Estimation & Prediction

A statistical test to compare the accuracy of genetic predictions

---

Abstract

AbstractThe accuracy of estimated breeding values (EBVs) is commonly assessed by their correlation with phenotypes in a validation set, and differences in these validation correlations are often used to compare prediction models or data sources. Classical tests for comparing two overlapping correlations, such as the Williams t-test, rely on independent and identically distributed observations across validation individuals and can be severely anti-conservative when validation individuals are related. We developed a generalized least squares wild bootstrap (GLS-WB) test that accounts for dependence in the validation phenotypes via a GRM-based working covariance model. The method decorrelates the validation phenotypes using a Cholesky decomposition and constructs an empirical null distribution for the difference in correlations by repeatedly perturbing null-model residuals with random sign flips and recomputing the correlation difference. Using realistic genotype data from a pig breeding population, we evaluated empirical Type I error and power across heritabilities of 0.3,0.6,and 0.9. The Williams test exhibited substantial Type I error inflation at nominal α=0.05, whereas GLS-WB remained close to the target level. When procedures were compared at matched empirical Type I error via size-corrected cutoffs, GLS-WB achieved power comparable to the Williams test. These results support GLS-WB as a dependence-aware paired test for comparison of validation correlations in breeding programs.

Keywords: 2026

How to Cite:

Ni, Z., Dekkers, J. & Fernando, R., (2026) “A statistical test to compare the accuracy of genetic predictions”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2293281. doi: https://doi.org/10.31274/wcgalp.24312