**Figure 2.**

Performance of the Gaussian model (GAUSS). The figure depicts the effect of the Gaussian scale parameter (θ in Eq. [6]) on the restricted log-likelihood (LL), the training population accuracy (*r*_{train}), and the cross-validation accuracy (*r*_{pred}) when predicting sets 5 or 6 in environment 1. For set 5 the restricted maximum likelihood solution for θ (θ_{REML}) = 0.5, and for set 6 θ_{REML} = 0.4. In both cases *r*_{train} approached 1 as θ → 0, but the trends for *r*_{pred} were different. For set 5 *r*_{pred} exhibited an interior maximum near θ_{REML}, while for set 6 *r*_{pred} increased monotonically with θ. Because GAUSS is approximately ridge regression (RR) when θ is large, the contrasting behavior in this figure illustrates why GAUSS had higher *r*_{pred} than RR for set 5 but vice versa for set 6 (see Table 2).