Answers are grounded in this paper and cite their sources.
Crop multi-environment trials (METs) commonly produce highly unbalanced datasets. This mainly arises from annual turnover of test genotypes and occasional changes in trial locations across years. The GGE biplot method is widely applied to interpret genotype-by-environment interactions (GEI) and carry out mega-environment (ME) classification for variety-testing datasets. Two practical approaches are available for handling missing entries in MET two-way matrices: iterative imputation built upon singular value decomposition (SVD), and best linear unbiased prediction (BLUP) derived from linear mixed models (LMM). We used two five-year (2021–2025) regional maize trial datasets from Henan Province, comprising 160 genotypes tested at 13 locations. The two datasets shared the same locations and field management, but contained largely different genotypes. Variance partitioning, ME investigation, and the selection of broadly adapted versus ME-specific genotypes at 10% and 20% selection intensities were compared. The SVD-completed matrices retained a markedly larger share of GEI variance than the LMM-BLUP matrices. A full variance component analysis revealed that variation arising from year-by-location interactions dominated total phenotypic variance, representing 54.3–65.3% of the total. By comparison, variance from the year-by-genotype interaction made up at most 2%. Part of the observed difference in GEI variance between the two analytical workflows stems from their intrinsic algorithm behavior. SVD-based imputation only estimates values for missing matrix cells and preserves all originally observed data points. In contrast, LMM-BLUP imposes shrinkage across every matrix entry. This shrinkage is a well-recognized statistical feature, which can enhance prediction performance when dealing with unbalanced field trial datasets. The SVD-based which-won-where view of the GGE biplot separated the trial locations into two MEs whose rankings were nearly uncorrelated or negatively correlated, and the resulting boundary lay at about 34°30′ N, broadly coinciding with the existing agroecological zoning of the province; this correspondence is reported as an empirical observation that still requires independent corroboration with environmental covariates. In a 100-run bootstrap resampling cross-validation based on field replicates, the two workflows differed little in prediction accuracy for the present datasets, with small differences favoring SVD imputation in some metrics. This outcome is specific to these data and to this resampling scheme; it does not establish the general superiority of either method, and prediction accuracy is not equivalent to the recovery of GEI patterns. Genotype selection outcomes differed between the workflows: more ME-specific and fewer broadly adapted genotypes were retained under SVD imputation, whereas BLUP-GGE biplots retained more broadly adapted genotypes and fewer region-specific elite materials. These findings indicate that, for unbalanced MET data, the choice of missing-data workflow can substantially change ME delineation and genotype selection. Analysts applying GGE biplots to unbalanced datasets should therefore evaluate the sensitivity of their conclusions to the treatment of missing data. Establishing the relative reliability of the two workflows will require independent validation across additional environments and datasets.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: