How to Write a Systematic Review and Meta-Analysis: Step-by-Step Guide

How to write a systematic review and meta-analysis: step-by-step guide covering PICO, PRISMA 2020, risk of bias, meta-analysis in R and GRADE, by Dr. Muhammad Aammar Tufail
Systematic review · Meta-analysis

How to Write a Systematic Review and Meta-Analysis: Step-by-Step Guide

A plain-English guide for BS, MS and PhD students and early researchers: choose the review type, write a PICO question, register a protocol, search, screen, extract data, judge risk of bias, run a meta-analysis in R, rate certainty with GRADE and report with PRISMA 2020.

Subscribe on YouTube
The short answer

To write a systematic review and meta-analysis, you follow a fixed, transparent method: ask one focused question, write and register a protocol, search several databases, screen with two reviewers, extract data, judge risk of bias, pool the results statistically, rate the certainty of evidence and report everything with PRISMA 2020. A systematic review is the method of finding and judging all studies on a question. A meta-analysis is the statistical step that combines their numbers into one pooled estimate. This guide walks you through every step with examples from agronomy, biology and health.

12steps from question to published review
2independent reviewers for screening and extraction
27items in the PRISMA 2020 checklist
10+studies before you test for funnel plot asymmetry
Learn the full method in Urdu/Hindi, step by stepMy Systematic Review & Meta-analysis course is a self-paced course on the Codanics learning platform. It takes you from a research question to a review you can submit to a good journal, with R, MS Excel and EndNote.
  • Protocol, searching, screening and PRISMA reporting
  • Data extraction, risk of bias and GRADE
  • Meta-analysis in R and MS Excel, with forest and funnel plots
See the course →
Key facts: systematic review and meta-analysis
What it isA systematic review finds, screens and appraises all studies on one question using a pre-planned method. A meta-analysis pools their results statistically.
Methods standardCochrane Handbook for Systematic Reviews of Interventions
Reporting standardPRISMA 2020: a 27-item checklist and a flow diagram (Page et al., BMJ 2021)
Where to registerPROSPERO for reviews with a health-related outcome; OSF Registries for any field
Free softwareR with metafor or meta, Rayyan and ASReview for screening, and free Codanics review tools
Typical time6 to 12 months for a first review with a small team

Many students are told, “Write a systematic review for your thesis.” Then they search Google Scholar, collect 40 PDFs and write a summary. That is a narrative review, not a systematic review, and journals reject it quickly. A real systematic review has rules. Once you know the rules, it is one of the best first papers you can publish: you do not need a lab, a field or a big grant, only a good question, a laptop and a careful method.

I teach this method to thesis students, MPhil and PhD scholars, lecturers and doctors. In this guide I explain each step the way I explain it in class: in simple words, with tables and real-looking examples. Wherever you see “Practise this step free”, you can open a free Codanics tool and do the step on your own topic.

Lesson · 22:19
What is a systematic review and meta-analysis?
What is a systematic review and meta-analysis?. The difference between a systematic review and a meta-analysis, explained simply.
Step 1

What is a systematic review and a meta-analysis, and how are they different from other reviews?

Quick answer

A systematic review answers one focused question by finding, screening and appraising all relevant studies with a pre-planned, reproducible method. A meta-analysis is an optional statistical step inside a systematic review: it combines the effect sizes of similar studies into one pooled estimate with a confidence interval. A narrative review is a flexible expert summary, and a scoping review maps what evidence exists without judging study quality.

Think of a farmer who asks: “Does inoculating wheat seeds with plant growth-promoting bacteria increase grain yield?” One field trial cannot answer this for everyone. Twenty trials in different countries can. A systematic review finds all twenty, checks how well each was done and summarises them fairly. A meta-analysis then calculates one average effect, for example “+0.6 t/ha”, and shows how much the trials disagree.

Review typeQuestion it answersMethodReporting guidelineTypical time
Narrative reviewWhat do experts know about a broad topic?Flexible reading, author chooses studiesJournal guidelines1–3 months
Scoping reviewWhat research exists on X, and where are the gaps?Systematic search, charting, usually no quality appraisalPRISMA-ScR3–6 months
Systematic reviewDoes X work, or is X linked to Y?Protocol, full search, dual screening, risk of biasPRISMA 20206–12 months
Meta-analysisHow big is the effect, on average?Effect sizes pooled from a systematic reviewPRISMA 2020Added to the systematic review
Umbrella reviewWhat do existing systematic reviews conclude?Systematic review of reviewsPRIOR4–8 months
Not a systematic review“We searched Google Scholar and selected relevant papers on biofertilisers.” No protocol, one database, one person, no criteria, no risk of bias. Anyone else would find different papers.
A systematic review“We searched Scopus, Web of Science and CAB Abstracts on 10 March 2026 with a registered protocol. Two reviewers screened 990 records independently against pre-set criteria and assessed risk of bias with RoB 2.”

The key word is reproducible. Another team should be able to repeat your search, apply your criteria and get almost the same set of studies. That is why every decision is written down before you start, and why the PRISMA 2020 statement asks you to report it all. If you are new to review writing in general, read my complete scientific writing guide first; it explains narrative reviews, research papers and theses.

Can you do a meta-analysis without a systematic review? Technically yes, but it is weak. If you pool only the studies you happen to know, the pooled number is biased. A meta-analysis is only as good as the systematic search behind it.
Key takeaways
  • A systematic review is a method; a meta-analysis is a statistical step inside it.
  • Narrative and scoping reviews have different goals and guidelines.
  • Reproducibility is what makes a review “systematic”.

Practise this step free: 🧭 Review Type Selector · 🔀 PRISMA Flow Diagram Generator · 🧠 Systematic Review & Meta-analysis Quiz All writing tools →

↑ Back to topNext: Review type and PICO question →

Step 2

Which review type should you choose, and how do you write a PICO question?

Quick answer

Let the question decide the review type. “Does X work?” needs a systematic review, with a meta-analysis if studies report similar numbers. “What is known about X?” needs a scoping review. Then write the question in PICO form (Population, Intervention, Comparison, Outcome), or PECO when you study an exposure, and check it with the FINER criteria: Feasible, Interesting, Novel, Ethical and Relevant.

Match the question to the review

Your questionChoose
Does seed priming with zinc increase wheat yield compared with no priming?Systematic review with meta-analysis
Is exposure to indoor smoke from biomass fuels linked to childhood pneumonia?Systematic review (observational studies), possibly meta-analysis
What research exists on microplastics in agricultural soils?Scoping review
How has thinking about soil health changed over 30 years?Narrative review
What do existing reviews conclude about probiotics for diarrhoea?Umbrella review

Write the question in PICO or PECO form

PICO turns a vague topic into a question you can search and test. Use PECO (Exposure instead of Intervention) when you cannot assign the factor, for example air pollution, diet or a soil contaminant. Add S for study design (PICOS) when the design is part of your criteria.

ElementAgronomy exampleHealth example (PECO)
PopulationWheat grown in field conditionsChildren under 5 years in low- and middle-income countries
Intervention / ExposureSeed inoculation with plant growth-promoting bacteriaHousehold use of biomass cooking fuel
ComparisonUninoculated seedsClean fuel (gas or electricity)
OutcomeGrain yield (t/ha)Acute lower respiratory infection
Study designField trials, at least one seasonCohort and case-control studies
Weak question“A review of biofertilisers in agriculture.”
Strong question“In field-grown wheat, does seed inoculation with plant growth-promoting bacteria, compared with no inoculation, increase grain yield? A systematic review and meta-analysis.”

