Table 3 presents risk prediction measures by ERS when outcomes were continuous (R2 and PRESS) and dichotomized (AUC). Base covariates and micronutrients explained approximately 13% of the variation for LDL, 26% for HDL, 33% for total cholesterol and 37% for triglycerides. ERS constructed with coefficients from single-pollutant models (ERS1) additionally explained variations from 0.33% for LDL to 0.72% for triglycerides. Addition of ERS1 decreased the PRESS by from 0.33% [(539.62–537.84)/539.62] for LDL to 1.1% [(967.24–956.76)/967.24] for triglycerides. When the dichotomous outcomes were used, the addition of the ERS1 only minimally modestly improved the AUC for each lipid outcome (Table 3 and Figure S2). Similar results were found with the ERS constructed with coefficients from multi-pollutant models (ERS2). Similar risk predictions were observed in the multi-phenotype approach although six new pollutants were identified in the multi-phenotype approach (Table S5).
