Software

DHGLMF90: Software for estimating genetic heterogeneity of residual variance using double hierarchical generalized linear models

---

Abstract

In animal breeding research and applications, there is increasing interest in selecting individuals that express consistent responses under environmental perturbations. For some traits, it is not only important to improve the population mean, but also to control the variation around it (i.e., dispersion). Dispersion represents the sensitivity of a genotype to the environmental conditions experienced by each individual. Such variability can be detected in the residual of the model, where genotypes with lower dispersion may show greater resilience to environmental perturbations and more consistent phenotypes. This can be modeled by allowing heteroskedastic residuals when estimating breeding values, using a hierarchical model that includes genetic effects for the mean and a dispersion component. Bayesian methods can estimate heteroskedastic residuals using the Metropolis-Hastings algorithm, which is slow and inefficient. Double hierarchical generalized linear models (DHGLM) provide a faster alternative; however, computational resources for large-scale applications remain challenging. We developed software named DHGLMF90, to implement DHGLM using a highly efficient algorithm for estimating genetic heterogeneity of residual variance and breeding values, particularly suited for large datasets. The method was extended to multiple-trait models and genomic information. We improved the Iteratively Reweighted Least Squares (IRWLS) algorithm to enhance convergence properties compared with previous implementations. IRWLS iteratively estimates variance components using a bivariate model that includes both mean and dispersion components. DHGLMF90 uses BLUPF90+ to estimate variance components using REML. Two datasets were used to test the software using single-trait and multiple-trait models. A simulated dataset with single and multiple traits models was used to assess the accuracy of the method, replicated 100 times. Whereas a large real dataset, including 80K records from two reproductive traits, 41K animals in the pedigree, and 1.5K genotyped animals was used to test the software performance. The inclusion of heterogeneous residual variances obtained from DHGLM for estimating breeding values using BLUP and ssGBLUP was also tested. Simulation results confirmed that DHGLMF90 converged an order of magnitude faster than Bayesian methods, with comparable parameter estimates. The feasibility of multiple-trait and inclusion of genomic information was confirmed. Variance components from pedigree-based and genomic analyses showed comparable performance. Intercepts and slopes from regressions of weighted (G)EBV on regular (G)EBV, calculated using all animals across both traits, ranged from 0.02 to 0.15 and from 0.59 to 1.18, respectively. The -2loglike and AIC criteria indicated improved model fit when heterogeneous residual variances were incorporated. Overall, DHGLMF90 provides an effective framework for estimating genetic effects on residual dispersion in large-scale data, demonstrating that selection for reduced variability is feasible and can be integrated into modern breeding programs.

Keywords: 2026

How to Cite:

àlvarez Múnera, A., Bermann, M., Casto-Rebollo, C., Chen, C., Holl, J., Ibáà±ez Escriche, N., Lourenco, D. & Misztal, I., (2026) “DHGLMF90: Software for estimating genetic heterogeneity of residual variance using double hierarchical generalized linear models”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2270184. doi: https://doi.org/10.31274/wcgalp.23405

Rights: 1