Check it with FINER, and check that it is new

  • Feasible: enough studies exist (run a quick scoping search) and you have time and a second reviewer.
  • Interesting: farmers, doctors or policy makers care about the answer.
  • Novel: search PROSPERO, OSF, the Cochrane Library and Google Scholar for an existing or ongoing review. If one exists, change the population, outcome or time period, or plan an update.
  • Ethical: reviews of published data rarely need ethics approval, but follow your university’s rules.
  • Relevant: the answer changes practice or the next study.

Finding the question usually starts with reading. In the videos below I show how to find a research gap from review articles, and why reading five good reviews is the fastest way to learn a field.

Workshop · 2:03:29
Find a research gap and write a narrative review
Find a research gap and write a narrative review. A full workshop on finding a research gap from review articles.
Short · 0:57
Master scientific writing: read 5 review articles
Master scientific writing: read 5 review articles. Why reading five good reviews is the fastest start.
Key takeaways
  • The question decides the review type, not the other way round.
  • Write PICO or PECO, and add study design when it matters.
  • Check FINER and search registries for existing reviews before you start.

Practise this step free: 🧭 Review Type Selector · 🧩 PICO Research Question Builder · 🎯 FINER Research Question Checker · 🧠 Research Question & PICO Quiz All writing tools →

↑ Back to topNext: Write and register the protocol →

Step 3

How do you write and register a systematic review protocol?

Quick answer

A protocol is your plan, written before you search: the question, eligibility criteria, databases, search strategy, screening and extraction process, risk-of-bias tool and analysis plan. Follow the PRISMA-P checklist (17 items) and register it publicly before screening starts: on PROSPERO if your review has a health-related outcome, or on OSF Registries for any field, including agriculture and scoping reviews.

Why write a plan first? Because it stops you from changing the rules after you see the results. If you decide on the outcome, subgroups and analysis before you look at the data, nobody can say you picked the method that gave the “best” answer. Registration also tells other teams that the review is in progress, so effort is not duplicated. Changes are allowed later, but you must report them as deviations from the protocol.

What goes into the protocol (PRISMA-P)

SectionWhat to writeExample
Title and registrationName the design in the title“… : protocol for a systematic review and meta-analysis”
Rationale and objectivesThe gap and the PICO questionMixed results of bacterial inoculants on wheat yield
Eligibility criteriaPICO plus design, language, yearsField trials, at least one season, any language, 2000–2026
Information sourcesDatabases, registers, grey literatureScopus, Web of Science, CAB Abstracts, theses repositories
Search strategyFull string for at least one databaseBoolean string with synonyms (see Step 4)
Selection and extractionWho screens, how conflicts are solved, which toolTwo reviewers in Rayyan; a third resolves conflicts
Risk of biasTool for each designRoB 2 for randomised trials
SynthesisEffect measure, model, heterogeneity, subgroupsMean difference, random effects, I², subgroup by soil type
CertaintyHow you will rate the evidenceGRADE for each main outcome

Write eligibility criteria that two people apply the same way

Vague“Studies about biofertilisers and wheat were included.”
Testable“We will include field trials (at least one growing season) comparing wheat inoculated with plant growth-promoting bacteria with uninoculated controls under the same fertiliser level, reporting grain yield with a measure of variance. We will exclude pot and laboratory studies, mixed inoculant-plus-fertiliser treatments without a matching control, and reviews.”

PROSPERO or OSF?

PROSPEROOSF Registries
FieldsReviews with at least one outcome of direct relevance to human or animal healthAny field: agriculture, ecology, education, health
Scoping reviewsNot acceptedAccepted
When to registerBefore screening startsBefore screening starts (recommended)
What you getA CRD registration numberA time-stamped registration with a DOI
CostFreeFree

You can also publish the full protocol as a paper, for example in a journal that accepts review protocols. A protocol is very similar to a research proposal, so my guides on writing a PhD research proposal and a postdoc research proposal will help you write the rationale and objectives.

Lesson · 18:34
Methodology of systematic review and meta-analysis
Methodology of systematic review and meta-analysis. The full method, from the question to the pooled result.
Common mistake: registering after you have already screened or extracted data. Registries check the review stage, and editors may ask why the registration date is after the search date. Register first, then run the final search.
Key takeaways
  • Write the protocol before you search, following PRISMA-P.
  • Make eligibility criteria specific enough for two people to agree.
  • Register on PROSPERO (health outcomes) or OSF (any field) before screening.

Practise this step free: 📝 Systematic Review Protocol Builder · 🧩 PICO Research Question Builder · 🗓️ Research Gantt Chart Maker · 🧠 Systematic Review & Meta-analysis Quiz All writing tools →

↑ Back to topNext: Build the search strategy →

Step 4

How do you build a systematic search strategy?

Quick answer

Split your PICO question into two or three concepts, list synonyms and controlled terms (such as MeSH) for each, join synonyms with OR inside brackets and join concepts with AND. Run the same logic in at least two or three databases, add grey literature and citation searching, and record every database, date, full string and number of results in a search log. Report the search with PRISMA-S.

From question to search string in five steps

  1. Step 1ConceptsPick 2–3 PICO parts: usually P and I. Leave the outcome out if it narrows too much.
  2. Step 2SynonymsSpellings, plurals, Latin names, abbreviations: maize, corn, Zea mays.
  3. Step 3Controlled termsMeSH in PubMed, Emtree in Embase, CAB Thesaurus in CAB Abstracts.
  4. Step 4CombineOR inside each concept block, AND between blocks, quotes and * where needed.
  5. Step 5TestCheck that 3–5 known key papers are found. If not, add terms.
OperatorWhat it doesExample
ANDNarrows: both terms must appearwheat AND inoculation
ORWidens: either termmaize OR corn
NOTExcludes (use with care, it can remove relevant records)malaria NOT vaccine
" "Exact phrase“seed priming”
*Truncation: word endingsinocula* finds inoculant, inoculated, inoculation
[MeSH]Controlled vocabulary in PubMed“Pneumonia”[MeSH]
Example search strings (illustrative)
Scopus:
TITLE-ABS-KEY ( ( wheat OR "Triticum aestivum" )
  AND ( "plant growth-promoting bacteria" OR pgpb OR pgpr OR bioinoculant* OR "bacterial inocula*" )
  AND ( yield* OR "grain yield" OR productivity ) )

Web of Science (Topic):
TS=( (wheat OR "Triticum aestivum")
  AND ("plant growth-promoting bacteria" OR PGPB OR PGPR OR bioinoculant* OR "bacterial inocula*")
  AND (yield* OR "grain yield" OR productivity) )

PubMed (health example, MeSH + free text):
("Pneumonia"[MeSH] OR pneumonia[tiab] OR "respiratory tract infection*"[tiab])
AND ("Biomass"[MeSH] OR biomass[tiab] OR "solid fuel*"[tiab] OR "indoor air pollution"[tiab])
AND (child*[tiab] OR infant*[tiab] OR "Child, Preschool"[MeSH])

Look up controlled terms in the free MeSH browser and test them in PubMed. Each database has its own syntax, so translate the string carefully; the free Search String Translator below does this for PubMed, Ovid, Embase, Scopus, Web of Science and CINAHL.

Which databases?

