Inspect study-level characteristics and subgroup or sensitivity analyses to find a plausible source of heterogeneity, then decide whether pooling remains defensible. If no explanation emerges, switch to a narrative synthesis and state why pooling was set aside.
Heterogeneity here means variation in effect estimates across the included studies beyond what chance alone would produce. The Cochrane Handbook treats heterogeneity assessment as part of interpreting pooled results, alongside the effect measure and the choice between fixed-effect and random-effects models.
- Recheck the data before interpreting the statistic. Confirm that each study's effect measure, direction of benefit, and variance are coded consistently, and that the same comparison and outcome are being pooled. A miscoded direction or a mixed outcome definition can produce apparent heterogeneity that disappears once corrected.
- Compare study-level characteristics across the estimates. List candidate modifiers such as population age or health status, winter exposure definition, intervention dose or timing, follow-up length, and risk of bias. Sort the studies by each characteristic and see whether effect sizes track it.
- Run subgroup analyses on the characteristics that plausibly modify the effect. Report the effect within each subgroup and the test for subgroup differences. Treat a subgroup result as a hypothesis about the source of heterogeneity, especially when subgroups are small or defined after looking at the data.
- Run sensitivity analyses by removing one study at a time and by excluding studies at high risk of bias or with imputed data. Note whether the pooled estimate and the heterogeneity measure change materially. A single influential study points to that study's design or population as a candidate explanation.
- Decide whether pooling remains defensible. If a credible study-level explanation accounts for the spread, present the subgroup findings alongside the overall estimate and describe the scope of the pooled result. If heterogeneity stays unexplained and the estimates conflict in direction or magnitude, present a narrative synthesis organized by the candidate modifiers and state the rationale for not pooling.
Report the heterogeneity measure with its confidence interval and the exact P value rather than a threshold-based label, following the Cochrane guidance on interpreting results. When judging how much certainty to place in the conclusion, GRADE's outcome-level criteria give a structured way to record concerns such as inconsistency across studies.