Skip to main content
ARS Home » Research » Publications at this Location » Publication #192197


item Kuhn, Melvin
item Hutchison, Jana

Submitted to: Journal of Dairy Science
Publication Type: Abstract Only
Publication Acceptance Date: 2/17/2006
Publication Date: 7/9/2006
Citation: Kuhn, M.T., Hutchison, J.L. 2006. Methodology for prediction of bull fertility from field data [abstract]. Journal of Dairy Science. 89(Suppl. 1):15-16(abstr. M26).

Interpretive Summary:

Technical Abstract: Simulated data were used to compare alternative statistical models for prediction of bull conception rate (CR) using field data. Two modeling aspects were investigated: an expanded service sire (SSR) effect vs a single term for SSR and linear vs threshold model. In practice, factors such as stud, inbreeding, or age may affect a bull's fertility. Estimating these factors as separate terms in the model (expanded SSR effect) and then adding them back to the bull solution may improve accuracy by utilizing more data to estimate the individual components. Simulated data included the effects of herd, SSR, and cow, where the SSR effect had 3 components. Each of 20 replicates had 100,025 cows with a maximum of 7 breedings per cow and an overall mean CR of 0.35. There were 250 sires per replicate and 595 herds, ranging in size from 35 to 5000 cows. The expected number of services per bull ranged from 270 to 5900. Accuracy (correlation between predicted and true SSR effect) and bias were used to compare models. The underlying variable was also analyzed with the true linear model to assess the maximum accuracy for this simulated data. The expanded SSR effect was superior to the single SSR term in all models. However, the additional terms must be fit as random effects to avoid bias. The linear and threshold models had the same accuracy (86.6%), which was only 4.6% lower than that for the true model for the underlying variable. However, the threshold model showed some bias while the linear model did not. Use of exact variances in the calculation of probabilities from threshold model solutions may improve the threshold model estimates. Furthermore, it is well known that the binomial distribution is approximated by the normal distribution and that this approximation improves as sample size increases. All herds in this simulation had a relatively large number of matings. Thus, further research will determine the effect of subclass sample size on the linear and threshold model comparisons.