DatabaseBest forAccess
PubMed / MEDLINEMedicine, public health, biologyFree
EmbaseDrugs, clinical research, European journalsSubscription
ScopusBroad science, agriculture, engineeringSubscription (HEC digital library in Pakistan)
Web of Science Core CollectionHigh-quality indexed journals in all fieldsSubscription
CAB AbstractsAgriculture, veterinary science, environmentSubscription
Cochrane Library (CENTRAL)Randomised trials in healthFree in many countries
Google ScholarSupplementary searching and citation chasing (not as a main database)Free

Google Scholar is useful for finding grey literature and checking “Cited by”, but its results are not fully reproducible and you can only export limited records. Use it as an extra source. If you are new to it, my Urdu/Hindi guide on how to use Google Scholar covers the basics.

Do not forget grey literature and citation searching

  • Grey literature: theses (for example the HEC Pakistan research repository and Shodhganga in India), conference proceedings, government and FAO reports, and trial registers such as ClinicalTrials.gov.
  • Backward citation searching: check the reference lists of included studies.
  • Forward citation searching: check who cited them (“Cited by” in Google Scholar or Scopus).
  • Contact experts: ask authors in the field about unpublished or ongoing studies.

Keep a search log (PRISMA-S)

PRISMA-S is a 16-item extension of PRISMA for reporting literature searches. In practice it means: for every source, report the name, platform, date searched, the full search string, limits and the number of records. Put the full strings in a supplementary file. A simple spreadsheet is enough:

DatabasePlatformDateLimitsRecords
Scopusscopus.com10 Mar 20262000–2026, articles520
Web of Science Core CollectionClarivate10 Mar 20262000–2026430
CAB AbstractsCAB Direct11 Mar 20262000–2026350
Total records identified (illustrative)1,300
Ask a librarian. Many universities have a subject librarian who will check your strings for free. The PRESS checklist (Peer Review of Electronic Search Strategies) is a good guide for this check. Mention it in your methods; reviewers like it.
Key takeaways
  • Build concept blocks: OR inside, AND between.
  • Use controlled terms (MeSH) plus free-text synonyms.
  • Search 2–3+ databases, grey literature and citations; log everything for PRISMA-S.

Practise this step free: 🔄 Search String Translator · 🧰 AI Research Tool Finder · 🔍 Reference Checker · 🧠 Literature Search & Boolean Operators Quiz All writing tools →

↑ Back to topNext: Screen studies with two reviewers →

Step 5

How do you screen studies for a systematic review?

Quick answer

Export all records, remove duplicates, then screen in two stages: titles and abstracts first (be inclusive), then full texts (be strict and record the reason for every exclusion). Two reviewers should screen independently, measure their agreement with Cohen’s kappa on a pilot sample, and solve conflicts by discussion or a third reviewer. Record the number of records at every stage for the PRISMA 2020 flow diagram.

The screening workflow

  1. Export the results of each database as RIS or CSV files.
  2. Deduplicate in EndNote, Zotero or Rayyan. The same paper often appears in Scopus and Web of Science. Count how many duplicates you remove.
  3. Pilot the criteria: both reviewers screen the same 50–100 records, compare decisions and sharpen any unclear criterion.
  4. Title and abstract screening: each reviewer decides include, exclude or maybe, without seeing the other’s decisions. When in doubt, include.
  5. Full-text screening: retrieve the PDFs and apply every criterion. Record one main reason for each excluded report.
  6. Resolve conflicts by discussion, or ask a third reviewer. Keep a record.

Measure agreement with Cohen’s kappa

Kappa shows how much two reviewers agree beyond what chance would give. Suppose two reviewers screen 200 pilot records: both include 30, both exclude 150, reviewer A includes 12 that B excludes, and B includes 8 that A excludes.

B includesB excludesTotal
A includes301242
A excludes8150158
Total38162200

Observed agreement = (30 + 150) / 200 = 0.90. Chance agreement = (0.21 × 0.19) + (0.79 × 0.81) = 0.68. Kappa = (0.90 − 0.68) / (1 − 0.68) ≈ 0.69, which is “substantial” on the common Landis and Koch scale (0.41–0.60 moderate, 0.61–0.80 substantial, above 0.80 almost perfect). If kappa is low, talk about the disagreements and clarify the criteria before you continue.

Tools for screening

ToolWhat it doesCost
RayyanWeb app for blind dual screening, labels and conflict resolutionFree plan
ASReviewOpen-source active learning: ranks the most likely relevant records firstFree
CovidenceFull review workflow, often licensed by universitiesPaid
EndNote / ZoteroDeduplication and full-text managementPaid / free

ASReview learns from your include and exclude decisions and moves likely relevant records to the top, so you often find most relevant studies after screening a fraction of the total. You still make every decision. If you stop screening early, pre-specify a stopping rule in the protocol (for example, stop after a set number of irrelevant records in a row, after screening a minimum share of records, and only when all known key papers have been found) and report it.

Record the PRISMA 2020 flow numbers

PRISMA stageCount (illustrative)
Records identified from databases1,300 (Scopus 520, Web of Science 430, CAB Abstracts 350)
Duplicates removed before screening310
Records screened (title and abstract)990
Records excluded905
Reports sought for retrieval85
Reports not retrieved3
Reports assessed for eligibility82
Reports excluded, with reasons54: pot study (24), no control group (18), no yield data (12)
Studies included in the review28

Check that the numbers add up: 1,300 − 310 = 990; 990 − 905 = 85; 85 − 3 = 82; 82 − 54 = 28. Reviewers check this, and mistakes here look careless.

Weak reporting“Irrelevant studies were excluded.”
Good reporting“Of 82 full-text reports, 54 were excluded: pot study (n = 24), no uninoculated control (n = 18) and no grain-yield data (n = 12).”
Key takeaways
  • Deduplicate first, then screen titles and abstracts, then full texts.
  • Screen independently in pairs and check agreement with kappa.
  • Count every stage and record full-text exclusion reasons for PRISMA.

Practise this step free: 👥 Dual-Screening Agreement Calculator · 🔀 PRISMA Flow Diagram Generator · 🧰 AI Research Tool Finder · 🧠 Systematic Review & Meta-analysis Quiz All writing tools →

↑ Back to topNext: Extract the data →

Step 6

How do you extract data for a systematic review and meta-analysis?

Quick answer

Build a data extraction form, pilot it on two or three studies, then let two reviewers extract the same items from every included study independently and compare. Collect study details, population, intervention, comparison, outcomes and the numbers you need for effect sizes (means, SDs and sample sizes, or event counts). Contact authors for missing data and convert medians, IQRs, SEs or CIs to means and SDs with standard formulas.

What to put in the extraction form

CategoryFieldsExample
Study detailsAuthor, year, country, design, fundingTrial C, 2021, Egypt, randomised field trial
Population / settingCrop, variety, soil, climate; or age, sex, diseaseWheat cv. Sakha 93, sandy loam, irrigated
InterventionType, dose, timing, durationAzospirillum seed coating, 10⁸ CFU per mL
ComparisonControl typeUninoculated seed, same NPK dose
OutcomesDefinition, unit, time pointGrain yield, t/ha, at harvest
ResultsMean, SD, n per group; or events and totals4.9 (1.0), n = 12 vs 3.8 (0.95), n = 12
NotesWhere in the paper, any conversion madeTable 3; SD calculated from SE

Build the form in Excel or Google Sheets, or in Covidence if your university has it. Add a codebook that defines each field, so both reviewers fill it the same way. Pilot it, then fix unclear fields before full extraction.

