Degree Type


Date of Award


Degree Name

Doctor of Philosophy


Animal Science


Animal Breeding and Genetics

First Advisor

Dorian J. Garrick


The genetic merit of livestock is now routinely evaluated using SNP genotype information on selection candidates to improve genetic gain. Improvement in genomic prediction accuracy will have a direct impact on genetic gains. The New Zealand dairy cattle population contains two major breeds: Holstein Friesian and Jersey; and KiwiCross, their admixed descendants are popular with many farmers. Genomic prediction models fitting haplotype alleles rather than SNPs have been shown to increase genomic prediction accuracy in simulated and purebred populations, but has not been assessed in admixed populations. This dissertation investigated whether prediction accuracy, and thus rate of genetic gain, can be improved in the admixed New Zealand dairy cattle population by fitting covariates for haplotype alleles rather than covariates for SNPs in genomic prediction models. Haplotype alleles were constructed from the phased genotypes at ~40,000 SNPs for ~58,000 New Zealand dairy cattle.

A measurement of the reliability of genomic prediction accuracy estimates is important for evaluating whether the performance of different genomic prediction models is significantly different. Chapter IV explored a method for calculating the posterior distribution of prediction accuracy from Bayesian genomic prediction models from calculating prediction accuracy in each iteration of the post-burn-in Markov chain Monte Carlo chain. Using 200 replicates of a simulated data set of 5,000 training and 1,000 validation individuals genotyped at 20,000 SNPs, our method appropriately captured the confidence in accuracy between true and estimated genetic merit but not between phenotype and estimated genetic merit. In practice the true genetic merit is not observed so the accuracy between true and estimated genetic merit cannot be calculated. Further research is needed to assess the reliability of prediction accuracy estimates when true genetic merit is unknown.

The use of genomic prediction models that fit covariates for haplotype alleles, which were constructed based on a fixed length (e.g. 250 kb) or based on a population parameter (e.g. recombination), has potential for increasing genetic gain in admixed populations compared to fitting covariates for SNPs. The best model explored for this data set was based on recombination (up to 7.7% improvement in prediction accuracy over the SNP model; p < 0.001); however, the best method for assigning SNPs to haplotype blocks may differ in admixed populations compared to purebred populations because patterns of linkage disequilibrium may be different between breeds within the population. The best method will also depend on the number of individuals genotyped and the relationships between them. Consistent with results in other populations, fixed-length haplotypes appear to perform well in the New Zealand dairy cattle population (up to 5.5% improvement over the SNP model; p = 0.002) as long as haploblock length is appropriate. Haplotype blocks generated using recombination events within the population may provide higher prediction accuracy if these recombination events can be accurately identified (i.e. many closely related animals). Removal of rare haplotype alleles from the data set reduces the computational demands of genomic prediction fitting haplotype alleles with no loss in prediction accuracy. Further reduction in computational demands and improvement in prediction accuracy could be obtained by fitting combined SNP and haplotype models. The increase in the number of individuals genotyped and sequenced will likely improve the benefits of fitting haplotype alleles in genomic prediction models.

Copyright Owner

Melanie Kate Hayr



File Format


File Size

190 pages