Abstract
BACKGROUND: HIV transmission and disease progression may be driven by associations HIV risk behaviors have with a constellation of alcohol, other substance, and mood-related conditions (CASM). However, observational study-based measures of these associations are often prone to unmeasured confounding. While meta-analysis offers a systematic approach to summarize effect sizes across studies, the validity of these estimates can be compromised if similar biases exist across studies. Our analysis assesses the likelihood that unmeasured confounding explains meta-analysis-derived measures of association between CASM and HIV risk behaviors, and provides bias-adjusted estimates. METHODS: We first conducted systematic reviews and meta-analyses to assess associations between CASM conditions and four HIV risk behaviors (medication non-adherence, unprotected sex, transactional sex, and multiple sexual partners). We then adjusted for potential unmeasured confounders using two methods designed for meta-analyses - Point Estimate and Proportion of Meaningfully Strong Effects methods. We selected "risk propensity" as an illustrative and potentially important unmeasured confounder based on the extant literature and mechanistic plausibility. RESULTS: In analyses unadjusted for unmeasured confounding, 89% (24/27) of odds ratios (ORs) show strong evidence of a positive association, with alcohol use and stimulant use emerging as dominant risk factors for HIV risk behaviors. After adjusting for unmeasured confounding by risk propensity, 81% (22/27) of ORs still showed strong evidence of a positive association. Associations between mood-related conditions and HIV risk behaviors were more robust to unmeasured confounding than associations between alcohol use and other substance use and HIV risk behaviors. CONCLUSION: Despite residual confounding present in constituent studies, there remains strong evidence of associations between CASM and HIV risk behaviors as well as the clustered nature of CASM conditions. Our analysis provides an example of how to assess unmeasured confounding in meta-analysis-derived measures of association.