Cross-Validation Assessment of Random Regression Specifications in a Single-Step Genomic Model for Dry Matter Intake

Abstract

Selection on feed efficiency traits can help to reduce costs and improve sustainability in the dairy cattle industry. Recent advances propose to use a random regression to derive breeding values for dry matter intake (DMI) from longitudinal models. In this study, we conduct a forward cross-validation of different random regression specifications of an animal model for DMI. The specifications combine basis functions for regression over days in milk with varying numbers of factors used in variance component estimation via a factor-analytic approach. Data from 10,766 predominantly Dutch and Belgian Holstein cows, comprising 21,008 lactations and 1,026,192 DMI records from 10 farms, were analyzed. Estimates obtained from partial data (pre-2020) were compared to those from the full dataset (up to early 2024). Multiple sets of focal individuals were used to estimate prediction errors for the models, decomposing global error summaries into intercept bias, slope bias, and correlation; for early, middle, and late lactation stages. The validation results identify random regression specifications that outperform the accuracy of a conventional repeatability model for DMI, in particular on the early and middle stages of lactation. This provides valuable insights for genomic prediction modeling of feed efficiency in cattle.