Effect sizes in child health and association with unreported outcomes

Article type
Authors
Hartling L1, Fernandes R1, Dryden D1, Vandermeer B1
1University of Alberta, Canada
Abstract
Background: Research has shown that half of randomized controlled trials (RCTs) included in Cochrane systematic reviews do not contribute to the meta-analysis of patient-important outcomes. Further, previous research showed an inverse relationship with effect size and the proportion of RCTs contributing to the meta-analysis. One potential explanation for this finding is selective outcome reporting.

Objectives: To describe effect sizes in pediatric RCTs and the association between the proportion of relevant RCTs contributing to a meta-analysis and the magnitude of effect.

Methods: Using the Cochrane Database of Systematic Reviews we identified systematic reviews relevant to child health. We identified the primary outcome for each systematic review and extracted data when a meta-analysis had been conducted (n = 432). We calculated an overall effect size for meta-analyses based on the proportion of relevant RCTs included using categories from previous research: < 20, 20 to < 40, 40 to < 60, 60 to < 80, ≥ 80%.

Results: The median effect size across all meta-analyses was 0.35 (IQR 0.15,0.59) which is considered small to moderate. The mean proportion of relevant studies contributing to a meta-analysis was 0.49 (95%CI 0.47, 0.52). There was no association between effect sizes and the proportion of RCTs contributing to a meta-analysis. However, effect sizes were associated with the size of the meta-analysis with smaller effect sizes seen with larger meta-analyses (i.e., smaller confidence intervals).

Conclusions: We found an association between effect sizes and the size of meta-analysis. This effect is similar to that seen for publication bias at the individual study level. Caution should be taken when interpreting meta-analyses based on small numbers of studies and patients, which is consistent with the GRADE recommendations for considering imprecision. Effect sizes were small to moderate overall. These data may provide a guide for sample size calculations in future pediatric trials.