External Validation

Environmental blocking produced less accurate predictions than those obtained through internal validation (Fig. 1). For DTB, model performance varied greatly depending on which planting was excluded and on the inclusion of GxE. For the G+E model, VF and HF had the lowest RMSE (RMSE = 27.928 days and RMSE = 30.648 days respectively), while r2 was highest for NSP (r2 = 0.497) and NF (r2 = 0.521). DTB was overpredicted in NSU06 and underpredicted in OF. Including GxE improved predictions for the two summer plantings NSU06 (RMSE from 46.783 to 10.462 days) and NSU07 (RMSE from 46.521 to 13.533 days), but offered no improvements in NF and OF. For HF and NSP, DTB was overpredicted such that RMSE increased despite a higherr2 . For the independent external validation in RS, the G+E model had a higher RMSE but higherr2 (RMSE = 35.574 days,r2 = 0.433) than the GxE model (RMSE = 18.790 days, r2 = 0.097).
Unlike DTB, environmental blocking results for SP did not differ between the two models (Fig. 2). Prediction accuracy was generally poor, with low r2 for all seven plantings. Interestingly, RMSE was weakly positively correlated with r2(Pearson’s ρ = 0.048). Models were either predicting SP closer to the observed value or better at ranking different genotypes, but not both.