BiostatisticsTier 1Medical Sciences concept

Interpreting forest plots and subgroup analyses

Core concept

One row per study or subgroup
- Square = point estimate; area proportional to weight (inverse of variance, so large/precise trials dominate)
- Horizontal line = 95% CI; crossing the line of no effect = non-significant
- Diamond = pooled estimate; width = its CI
Line of no effect: 1 for ratios (RR, OR, HR), 0 for differences (mean difference, risk difference)
- Check which scale before reading anything - ratio plots are drawn on a log axis so the CI looks symmetrical

Heterogeneity

Statistic	Meaning
I-squared	% of total variability due to between-study difference rather than chance
	<25% low, 25-50% moderate, >50-75% substantial
Cochran's Q / chi-square	Test for heterogeneity - underpowered when few studies, so p<0.10 is used
tau-squared	Between-study variance, used in random-effects weighting

Fixed-effect model - assumes one true effect; narrow CI
Random-effects model - assumes a distribution of true effects; wider CI, small studies weighted relatively more
- Use when I-squared is substantial; random effects does not fix heterogeneity, it just acknowledges it
Funnel plot asymmetry -> publication bias (small negative trials missing); Egger's test

Subgroup analyses

Credibility checklist
- Pre-specified, not post-hoc
- Small number of subgroups
- Test for interaction significant - not "significant in one subgroup, not in the other"
- Consistent across related endpoints and other trials
- Biologically plausible; direction predicted in advance
A within-subgroup p-value is the wrong test. Only the interaction p answers "does the effect differ?"
With 10 subgroups at alpha 0.05, ~40% chance of at least one spurious interaction
- ISIS-2: aspirin appeared not to work in Gemini and Libra - the deliberate parody that made the point

Subgroups are hypothesis-generating; they change practice only when a dedicated trial confirms them
Qualitative interaction (benefit in one group, harm in another) is rare and demands scepticism; quantitative (more vs less benefit) is common and usually real
Meta-analysis of small trials can be overturned by one large RCT (magnesium in MI: positive meta-analysis, null ISIS-4)
Individual patient data meta-analysis > study-level, especially for subgroups
Watch the axis: an HR of 0.8 and 1.25 are the same size effect in opposite directions

3 more sections, plus exam facts

Premium unlocks every note across every specialty, and the full exam fact library behind it.