The 5,000 genes with the highest correlation to age were used as potential regressors. Cross-validation was used to select the number of steps in the model selection procedure. The number of predictors in the model was between 67 and 76 for the four different models. The percentage of variation explained in the training set is quite high (97%–99%) for three of the models. For the fourth, the model for Vanderbilt females, the percentage of variation explained was slightly lower, 0.92. This is a vast improvement over more naïve imputation methods that are used when adjusting for covariates with missing data, where mean values of the nonmissing data are used to fill in the missing values. Very few of the predictors we constructed were common between the different models. Given the number of predictors with high correlation to age, this is not surprising. Nonetheless, within a given data source (i.e., Pittsburgh or Vanderbilt samples), the male model is a reasonable predictor for the ages of the females and vice-versa. This same trend did not hold for predicting the ages of same-sex individuals across data sources.
