*2.4. Analyses*

Statistical analyses were conducted using commercially-available statistical software, SAS v. 9.4 (SAS Institute Inc., Cary, NC, USA) [31] and Mplus v. 8.0 (Muthén & Muthén, Los Angeles, CA, USA) [32], and tests were two-sided with significance set at *p* = 0.05. Workplaces represented by OLIP questionnaire responses were statistically weighted to permit inferences from the sample to a comparable population of Ontario organizations based on strata of workplace size, region, and industry sector. For RQ1, mean values of OHS and wellness activities were compared separately according to workplace size, union status, industry sector, health and safety leadership, and people-oriented culture. Analysis of variance was used to examine differences in the mean OHS and wellness activities scores. Workplaces were assigned to 'profiles' based on the probability that they had similar numbers of OHS and wellness activities to other workplaces using Mplus's latent profile analysis function. The latent profile analysis statistical technique aims to recover hidden groups from observed data, similar to clustering techniques, but is more flexible because the approach is based on an explicit model of the data, and accounts for the fact that recovered groups are uncertain [33]. Data on OHS performance scores and number of wellness activities were transferred from SAS to Mplus and analyzed as continuous variables in a mixture model with sample weights. Several models were fit with increasing numbers of profiles (one profile, two profiles, three profiles etc.). A decision on the most suitable number of profiles fitting the data was made by inspecting model-fit statistics for the Lo–Mendell–Rubin adjusted likelihood ratio test. The Lo–Mendell–Rubin test had a *p*-value of 0.58 when comparing four profiles to three profiles, suggesting that three profiles sufficiently modelled the data. For RQ2, associations between the latent profile groups and workplace demographic characteristics (independent variables) were estimated using multinomial logistic regression by transferring latent profile probability data generated from Mplus back into SAS and matching them to corresponding data from individual survey respondents. The odds of a co-occurrence profile associated with a workplace characteristic of interest compared to the odds of the lowest co-occurrence profile

and a reference workplace characteristic (e.g., a small workplace without a JHSC) were described as odds ratios (OR) and 95% confidence intervals (CIs).
