*2.10. Data Analysis*

All data analyses were completed in SAS Version 9.3 (SAS Institute Inc., Cary, NC, USA). First, we compared worker demographics between control and intervention sites using chi-squared tests of homogeneity for categorical variables and *t*-tests for continuous variables. A priori power calculations were conducted for the primary outcome, pain and injury, adjusting for potential intra-class correlation, ICC = 0.05 due to the cluster-based design, and using a two-sided test at α = 0.05. We have sufficient power (>0.8) to detect at effect size greater than 0.6 with an estimated sample size of 176.

As pain and injury outcomes were binary measures, we first performed logistic regression models accounting only for the baseline level of the outcome variables. Each model also utilized cluster robust standard errors to account for individual correlation within worksites. Second, we included fixed effects for the matched pairs within each company, and adjusted the models for age, race, and job title.

All other variables were continuous. We conducted linear regression models on the change scores between baseline and FU1 and, baseline and FU2 as the dependent variables and treatment status (intervention; control) as the independent variable. We used cluster robust standard errors to account for individual clustering within worksites. We then adjusted for matched pairs within the companies through the addition of a fixed effect and also accounted for the possibility of post-randomization, and residual confounding by adjusting for age, sex, race, job title, and trade. No analyses were conducted for smoking as there were too few smokers who changed their smoking status over the course of the intervention on the sites.

We conducted sensitivity analyses to observe initially whether the removal of the one matched pair for the site that did not perform the intervention activities for the soft tissue ergonomics program, resulted in any differences in effect of the intervention on the primary and secondary outcomes. We then sequentially removed each pair per analysis to evaluate whether removal of any pair resulted in differences in the effectiveness evaluation.
