. 2020 Jan 13;36(9):2770–2777. doi: 10.1093/bioinformatics/btaa024

Table 1.

Properties of standard Relief-based methods and generalizations STIR and NPDR

	Standard Relief-based	STIR	NPDR
Importance score (dichotomous)	${\bar{M}}_{a} - {\bar{H}}_{a}$	$\frac{{\bar{M}}_{a} - {\bar{H}}_{a}}{S_{p} (M, H) \sqrt{\frac{1}{\| M \|} + \frac{1}{\| H \|}}}$	$β'_{a}$ coefficient
Score has a null distribution	No	Yes	Yes
Supports continuous outcome	Yes	No	Yes
Supports covariates	No	No	Yes
Supports regularization	No	No	Yes

Note: The NPDR score is the standardized regression coefficient $β'_{a}$ from a logistic model of projected distances [Equation (14)] for dichotomous outcomes and from a linear model [Equation (8)] for continuous outcomes. The quantity S_p in STIR is the pooled SD of the hit and miss means. Only the score for dichotomous (hit/miss) Relief is shown and STIR is limited to dichotomous outcomes (Le et al., 2018b).