Data analyses

We estimated sample size using the suggested calculation from (Gupta, Attri, Singh, Kaur, & Kaur, 2016), based on the following parameters: (1) on the SESMQ, an expected average score of 84, a minimal difference considered real (MDR) score of 26 points, and standard deviation of 35.8 (Jennings et al. 2014); and (2) a conservative 20 percent loss to follow-up. A total enrollment of 30 participants (15 per group) was estimated to produce 80 percent power for a two-tailed test with an alpha error of 5 percent.

To test for differences between groups on demographic data, analyses were done using independent sample t-tests the chi-square test or Fisher’s exact test. Linear mixed effect models were used to test differences between groups (CHW and non-CHW facilitator) across the timepoints. Fixed effects for time, group, and their interaction were included. A random intercept per subject was included to account for the possible correlation among data points for the same subject. Models were fitted using the lmer function in the statistical software R (R Core Team, 2017). P-values were obtained using Satterthwaite approximations for degrees of freedom, and were considered statistically significant at p < 0.05 (Bates et al., 2015). An intent-to-treat (ITT) approach was used, meaning all randomized participants were included in the analyses (Fisher et al., 1990). Post hoc analyses were carried out if significant main effects were found, using Bonferroni corrections for multiple testing. We used Cook’s Distance to compute the estimated degree of influence exerted by each data point on each of the predicted outcomes (Cook, 1977). Data points with large Cook’s Distance (a percentile greater than 50 using the F-distribution) were investigated further for validity, and models were fit without those points to identify if there was a change in significance (Foley, 2019). A linear mixed effect model is robust in the presence of data that are missing completely at random (MCAR) due to likelihood-based estimation (Ibrahim et al., 2005). To investigate missingness, potential relationships between missing data and other variables (age, degree of hearing loss, sex) were conducted using t-tests and Fisher’s exact test.