Use of direct and iterative solvers for estimation of SNP effects in genome-wide selection

The aim of this study was to compare iterative and direct solvers for estimation of marker effects in genomic selection. One iterative and two direct methods were used: Gauss-Seidel with Residual Update, Cholesky Decomposition and Gentleman-Givens rotations. For resembling different scenarios with respect to number of markers and of genotyped animals, a simulated data set divided into 25 subsets was used. Number of markers ranged from 1,200 to 5,925 and number of animals ranged from 1,200 to 5,865. Methods were also applied to real data comprising 3081 individuals genotyped for 45181 SNPs. Results from simulated data showed that the iterative solver was substantially faster than direct methods for larger numbers of markers. Use of a direct solver may allow for computing (co)variances of SNP effects. When applied to real data, performance of the iterative method varied substantially, depending on the level of ill-conditioning of the coefficient matrix. From results with real data, Gentleman-Givens rotations would be the method of choice in this particular application as it provided an exact solution within a fairly reasonable time frame (less than two hours). It would indeed be the preferred method whenever computer resources allow its use.

Use of ridge regression for the prediction of early growth performance in crossbred calves

The problem of multicollinearity in regression analysis was studied. Ridge regression (RR) techniques were used to estimate parameters affecting the performance of crossbred calves raised in tropical and subtropical regions by a model including additive, dominance, joint additive or "profit heterosis" and epistatic effects and their interactions with latitude in an attempt to model genotype by environment interactions. A software was developed in Fortran 77 to perform five variant types of RR: the originally proposed method; the method implemented by SAS; and three methods of weighting the RR parameter lambda. Three mathematical criteria were tested with the aim of choosing a value for the lambda coefficient: the sum and the harmonic mean of the absolute Student t-values and the value of lambda at which all variance inflation factors (VIF) became lower than 300. Prediction surfaces obtained from estimated coefficients were used to compare the five methods and three criteria. It was concluded that RR could be a good alternative to overcome multicollinearity problems. For all the methods tested, acceptable prediction surfaces could be obtained when the VIF criterion was employed. This mathematical criterion is thus recommended as an auxiliary tool for choosing lambda.