Combining multiple pure breeds or admixed breeds into one evaluation can be detrimental if the accuracy of prediction for one is lower than within-breed. Prediction accuracy was compared when considering SNP effects as different across breeds (non-shared), or the same in all (shared) for 5 simulated breeds. The non-shared approach prevented changes in accuracy, while the shared method only maintained accuracy when the SNP density was high and effective population size was large. Imputation accuracy of Holstein-Jersey crossbred genotypes when using different reference populations – crossbreds, Jersey, Holstein, or all combined. Accuracy was higher when using Jersey than Holstein. The best results were achieved with a crossbred reference population. The accuracy and inflation of indirect genomic predictions (IP) for milk yield were evaluated for Holstein-Jersey crossbred animals. Different reference populations were used to calculate SNP effects – ~80k Holstein (HO), ~40k Jersey (JE), ~22k crossbreds (CROSS), Holstein and Jersey combined (JE_HO), or equal proportions of each pure breed and crossbred animals (MIX). While JE, CROSS, and JE_HO gave the same accuracy (0.50), HO and MIX were slightly lower (0.47 and 0.46). An additional method that used breed proportion in combination with SNP effects based on pure breeds showed to have the lowest accuracy (0.32). Inflation was best when using the MIX scenario (1.00), and worst when using HO (0.55). Diversity within 20,990 US Holstein cattle was evaluated by using k-means clustering on the genomic relationship matrix. Each of the 5 clusters were traced back for 10 generations - G0 (oldest) to G10 (youngest) to form 5 families (F1 to F5). Allele frequency changes over time was observed for specific SNP based on different criteria – key genes of known importance, markers associated with time, a population diversity parameter (Fst), markers that changed the most in the whole population, and markers that have changed differently across families (based on greatest variance and range). Non-parallel changes were observed across families, showing genetic redundancy and divergent selection. The Replicate Frequency Spectrum (RFS) was used to measure the similarity of change across families. Results show that populations have changed differently, supporting the presence of genetic redundancy.