Transforming estimated breeding values from observed to probability scale: how to make categorical data analyses more...

Area of Science:

• Animal genetics
• Quantitative genetics
• Statistical genetics

Background:

• Threshold models are standard for categorical trait genetic analysis (e.g., calving ease), publishing estimated breeding values (EBVs) as probabilities for easy interpretation and selection.
• Implementing threshold models is complex due to nonlinear equations and probability functions, leading to long computing times and convergence issues, especially with genomic data.
• Linear models offer a computationally efficient alternative, yielding EBVs highly correlated with threshold models, but lack a direct transformation to the probability scale.

Purpose of the Study:

• To propose and validate transformations for converting EBVs from linear models to the probability scale, analogous to the liability scale transformation in threshold models.
• To assess the impact of using linear models with proposed transformations on computing time, memory usage, EBV correlations, and genetic trends compared to threshold models.
• To evaluate the potential of linear models to improve the efficiency and scope of genetic evaluations for categorical traits.

Main Methods:

• Developed and applied transformations to map EBVs from linear models to the probability scale.
• Compared linear and threshold models using a large dataset (11M pedigree, 965k genotyped animals) assessing convergence time, peak memory usage, and EBV correlations.
• Analyzed estimated genetic trends derived from both model types to ensure consistency in genetic progress.

Main Results:

• Linear models converged 5x faster (32 vs. 145 hours) than threshold models, with comparable peak memory usage.
• The proposed transformations yielded highly correlated probabilities between linear and threshold models (≥0.99 for direct EBVs, ≥0.97 for maternal EBVs).
• Estimated genetic trends were analogous between models, indicating no loss of genetic progress.

Conclusions:

• The proposed transformations effectively enable the use of linear models for categorical trait genetic analysis on the probability scale, maintaining accuracy.
• Linear models offer a significant computational advantage, facilitating faster genetic evaluations (e.g., weekly) and enabling multi-trait analyses for enhanced selection.
• Adopting linear models with these transformations can improve breeding program efficiency without compromising genetic gain.