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

If you want to learn how to do a meta-analysis, the most important thing to internalize first is that the statistics come last. A meta-analysis is the pooling step at the end of a systematic review, and its credibility is inherited entirely from the review process underneath it: the question, the search, the screening, the extraction, and the bias assessment. Pool sloppily collected studies and the forest plot will be precise, confident, and wrong.

This guide walks the seven steps in execution order. It reflects how reviews are reported under PRISMA 2020 [1] and how the Cochrane Handbook frames the analysis decisions [2] - and it is the same order the Arakis pipeline executes them, with a human governing each judgment call.

What you need before starting

Three things, none of them statistical. First, a team: the reporting standard assumes screening and selection are performed by more than one person, and several judgment calls in this guide explicitly require a second reviewer [1]. Second, access: a meta-analysis lives on full-text articles, so know before you begin how you will retrieve papers behind paywalls - institutional access, interlibrary loan, or author contact - because abstract-only extraction is an evidence policy of last resort, not a convenience. Third, tooling for bookkeeping: a reference manager for deduplication, a screening log that records every decision with its reason, and an extraction workspace you can audit later. The recurring failure mode in review after review is not a wrong formula; it is a record that silently disappears between stages with nobody able to say why.

Step 1: Frame the question and register a protocol

Write the question in PICO form - population, intervention, comparator, outcomes - and be concrete about each slot: which doses or variants count as the intervention, which outcome definitions and timepoints count as the outcome. Vague PICO slots become screening disagreements and extraction ambiguity three steps later.

Then write the protocol before touching the data: eligibility criteria, planned outcomes, planned model and subgroup analyses. Register it prospectively on PROSPERO so the plan is timestamped and deviations are visible [3]. PRISMA 2020 expects the report to say where the protocol is and what changed [1].

Step 2: Design and run the search

A meta-analysis is only as unbiased as the sample of studies feeding it, so the search must be systematic and documented. Search multiple databases rather than one, combine controlled vocabulary with free-text terms for each PICO concept, and add trial registries to catch completed-but-unpublished studies. PRISMA 2020 requires reporting the full search strategy for every source and the date each was last searched [1] - which means you must record both at execution time, not reconstruct them at write-up.

Step 3: Screen in duplicate

Screen titles and abstracts against the eligibility criteria with two independent reviewers, then repeat in duplicate at full text, recording the reason each full-text article is excluded [1]. Dual review is not a formality: single-reviewer screening trades recall for speed exactly where errors are invisible, because a wrongly excluded study simply vanishes. Disagreements go to discussion or a third reviewer, and every unresolved judgment should be adjudicated by a human before screening is declared final. The PRISMA flow diagram - records identified, screened, excluded, assessed at full text, included - is assembled from these counts [1].

Step 4: Extract data with a piloted form

Build an extraction form that captures study characteristics, arm-level numbers (events and totals for binary outcomes; means, standard deviations, and group sizes for continuous ones), and everything your planned subgroups need. Pilot it on a few studies before running it across all of them. Extract in a way you can audit later: every number in the analysis should trace to a location in the source paper, and papers whose numbers cannot be recovered should be recorded as such rather than silently dropped - a lost full text is often the difference between a poolable outcome and a narrative summary.

Step 5: Assess risk of bias

Assess each included study with a structured tool - RoB 2 for randomized trials, which walks five domains from the randomization process to selective reporting and yields a judgment per domain with supporting evidence [4]. Anchor every judgment to a quote or specific feature of the study report, not a gestalt impression. These assessments are not decoration: they feed the sensitivity analyses in the next step and the certainty rating in the last one [4].

Step 6: Pool the effects

Now, and only now, the statistics. Convert each study's result to a common effect measure - risk ratio or odds ratio for binary outcomes, mean difference or standardized mean difference for continuous ones - and pool only studies that share a compatible measure and outcome construct [2]. Choose the model from your protocol: in clinical evidence the random-effects model is usually the defensible assumption, because true effects vary across settings [2].

Quantify heterogeneity with tau-squared and I-squared, but interpret them with their uncertainty, especially in small pools [2]. Run the prespecified sensitivity analyses - excluding high-risk-of-bias studies, leave-one-out - and check funnel plot asymmetry only when the pool is large enough for the test to mean anything, which the standard guidance places at a minimum of 10 studies [2]. If fewer than two compatible studies survive, the correct output is a synthesis without meta-analysis and a recorded reason, not a forced pool.

Step 7: Grade the certainty and report

Rate the certainty of each pooled outcome with GRADE, which starts randomized evidence at high certainty and downgrades across five domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias [5]. Every certainty statement in your abstract and conclusions should trace to that profile - no profile, no certainty claim. Then report the whole review against the PRISMA 2020 checklist, including the flow diagram, per-source search strategies, and per-study risk-of-bias judgments [1].

Key takeaways

• The statistics are Step 6 of 7; credibility is built in the question, search, screening, extraction, and bias steps before them [1].
• Register the protocol prospectively on PROSPERO so deviations are visible [3].
• Screen and select in duplicate, and record a reason for every full-text exclusion [1].
• Pool only compatible effect measures, choose the model in the protocol, and report heterogeneity with its uncertainty [2].
• Rate certainty with GRADE across its five downgrading domains, and let that rating govern the strength of your conclusions [5].

FAQ

How long does a meta-analysis take?

Traditional timelines for a full systematic review run months to a year or more, dominated by screening and data extraction rather than the analysis itself. Automation compresses the mechanical stages dramatically, but the governing judgments - eligibility, bias, certainty - still require human review.

Can I skip the protocol if my question is simple?

No. The protocol is what makes the difference between testing a hypothesis and fishing for one, and prospective registration is what makes that difference verifiable [3]. Simple questions produce short protocols; they do not remove the need for one.

What software can run the pooled analysis?

Any tool that implements inverse-variance pooling correctly: R's metafor package, RevMan, Stata, or an automated pipeline such as Arakis. The model and estimator choices matter far more than the tool [2].

What if my studies are too different to pool?

Then do not pool. Report a structured synthesis without meta-analysis, state the reason pooling was infeasible, and be specific about what evidence would change that [2]. A narrative result with a recorded reason beats a meaningless average.