Research & EvidenceEvidence SynthesisAdvanced LevelProfessionals
Meta-Analysis Heterogeneity: I², Tau² and Prediction Intervals Explained
A practical, evidence-informed guide to meta-analysis heterogeneity: i², tau² and prediction intervals explained, with safety screening, clinical reasoning, intervention progression and outcome-based reassessment.
11 October 20269–12 min readEvidence sources included
What You Should Be Able to Explain After Reading
- Differentiate clinical, methodological and statistical heterogeneity.
- Interpret I² and Tau² without rigid cutoffs.
- Explain why a pooled confidence interval is not a prediction interval.
Start with the clinical question
- Before pooling rehabilitation trials, compare populations, diagnosis definitions, treatment dose, comparator, follow-up and outcome scales. Statistical software cannot fix a clinically incoherent comparison.
- Heterogeneity can be clinical (patients/interventions), methodological (study design or bias) or statistical (effects differ more than expected from sampling error).
What I² means
- I² estimates the proportion of observed variability in study effect estimates attributed to heterogeneity rather than chance, under the model assumptions.
- It is not 'the percentage of patients who differ' and not the size of the treatment effect. With few studies its uncertainty is often substantial.
What Tau² means
- Tau² estimates between-study variance in true effects on the analysis scale; Tau is its square root and has the same units as the effect measure.
- Tau² values cannot be interpreted meaningfully without considering the outcome scale and clinical context.
Confidence interval versus prediction interval
- A 95% confidence interval around the random-effects pooled mean describes uncertainty about the average effect.
- A 95% prediction interval aims to describe the range of underlying effects expected in a similar new study, incorporating between-study variation; it may be wide and unstable with few studies.
Worked clinical interpretation
- Suppose pooled pain improvement favors treatment, but studies range from small to large effects and a prediction interval crosses no effect. The average may be beneficial while a new setting could experience little benefit.
- Ask whether intervention dose, chronicity or comparator differences plausibly explain the variation before applying the pooled estimate to an individual patient.
Analytic workflow
- Pre-specify eligible comparisons and subgroup hypotheses. Check extraction units, standard deviations versus standard errors, outcome direction and missing data before attributing differences to biology.
- Consider random-effects analysis when justified, sensitivity analyses for risk of bias, and narrative synthesis if pooling would mislead.
Reporting and limitations
- Report effect estimate, 95% confidence interval, I², Tau² where available, prediction interval when appropriate, number of studies and clinical differences.
- Avoid rigid 'low/moderate/high' I² labels, especially with only two or three trials; heterogeneity statistics are imprecise in small meta-analyses.
Common mistakes and summary
- A significant pooled P value does not guarantee consistent benefit. Good evidence synthesis explains the spread of plausible effects and the conditions under which the result applies.
Evidence & Further Reading
These sources are provided so learners can trace the clinical recommendations and read the original guidance.