To further characterize the nominally significant findings (either main genotype effects or the combined effects of GxE interactions) we applied Bayesian relevance analysis [26,27] based on Bayesian networks [28]. This method applies Bayesian statistics [29] to quantify the strong relevance of predictors with respect to a selected target as probability scores (posterior probability of relevance) and allows the detailed investigation of possible effect size of predictors (i.e. odds ratios). The method performs Bayesian model averaging [30,31], both at structural and parametric levels, thus handling the multiple hypothesis problem. This approach provides full Bayesian Odds Ratio measures for the effect size of a predictor, which is a more realistic measure than e.g. a single model-based confidence interval. To cope with heterogeneity of effects in various subpopulations, suggested by the scientific literature and our PLINK analysis, we performed separate analyses in subpopulations defined by the recent life event categories, childhood adversity categories and/or age (equal or <30 and >30 years of age), respectively. All odds ratios were estimated using the ll genotype of 5-HTTLPR as a basis. (For details see Supporting Information in S1 File).
