The continuous growth in both the number of phenotypic records and the range of traits included in beef cattle genetic evaluations poses substantial statistical and computational challenges for the estimation of genetic and residual (co)variance matrices required for breeding value estimation. Structural equation models (SEM), implemented using either factor analysis (FA) or recursive model (REC) structures, provide a flexible framework to model genetic and residual (co)variance matrices while yielding more parsimonious and computationally efficient parameterizations. Here, SEM was applied to estimate parameters for growth and ultrasound-measured carcass traits in beef cattle. The dataset comprised 2,942 animals, and six traits were evaluated using standard multiple-trait mixed models (SMTM) and SEM. We considered FA and REC models implemented with six alternative parameterizations, in which random effects were represented as linear combinations of fewer unobservable random variables. Relative to the SMTM, both the model with two factors in the genetic covariance matrix (FA2G) and the model in which six recursive effects were constrained to zero in the residual covariance matrix (REC1) demonstrated a strong ability to capture genetic variability, as reflected by comparable heritability estimates. Correlations between estimated breeding values (EBV) for the same traits across models were consistently high, ranging from 0.94 to 1.00, indicating strong agreement among model estimates. The FA2G model was the most parsimonious in terms of the effective number of parameters (pD), with 431.2 pD, corresponding to a reduction of 25.3 parameters relative to the SMTM. The REC1 model also emerged as a competitive alternative for this dataset, exhibiting a lower pD (443.6) than the SMTM (456.5) and the most favorable deviance information criterion among all models evaluated (e.g., 37,868.6 for REC1 versus 37,874.7 for SMTM). Overall, these results demonstrate that mixed-effects multi-trait models for beef cattle genetic evaluation can be effectively implemented using FA or REC structures, which provide parsimonious representations of the underlying covariance patterns while maintaining high agreement in EBV.
Yokoo et al. (Thu,) studied this question.