Impact of Phenotypic Correction Method and Missing Phenotypic Data on Genomic Prediction of Maize Hybrids

Phenotypic datasets in plant breeding are commonly incomplete due to missing phenotypic information. The best approach for correcting these datasets for a stage-wise genomic prediction (GP) is not unanimous in the scientific community. Therefore, this study evaluates a two-step GP based on different methods of phenotypic correction considering complete and incomplete datasets of maize (Zea mays L.) single crosses. The dataset consists of 325 hybrids evaluated for grain yield and plant height in four sites. Sequential levels of data loss were simulated to the original dataset (from 0 to 30%) to assess the impact of missing information. The prediction was performed by an additive genomic best linear unbiased prediction model (GBLUP) using best linear unbiased estimations (BLUEs), best linear unbiased predictions (BLUPs), and deregressed BLUPs as the response variable. Mean reliability and predictive ability slightly decreased as missing phenotypic information increased, irrespective of the response variable. Regarding phenotypic correction, all methods yielded similar results for these parameters over most missing information percentages. The coincidence of selection between single- and two-stage GP was not systematically affected by response variable across multiple selection intensities, and missing data only led to a minor decrease in coincidence. Therefore, from a breeding standpoint, regardless of phenotypic correction method and missing data level, a similar set of genotypes tend to be selected.