Estimates of effectiveness of hand sanitiser on outcome measures.

Estimates obtained from marginal models using GEEs with an exchangeable correlation structure and robust variance estimation. All models include the stratification variable “city” (Invercargill, Dunedin, or Christchurch).*ICC point estimates are those resulting from the GEE models with no adjustment for the stratification variable. Confidence intervals for the ICCs were bootstrapped using the combination of the bootstrap and xtgee commands in Stata. Bootstrapping allowed for the clustering of observations within schools (using both the cluster() and idcluster() options). Bias-corrected 95% bootstrap confidence intervals were calculated from 5,000 replicates.‡The exposure period was the number of school days.↑The exposure period was the period the child was enrolled in the study minus the length of the school holidays.†The ICC point estimate resulting from the GEE model for these outcomes was negative. In this circumstance, the model GEE model was refitted with an independent correlation structure, making the assumption that in the context of a cluster-based evaluation such as this, negative ICCs are more likely to occur through sampling error than because of a true negative ICC [20],[21]. Assuming an independent correlation provides more conservative estimates of the estimated standard errors.**Data aggregated to the level of the school, and analysed at the school level. The exposure period was calculated as the average school roll over the period of the trial multiplied by 100 (the number of school days encompassed by the trial period).