Estimation & Prediction

Comparison between single lactation models and ensemble models for the estimation of variance components and genomic breeding values for resilience indicators in dairy sheep

---

Abstract

Resilience can be defined as the capacity to endure or rapidly recover from disturbances. In farm animals, resilience is a valuable trait associated with reduced handling and veterinary costs and greater longevity. However, resilience cannot be measured directly; instead, it requires the estimation of resilience indicators (RIs). In dairy animals, RIs can be derived from deviations between observed and predicted daily milk yields (dMY). The natural logarithm of the residual variance (LnVar), the residual lag-1 autocorrelation (Lag1), and the residual skewness (Skew) are examples of RIs estimated from lactation curves. Therefore, the efficient use of these RIs depends on the quality of the fitted lactation curve. Despite their importance, there is no consensus on which lactation curve model should be preferred, especially in species with a high frequency of atypical lactation patterns, such as dairy sheep. Ensemble models, which combine predictions from multiple individual models to reduce variance and bias, emerge as an attractive alternative for modeling lactation curves. In this study, dMY records up to 210 days in milk (DIM) from 1,458 Spanish Assaf ewes from two flocks were used to estimate LnVar, Lag1, and Skew. Two approaches were compared: (1) a 5th-percentile cubic spline with knots at 36, 68, 100, 132, and 165 DIM (spline), and (2) an ensemble model integrating 47 individual lactation curve models implemented in the R package EMOTIONS. Variance components were estimated using Gibbsf90, fitting flock-year-season of lambing, number of lambs born alive, and lambing age as fixed effects. Based on genotypes for 42,720 SNPs, genomic estimated breeding value (GEBV) accuracy was calculated as the square root of reliability, defined as 1 − (se²/σₐ²), where se is the GEBV standard error and σₐ² the additive genetic variance. Kendall's tau rank correlations (τ) between GEBVs obtained with the spline and ensemble models were estimated for all three RIs. The ensemble model produced slightly higher heritabilities despite similar σₐ² values across RIs (Table 1). Average GEBV accuracies were lower for Lag1 (0.23 ± 0.15 for ensemble; 0.22 ± 0.15 for spline) and Skew (0.13 ± 0.11 and 0.12 ± 0.11) than for LnVar (0.35 ± 0.22 for both). The τ values for GEBV accuracies between ensemble and spline were 0.86, 0.74, and 0.52 for LnVar, Lag1, and Skew, respectively. The limited sample size resulted in few GEBVs with accuracy >50% for Lag1 (79 for ensemble; 66 for spline) and Skew (6 and 4, respectively). τ estimated exclusively for LnVar GEBVs with accuracy >50% was 0.93. In addition to the abovementioned results, ensemble models yielded smaller root mean squared errors and better agreement between observed and predicted dMY, suggesting they may provide more reliable variance component estimates and GEBVs for RIs in dairy sheep.

Keywords: 2026

How to Cite:

Gutiérrez-Gil, B., Arribas-Gonzalo, R., Pelayo, R., Arranz, J., Suárez-Vega, A. & Fonseca, P., (2026) “Comparison between single lactation models and ensemble models for the estimation of variance components and genomic breeding values for resilience indicators in dairy sheep”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286654. doi: https://doi.org/10.31274/wcgalp.24081

Rights: 1