Fig. 3

Impact of clustering individuals to groups on model fit and prediction accuracy. The impact was evaluated for three real traits in three species: short-day vernalization (SDV), weight growth intercept, and plant height for Arabidopsis, mice, and maize, respectively. The entire population in each species was randomly divided into reference (80%) to exam model fit and the rest as inference to exam prediction accuracy. Clustering individuals into groups was based on kinship among individuals from both references and inferences; however, the phenotypes of inferences were masked. When number of groups equaled number of individuals, the prediction was equivalent to the prediction resulting from the conventional genomic BLUP (gBLUP) method. When a group contained more than one individual, the prediction for a group was used as the prediction for each individual within the group. The phenotypes of references were used to evaluate the model fit (top panel: a–c), as indicated by twice the negative log likelihood (−2LL). Prediction accuracies for cBLUP were evaluated as the correlations between predicted and observed phenotypes of inferences (bottom panel: d–f). Initially, both model fit and prediction accuracy increased with increasing number of groups, but decreased after reaching an optimum peak. The optimum peak of model fit corresponded to the optimum peak of prediction accuracy for compressed BLUP (cBLUP). This trend was consistent for all the traits in the three species we examined, as demonstrated on a replicate for Arabidopsis (a, d), mice (b, e), and maize (c, f). The trend was also consistent across replicates of randomly assigning reference and inference, however, the magnitude of likelihood, accuracy, and number of groups at the peaks varied slightly among replicates