When papers do not report means and SDs

This is the most common problem in meta-analysis. First, look in tables, figures and supplements. Then email the corresponding author (send one polite reminder after two weeks). If the data are still missing, use a standard conversion and record it in your notes column:

ReportedConversionWorked example
Standard error (SE)SD = SE × √nSE 0.15, n = 36 → SD = 0.15 × 6 = 0.90
95% CI of a mean (large sample)SD = √n × (upper − lower) / 3.92n = 60, CI 4.1 to 4.6 → SD ≈ 0.99
Median and IQRMean ≈ (Q1 + median + Q3) / 3; SD ≈ IQR / 1.35 (assumes roughly normal data)Median 4.2, IQR 3.6–4.9 → mean ≈ 4.23, SD ≈ 0.96
Median and rangeMethods of Hozo (2005), Wan (2014) and Luo (2018)Use a converter, and run a sensitivity analysis without these studies
Data only in a figureDigitise the graph with a tool such as WebPlotDigitizerReport that values were estimated from figures

For small samples, replace 3.92 with 2 × the t value for n − 1 degrees of freedom (the Cochrane Handbook explains this). The free Meta-Analysis Data Converter below does all of these conversions, including the sample-size-adjusted Wan, Luo and Shi formulas.

Watch for SE reported as SD. Many agronomy papers write “mean ± 0.15” without saying whether it is SD or SE. Check the methods or the figure legend. Treating an SE as an SD makes the study look far more precise than it is and gives it too much weight.
Key takeaways
  • Pilot the form, then extract in pairs and compare.
  • Collect the exact numbers needed for effect sizes, with their location in the paper.
  • Contact authors, use standard conversions and record every conversion.

Practise this step free: 🗂️ Data Extraction Form Builder · 🔁 Meta-Analysis Data Converter · 📏 Effect Size Calculator · 🧠 Results Section & APA Statistics Reporting Quiz All writing tools →

↑ Back to topNext: Assess risk of bias →

Step 7

How do you assess risk of bias in included studies?

Quick answer

Use a validated tool that fits each study design: RoB 2 for randomised trials, ROBINS-I for non-randomised studies of interventions, the Newcastle–Ottawa Scale for cohort and case-control studies and QUADAS-2 for diagnostic accuracy studies. Two reviewers judge each domain independently, support each judgement with a quote from the paper and show the results in a traffic-light plot.

Risk of bias is about whether a study’s design or conduct could make its result wrong. It is different from reporting quality: a well-written paper can still have high risk of bias. A study is not “good” because it is published in a high-impact journal.

Choose the right tool

Study designToolJudgements
Randomised controlled trial (including randomised field trials)RoB 2Low risk, some concerns, high risk
Non-randomised study of an interventionROBINS-ILow, moderate, serious, critical
Cohort or case-control studyNewcastle–Ottawa Scale (NOS)Stars, up to 9
Diagnostic test accuracy studyQUADAS-2Low, high, unclear (risk of bias and applicability)

RoB 2: the five domains for randomised trials

RoB 2 (Sterne et al., BMJ 2019) is the Cochrane tool for randomised trials. You answer signalling questions in each domain, and they lead to a judgement.

DomainMain questionField-trial example of high risk
1. Randomisation processWas allocation random and concealed?Plots assigned “by convenience” next to the irrigation channel
2. Deviations from intended interventionsWere groups treated as planned?Control plots also received a commercial biofertiliser
3. Missing outcome dataWere results available for nearly all units?A third of control plots lost to flooding, not reported why
4. Measurement of the outcomeCould measurement differ between groups?Yield estimated visually by a person who knew the treatment
5. Selection of the reported resultWas the result chosen from many analyses?Only the best of several harvests reported

The overall judgement is low only if all domains are low, high if any domain is high (or several domains have some concerns), and some concerns otherwise.

Unsupported judgement“Study looks good. Low risk.”
Supported judgement“Randomisation: low (‘plots were allocated using a computer-generated list’). Missing data: high (28% of control plots lost, no reason given, no sensitivity analysis).”

ROBINS-I, Newcastle–Ottawa and QUADAS-2 in brief

  • ROBINS-I has seven domains: confounding, selection of participants, classification of interventions, deviations from intended interventions, missing data, measurement of outcomes and selection of the reported result. Confounding is usually the biggest problem in non-randomised studies.
  • Newcastle–Ottawa Scale gives stars in three areas: selection (up to 4), comparability (up to 2) and outcome for cohort studies or exposure for case-control studies (up to 3), for a maximum of 9. Report the stars per area, not only the total.
  • QUADAS-2 covers patient selection, index test, reference standard, and flow and timing, with separate judgements for applicability.

Show it in a traffic-light plot

Journals expect a figure: a traffic-light plot (one row per study, one column per domain, green, yellow or red) and a summary bar plot (percentage of studies at each level per domain). Make them with the free robvis web app, the robvis R package, or the Codanics Risk of Bias Plot Generator below.

Then use your judgements: run a sensitivity analysis without high-risk studies (Step 8), and downgrade certainty for risk of bias in GRADE (Step 9).

Key takeaways
  • Match the tool to the design: RoB 2, ROBINS-I, NOS or QUADAS-2.
  • Two reviewers, independent judgements, a quote for each.
  • Show results in traffic-light and summary plots, then use them in analysis and GRADE.

Practise this step free: 🚦 Risk of Bias Plot Generator · 🧭 GRADE Summary of Findings Table Maker · 🧠 Risk of Bias & GRADE Quiz All writing tools →

↑ Back to topNext: Run the meta-analysis →

Step 8

How do you do a meta-analysis step by step?

Quick answer

Calculate one effect size per study (for example mean difference, standardised mean difference, odds ratio or risk ratio), pool them with inverse-variance weights in a random-effects model, and report the pooled effect with its 95% confidence interval. Then measure heterogeneity (Q, I², tau² and the prediction interval), draw a forest plot, explore differences with planned subgroup analyses or meta-regression, run sensitivity analyses and check publication bias with a funnel plot and, with 10 or more studies, Egger’s test.

Do not be afraid of the statistics. A meta-analysis is a weighted average: bigger, more precise studies count more. Everything else is checking whether that average makes sense. If you need a refresher on means, SDs and confidence intervals, start with my free ABC of Statistics series.

8.1

Choose the effect size

Effect sizeUse whenExampleNo effect =
Mean difference (MD)Continuous outcome, same unit in all studiesGrain yield in t/ha0
Standardised mean difference (SMD, Hedges’ g)Same outcome measured on different scalesAnxiety on different questionnaires0
Log response ratio (lnRR)Ratio of means; common in ecology and agronomyYield in treated plots / yield in control plots0 (ratio of 1)
Risk ratio (RR)Binary outcome, riskPneumonia: yes or no1
Odds ratio (OR)Binary outcome, odds; case-control studiesDisease in exposed vs unexposed1
Proportion, correlationPrevalence; association between two variablesPrevalence of anaemia; r between soil N and yieldNot a comparison

Hedges’ g is Cohen’s d with a small-sample correction, so use it for the SMD. Ratios (OR, RR, lnRR) are analysed on the log scale and back-transformed for reporting. Proportions are usually analysed after a logit or Freeman–Tukey transformation, and correlations after Fisher’s z transformation.

8.2

Fixed effect or random effects?

