For all traits, we show that pPS calculated with even a small portion of available genome are capable of significantly improving the phenotype prediction accuracy, to an extent that depends on the SNPs included in that portion and its size (Fig. 3, likelihood ratio test). Moreover, pPS calculated using certain local subsetting patterns tend to perform better than others, regardless of the amount of SNPs included. In essence, on the one hand we observe a macroscopic trend which broadly agrees with the hypothesis that a given genomic subset contributes additively to the phenotypes studied. On the other hand, each subsetting pattern has its own phenotype prediction performance, which we can estimate by using that particular pattern on a set of control individuals as shown above. These results suggest that in individuals where a fraction of the genome is missing or assigned to another ancestry (as in the case of Fig. 2), the pPS can be used as a proxy for the actual PS.
