Figure 1.

True prediction accuracy of single-trait and multi-trait prediction methods in simulated data. 500 simulations were run for each heritability of the secondary trait ($h_2^2=\left\{0.2,0.6\right\}$), and each combination of genetic and non-genetic correlation between the two traits (${\rho }_g=\left\{0,0.3,0.6\right\},{\rho }_R=\left\{-0.6,-0.4,-0.2,0,0.2,0,4,0.6\right\}$), all with $h_1^2=0.2$. For each simulation, we used 90% of the individuals as training to fit linear mixed models (either single or multi-trait), predicted the genetic values of the remaining validation individuals, and then measured the Pearson’s correlation between the predicted (${\widehat{u}}_{n1}$) and true ($u_{n1}$) genetic values. In the CV1 method, we used only information on the training individuals to calculate ${\widehat{u}}_{n1}$. In the CV2 method, we used the training individuals to calculate ${\widehat{u}}_o$ and combined this with the observed phenotypes for the secondary trait on the validation individuals ($y_{n2}$). Curves show the average correlation for each method across the 500 simulations. Ribbons show $\pm 1.96\times SE$ over the 500 simulations.