How to write a systematic review and meta-analysis: step by step

How do you write a PICO question for a systematic review?

Write the question as a PICO statement: the Population, the Intervention (or exposure), the Comparison, and the Outcomes that matter. The Cochrane Handbook treats this as the foundation of the review, because the PICO you set in the protocol is the one your eligibility criteria are built from, and the objectives must be defined in advance.

A good first question is narrow enough to finish but broad enough to find studies. For example: in adults with type 2 diabetes (P), does a structured exercise program (I) compared with usual care (C) lower HbA1c at 12 weeks or later (O)? The population, intervention, and comparator then become eligibility criteria, together with the study designs you will accept, such as randomized trials only, or trials plus cohort studies. Outcomes are usually not used to decide eligibility, because Cochrane advises that whether a study reported an outcome should rarely determine whether it is included.

Before committing, search PROSPERO and the Cochrane Library for an existing or registered review on the same question. If one was published recently, find a different angle (another population, comparator, or outcome) rather than duplicating it. Cochrane notes that every PICO element does not need equal precision, so you can define the population tightly and the outcomes more broadly, as long as you decide before you start.

When should you register a systematic review on PROSPERO?

Register on PROSPERO after your protocol is written and before data extraction begins. Since October 2019, PROSPERO has required submissions to be made before data extraction starts. Its guidance defines a review's start date as a point after the protocol is complete but before formal screening against the eligibility criteria begins, so registering before screening is the safer target.

Write the protocol first. It should state the PICO question, eligibility criteria, databases and draft search strategy, how screening and extraction will work, the risk of bias tool, planned effect measures, the meta-analysis model, how heterogeneity and subgroups will be explored, and how certainty will be rated. PRISMA-P (2015) is the reporting guideline for protocols.

PROSPERO is a free registry run by the Centre for Reviews and Dissemination at the University of York for systematic reviews with a health-related outcome. A redesigned PROSPERO launched in 2025, and records are now published automatically once every record author has checked and approved them. Put your registration number in the manuscript. If your plans change later, amend the record and explain the change in the paper, because PRISMA 2020 asks authors to describe amendments to the registered protocol.

Which databases should you search for a systematic review?

Search at least MEDLINE (for example through PubMed), Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL), plus the trial registers ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform (ICTRP). Those are the Cochrane Handbook's core sources for reviews of interventions; Embase is expected when your team has access, so check your library. Add subject databases when your topic calls for them, such as CINAHL for nursing or PsycINFO for mental health.

Build the strategy with a medical librarian or information specialist, ideally from the protocol stage. For each concept, combine controlled vocabulary (MeSH in MEDLINE, Emtree in Embase) with free-text synonyms joined by OR, then link the concepts with AND. Most strategies search the population and intervention concepts and leave outcomes out, because outcomes are often not described in titles and abstracts.

Do not restrict by language; Cochrane says language limits risk introducing bias. Have the strategy peer reviewed before you run it (the PRESS checklist is the usual tool), merge results and remove duplicates in a reference manager, and save the exact strategy and search date for every source, because PRISMA 2020 and its search extension, PRISMA-S, expect full strategies in the report. Rerun the searches if more than 12 months will pass before publication; Cochrane prefers six months or less.

How do two reviewers screen studies and extract data?

Two reviewers work independently at each decision point, then compare results and resolve disagreements by discussion or, when needed, with a third person. Cochrane accepts one person doing the first pass of titles and abstracts, although two is preferred, but final full-text eligibility decisions must be made by two people working independently.

Screening runs in two rounds. First, titles and abstracts against your eligibility criteria, keeping anything uncertain. Second, full texts, recording at least one specific reason for every excluded article, such as wrong population, wrong comparator, or wrong study design. Those counts become your PRISMA flow diagram, so log them as you go. Screening tools such as Covidence and EPPI-Reviewer, which the Cochrane Handbook describes as Cochrane-preferred tools, track decisions and conflicts for you.

For data extraction, build a form and pilot it on a few included studies before using it on all of them. Capture the design, setting, participants, intervention and comparator details, outcome definitions and time points, and the numbers the analysis needs: events and totals for binary outcomes, or means, standard deviations, and group sizes for continuous ones. Cochrane requires at least two people to extract outcome data independently and strongly encourages contacting study authors when key results are missing.

Which risk of bias tool should you use: RoB 2, ROBINS-I, or Newcastle-Ottawa?

Choose the tool by study design: RoB 2 for randomized trials, ROBINS-I for non-randomized studies of interventions, and the Newcastle-Ottawa Scale (NOS) for cohort and case-control studies in observational reviews. As with screening, Cochrane requires two people to apply the tool independently to each included study.

RoB 2 rates a specific result rather than a whole paper, across five domains: the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Signalling questions feed an algorithm that proposes low risk, some concerns, or high risk of bias, which you can override with a written justification.

