Table 5.

Performance summary of full and reduced variable random forest models on simulated threenorm data sets with d relevant predictors and k noise predictors. Performance metrics were computed using independent test sets of size 1000 (Test) and the out-of-bag data (OOB), and averaged over 20 simulation runs. PCC is the percent of observations correctly classified, sensitivity is the percent of observations in class 1 correctly classified, specificity is the percent of observations in class 2 correctly classified, and AUC is the area under the receiver operating characteristic curve

	d	k	Model	PCC	Sens.	Spec.	AUC
Test	50	150	Full	0.805	0.807	0.806	0.893
OOB	50	150	Full	0.794	0.793	0.793	0.877
Test	50	150	Reduced	0.828	0.829	0.828	0.909
OOB	50	150	Reduced	0.840	0.838	0.842	0.909
Test	150	50	Full	0.768	0.763	0.777	0.859
OOB	150	50	Full	0.748	0.739	0.753	0.833
Test	150	50	Reduced	0.755	0.755	0.757	0.838
OOB	150	50	Reduced	0.795	0.798	0.790	0.865