Upon the removal of outliers, the statistical parameters for all models improved with the largest improvement being obtained for V OC ((Q_{train}^{2}) between 0.78 0.82, 0.65 0.73 (excluding the (TiO_{2} left| {Co_{3} O_{4} } right|MoO_{3})) and 0.78 0.85 for J SC, V OC and IQE, respectively).
It was found that the prediction accuracy of the two models improved with time scale length.
Logistic-regression models were unbiased and precise when the area in which use locations were generated and the area defined to be available were the same size; the fit of these models improved with increased numbers of random locations.
A distance-based analysis of the applicability domain (AD) for WOMBAT compounds showed that the reliability of the predictions from the models improved with the increasing similarity between the training and WOMBAT test set.
As expected, prediction accuracies for GWP with SE and ME models improved with increasing marker density in the example using the NAM population B73×CML322.
Prediction ability of single-family models improved with less stringent thresholds and were optimal at P=0.01, but then declined when the threshold was relaxed further to P=0.05, whereas joint-family models were optimal at P=0.0001.
