To test the discriminative accuracy of models to predict obesity classification, ROC curves were plotted and the corresponding AUCs were calculated. Three sets of nested models were tested: 1) covariates (PCs, sex, age, ancestry by sex interaction), 2) covariates, SNP-GRSS and interaction with sex and 3) covariates, SNP-GRSS and three obesity-associated CNVs (the 21 kb deletion on 16p.12.3, the 66 kb duplication on 1p36.1, and the 1440 kb deletion on 5q13.2). Table 4 displays fit statistics from ROC curve analysis by BMI category (Additional file 6: Table S6 displays by ancestry). AUC estimates indicated the models significantly predicted overweight and obesity classification with maximum discriminative ability when employing model 3 to predict class III obesity (AUC = 0.750, 95% CI = [0.702, 0.797]). Models that included genetic information had significantly greater AUCs than models only including covariates (Table 4).Table 4 Discriminative accuracy of covariates, SNP-GRSS and CNV predicting BMI category in European- and African-Americans ModelAUC95% CIAsy. Sig. of ModelΔ AUC% Δ AUC p Δ AUC Overweight: n = 1443 (61.4%) 1. Covariates 0.679[0.657,0.700]2.68×10−48 --- 2. Model 1 + SNP-GRSS 0.692[0.671,0.714]9.23×10−56
