standard route to estimating that between-study variance.
Skip the question of which model is “better.” Ask which one matches the evidence actually sitting in front of you. Get that right and your meta-analysis becomes more accurate, more transparent, and far easier to trust.
Systematic Review
The model choice happens in the protocol, months before the data arrives.
By the time you’re staring at five disagreeing risk ratios, the decision should already be written down. Learn to build a review from question to protocol to publication, with a mentor who has taken one through the whole arc.
Frequently Asked Questions:
1. What is the difference between fixed-effect and random-effects meta-analysis?
Fixed-effect assumes every study shares one true effect and blames the differences on sampling error. Random-effects assumes the true effect varies across studies and adds between-study variance to account for it.
2. When should I use a random-effects model?
Reach for it when your studies differ in patients, doses, settings, or protocols, or when heterogeneity is moderate to high. Since real trials rarely match up neatly, plenty of reviewers treat random-effects as the safer default.
3. What is heterogeneity in meta-analysis?
It means the results disagree more than chance explains. The causes are clinical (age, disease severity, dose) or methodological (design, blinding, risk of bias), and usually both.
4. What does the I² statistic tell you?
It estimates what percentage of the total variation comes from real differences between studies rather than chance. Under 25% is low. Over 75% is high.
5. What is the DerSimonian-Laird method?
The most widely used method for random-effects meta-analysis. It estimates between-study variance (τ²) from how much the results spread out, then rebalances each study’s weight to reflect it.
6. What is tau squared (τ²)?
The estimated between-study variance. It tells you how far apart the true effect sizes are. A τ² of zero suggests your studies share a common effect.
7. Does a high I² always mean I should use a random-effects model?
No. Treat it as a signal, not an instruction. I² is shaky with few studies, and the clinical and methodological differences between your trials deserve at least as much weight in the decision.
Disclaimer:
Articles published by American Academy of Research & Academics are prepared by our team using information from direct experience, publicly available resources, and educational references. AI tools may be used to assist with drafting, proofreading, and formatting; however, all content undergoes review and approval before publication.
The information provided is intended for educational purposes only. Requirements, policies, and processes may change over time. Readers should consult official sources for the most current information.