. 2013 Jun;194(2):473–484. doi: 10.1534/genetics.113.150201

Table 2. Different statistics and their standardized coefficients estimated via a linear model selected by boosting.

Statistics	Meaning	coef
H	Number of different haplotypes	0.00
$R_{h^{*}}$	H − S (Myers and Griffiths 2003)	50.18
R_m	Minimum number of recombination events inferred from TGT (Hudson and Kaplan 1985; Gusfield 1991)	0.72
$S_{k}^{2}$	Variance of average pairwise differences for the whole region (Sved 1968; Hudson 1987)	1.78
Hetero	Haplotype heterozygosity (the probability that two segments are different in haplotype)	−0.58
pR	Proportion of SNP pairs that TGT indicates a recombination event	0.55
Mean $S_{k}^{2}$	For each SNP pair, $S_{k}^{2}$ was computed, and then the mean was computed over all SNP pairs	−3.19
Mean r²	Similar to mean $S_{k}^{2}$ , but r² was computed instead of $S_{k}^{2}$ (Hudson 1985)	−3.03
Mean LD₁	Similar to mean $S_{k}^{2}$ , but LD₁ was computed instead of $S_{k}^{2}$ . LD₁ is a measure of LD proposed by Batorsky et al. (2011), defined as the normalized frequency of LRH (least-represented haplotype) for a pair of SNPs.	−2.01
LD₂	LD₂ is a measure of LD proposed by Batorsky et al. (2011), defined as the proportion of the SNP pairs for which the frequency of LRH is below 0.04.	−0.26

TGT, three-gamete test (Gusfield 1991). We chose 10 summary statistics that are directly or indirectly related to ρ. Sample parameters: n = 200, S = 59. In the original article of Myers and Griffiths (2003), R_h is the maximum of H − S over some subsets of segregating sites. Here we computed H − S only once for the whole region, so we denoted this value as $R_{h^{*}}$ rather than R_h. coef shows the standardized linear coefficient for each statistic obtained via a linear model (glm) where we fitted all 10 statistics to predict ρ. Training was done with 5 ρ’s (20, 60, 100, 140, and 180). A large absolute value of a standardized coefficient indicates a large influence of the corresponding statistic. When fixing S, $R_{h^{*}}$ contains the same information as H. Thus boosting only selects one of them and discards the other (by assigning coefficients equal to 0).