Home / Guides / How to do a meta-analysis

How to Do a Meta-Analysis: A Step-by-Step Guide

By Bara M. Hammadeh, MD'28 · Updated June 2026 · 10 min read

A meta-analysis is a research method that statistically combines the results of multiple studies asking the same question, producing one pooled estimate that is more precise than any single study on its own. You do it in a fixed sequence: frame a focused question, register a protocol, search the literature systematically, screen the results, extract the data, pool the numbers, test how robust that pooled number is, and interpret it honestly. This guide walks through each stage so you can see the whole path before you start.

One clarification first, because it trips up almost everyone new to this.

Systematic review vs meta-analysis: which is which

A systematic review is the entire process of finding, appraising, and summarising every study relevant to a question using a transparent, pre-specified method. A meta-analysis is one optional step at the end of that process: the statistical pooling of the studies' numerical results into a single combined effect. Every meta-analysis lives inside a systematic review. A systematic review without poolable data simply stops before the pooling step and reports its findings narratively. So when people say "I am doing a meta-analysis," they almost always mean "a systematic review that ends in a pooled estimate."

The stages, start to finish

Here is the full sequence. Each stage feeds the next, and skipping ahead is the most common reason first projects stall or get rejected.

1. Frame a focused question (PICO)

A poolable question is narrow and structured. The standard scaffold is PICO: Population, Intervention, Comparison, Outcome. "Does this drug help patients" is too vague to pool. "In adults with condition X (Population), does drug A (Intervention) versus placebo (Comparison) reduce 30-day mortality (Outcome)" is a question you can actually search for and combine studies around. A sharp PICO question is what makes every later stage feasible, so it is worth real time.

2. Write and register a protocol

Before you collect anything, you write a protocol: your eligibility criteria, your search plan, your outcomes, and your planned analysis. You then register it publicly, usually on PROSPERO, so the world can see what you committed to before you saw the results. This is not bureaucracy. A registered protocol is what protects your review from the accusation that you changed the rules once you saw which answer you liked. It is also the single biggest time-saver, because every later decision is already written down.

3. Search the literature systematically

A meta-analysis search is meant to be reproducible and as complete as practical. You translate your PICO terms into a structured search string with Boolean operators and controlled vocabulary (such as MeSH terms in PubMed), then run it across several databases, not just one. The goal is to find every eligible study, not a convenient sample, so you document the exact strings and the date you ran them.

4. Screen the results (PRISMA)

Your search will return far more records than you need. Screening removes the irrelevant ones in two passes: first by title and abstract, then by full text for the survivors. To reduce error, two reviewers screen independently and resolve disagreements together. You log how many records were excluded and why, which becomes your PRISMA flow diagram, the standard picture of how you got from thousands of hits to a handful of included studies.

5. Extract the data and assess risk of bias

From each included study you pull the numbers you need to pool: sample sizes, event counts, means, standard deviations, or effect estimates. You also judge how trustworthy each study is using a risk-of-bias tool appropriate to its design. This matters because a pooled estimate is only as honest as the studies feeding it, and a meta-analysis of biased studies produces a precise but misleading number.

6. Choose your effect measure

You cannot pool studies until they speak the same numerical language. For binary outcomes (event happened or not) that is usually an odds ratio or risk ratio. For continuous outcomes (a score, a level) it is a mean difference, or a standardised mean difference when studies used different scales. Picking the right effect measure for your outcome type is the bridge between your extracted data and the pooled analysis.

7. Pool the studies (and read heterogeneity)

Now you combine the studies. Each study is weighted, usually by how precise it is, and the weighted average becomes your pooled effect, displayed as a forest plot. Two decisions shape this step: whether to use a fixed-effect or a random-effects model, and how to handle heterogeneity, the degree to which the studies disagree with each other (commonly summarised with the I-squared statistic). Heterogeneity is not a failure; it is information about whether these studies are even measuring the same underlying effect.

8. Test how robust the result is

A single pooled number is not the finish line. You probe it: sensitivity analyses (does the result hold if you drop one study at a time?), subgroup analyses (does the effect differ by population or dose?), and checks for publication bias (are small negative studies missing?), often using a funnel plot and Egger's test. These checks are what separate a credible review from a fragile one.

9. Interpret and write it up

Finally you translate the statistics into a defensible conclusion, rating your overall certainty in the evidence (GRADE is the common framework) and writing the manuscript in the standard structure. Interpretation is where statistical significance and clinical meaning have to be separated honestly: a tiny effect can be statistically significant and still not matter to a patient.

The mistakes that sink first projects

• Too broad a question. If your PICO is vague, nothing downstream works. Narrow ruthlessly.
• Skipping the protocol. Deciding your methods after seeing the data is the fastest route to reviewer rejection.
• Searching one database. Relying on PubMed alone misses eligible studies and weakens your claim of completeness.
• Ignoring heterogeneity. Pooling wildly different studies into one number hides more than it reveals.
• Confusing significance with importance. A narrow confidence interval around a trivial effect is still a trivial effect.

Frequently asked questions

What is the difference between a systematic review and a meta-analysis?

A systematic review is the full process of finding and appraising all the studies on a question using a pre-set method. A meta-analysis is the optional statistical step at the end that pools the numerical results of those studies into a single combined estimate. Every meta-analysis sits inside a systematic review, but not every systematic review includes a meta-analysis.

Do I need to know statistics to do a meta-analysis?

You need to understand the concepts: what an effect measure is, the difference between fixed and random effects, and how to read heterogeneity. You do not need to perform the calculations by hand. Software does the math once you understand what to ask it for.

How long does a meta-analysis take?

A focused first project typically takes three to six months of part-time work. Searching and screening are usually the slowest stages. Registering a clear protocol up front is the single biggest time-saver, because it prevents rework later.

Can I do a meta-analysis on my own?

You can lead one on your own, but screening and data extraction should be done by at least two people independently to reduce error and bias. Many first-time authors pair up with a colleague for those two stages and work solo on the rest.

Keep learning

• Browse all meta-analysis guides
• The full video course (Module 0 free)
• Run your analysis with the free tool