Table 1.

Prediction results with and without erroneous exploitation of identifiability. Best Prediction for UKB is reported for 1350 subjects in the training set, whereas for all other datasets it is reported with the maximum number of training subjects available.

Dataset	Task	Null Model	Best Prediction	Double-Dipping Prediction	Unintentional Improvement
				UKB	Sex (Accuracy)	0.528	0.82 ± 0.02	0.89 ± 0.006	7%
UKB	Age (RMSE)	7.68	5.92 ± 0.03	4.97 ± 0.50	12.4%
PNC	WRAT (RMSE)	14.6	14.45 ± 0.92	11.0 ± 0.28	23.6%
Fibro	Diagnosis (Accuracy)	0.51	0.61 ± 0.17	0.86 ± 0.038	25%
BSNIP	Diagnosis (Accuracy)	0.5	0.78 ± 0.05	0.89 ± 0.04	11%