A linear mixed model fitting both genome-wide cosegregation (CS) and linkage disequilibrium (LD) was developed to improve accuracy of genetic prediction for pedigreed populations of unrelated families that have half sibs represented in both training and validation. Cosegregation was modeled as the effects of genome-wide1-centimorgan haplotypes that one individual inherits from pedigree founders through identity-by-descent, while LD was modeled as allele substitution effects of all marker genotypes. Prediction accuracy of the LD-CS method was compared to the accuracy of three LD methods – GBLUP, BayesA and BayesB, using simulated datasets of varying numbers of paternal half sib families. Results show that the LD-CS method tended to have higher accuracy than any of the LD methods. With an increase in the number of families, the accuracy of the LD-CS method persisted, while the accuracy of the LD methods dropped. The results indicate that by fitting CS explicitly, the LD-CS method has higher and more consistent prediction accuracy than LD methods.