Meta-analysis is a method, not a study design by itself
A meta-analysis should normally sit inside a systematic review with a prespecified question, eligibility criteria, comprehensive search, selection process and risk-of-bias assessment. Statistical pooling cannot repair a biased search or incompatible studies.
Approval and governance
Usually needs formal review / authorisation
- Institutional research-office approval if required for research conducted under local policy.
- Permission for any non-public individual-participant data or licensed datasets used in an individual-participant-data meta-analysis.
May follow a lighter or different route
- Aggregate-data meta-analysis of published studies often does not involve new human participants; human-subject ethics review may not be required in many systems, but obtain the institutional determination.
Do not do this
- Do not pool just because software can produce a forest plot.
- Do not choose fixed versus random effects solely based on whether an I² p-value is “significant.”
- Do not switch effect measures/models until the result becomes statistically significant.
Mediclinic Middle East publicly states that research projects carried out at MCME are to receive approval from its internal Research and Ethics Committee and applicable local regulatory authorities before initiation. The exact route varies by project, facility and emirate. Dubai projects may involve DSREC depending on applicability. This hub must therefore route users to the Research Office and current local forms rather than declaring a project “ethics exempt.” Institution-specific forms and contacts will be inserted after verification.
Step-by-step workflow
Finish the systematic-review steps first
Complete eligibility, searching, selection, extraction and risk-of-bias assessment before statistical pooling. Predefine which studies/outcomes/time points are eligible for each synthesis.
Define the estimand and effect measure
- Binary outcomes: risk ratio, odds ratio, risk difference or another justified measure.
- Continuous outcomes: mean difference when scales match; standardised mean difference when conceptually equivalent constructs use different scales.
- Time-to-event outcomes: hazard ratio when appropriate.
- Rates: rate ratio/difference where person-time is relevant.
Harmonise direction and units
Ensure higher/lower scores mean the same thing, convert units correctly, and avoid double-counting participants from multi-arm studies. Document all transformations.
Choose one outcome time point strategy
Prespecify clinically meaningful time points or windows. Do not select the time point with the strongest treatment effect from each study.
Recover missing statistics transparently
When standard deviations or effect estimates are missing, use validated transformations or contact authors when feasible. Record assumptions and test them in sensitivity analysis.
Assess clinical and methodological heterogeneity first
Compare populations, interventions/exposures, comparators, outcome definitions, follow-up, design and risk of bias. Statistical heterogeneity metrics are secondary to this judgement.
Choose the statistical model for scientific reasons
Fixed-effect models estimate a common-effect assumption; random-effects models allow true effects to vary. State the intended inference, estimator and any small-sample adjustments.
Quantify uncertainty and heterogeneity
Report pooled effect with confidence interval and measures such as tau-squared/I-squared when relevant, but do not use thresholds mechanically. Prediction intervals can help communicate between-study variation when enough studies are available.
Investigate heterogeneity only as prespecified and plausible
Use subgroup analysis or meta-regression for a small set of hypotheses grounded in clinical reasoning. Interaction tests are more appropriate than comparing significance within separate subgroups.
Handle dependent estimates
If one study contributes multiple correlated outcomes, time points or intervention arms, select/aggregate according to protocol or use methods that account for dependence. Never treat correlated estimates as independent simply to increase sample size.
Assess small-study/publication bias cautiously
Funnel-plot asymmetry and statistical tests are unreliable with few studies and can reflect heterogeneity rather than publication bias. Interpret them in context and follow current guidance.
Run sensitivity analyses
Examples include excluding high-risk-of-bias studies, changing reasonable correlation assumptions, alternative estimators, excluding imputed statistics or analysing adjusted versus unadjusted estimates. Prespecify key analyses when possible.
Do not confuse statistical significance with importance
Interpret effect size, confidence interval, baseline risk, clinical importance, heterogeneity, risk of bias and certainty together. A narrow p-value threshold is not the conclusion.
Report within PRISMA
Describe software/version, effect measures, model, heterogeneity methods, subgroup/sensitivity analyses, study contributions, risk of bias and any deviations from protocol.
Final checklist
- Pooling question prespecified.
- Studies clinically/methodologically compatible enough to combine.
- Effect measure appropriate and direction consistent.
- Model justified before seeing preferred result.
- Multi-arm/dependent data handled correctly.
- Heterogeneity interpreted clinically and statistically.
- Sensitivity analyses reported.
- Risk of bias and certainty considered in interpretation.