We obtained ethics for this study from the Ottawa Health Science Network Research Ethics Board (No 20200286-01H, CRRF ID 2228), and the protocol was registered on the Open Science Framework (OSF) website.8 Inclusion criteria were: ≥16 years; Ottawa resident for the past six months or incarcerated at the Ottawa Carleton Detention Centre; being homeless or at risk for homelessness, low-income (spending ≥ 30% of income on rent); history of substance misuse; and racialization, including being Indigenous. Based on previous studies and the target population size, we determined that a sample size of 377 would yield a 95% confidence interval and 5% margin of error, but we aimed for 416 participants since approximately 10% of potential participants refuse to participate, are ineligible, or have missing data.
At The Bridge Engagement Centre (The Bridge), community peer researchers administered a one-time survey that they co-created with academic researchers. The survey captured demographics, socioeconomic characteristics, and COVID specific knowledge including risk perception and sources of SARS-CoV-2 transmission and infection. Demographic information included ethnicity, which was determined through self-identification by each survey participant and recorded to assess the impact of COVID on employment and other socioeconomic privileges that have historically been unevenly distributed in Canada. The categories of ethnicity were white, African Caribbean Black, Indigenous, and other, where other included Arab, Latino, South and East Asian, mixed, and any other ethnicity(s) not included in the main categories. Participants compared income, employment, food insecurity, social assistance, housing, substance misuse, the justice system, health, and healthcare access before and during COVID. We defined the period prior to March 11, 2020, as being before COVID-19, and all dates from then onward being during COVID-19.9 Surveys occurred from October 2020 to November 2021.
We constructed a directed acyclic diagram to test the causality between selected demographics and employment income and used these relationships to perform Chi-squared tests to assess correlations between these variables by employing the function localTests in the R library dagitty.
We then created Bayesian mixed-effects models with the BRMS package using the following relationships:
Yi ~ Binomial(Pi, σe)
where
logit(Pi) = β0 + bParticipant[i] + βx∙Predictori
and
β x ∙Predictorsi = β1∙Time Periodi + β2∙Sexi + β3δ ∙Ethnicityi + β4∙Sex∙ Periodi + β5δ∙Ethnicity∙Periodi + β6∙Educationi
We derived informed priors from Ontario specific data of the Canadian Labor Force Survey (LFS) for September 2019 to December 2021.10 We estimated three separate unemployment models (no-work, full-time work, and part-time work). We decided standard deviations for informed priors using prior predictive checks (SD = 0.05 for the “no work” model and SD = 1 for the others). We used the Student’s t-distribution (3, 0, 2.5) for variance parameters. We lacked data for “multiple responses” and “other type of work,” so estimates were used by assuming these categories mimicked the distribution for part-time employment.
Pre-pandemic monthly labor force participation rates for Indigenous populations was not available. For these priors, we relied on LFS-2019 one-year data.11 For \({\beta }_{3Indigenous}\), we calculated odds ratios of unemployed vs. not-unemployed and employed vs. not-employed for Indigenous vs. non-Indigenous populations using the epiR software package. These ratios were then back transformed into logit estimates. The full-time or part-time status is not reported in LFS-2019. Therefore, we used the logit coefficient for full-time, part-time, multiple responses, and other work models.
We used unemployment, full-time, and part-time data for visible minorities from Census 2016H to estimate \({\beta }_{3Black}\) and \({\beta }_{3Other}\) priors as monthly/yearly LFS data was not available for our specified survey period. As with Indigenous priors, odds ratios for the three work categories were derived for African Caribbean Black and other ethnicities separately and then back transformed to logits. Estimates for the part-time category were used to define visible minority priors for “multiple responses” and “other types of work” models.
To derive stable estimates, we set iterations at 10,000 and warmup samples to 5,000. We conducted graphical prior and posterior predictive checks to assess the efficiency of our prior specifications at predicting model estimates.12 Ȓ was assessed, whereby Ȓ=1 indicated convergence. We derived multiple models for job category.13 We employed logical explanations & visualizations of estimates in linear mixed models and Bayesian reporting guidelines.14,15 We report mean values from the posterior distributions and 95% high density intervals and other measures of model fit.16 Interaction effects (odds ratios and associated 95% credible intervals (CI)) were adjusted for participant education. We also predicted probabilities averaged over educational status through post-hoc analysis. We used gtsummary, flextable, ggplot2 grafify, ggeffects emmeans, ggthemes, sjPlot, ggtext, ggpattern, dagitty, and dplyr software packages for R (v.4.1.2) for data wrangling, visualization, and reporting. We assumed the prior distributions of the intercept \({\beta }_{0}\) and fixed effects coefficients \({\beta }_{x}\) followed normal distributions with the mean (SD), \({\mu }_{0}\)(\({SD}_{0}\)), and \({\mu }_{x}\left({SD}_{x}\right)\), respectively. For odds ratios, the baseline was white and male and above high school education. In lieu of sufficient data, we assumed interaction coefficient \({\beta }_{5\delta }\) followed the same distribution as \({\beta }_{1}\). Thus, we assumed the odds of outcome of interest during COVID for men with a high school education or more was the same irrespective of ethnicity.