Strata Academy
Effect Measures for Meta-Analysis: RR, OR, HR, MD & SMD
Choosing and interpreting risk ratios, odds ratios, hazard ratios, mean differences, and standardised mean differences in systematic reviews
Quick answer
Choose effect measures that match your outcome type: RR or OR for dichotomous events (RR preferred when events are common), HR for time-to-event, MD for same-scale continuous outcomes, SMD when scales differ. Meta-analysis pools on a log scale for ratios; forest plot null lines are at 1 for RR/OR/HR and 0 for MD/SMD. Never mix incompatible measures in one forest plot.
- RR is usually preferred over OR when events are not rare — OR exaggerates effects.
- MD preserves original units; SMD is for different instruments measuring the same construct.
- HR requires time-to-event data — do not substitute log-rank p-values without estimates.
- One forest plot = one effect measure on one scale — check labels before interpreting.
1. Why effect measure choice matters
Meta-analysis combines estimates that measure the same treatment effect on the same scale. Choosing the wrong effect measure distorts pooled results, confuses readers, and triggers GRADE indirectness or inconsistency concerns.
The effect measure should be pre-specified in PROSPERO or the review protocol. Post-hoc switching from OR to RR after seeing results invites selective reporting criticism.
Software (RevMan, R meta, Stata metan) will pool numbers you provide — your methodological job is to ensure those numbers answer a coherent clinical question.
- Dichotomous outcomes → RR, OR, risk difference, or Peto OR (rare events)
- Time-to-event → hazard ratio (log HR)
- Continuous same scale → mean difference (MD)
- Continuous different scales → standardised mean difference (SMD)
2. Risk ratio (RR) vs odds ratio (OR)
Risk ratio compares event probabilities between groups: (events in intervention / n intervention) ÷ (events in control / n control). It is intuitive and aligns with absolute risk reduction when baseline risk is known.
Odds ratio compares odds of events, not probabilities. When events are common (>10%), OR diverges from RR and can exaggerate apparent benefit or harm. Cochrane prefers RR for interpretability when data allow.
Both are pooled on the log scale in meta-analysis. Forest plots display OR or RR with null line at 1.0 — values left of 1 favour control; right favour intervention (for beneficial outcomes framed as reduction).
| Measure | Best when | Null value | Common pitfall |
|---|---|---|---|
| RR | Events not extremely rare; RCTs with complete follow-up | 1 | Ignoring baseline risk for absolute effects |
| OR | Case-control studies; logistic regression outputs | 1 | Treating OR as RR when events are common |
| Risk difference | Absolute effects needed directly | 0 | Harder to pool when baselines differ widely |
| Peto OR | Rare events in fixed trials | 1 | Not for large effects or imbalanced arms |
Note: If included papers report only OR but events are common, re-analyse from 2×2 tables to RR where possible — or clearly justify OR and avoid overstating clinical importance.
3. Hazard ratio (HR) for time-to-event outcomes
Hazard ratio compares event rates over time between groups, accounting for censoring. Use HR when the outcome is time-to-death, time-to-relapse, or similar survival endpoints.
Meta-analysis pools log HRs with their standard errors — usually extracted from published Kaplan–Meier analyses, Cox regression tables, or via digitisation methods when only figures exist.
Do not substitute a significant log-rank p-value without a hazard ratio estimate. Do not pool HRs with RRs on the same forest plot.
- HR < 1 → lower hazard in intervention group (for adverse outcomes framed as events)
- Requires comparable follow-up and censoring assumptions across trials
- Check whether HR is from Cox model adjusted for covariates vs unadjusted
- Immature survival data (few events) → imprecise HRs and GRADE downgrades
4. Mean difference (MD) vs standardised mean difference (SMD)
Mean difference pools continuous outcomes measured on the same scale (e.g. mm Hg, points on the same questionnaire). The pooled MD is in original units — directly comparable to MCID thresholds.
Standardised mean difference (Cohen's d or Hedges' g) pools trials using different instruments measuring the same construct (e.g. multiple depression scales). SMD is unitless — harder to interpret clinically without back-translation.
Rule of thumb: if all trials use the same validated scale, prefer MD. If scales differ but construct is shared, use SMD and explain clinical meaning cautiously.
- MD null value on forest plot → 0
- SMD null value → 0; |0.2| small, |0.5| medium, |0.8| large — context-dependent
- Change scores vs follow-up scores — extract consistently
- Correlated outcomes in same trial — avoid double-counting in synthesis
5. Pooling rules and heterogeneity
Pool only studies estimating the same contrast on the same scale: intervention vs comparator, same time point, same outcome definition.
Random-effects models are Cochrane's default when clinical heterogeneity is plausible. Fixed-effect assumes one true effect — large trials dominate.
High I² does not automatically forbid pooling but demands investigation. If studies measure different constructs, switch to narrative synthesis rather than forcing SMD.
6. Reading effect measures on forest plots
Before interpreting any diamond, read the axis label: OR, RR, HR, MD, or SMD. The null line position follows the measure — 1 for ratios, 0 for differences.
Study weights reflect precision (and model choice), not study quality. A biased but large trial can pull the pool — pair forest reading with ROB 2.
Use our interactive forest plot to practise identifying measure type, null line, and pooled diamond crossing.
- Read effect measure label and null value
- Confirm all studies use compatible contrasts
- Check whether model is random or fixed effects
- Note whether diamond crosses null — statistical significance
- Translate ratio measures to absolute effects for clinical significance
- Read heterogeneity statistics below the plot
7. Extraction tips for student dissertations
Build your extraction spreadsheet with columns for events and totals (for RR/OR), means and SDs (for MD), or HR and log SE. Record whether estimates are adjusted or ITT.
When papers report multiple time points, extract the pre-specified primary time point from your protocol — not the most favourable post-hoc time point.
Contact authors for missing SDs or incomplete 2×2 tables before excluding studies. Document imputation rules if using published methods (e.g. Wan et al. for medians).
Tip: Pilot extraction on two papers and run a test meta-analysis in RevMan or R before extracting the full set — catches scale mismatches early.
Frequently asked questions
When should I use RR instead of OR in meta-analysis?
Prefer RR for RCTs with common events — it is more interpretable and aligns with absolute risk. OR is acceptable for rare events or when only OR is reported, but do not interpret OR as RR when events exceed roughly 10%.
What is the null value on a forest plot?
For RR, OR, and HR the null is 1 (no relative effect). For mean difference and SMD the null is 0 (no difference). Misreading the null line is a common student error.
Can I pool HRs and RRs together?
No. They measure effects on different scales and from different analysis frameworks. Choose the outcome type appropriate to your data and extract consistently.
When should I use SMD instead of mean difference?
Use SMD when trials measure the same construct with different instruments or units and MD pooling is impossible. Prefer MD when all trials use the same scale — it is easier to interpret against MCIDs.
What if studies report median instead of mean?
Use established imputation methods to estimate mean and SD, or contact authors. Do not enter medians as means without transformation — it biases the pool.
Interactive walkthroughs and quizzes load when JavaScript is enabled — the checklist and tables above are fully readable without it.