Figure 8. Prediction performance (Pearson’s correlation coefficient) of fluid intelligence in the UK Biobank dataset with different number of training and validation subjects.

The performance of all algorithms generally increased with more training and validation subjects. In the case of 100, 500, 1000 and 2000 subjects, 3/4 of the subjects were used for training and 1/4 of the subjects were used for validation. In the remaining cases, 1000 subjects were used for validation, while the remaining subjects were used for training. For all cases, test set comprised the same set of 1000 subjects. Kernel regression was highly competitive across all sample sizes. See Figure S11 for MAE results.