Differences in presence/absence of complaints between groups, at the end of the trial, were analysed using exact logistic regression, because the endpoints were binary.
Differences in average number of complaints between groups, at the end of the trial, were analysed using linear regression, including the number of complaints at the start of the trial as covariate.
Fasting plasma glucose at the end of the trial was analysed using a similar model, replacing baseline HbA1c with baseline fasting plasma glucose.
The quality of the trials was analysed using the Jadad score.
The significance of the variability among the trials was analysed using GraphPad Prism (version 5.0) software.
We conducted intention to treat analyses, and numerical outcomes of the randomised controlled trial were analysed using a multilevel linear model, accounting for possible cluster effects caused by the cluster randomisation.
Error rate and latency of correct trials were analysed using randomized block ANOVAs with Group ASDD versus controls) as the between-subjects factor, Task (antisaccade versus prosaccade) as the within-subjects factor and subjects nested within group as the random factor.
Root data in both trials were analysed using a two-step method, similar to that used previously (Douglas et al. 2010), comprising (1) analysis of the proportion of width × depth × year combinations (cells) with roots and (2) analysis of data for cells where roots were present.
Differences in bat activity (bat counts and bat passes), bat foraging activity (feeding buzz ratios) and insect abundance between experimental and control trials were analysed using paired t tests.
In order to evaluate the potential utility of SMS depression scores to detect the group differences over the same time period among those opting in to the SMS messaging, trajectories of change across the three ACUDep trial arms were analysed using a random slope linear mixed model.
