Read I-squared, tau-squared, and the overlap of study confidence intervals together, then decide whether a pooled winter-health estimate is meaningful or should be described narratively. Cochrane Handbook Chapter 10 treats heterogeneity as part of interpreting pooled results, and GRADE rates certainty at the outcome level, so the statistics and the clinical picture belong in the same judgement.
What each statistic tells you
- I-squared describes the proportion of variability in effect estimates attributed to heterogeneity rather than chance. A high value signals that the studies disagree more than sampling error explains.
- Tau-squared estimates the between-study variance in a random-effects model. A large tau-squared means the true effect may differ across settings, which widens the prediction interval around the pooled estimate.
- Overlap of confidence intervals shows whether individual study estimates are compatible with each other. Wide, overlapping intervals suggest uncertainty; intervals that barely overlap or point in opposite directions suggest genuine conflict.
A short heterogeneity paragraph
- Report I-squared and tau-squared with the pooled estimate, and state whether the confidence intervals of the included studies overlap or conflict. If they conflict or tau-squared is large, describe the range of study findings narratively and explain why pooling would obscure that variation.
- Check the forest plot for studies that drive the heterogeneity, and note whether removing them changes the direction or size of the pooled effect. If the pooled estimate depends on one or two studies, present the results study by study.
- Rate the certainty of the pooled outcome using GRADE criteria, including inconsistency, and state the resulting judgement in the paragraph. If inconsistency lowers certainty, say so explicitly and let that rating guide whether a pooled number is reported.
A worked example: if five winter-health trials show I-squared of 70% and tau-squared that produces a wide prediction interval, and two trials report benefits while three report none, the paragraph should describe the split findings and the settings that differ, then state that a single pooled estimate would misrepresent the evidence.