Fixed (common) effectAssumes one true effect; all differences between studies are chance. Narrower CI. Reasonable only when studies are very similar (same population, same intervention, same methods).
Random effectsAssumes true effects vary between studies (different soils, climates, patients) and estimates the average of that distribution. Wider, more honest CI. The usual choice in agronomy, ecology and most health reviews.

For random effects, the between-study variance (tau²) must be estimated. REML is a good default and is the default in metafor; the older DerSimonian–Laird method is still common. With few studies, many methodologists also recommend the Hartung–Knapp (Knapp–Hartung) adjustment, which gives wider, more reliable confidence intervals.

8.3

Measure heterogeneity

StatisticWhat it tells youHow to read it
Cochran’s Q (with p-value)Is there more variation than chance?Low power with few studies; do not rely on p alone
I²Share of variation due to real differences, not chance0–40% might not be important; 30–60% moderate; 50–90% substantial; 75–100% considerable (Cochrane)
tau² (tau)Variance (SD) of true effects, in the effect-size unittau = 0.3 t/ha means true effects typically differ by about that much
95% prediction intervalWhere the true effect of a new, similar study is likely to fallThe most practical number for farmers, doctors and policy makers

I² was introduced by Higgins et al. (BMJ 2003). It is a relative measure, so always look at the forest plot, tau and the prediction interval too. A high I² does not by itself mean you must not pool; it means you must explain the differences.

8.4

A worked example

Here is a small illustrative dataset (made-up teaching numbers, not real studies): five field trials of bacterial seed inoculation on wheat grain yield. Effect size: mean difference in t/ha, random-effects model with REML.

Trial (illustrative)n per groupInoculated, mean (SD)Control, mean (SD)MD (95% CI), t/haWeight
Trial A 2019 (Pakistan)204.6 (0.90)3.8 (1.00)0.80 (0.21 to 1.39)17.3%
Trial B 2020 (India)305.1 (1.00)4.6 (0.90)0.50 (0.02 to 0.98)24.5%
Trial C 2021 (Egypt)124.9 (1.00)3.8 (0.95)1.10 (0.32 to 1.88)10.5%
Trial D 2022 (China)406.0 (0.90)5.7 (0.90)0.30 (−0.09 to 0.69)33.6%
Trial E 2023 (Iran)154.2 (0.95)3.5 (0.90)0.70 (0.04 to 1.36)14.1%
Random-effects pooled (REML)0.58 (0.31 to 0.84)100%
Forest plot of five illustrative wheat field trials: pooled mean difference 0.58 t/haTrial (illustrative)MD, t/ha (95% CI)WeightTrial A 2019 (Pakistan)0.80 [0.21, 1.39]17.3%Trial B 2020 (India)0.50 [0.02, 0.98]24.5%Trial C 2021 (Egypt)1.10 [0.32, 1.88]10.5%Trial D 2022 (China)0.30 [−0.09, 0.69]33.6%Trial E 2023 (Iran)0.70 [0.04, 1.36]14.1%Random effects (REML)0.58 [0.31, 0.84]100%−0.5000.5011.502Favours controlFavours inoculationI² = 14%, tau² = 0.013
Figure: forest plot of the illustrative data (made-up teaching numbers). Squares = trials (size = weight), lines = 95% CI, diamond = pooled mean difference. 95% prediction interval 0.23 to 0.92 t/ha.

How to read it: each square is one trial (bigger square = more weight), the line is its 95% CI, and the diamond is the pooled effect. Inoculation increased yield by 0.58 t/ha on average (95% CI 0.31 to 0.84). Heterogeneity was low (Q = 4.38, df = 4, p = 0.36; I² = 14%; tau² = 0.013), and the 95% prediction interval (0.23 to 0.92 t/ha) says that a new, similar trial would probably also show a gain. With only five trials, these heterogeneity estimates are imprecise, so say so in your discussion. A fixed-effect model gives almost the same answer here (0.56 t/ha, 95% CI 0.32 to 0.80) because heterogeneity is low. I checked these numbers with the R code below and metafor.

meta_analysis.R: random-effects meta-analysis with metafor
# install.packages("metafor")   # run once
library(metafor)

# Illustrative data (made-up teaching numbers): wheat grain yield, t/ha
dat <- data.frame(
  trial = c("Trial A 2019", "Trial B 2020", "Trial C 2021", "Trial D 2022", "Trial E 2023"),
  m1i = c(4.6, 5.1, 4.9, 6.0, 4.2), sd1i = c(0.90, 1.00, 1.00, 0.90, 0.95), n1i = c(20, 30, 12, 40, 15),
  m2i = c(3.8, 4.6, 3.8, 5.7, 3.5), sd2i = c(1.00, 0.90, 0.95, 0.90, 0.90), n2i = c(20, 30, 12, 40, 15))

# 1. Effect size per trial: mean difference (use measure = "SMD" for Hedges' g)
dat <- escalc(measure = "MD", m1i = m1i, sd1i = sd1i, n1i = n1i,
              m2i = m2i, sd2i = sd2i, n2i = n2i, data = dat)

# 2. Random-effects model with REML
res <- rma(yi, vi, data = dat, method = "REML")
summary(res)      # pooled MD 0.58 (0.31 to 0.84), tau^2 = 0.013, I^2 = 14%, Q(4) = 4.38
predict(res)      # adds the 95% prediction interval (0.23 to 0.92)

# 3. Forest and funnel plots
forest(res, slab = dat$trial, xlab = "Mean difference in grain yield (t/ha)")
funnel(res)

# 4. Robustness
leave1out(res)                 # leave-one-out sensitivity analysis
# regtest(res)                 # Egger's test: only with 10 or more studies
# rma(yi, vi, data = dat, test = "knha")   # Knapp-Hartung adjustment

metafor (Viechtbauer, Journal of Statistical Software 2010) is free and used in thousands of published meta-analyses. If you prefer not to code yet, the free Codanics Meta-Analysis Calculator gives the same kind of output (pooled effect, I², forest and funnel plots, Egger’s test) in your browser. In my Systematic Review & Meta-analysis course I run this whole workflow on screen, in R and in MS Excel, and explain every line of output in Urdu/Hindi.

📘 Book pick: Doing Meta-Analysis with R: A Hands-On Guide by Mathias Harrer, Pim Cuijpers, Toshi A. Furukawa and David D. Ebert (Chapman & Hall/CRC, 2021). The full book is free online as Doing Meta-Analysis in R. It teaches meta-analysis in R with the meta and metafor packages, step by step and with code you can copy. Pair it with this guide:
  • Step 6 and 8.1: the chapters “Effect Sizes” and “Effect Size Calculation & Conversion”
  • Step 8.2–8.4: “Pooling Effect Sizes”, “Between-Study Heterogeneity” and “Forest Plots”
  • Step 8.5: “Subgroup Analyses”, “Meta-Regression” and “Publication Bias”
  • Step 7: “Risk of Bias Plots”; Step 10: “Reporting & Reproducibility”
I recommend it because it is free, practical and written by meta-analysis researchers. No affiliate link.
Read it free →
8.5

Subgroups, meta-regression, sensitivity analysis and publication bias

AnalysisQuestionRules of thumb
Subgroup analysisDoes the effect differ between groups (for example irrigated vs rain-fed)?Plan subgroups in the protocol; report the test for subgroup differences, not separate p-values
Meta-regressionDoes the effect change with a continuous variable (for example nitrogen dose)?Cochrane advises against it with fewer than about 10 studies; results are observational, not causal
Sensitivity analysisIs the result robust?Remove high-risk studies, studies with converted data, or one study at a time (leave-one-out)
Funnel plotAre small studies with null results missing?Should look like a symmetric, upside-down funnel; contour-enhanced versions help
Egger’s regression testIs the funnel asymmetric?Only with 10 or more studies; asymmetry can also come from heterogeneity, not only publication bias
Trim-and-fillWhat if “missing” studies were added?A sensitivity analysis only; never report the filled estimate as the main result
Common meta-analysis mistakes: pooling studies that ask different questions (apples and oranges); using the same control group twice for two treatment arms without adjusting (double counting); treating SE as SD; running Egger’s test on four studies; testing ten unplanned subgroups and reporting the one that is “significant”.
Key takeaways
  • Pick the effect size that fits your outcome; pool with random effects (REML).
  • Report Q, I², tau² and the prediction interval, and show a forest plot.
  • Plan subgroups, run sensitivity analyses, and test funnel asymmetry only with 10+ studies.

