We performed this meta-analysis following the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines20.
Literature Search and Selection
The relevant articles were systematically searched in PubMed, Embase, Wed of Science, Chinese National Knowledge Infrastructure, China Biology Medicinedisc, Chongqing VIP database for Chinese Technical Periodicals, and Wanfang Databases from inception to April 1, 2019. A combination of the following search terms was used: ((( ‘man’ OR ‘male’ OR ‘men’) AND (homosexual* OR bisexual* OR androphilia OR ‘same-sex’)) OR ‘men who have sex with men’ OR ‘gay’) AND (suicidal OR suicide) AND (Chinese OR China).Additionally, a manual search was performed to screening reference lists of the previous systematic reviews and other relevant articles to find potential relevant studies.
All retrieved literature were imported into Endnote X8.2 software to remove duplicates. Eligible studies were included in the review if they: 1) were conducted among MSM in mainland China; 2) were conducted after 2005; 3)were published either in English or Chinese; 4) provided adequate information on the prevalence estimation of suicidal ideation or suicide attempt. On the contrary, studies that met any of the following criteria were excluded: 1) studies that were systematic reviews or qualitative interview studies; 2) studies that were duplicate publications; 3) studies that didn’t provide the prevalence estimation or sufficient data for the prevalence calculation; 4)studies that were conducted before 2005 or didn’t provide survey dates. Additionally, for multiple studies using the same sample, only studies that presented detailed and comprehensive information were retained.
All the work above was independently completed by two researchers. Any discrepancy was resolved by discussion to reach a consensus.
Using a standardized form, two researchers independently extracted the following information from each eligible study: first author, publication year, survey dates, study location (city), sampling method, method of data collection, average age of subjects, sample size, number of events, survey point (i.e. past one month, past six months, past one year or lifetime), characteristic of subjects (i.e. status of HIV infection). Inconsistent data extraction were resolved by discussion or reconciled by the primary author if necessary.
Due to the methodological limitations concerned with the special population, traditional quality assessment tools were not well-suited to appraise individual articles. For example, the representativeness of sample, an important indicator of quality in the assessment tools, was poor as a result of non-probability sampling used in nearly all included studies. Therefore, following the Meta-analysis of Observational Studies inEpidemiologyguidelines21, we instead analyzed the effect of risk-of-bias variables (i.e. sampling method and method of date collection) on the pooled prevalence estimates, as did in the previous studies9 22.
All statistical analyses were performed using the meta package in R 3.5.0 software. In view of the potential between-study heterogeneity, the random-effects meta-analyses were adopted to calculate the pooled prevalence estimates of suicidal ideation and suicide attempts with their corresponding95% confidence intervals (CIs). Heterogeneity between studies was assessed using Cochran’s chi-squared test(Cochran’sQ) and I2values, in the case of which P< 0.1 or I2≥ 50% was considered moderate or high heterogeneity.
Subgroup analyses were conducted to examine whether the pooled prevalence estimates differed by survey dates (2010 before vs 2010 after), study location(northern city vs southern city), sampling method (respondent-driving sampling vs non respondent-driving sampling), method of data collection (face-to-face interview and self-administered), survey point, HIV infection (HIV-positive vs HIV-negative). A random-effects meta-regression was performed to identify the impact of covariates (i.e. average age and sample size) on the effect sizes.
Besides, sensitivity analysis was conducted by serially excluding each included study to detect the effects of single studies on the robustness of the combined prevalence. For the assessment of publication bias, a visual inspection of funnel plots along with Egger’s tests was used. All tests were two-sided and a P-value<0.05 was considered statistically significant.