Extended Data Fig. 3: Bias and 95% CIs of full prediction models, as a function of measured VAT.

a,b, VAT^, as predicted from the leave-one-out cross-validation, plotted against measured VAT mass for females (a; black circles) and males (b; red circles). Also plotted is VAT^ against measured VAT mass for the out-of-sample data (green dots in a and blue dots in b). The out-of-sample datasets constitute Irish and other white individuals from the UKBB, excluding white British. The sample sizes are n = 2,010 females and n = 2,188 males for the training datasets, and n = 119 females and n = 102 males for the out-of-sample datasets. The long-dashed, gray lines denote the linear fits (ordinary least squares regression) to the leave-one-out cross-validation data, and the gray (a) and red shaded areas (b) denote the corresponding CIs of the estimated slopes. Green (a) and blue shaded areas (b) denote the CIs of the linear fits (not plotted) to the out-of-sample data. Thin black lines denote the one-to-one relation. A slope below the one-to-one relation indicates that a small bias is present in the data. However, note that the attenuation is exaggerated due to measurement errors also in the measured VAT mass. c,d, VAT prediction residuals plotted against measured VAT mass for females (c) and males (d). The long-dashed lines correspond to the fitted regression lines in a and b. The gray, solid lines denote the conditional 95% CIs. These lines become dashed at high VAT mass, to indicate an increasing uncertainty in the CIs. Otherwise, symbols are as in a and b.