Figure 4

Robustness check via bootstrapping to compare the performance of the three models. Plot a shows boxplots of the bootstrap training and testing errors for the train-test split dataset. It reports performance for 500 bootstrap samples of the training dataset and prediction with the test dataset. Plot b shows kernel densities of the prediction errors from the original dataset by bootstrap training of the training dataset (\(N=2590\)) or original dataset (\(N=3234\)). It reports the impact of sample size on model performance, in which the blue line is modeled with random forests, the red line is modeled with cubic regression, and the green line is modeled with linear spline regression; the solid line is trained with the training dataset in bootstrapping, and the dashed line is trained with the original dataset in bootstrapping. Plot c is a schematic flow diagram of the bootstrap performance test