Data synthesis and statistical analysis

We used the Meta-analysis of Observational Studies in Epidemiology (MOOSE) statement23 to guide the reporting of this review.

When studies provided appropriate data, we pooled the results using a random-effects (RE) model, thus reporting the aggregate prevalence, corresponding p-value, and 95% confidence interval (95%CI). We used double arcsine transformation and normalized prevalence data after pooling and back-transformation.24 We presented the results in forest plots. We also performed a sensitivity analysis to examine whether use of a QE model22 produced a substantial difference in the results. We investigated the QE model because it accounts for study quality and leads to a distinctly conservative confidence interval when heterogeneity exists.22 When two or more studies reporting the same mental health problem were based on the same database, we selected only one for the quantitative synthesis, favoring the study that was first published. We selected this criterion because additional studies have focused on particular subgroups, which could augment their contribution to the meta-analysis results.

We assessed heterogeneity using the I² statistic. We considered an I² value of 75 to 100% to represent high heterogeneity.25,26 When at least 10 studies25 were available for a meta-analysis, we investigated heterogeneous results through subgroup analysis and meta-regression. For subgroup analyses, we considered the following characteristics: 1) gender; 2) study cycle (the Brazilian medical school years are divided into three cycles of 2 years each); 3) country region where the school is located; 4) cutoff scores (when we noted variation between studies); 5) symptom severity; and 6) risk-of-bias score. For the meta-regression, we considered: 1) proportion of male students; 2) age; and 3) risk-of-bias score. We divided studies into low ( ≥ 0.9) and high ( < 0.9) risk of bias. We assessed evidence of publication bias by Egger’s regression method,27 when at least 10 studies were available.25 We performed meta-analyses using MetaXL version 5.3 (EpiGear International, Sunrise Beach, Queensland, Australia), and carried out meta-regression and Egger’s regression method using the “metafor” function in R software version 3.2.0 (R Foundation for Statistical Computing, Vienna, Austria).