Practise this step free: 🌲 Meta-Analysis Calculator · 📏 Effect Size Calculator · 🔻 Publication Bias Calculator · 📈 Subgroup Analysis and Meta-Regression Calculator · 🧠 Systematic Review & Meta-analysis Quiz All writing tools →

↑ Back to topNext: Rate certainty with GRADE →

Step 9

How certain is the evidence? Using GRADE and a summary of findings table

Quick answer

GRADE rates the certainty of evidence for each important outcome as high, moderate, low or very low. Evidence from randomised trials starts at high and evidence from observational studies starts at low. You downgrade for risk of bias, inconsistency, indirectness, imprecision and publication bias, and for observational evidence you can upgrade for a large effect, a dose-response gradient or when plausible confounding would reduce the effect. Present the results in a summary of findings table.

A pooled effect alone can mislead. “Yield increased by 0.58 t/ha” sounds certain, but if all the trials had high risk of bias, readers should be careful. GRADE, developed by the GRADE Working Group, makes that judgement transparent.

CertaintySymbolMeaning
High⊕⊕⊕⊕We are very confident the true effect is close to the estimate.
Moderate⊕⊕⊕◯The true effect is probably close, but it may be substantially different.
Low⊕⊕◯◯The true effect may be substantially different.
Very low⊕◯◯◯The true effect is likely to be substantially different.

The five reasons to downgrade

DomainDowngrade when…Where you find it
Risk of biasMost of the weight comes from studies at high risk or with some concernsYour RoB 2 / ROBINS-I results (Step 7)
InconsistencyEffects point in different directions, high I², wide prediction intervalForest plot and heterogeneity (Step 8)
IndirectnessPopulation, intervention or outcome differ from your questionCharacteristics table, for example pot trials used to answer a field question
ImprecisionWide CI that includes both benefit and no effect, or few participants or plotsPooled 95% CI
Publication biasFunnel asymmetry, only small positive industry-funded studiesFunnel plot and search for unpublished data

Example summary of findings table (illustrative)

OutcomeEffect (95% CI)Studies (plots)CertaintyComment
Grain yield, t/haMD +0.58 (0.31 to 0.84)5 (234)⊕⊕⊕◯ ModerateDowngraded once for risk of bias (unclear randomisation in two trials)
Grain protein, %MD +0.3 (−0.1 to 0.7)3 (120)⊕⊕◯◯ LowRisk of bias and imprecision (CI includes no effect)
Soil available nitrogenNot pooled; mixed directions2 (80)⊕◯◯◯ Very lowInconsistency, imprecision and indirectness

Keep the table to the most important outcomes (Cochrane suggests no more than seven). For binary outcomes, add absolute effects per 1,000 people, which make the result easy to understand: for example “45 fewer cases of pneumonia per 1,000 children”. You can build the table in GRADEpro GDT or with the free Codanics GRADE Summary of Findings Table Maker.

Write conclusions that match the certainty. High: “X increases Y.” Moderate: “X probably increases Y.” Low: “X may increase Y.” Very low: “We are uncertain whether X increases Y.” This simple wording rule makes your abstract honest and is what reviewers expect.
Key takeaways
  • Rate certainty per outcome: high, moderate, low or very low.
  • Downgrade for risk of bias, inconsistency, indirectness, imprecision and publication bias.
  • Use a summary of findings table and match your wording to the certainty.

Practise this step free: 🧭 GRADE Summary of Findings Table Maker · 🚦 Risk of Bias Plot Generator · 🧠 Risk of Bias & GRADE Quiz All writing tools →

↑ Back to topNext: Write the review with PRISMA 2020 →

Step 10

How do you write a systematic review paper using PRISMA 2020?

Quick answer

Use the PRISMA 2020 checklist as your outline. Name the design in the title, write a structured abstract with the 12-item PRISMA abstract checklist, then report methods (protocol, criteria, sources, search, selection, extraction, risk of bias, synthesis, certainty), results (PRISMA flow diagram, study characteristics, risk of bias, syntheses), and a discussion of the evidence, limitations and implications. Submit the completed checklist with page numbers.

The PRISMA 2020 statement (Page et al., BMJ 2021;372:n71) replaced PRISMA 2009. It has a 27-item checklist, an abstract checklist and new flow diagram templates for new reviews and updates, all free on the PRISMA website. Other reporting guidelines are listed on the EQUATOR Network.

Section by section

SectionWhat to includeTip
TitleTopic and “systematic review and meta-analysis”“Effect of bacterial seed inoculation on wheat yield: a systematic review and meta-analysis”
AbstractBackground, objective, methods, results, limitations, conclusion, funding, registration numberFollow the 12-item PRISMA 2020 abstract checklist
IntroductionWhy the question matters, what is unknown, the PICO objectiveExplain why a new review is needed if others exist
MethodsRegistration, eligibility, sources and dates, full search (supplement), selection, extraction, risk of bias, effect measures, synthesis model, heterogeneity, sensitivity, certaintyWrite it in the past tense, closely following your protocol; report deviations
ResultsPRISMA flow diagram, characteristics table, risk-of-bias figure, forest plots, heterogeneity, subgroup and sensitivity results, publication bias, summary of findings tableEvery number in the text must match the tables and figures
DiscussionMain findings with certainty, comparison with other reviews, limitations of the evidence and of the review process, implicationsSeparate limitations of the studies from limitations of your review
Other informationRegistration and protocol, support, competing interests, data and code availabilityShare your extraction sheet and R code (for example on OSF)

Synthesise, do not list

Listing“Trial A found a yield increase. Trial B found a yield increase. Trial D found no significant effect.”
Synthesising“Four of five trials reported yield gains of 0.5 to 1.1 t/ha. The smallest effect came from the largest, irrigated trial with high baseline yield, suggesting that inoculation helps most where yields are limited.”

If you cannot pool studies, use a structured narrative synthesis and follow the SWiM (Synthesis Without Meta-analysis) reporting guideline: group studies, tabulate results, describe the direction and size of effects and explain differences. Never use “vote counting” of significant p-values as your main result.

Common writing mistakes reviewers flag

  • No registration number, or registration after the search date.
  • Only one database searched, or the full search string missing.
  • PRISMA flow numbers that do not add up.
  • Risk of bias assessed but never used in the analysis or conclusions.
  • I² reported but heterogeneity never explained.
  • Conclusions stronger than the GRADE certainty allows.
  • The PRISMA checklist not submitted, or submitted without page numbers.

