Our base model included age, gender, race/ethnicity, education and BMI as was also done by Patel et al. [5], [6]. Then we selected important micronutrients corresponding to each phenotype using the full data (stage 1 and 2 samples combined). Specifically, we first regressed each phenotype on the set of covariates in the base model to obtain the residuals, and then used the residuals as the outcome to select the micronutrients. For micronutrient selection we applied the Bayesian model averaging technique (BMA) to jointly analyze all micronutrients and select the ones with posterior inclusion probability greater than 0.8 (see Sun et al. [10] for details). Other simpler methods (e.g., best subset regression) may also be used at this step.
