Estimating on-farm genotypic performance and variability using ranking data

Abstract

Our scalable two-step method estimates genotypic performance and genetic parameters from ranking data, producing reliable results comparable to quantitative analyses, enabling the integration of ranking data into breeding pipelines. Plant breeding research has chiefly relied on on-station experiments to evaluate varietal performance. Nevertheless, these trials often fail to represent on-farm growing conditions and farmers’ preferences, potentially leading to poorly defined breeding targets. Recent work has demonstrated the potential of using on-farm verification trials combined with ranking data to support farmers in evaluating varieties while providing information that is representative of farmers’ needs. Despite this potential, scalable methods for quantifying genetic differences and assessing the strength of the genetic signal in such trials remain limited. Here, we present a two-step procedure for analyzing trials based on ranking data, allowing the estimation of genetic parameters. The approach follows a common strategy in quantitative genetics, in which parameters are estimated from tables of genotypic means and their variances. In our framework, these estimates are obtained from Thurstonian and/or Plackett-Luce models, which treat rankings as observations of an underlying continuous trait associated with genotypic performance. Using simulated data, we showed that genotypic mean estimates derived from ranking analyses are linearly related to those obtained from quantitative trait analyses and that their variances adequately capture estimation uncertainty. We further demonstrated that incorporating these estimates and their variances into a second-step mixed-effects model yields accurate estimates of variance components. Analyses of groundnut, maize, and sweetpotato datasets confirmed the applicability of the approach and showed that ranking data can provide reliable estimates of genetic parameters. We argue that this framework can be scaled to obtain genotypic performance estimates from multi-trial on-farm data.