Meta-Analysis Basics
On this page
Direct answer
A forest plot with eleven squares, a vertical line of no effect and a diamond at the foot of the figure is the public face of meta-analysis — the statistical pooling of results from two or more studies asking the same question, performed as the final step inside a systematic review. Each study contributes a point estimate weighted by its precision (inverse-variance weighting; larger squares carry larger weights), and the model is either fixed-effect, assuming one true effect shared by all studies, or random-effects, allowing real between-study variation through the tau-squared term. Heterogeneity is quantified by Cochran's Q and expressed as I-squared, with 25, 50 and 75 per cent marking low, moderate and high thresholds; funnel-plot asymmetry and Egger's test probe the suspicion that small negative studies went unpublished. Reporting follows the PRISMA 2020 statement.
What you must remember
- Definition and hierarchy: meta-analysis is the statistical step; the systematic review is the method that earns it — a review can exist without pooling, never the reverse.
- Fixed versus random: fixed assumes a single common effect; random incorporates between-study variance (tau-squared), widening confidence intervals, and is chosen when studies genuinely differ.
- Heterogeneity numbers: Cochran's Q chi-square with p < 0.10 raises the flag; I² = (Q − df)/Q × 100, graded 25/50/75 per cent low, moderate, high (Higgins benchmark).
- Forest plot anatomy: square = point estimate, square size = weight, horizontal line = confidence interval, diamond = pooled estimate with its tips as the CI; the vertical null sits at 1 for OR/RR and 0 for risk difference or standardised mean difference.
- Publication bias toolkit: funnel plot asymmetry, Egger's regression test, trim-and-fill correction; searching trial registries is the design-level antidote.
- Effect measures: OR or RR for dichotomous outcomes, risk difference for absolute impact, mean difference when scales match and standardised mean difference when they do not.
- Reporting and certainty: PRISMA 2020 governs reporting; GRADE grades the certainty of the pooled evidence — pooling weak trials yields a precise but wrong answer.
Reading a forest plot from left to right
Begin at the left column: study identifiers, then events and totals for each arm. The middle panel plots each study's odds ratio as a square whose area mirrors its statistical weight — one multicentre trial may carry a fifth of the total weight while three small pilots carry a twentieth each — with its 95% confidence interval as a horizontal line. The pooled diamond at the bottom centres on the combined estimate; if both lateral tips stay to the right of 1, the pooled association is significant, and a diamond touching or crossing the line is a null pooled result however many studies were combined.
Then the decision layer. An I² of 70% means most of the variation across studies is real rather than chance, so before accepting the diamond you look for subgroup analyses (dose, region, risk stratum), sensitivity analyses (drop the outlier and see whether the estimate moves) or meta-regression. If the funnel plot looks like a lopsided triangle with small studies missing from the null side, publication bias is the working diagnosis, and trim-and-fill estimates how far the pooled result would move once the missing studies are imputed.
Viva favourites
The safest distinction is systematic review versus meta-analysis: the review is the protocol-driven search, selection and appraisal machinery; the meta-analysis is only the arithmetic stitched on when studies are similar enough to pool. The second favourite is Q versus I² — Q is underpowered with few studies, which is precisely why I² was introduced as a quantity, not a significance test. Third: why the diamond can be non-significant when most individual squares are significant — pooling averages heterogeneous estimates and the interval widens under random effects. Asked whether a meta-analysis tops the evidence pyramid, the nuanced answer is yes only when it pools good, homogeneous studies; one large well-conducted trial outranks a meta-analysis of biased ones — garbage in, garbage out is an accepted viva answer.
Frequently asked questions
What does an I-squared of 60 per cent mean?
Moderate-to-substantial heterogeneity — 60% of the observed variation between studies reflects real differences in effect rather than chance, prompting subgroup and sensitivity analysis.
When is a random-effects model chosen over fixed-effect?
When studies differ in populations, interventions or methods, or when heterogeneity is appreciable, since random effects assume each study estimates its own true effect around a mean.
What does an asymmetric funnel plot suggest?
Small-study effects, most often publication bias — small negative studies failing to reach print, leaving the bottom of the funnel sparse on one side.
How does a meta-analysis differ from a systematic review?
The systematic review is the transparent, reproducible search and appraisal process; the meta-analysis is the optional statistical pooling of extracted results within it.
What does a diamond crossing the line of no effect indicate?
The pooled estimate's confidence interval includes the null value, so the combined evidence does not demonstrate a significant effect.