Performance of random forests and logic regression methods using mini-exome sequence data

BMC Proceedings

Table 1 Percentage of CVs and UNCVs that were in the top-ranked 10% of predictors (RF and ULR) or that were included in the final model (LR) in at least 5%, 10%, and 20% of the 200 simulated replicates

Data set	Total number^a		Random forest, % variants ranked in the top 10% of predictors			Univariate linear regression, % variants ranked in the most significant 10%			Logic regression, % variants included in final model
			In ≥5 PoR	In ≥10 PoR	In ≥20 PoR	In ≥5 PoR	In ≥10 PoR	In ≥20 PoR	In ≥5 PoR	In ≥10 PoR	In ≥20 PoR
Uncollapsed	CVs	36	100	72	33	94	72	39
	UNCVs	12,485	76	46	8	90	25	3
Gene-collapsed	CVs	15	100	100	50	87	73	53	91^b	64	45
	UNCVs	6,642	98	81	24	90	26	3	63	32	9
Pathway-collapsed	CVs	167	99	95	72	93	50	26	1^c	0	0
	UNCVs	2,249	91	68	38	88	32	8	0	0	0

UNCVs are noncausal variants that were not used in the simulation to determine Q2, were not in one of the causal genes, and did not display correlation of at least 0.6 with any CVs.
^a Total number of nonmonomorphic variants in Asians. For LR we excluded 4 common CVs and 4,725 common NCVs.
^b Final LR model: 3 trees with 10 leaves.
^c Final LR model: 2 trees with 20 leaves.

ISSN: 1753-6561