Evaluating design features and analysing the intra-cluster correlation coefficients for pupil health outcomes in school-based cluster randomised controlled trials

Abstract

Cluster randomised trials (CRTs) are used in schools to evaluate interventions for improving pupil health outcomes. Little is known about the methodological practices of these studies and plausible values of the intra-cluster correlation coefficient (ICC) of pupil outcomes to inform sample size calculation for CRTs.

Systematic reviews were undertaken to identify the practices of definitive and feasibility CRTs. ICC estimates for pupil health outcomes were collated from published reports of school-based CRTs worldwide, and the relationships between these and the design and contextual characteristics of the studies were examined. A secondary analysis of raw data from five UK school-based CRTs explored patterns in ICCs for pupil social emotional functioning outcomes.

The rate of publication of school-based CRTs is increasing. Estimates of the ICC are poorly reported in such studies. Better use could be made of feasibility CRTs to assess challenges that are specific to studies that allocate school-based clusters.

The median (interquartile range; range) ICC for pupil health outcomes worldwide was 0.031 (0.011 to 0.08; 0 to 0.47) at the school level and 0.063 (0.024 to 0.1; -0.009 to 0.262) at the class level. There were no clear associations between study characteristics and the ICC, other than estimates being larger in definitive trials than feasibility CRTs.

School-level and class-level ICCs for pupil social emotional functioning outcomes reported by the same teacher for all pupils in the same class were larger than ICCs for the parent and pupil-reported versions of the same outcomes. School-level ICCs were larger in the study that sampled only one class from each school compared to the other studies that included pupils from multiple classes in each school.

When specifying an ICC for the sample size calculation for school-based CRTs, the potential impact of the different levels of clustering in the data and the outcome reporter need to be considered.