Key message <p>Using the full weight matrix and deregressed pedigree-based best linear unbiased predictions in second-stage models lead to selections and genomic predictions closer to those obtained using a single-stage model.</p> Abstract <p>In multi-environment genomic selection, although single-stage (SS) models are generally more efficient (no loss of information), there are contexts where they are difficult to fit, making two-stage models the most practical alternative. An example is the evaluation of early-stage observational trials (OTs) of sweetpotato breeding, where several clones are tested in unreplicated trials. In this study, 1,138 clones derived from partial diallels within two gene pools had their storage root yield evaluated across six OTs. Using this scenario, we compared the selection and prediction performances of models under different two-stage strategies against the SS benchmark. We also tested whether pool-specific genomic prediction models offered advantages over models trained with the complete dataset. Given the lack of replication in OTs, we hypothesized that deregressed best linear unbiased predictions (dBLUPs) or pedigree-based dBLUPs (dABLUPs) would work more appropriately as inputs for second-stage models than best linear unbiased estimates (BLUEs). These comparisons were conducted within weighted models using either a diagonal weight matrix or the full weight matrix. For selection, differences among second-stage models were minor, with a slight advantage for those using dABLUPs as entries, combined with the full weight matrix. For prediction, however, the choice of weighting scheme had a greater impact on performance than the choice of entry. Using the complete dataset, differences between entries were marginal, but for pool-specific predictions, dABLUPs provided the best performance. Overall, if adopting a two-stage strategy for the analysis of augmented trials, we recommend using dABLUPs together with the full weight matrix.</p>