Effects of weighting schemes and base adjustment on genomic validation outcomes in the Interbull GEBVtest framework

Abstract

Genomic validation is required to ensure the reliability and international acceptance of genomic estimated breeding values (GEBV). Recent updates to the Interbull GEBVtest software introduced alternative options for checking, including different regression equations, weights, and base-adjustment procedures. The objective of this study was to evaluate the effects of these options on genomic validation outcomes in U.S. dairy cattle populations. Genomic validations were performed using the Interbull GEBVtest software and August 2025 U.S. genomic evaluations. A reduced dataset was generated by truncating domestic information after 2021 and combining the resulting national evaluations with foreign evaluations from Truncated Multiple Across Country Evaluation (TMACE) information from Interbull. Five dairy breeds and nine production, health, longevity, and fertility traits were evaluated. All validation scenarios used predicted deregressed genomic estimated breeding values (dGEBV) as the validation target. Four scenarios were compared, differing in regression methodology [Interbull weighted regression (ITB) or unweighted Legarra-Reverter (LR) regression] and application [yes/no] of base adjustment. A total of 42 breed-trait combinations were evaluated under four validation scenarios, resulting in 168 scenario-specific validation analyses. Across all analyses, 91 resulted in PASS outcomes, 18 in FAIL outcomes, 12 in WARN outcomes, and 47 in HiSE outcomes. WARN indicates validation statistics close to predefined acceptance thresholds, whereas HiSE indicates inconclusive validation results due to high slope standard errors. Validation results were generally consistent across scenarios, indicating that the genomic validation framework was robust for most breed-trait combinations. However, a limited number of analyses produced different validation outcomes depending on the methodological configuration applied. These differences were associated with interactions among regression methodology, base adjustment, and validation-group definition rather than any single factor alone. However, for traits that failed validation in some scenarios, validation statistics classified as WARN remained close to the threshold for FAIL. The results demonstrate that genomic validation outcomes are sensitive to details when validation statistics are close to predefined acceptance thresholds. A WARN result close to a FAIL threshold should prompt further investigation of the factors contributing to potential bias. The observed FAIL outcomes may be associated with limitations of the current multiple-step evaluation framework. CDCB is transitioning to single-step evaluation, which should impact validation performance.