Models showing the largest variance about the mean were either rejected or down-weighted, and the arithmetic mean model was recalculated from the remaining weighted models until some kind of convergence.
When the K-means model was applied on the dataset with feature selection filtering, again five clusters (with starting iteration of 6.44) were generated, with 173, 13, five, two, and one records in each cluster, respectively.
This large difference implies that the simple arithmetic mean model is probably biased towards model C for the dipole terms.
Conversely, none of the candidate models that compared well to the arithmetic mean model are allocated full weight for all three components.
Interestingly, the residual map for the SV candidate model A when compared to the mean model is similar in shape to the residual map that was observed between the IGRF-2015 candidate model A and the mean model (Fig. 7).
However, the results for the pairwise comparisons and interaction from the cell-means model are identical to those calculated from a 2-way ANOVA [77].
The mean model is suitable for analyzing datasets with more uncertain prior knowledge, which calculates the average gene expression level.
Thus, a sire-dam variance model implies that genetic heterogeneity of residual variance in the mean model is completely explained by sire and dam effects.
Given the mean model is slightly better in terms of fit and numbers of inconsistencies this is the one recommended for use.
Equivalence between the time varying and mean models is shown through averaging theory.
Similar correlations between the differences of IGRF-2015 and SV-2015-2020 candidate models to the arithmetic mean models are found for candidate models C, E, and I.
