First, we calculated summary statistics, such as frequency distributions of categorical variables or means and standard deviations of numeric variables, to describe the sample and characterize their COVID-19 experiences. In addition to overall frequencies and summaries, we produced stratified estimates by gender and tested for bivariate associations. Where possible, analyses were adjusted using the geodemographic survey weights described above in the survey package (Lumley, 2021) for R version 4.0.3 (R Core Team, 2020). In circumstances where small expected cell counts made survey adjustment unreliable, unadjusted tests used Fisher’s exact test. We then performed a survey weighted logistic regression, modelling the log-odds of relapse overall and stratified by gender. Collinearity was assessed using (generalized) variance inflation factors provided by the car package (Fox and Weisberg, 2019), which were all less than 2.1 for the largest model, compared to a typical cutoff for problematic multicollinearity of 10. Final estimated coefficients were also graphically assessed for stability through comparison to the result from unweighted penalized logistic regression (specifically, a LASSO model) implemented with the glmnet package (Friedman et al., 2010). Penalized regression techniques
