Bayesian genomic-enabled prediction as an inverse problem.
Bottom Line: Many prediction models address the collinearity that arises when the number (p) of molecular markers (e.g. single-nucleotide polymorphisms) is larger than the sample size (n).Here we propose four Bayesian approaches to the problem based on commonly used data reduction methods.Because shrinkage of estimates is affected by the prior variance of transformed effects, we propose four structures of the prior variance as a way of potentially increasing the prediction accuracy of the models fitted.
Affiliation: Colegio de Posgraduados, 56230, Montecillo, Texcoco, Edo. de México.Show MeSH
Related in: MedlinePlus
Mentions: Comparing Figure 1C (OLS estimates for MFL-WW) and Figure 4B (OLS estimates for GLS-3), instability of OLS estimates is observed for the last 50 OLS values in MFL-WW but this occurred in only a few of the last OLS estimates for GLS-3. This indicates that trait MFL-WW required more shrinkage than that needed for trait GLS-3. This may explain why, for MFL-WW, model BIR2, which showed more shrinkage than BIRR, had a better predictive correlation than BIRR. For the same reasons, model BIRR was a better predictor than BIR2 for trait GLS-3. BIR1 had a good predictive correlation for most traits because the decay of the variance was smoothed out by (see Equation 17), which allowed “intermediate” shrinkage.
Affiliation: Colegio de Posgraduados, 56230, Montecillo, Texcoco, Edo. de México.