# Randomization across breeding cohorts improves the accuracy of conventional and genomic selection

## TL;DR

Randomizing breeding groups across trials improves selection accuracy, especially when genomic data is limited.

## Contribution

The study introduces a novel comparison of randomization strategies in breeding trials using simulations.

## Key findings

- Complete randomization improved BLUP accuracy by 11.7% under low genetic correlation.
- Genomic data reduced the impact of randomization design on selection accuracy.
- Sparse genomic testing showed a 5.5% accuracy improvement under challenging conditions.

## Abstract

Breeding programs conventionally evaluate cohorts in separate trials; however, environmental differences across testing areas can be confounded with genetic differences between cohorts, potentially reducing the accuracy of breeding value estimation. We test whether the conventional approach of restricting randomization of cohorts to within trials reduces genomic and conventional selection accuracy when compared to the complete randomization of all cohorts across a trial, using in silico simulation with marker data from University of Illinois winter wheat breeding lines. We evaluated selection accuracy for conventional best linear unbiased prediction (BLUP), genomic BLUP (GBLUP), and genomic‐enabled sparse testing across a comprehensive simulation space spanning narrow‐sense heritabilities of 0.2–0.8, genetic correlations between testing areas from 0.2 to 1.0, and three replication levels. Difference‐in‐differences (DiD) analysis established causal inference by comparing design performance as conditions deteriorated from an optimal baseline where both designs performed equivalently. Complete randomization improved BLUP accuracy by 11.7%, reaching 15.7% under low replication and low genetic correlation between areas. Genomic data largely eliminated this design effect, with GBLUP showing no significant DiD interaction effect. However, genomic‐enabled sparse testing revealed a significant DiD effect and an improvement in selection accuracy of 1.5% that increased to a 5.5% advantage under challenging conditions. While heritability had the strongest main effect on selection accuracy, genetic correlation between areas showed the largest interaction with randomization scheme, with design performance diverging significantly only as this parameter decreased. Programs with genomic data and balanced phenotypic data can use either restricted or complete randomization, but those with other circumstances can benefit from complete randomization.

In breeding programs, groups of related breeding materials, called cohorts, are often tested in separate trials. As a result, confounding of genetic and non‐genetic effects occurs. This confounding is expected to reduce the accuracy of selection when cohorts are analyzed together for selection or predictive modeling. We evaluated the extent to which randomizing all cohorts together for testing improves selection accuracy relative to randomization solely within cohorts. To do this, we simulated a range of scenarios and assumptions and determined the accuracy gained from complete randomization across cohorts in each case. We found that randomization across cohorts provided large benefits when genomic relationship information was absent and modest benefits when data were unbalanced but genomic relationship information was present. Breeding programs should strongly consider randomizing cohorts together whenever feasible.