ROBINS-I compares each non-randomized study with a hypothetical target trial. The version in the Cochrane Handbook covers seven domains, starting with confounding and selection of participants, with judgements of low, moderate, serious, or critical risk of bias, or no information. A draft ROBINS-I V2, last posted in November 2025, drops the domain on deviations from intended interventions and is still subject to change, so check riskofbias.info before you start. The NOS, developed by the Universities of Newcastle (Australia) and Ottawa, awards stars across selection, comparability, and exposure or outcome; it is simpler to apply than ROBINS-I but does not judge each study against a target trial. Whatever you use, carry the judgements into sensitivity analyses and your GRADE rating.

How do you do a meta-analysis: effect measures, fixed vs random effects, and I²?

Pick an effect measure that fits the outcome, choose a fixed-effect or random-effects model on clinical and methodological grounds, and then quantify heterogeneity with I². Only pool studies that are similar enough in population, intervention, comparator, and outcome that one average makes sense; otherwise, write a structured narrative synthesis instead.

For binary outcomes, Cochrane generally recommends pooling risk ratios or odds ratios and usually avoiding meta-analysis of risk differences. For continuous outcomes, use the mean difference when every study uses the same scale and the standardized mean difference when studies measure the same outcome on different scales, after checking that higher scores mean the same thing in each one. For time-to-event outcomes, extract hazard ratios.

A fixed-effect model assumes every study estimates the same true effect, so differences are due to chance. A random-effects model assumes the true effects vary between studies and estimates their average; when heterogeneity is present it gives relatively more weight to smaller studies. The Cochrane Handbook makes no universal recommendation between them but says the choice should never be based on a statistical test for heterogeneity, so prespecify it in your protocol. In Cochrane's RevMan software, random-effects analyses default to the REML estimator for between-study variance.

I² is the percentage of variability in effect estimates that is due to heterogeneity rather than chance. Cochrane's rough, overlapping guide is 0 to 40% might not be important, 30 to 60% may be moderate, 50 to 90% may be substantial, and 75 to 100% is considerable, with a warning that such thresholds can be misleading. A P value of 0.10 rather than 0.05 is sometimes used for the chi-squared test of heterogeneity. When you have a reasonable number of studies (for example, five or more), report a prediction interval, and explore heterogeneity with prespecified subgroup and sensitivity analyses.

How do you assess publication bias and certainty of evidence with GRADE?

Look for publication bias with a funnel plot and an asymmetry test, but only when the meta-analysis includes at least 10 studies, which is the Cochrane Handbook's threshold; with fewer, the tests have too little power to be useful. Asymmetry does not prove publication bias. It can also come from flawed small studies, genuine differences between small and large studies, measures such as odds ratios and standardized mean differences that correlate with their standard errors, or chance. Searching trial registers for completed but unpublished studies is a more direct check, and the ROB-ME tool (Risk Of Bias due to Missing Evidence), described in the Cochrane Handbook, structures the judgement.

Then rate the certainty of the evidence for each important outcome with GRADE, which uses four levels: high, moderate, low, and very low. A body of randomized trials starts at high and a body of non-randomized studies usually starts at low; when non-randomized studies are assessed with ROBINS-I they may start at high, but are then generally rated down for confounding and selection bias. You rate down for risk of bias, inconsistency, indirectness, imprecision, and publication bias, and you can rate up, mainly for non-randomized evidence, for a large effect, a dose-response gradient, or plausible confounding that would work against the observed effect. Present the results in a summary of findings table, which Cochrane caps at seven outcomes.

How long does a systematic review take, and how do you report it with PRISMA 2020?

Plan for more than a year from start to publication, and report the review against the PRISMA 2020 statement. In a 2017 analysis of 195 completed reviews registered in PROSPERO, the mean time from the registered start date to publication was 67.3 weeks, with a range of 6 to 186 weeks, and the reviews had a mean of 5 authors. That span runs to publication, so it includes journal submission, peer review, and revisions, which authors do not control; reaching a submission-ready manuscript takes less time. In practice, a broad question brings more search results to screen and more studies to extract, so a narrower question usually moves faster.

PRISMA 2020 replaced the 2009 statement. It has a 27-item checklist, a separate 12-item checklist for abstracts, and flow diagram templates for new and updated reviews. The flow diagram shows records identified from each source, duplicates removed, records screened and excluded, full texts assessed with reasons for exclusion, and studies included. Fill in the checklist as you write, noting where each item appears, since journals may ask for it at submission.

If you want to work through these steps with a research mentor while doing the work yourself, ProMed United's Research Accelerator is built around this same sequence.

1. Question and protocol: PICO, eligibility criteria, and an analysis plan (PRISMA-P).
2. Registration: PROSPERO before data extraction, ideally before screening.
3. Search: peer-reviewed strategies across databases and trial registers.
4. Screening: titles and abstracts, then full texts, with exclusion reasons logged.
5. Extraction and risk of bias: piloted form, two independent reviewers.
6. Analysis: effect measures, model, heterogeneity, and publication bias.
7. Write-up: GRADE certainty, PRISMA 2020 checklist and flow diagram, and a search rerun if more than 12 months have passed.