Statistical analysis

Observations from participants who refused to respond to at least one item on either the HSCL-25 or HTQ-16 were excluded from the analysis. Data analysis was conducted in Stata V.13.1 (StataCorp, Texas, USA).³¹ We conducted descriptive analysis on respondent characteristics using frequencies, means, SD and 95% CI. Data were then weighted to account for the sampling design and for the over-representation of females in the sample, and prevalence estimates were calculated for probable anxiety, depression and PTSD using the prior stated cut-points. We used the complex survey design command in Stata to apply probability weights. Sensitivity tests were conducted on sample weights to assess differences in prevalence estimates after poststratification weighting on gender. SEs were estimated using the Taylor series linearisation method to adjust for design effects.³² The Wilcoxon-Mann-Whitney test was utilised to compare continuous variables. The Χ^{2 2}was used to test for associations between categorical variables. The psychometric properties of the instruments were evaluated using Cronbach’s alpha for internal reliability, where an alpha of 0.7 was considered to be an acceptable reliability coefficient and 0.9 or larger, excellent reliability. ³³

Univariate logistic analysis was used to calculate crude ORs with 95% CIs, identifying evidence of association with study outcomes (depression, anxiety and PTSD) and a priori risk factors, that is, sex, age group, marital status, education, main daily activity, rural or urban residence and exposure to traumatic events. Variables that showed a significant association with the outcome of interest, with a p value of less than 0.25 were included in the multivariable analysis; with separate models created for each study outcome. ^34–36 Adjusted ORs and adjusted Wald tests were calculated.³⁵ Forward step-wise regression was utilised to build the final main effects model.³⁵ Backward step-wise regression was then used to remove variables with an adjusted Wald test > 0.05. In the final model, only exposure factors which had adjusted ORs >1.0 with a p-value <0.05 were considered statistically significant predictors of the study outcomes.³⁵ Variance inflation factors were estimated to check for collinearity.³⁵