For general writing skills (titles, abstracts, paraphrasing, journal choice), see my complete scientific writing guide. If you want structured daily practice in writing papers and reviews, the Scientific Writing Marathon covers the whole paper, including a full block on reviews and meta-analysis.

Key takeaways
  • Write with the PRISMA 2020 checklist open; it is your outline.
  • Report the flow diagram, characteristics, risk of bias, syntheses and certainty.
  • Synthesise patterns, share data and code, and submit the checklist with page numbers.

Practise this step free: ☑️ CONSORT and PRISMA Checklist Tool · 🔀 PRISMA Flow Diagram Generator · 📝 APA Results Sentence Generator · ✅ Pre-Submission Manuscript Checklist · 🧠 Results Section & APA Statistics Reporting Quiz All writing tools →

↑ Back to topNext: Use AI tools responsibly →

Step 11

How can you use AI tools responsibly in a systematic review?

Quick answer

Use AI to save time on ranking records (ASReview), drafting extraction tables (Elicit), suggesting search terms and improving language, but keep humans in charge of every inclusion decision, risk-of-bias judgement and extracted number. Verify all AI output against the original papers, state planned AI use in the protocol, and report the tool, version, task and human verification in your methods.

AI tools can make a review faster, but they can also miss studies, misread numbers and invent references. In a systematic review, an error in one number can change the pooled result. So the rule is simple: AI can suggest; people decide and verify.

Review stepAI can help withRiskHuman check
Question and searchSynonyms, MeSH ideas, draft Boolean stringsMissing terms, wrong syntaxTest strings against known papers; librarian check
ScreeningRanking records with active learning (ASReview)Stopping too earlyPre-set stopping rule; two human reviewers decide
Data extractionFirst-draft tables (Elicit, SciSpace)SE read as SD, groups swapped, invented valuesCheck 100% of cells against the PDF
Risk of biasFinding the relevant sentencesOver-confident judgementsTwo reviewers judge each domain
AnalysisWriting and explaining R codeWrong model or data errorsRun and check the code yourself
WritingLanguage editing, structureFake citations, overclaimingCheck every reference; follow the journal’s AI policy
Weak AI reporting“AI tools were used to help with the review.”
Transparent AI reporting“Title and abstract screening was prioritised with ASReview LAB (version X, default model); two reviewers screened all records shown and stopped after the pre-registered rule was met. Data were pre-extracted with Elicit and verified in full against the PDFs by two reviewers.”

Tools such as Consensus are useful for a quick look at the evidence, but they are not a systematic search and cannot replace your database searches. AI tools also cannot be authors: authors take responsibility for the work, and journals ask you to disclose AI use. Check your target journal’s latest policy.

To learn these tools properly, take my free course Using AI tools in Scientific Writing (with a certificate), and read my guides on AI tools for students and academics and prompt engineering.

Short · 1:07
5 best AI research tools for scientific writing
5 best AI research tools for scientific writing. Five AI research tools in about a minute.
Lesson · 15:34
Ethical use of AI tools for scientific writing
Ethical use of AI tools for scientific writing. Best practices and guidelines for using AI ethically.
Copy a safe AI prompt for drafting extraction (you verify every value)
Act as a careful systematic review assistant.
Below is the methods and results text of one included study.
Extract only what is written; do not guess or calculate.
Return a table with: study design, country, population, intervention,
comparison, outcome name, unit, time point, and for each group the
mean, SD (or SE, say which) and n. Add the page or table where each
value appears. Write "not reported" when a value is missing.

[paste the text here]
Key takeaways
  • AI ranks, drafts and suggests; humans decide, judge and verify.
  • Check every AI-extracted value and every reference.
  • Plan AI use in the protocol and report tool, version, task and verification.

Practise this step free: 🧰 AI Research Tool Finder · 🧠 Research Prompt Builder · 🔍 Reference Checker · 🧠 AI Ethics in Academic Writing Quiz All writing tools →

↑ Back to topNext: Timeline, team and publishing →

Step 12

How long does a first systematic review take, and how do you publish it?

Quick answer

A first systematic review with meta-analysis usually takes 6 to 12 months with a small team: a lead reviewer, a second reviewer, a supervisor or topic expert, and help from a statistician and a librarian. Plan the work month by month, choose a journal that publishes reviews in your field, submit the PRISMA checklist and flow diagram with your manuscript, and share your data and code.

A realistic timeline for a first review

  1. Month 1Question and protocol. Check for existing reviews, write PICO, draft and register the protocol.
  2. Month 2Search. Build and test strings, run the final searches, export and deduplicate.
  3. Months 3–4Screening. Pilot, dual title/abstract screening, full-text screening, PRISMA counts.
  4. Month 5Extraction and risk of bias. Pilot the form, dual extraction, contact authors, RoB judgements.
  5. Month 6Analysis. Meta-analysis, heterogeneity, subgroups, sensitivity, publication bias, GRADE.
  6. Months 7–8Writing. Methods and results first, then discussion, introduction and abstract; PRISMA checklist.
  7. ThenSubmit. Rerun the search if more than about 12 months old, then submit.

If you are doing your first review alone with a supervisor, a guided course saves months of trial and error. My Systematic Review & Meta-analysis course follows exactly this order, so you can work on your own review while you learn.

Who is on the team?

RoleJob
Lead reviewer (you)Protocol, search, screening, extraction, analysis, first draft
Second reviewerIndependent screening, extraction and risk of bias
Supervisor or topic expertQuestion, criteria, interpretation, third reviewer for conflicts
StatisticianChecks the effect sizes, model and code
LibrarianChecks the search strategy

Publishing your review

  • Choose the journal: pick a journal in your field that publishes systematic reviews and meta-analyses (check its recent issues). Options include field journals and specialist journals such as Systematic Reviews and, for environmental topics, Environmental Evidence. Avoid predatory journals.
  • Prepare the files: manuscript, PRISMA 2020 checklist with page numbers, flow diagram, supplementary search strings, extraction data and R code.
  • Write a short cover letter: the question, why it matters, the registration number and the main finding with its certainty.
  • Respond to reviewers point by point, and update the search if the editor asks.
  • Share it: present it as a conference poster (see my conference poster guide) and add it to your CV. A published review also makes a strong base for a PhD proposal or a postdoc proposal.
Key takeaways
  • Plan 6–12 months and a team of at least two reviewers.
  • Submit with the PRISMA checklist, flow diagram, search strings, data and code.
  • Rerun an old search before submission, and turn the review into posters and proposals.

Practise this step free: 🗓️ Research Gantt Chart Maker · ✅ Pre-Submission Manuscript Checklist · 🔍 Reference Checker · 🧠 Systematic Review & Meta-analysis Quiz All writing tools →

↑ Back to topNext: Learn it all in one course →

Learn systematic reviews and meta-analysis in one course

Self-paced, in Urdu/Hindi, by Dr. Muhammad Aammar Tufail

This guide gives you the map. If you want me to walk you through every step on screen, with real files and real software, join my Systematic Review & Meta-analysis course. It is a self-paced course on the Codanics learning platform, taught in Urdu/Hindi with English scientific terms, for thesis students, MPhil and PhD scholars, lecturers, professors and medical doctors.

PlanQuestion to protocolReview types, PICO, eligibility criteria and protocol registration.
FindSearch and screenSystematic searching, EndNote, deduplication, screening and the PRISMA flow diagram.
AnalyseExtract and poolData extraction, risk of bias and meta-analysis in R and MS Excel.
Go from a research question to a publishable reviewThe Systematic Review & Meta-analysis course follows the same steps as this guide, with R, MS Excel, MS Word and EndNote, so you finish with a review you can submit to a good journal.
  • Self-paced lessons on the Codanics LMS, in Urdu/Hindi
  • Searching, screening, extraction, risk of bias and GRADE
  • Meta-analysis, forest plots and funnel plots in R and Excel
  • Writing and reporting with PRISMA 2020
Join the course →

Free options to start today: the free Scientific Writing Master Class on YouTube, the free course Using AI tools in Scientific Writing with a certificate, and the playlists below. If you want to improve your whole paper, not only reviews, look at the Scientific Writing Marathon.

Videos and playlists on systematic reviews

Free lessons in Urdu/Hindi on the Codanics YouTube channel

Systematic review and meta-analysis complete coursePress play to start from the first lesson.
Open on YouTube
Scientific writing: reviews and meta-analysisNarrative reviews, systematic reviews and meta-analysis in scientific writing.
Open on YouTube

All free tools and quizzes for reviews

Free, no login, and everything runs in your browser

Updated automatically from codanics.com · All scientific writing tools → · All free tools → · All quizzes →

Latest from Codanics on research writing

New guides on systematic reviews and scientific writing

Updated automatically from codanics.com · All guides on the Codanics blog →

Links and sources

Related guides on Codanics

More research, AI and career guides on the Codanics blog.

Key terms explained

Short definitions of the words used in this guide

Systematic review
A review that uses a pre-planned, reproducible method to find, select and appraise all studies on one question.
Meta-analysis
A statistical method that combines the results of several studies into one pooled estimate.
Scoping review
A review that maps what evidence exists on a topic, usually without appraising study quality; reported with PRISMA-ScR.
PICO / PECO
Population, Intervention (or Exposure), Comparison, Outcome: a format for a focused review question.
FINER
Feasible, Interesting, Novel, Ethical, Relevant: criteria for a good research question.
Protocol
The written plan of a review, prepared and registered before screening starts.
PRISMA 2020
The reporting guideline for systematic reviews: a 27-item checklist, an abstract checklist and a flow diagram.
PRISMA-P
The 17-item checklist for writing systematic review protocols.
PRISMA-S
A 16-item PRISMA extension for reporting literature searches.
PROSPERO
A free international register for systematic review protocols with health-related outcomes.
MeSH
Medical Subject Headings: the controlled vocabulary used to index PubMed records.
Grey literature
Research not published in commercial journals, such as theses, reports and conference papers.
Cohen's kappa
A measure of agreement between two reviewers beyond what chance would give.
Risk of bias
The chance that a study's design or conduct makes its result wrong.
RoB 2
The Cochrane risk-of-bias tool for randomised trials, with five domains.
ROBINS-I
A risk-of-bias tool for non-randomised studies of interventions, with seven domains.
Newcastle–Ottawa Scale
A star-based quality tool for cohort and case-control studies (up to 9 stars).
Effect size
A number that shows the size and direction of an effect, such as MD, SMD, OR or RR.
Hedges' g
A standardised mean difference with a small-sample correction.
Random-effects model
A meta-analysis model that assumes true effects vary between studies and estimates their average.
Heterogeneity (I², tau²)
Variation in results between studies beyond chance; I² is the share, tau² the variance of true effects.
Prediction interval
The range in which the true effect of a new, similar study is expected to lie.
Forest plot
A chart of each study’s effect and confidence interval with the pooled effect as a diamond.
Funnel plot
A plot of effect size against precision used to look for small-study effects and publication bias.
GRADE
A system to rate the certainty of evidence as high, moderate, low or very low.

Frequently asked questions

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

A systematic review is the method of finding, screening and appraising all studies on a focused question with a pre-planned protocol. A meta-analysis is the statistical step that combines the numerical results of similar studies into one pooled estimate. A meta-analysis is usually done inside a systematic review, but a systematic review does not always include a meta-analysis.

How do I write a systematic review step by step?

Ask a focused PICO question, write and register a protocol, search at least two or three databases plus grey literature, remove duplicates, screen titles and abstracts and then full texts with two reviewers, extract data, assess risk of bias, synthesise the results (with a meta-analysis if possible), rate certainty with GRADE and report with the PRISMA 2020 checklist and flow diagram.

How long does a systematic review take?

A first systematic review with a meta-analysis usually takes 6 to 12 months with a small team. Narrative reviews are faster (1 to 3 months) and large reviews with many studies can take longer.

How many studies do I need for a meta-analysis?

You can pool as few as two studies if they are similar enough, but estimates of heterogeneity are very imprecise with few studies. Tests for funnel plot asymmetry such as Egger's test should only be used with 10 or more studies, and meta-regression is generally not advised with fewer than about 10 studies.

Do I need to register my protocol on PROSPERO?

Registration is strongly recommended and many journals expect it. Register on PROSPERO if your review has a health-related outcome, or on OSF Registries for any field, including agriculture and scoping reviews. Register before screening starts.

How many databases should I search for a systematic review?

Search at least two or three relevant databases, for example PubMed, Scopus and Web of Science for health, or Scopus, Web of Science and CAB Abstracts for agriculture. Add grey literature and citation searching, and record the full search string, date and results for each source.

Can I do a systematic review alone?

Not properly. Screening, data extraction and risk-of-bias assessment should be done by two reviewers independently to reduce errors and bias. Ask a classmate, colleague or co-supervisor to be the second reviewer.

Which software is used for meta-analysis?

Free options include R with the metafor or meta packages, Cochrane's RevMan, and JASP or jamovi. For screening use Rayyan or ASReview, and use EndNote or Zotero to manage references. The free Codanics Meta-Analysis Calculator runs in your browser.

What is a good I² value in a meta-analysis?

There is no single good value. The Cochrane Handbook gives overlapping guides: 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. Always look at the forest plot, tau² and the prediction interval as well, and explain heterogeneity with planned subgroup or sensitivity analyses.

Should I use a fixed-effect or random-effects model?

Use a random-effects model in most reviews, because true effects usually differ between studies in different places, populations and conditions. A fixed-effect model is reasonable only when studies are very similar. Choose the model in your protocol, not after seeing the results.

Can I use ChatGPT or other AI tools for a systematic review?

Yes, for support tasks such as suggesting search terms, ranking records with ASReview or drafting extraction tables with Elicit. Humans must make every inclusion decision and risk-of-bias judgement, check every extracted value and reference, and report the AI tools, versions and verification in the methods.

Is there a course on systematic review and meta-analysis in Urdu or Hindi?

Yes. Dr. Muhammad Aammar Tufail's Systematic Review & Meta-analysis course on Codanics is a self-paced course in Urdu/Hindi that takes you from a research question to a publishable review, with R, MS Excel and EndNote. Free lessons are also available on the Codanics YouTube channel.

About the author. Dr. Muhammad Aammar Tufail is the founder of Codanics and completed his PhD in a Marie Skłodowska-Curie programme. He has trained hundreds of students in scientific writing, systematic reviews and meta-analysis, and teaches research, AI, data science and bioinformatics in Urdu and Hindi.

See his research on Google Scholar and ResearchGate, and follow him on LinkedIn.

Last updated 22 November 2026. Methods follow the Cochrane Handbook, PRISMA 2020 and GRADE. Worked examples use illustrative teaching data, not real studies. Journal and registry policies change; always check the latest guidance.

Leave a Reply

Your email address will not be published. Required fields are marked *