Table 1.
Kruskal–Wallis test for the effect of LD and MAF on the variable importance of null SNPs
| Effect | Condition | df | RF Gini |
RF MDA |
GBM |
GBM + subsetting |
GBM + PCVs |
|||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| x2 | P | x2 | P | x2 | P | x2 | P | x2 | P | |||
| MAF | ρ = 0.0 | 4 | 962.4 | <1 × 10−10 | 2.1 | 7.1 × 10−1 | 148.0 | <1 × 10−10 | 381.7 | <1 × 10−10 | 18.5 | 9.8 × 10−4 |
| ρ = 0.3 | 4 | 1008.6 | <1 × 10−10 | 3.5 | 4.8 × 10−1 | 222.6 | <1 × 10−10 | 399.7 | <1 × 10−10 | 8.8 | 6.6 × 10−2 | |
| ρ = 0.6 | 4 | 1074.4 | <1 × 10−10 | 3.5 | 4.8 × 10−1 | 206.1 | <1 × 10−10 | 355.1 | <1 × 10−10 | 22.7 | 1.5 × 10−4 | |
| LD | m = 0.05 | 2 | 342.8 | <1 × 10−10 | 21.2 | 2.5 × 10−5 | 2.8 | 2.5 × 10−1 | 0.5 | 7.6 × 10−1 | 0.1 | 9.5 × 10−1 |
| m = 0.10 | 2 | 195.7 | <1 × 10−10 | 5.4 | 6.8 × 10−2 | 3.1 | 2.1 × 10−1 | 6.4 | 4.2 × 10−2 | 0.7 | 7.1 × 10−1 | |
| m = 0.20 | 2 | 162.0 | <1 × 10−10 | 1.8 | 4.2 × 10−1 | 9.3 | 9.7 × 10−3 | 0.2 | 9.1 × 10−1 | 0.7 | 7.2 × 10−1 | |
| m = 0.30 | 3 | 596.0 | <1 × 10−10 | 26.3 | 8.3 × 10−6 | 52.4 | <1 × 10−10 | 1.6 | 6.6 × 10−1 | 1.1 | 7.7 × 10−1 | |
| m = 0.50 | 2 | 75.2 | <1 × 10−10 | 0.9 | 6.5 × 10−1 | 0.1 | 9.6 × 10−1 | 2.7 | 2.6 × 10−1 | 2.7 | 2.6 × 10−1 | |
Significance test results are given for uncorrected variable importances in RF and GBM, as well as for the GBM importance with LD subsetting and with PCVs. Results show the similar impact of MAF and LD on the RF Gini and GBM importances, the reduced effect of LD after subsetting, and the reduced effect of MAF after correction with PCVs. Tests are performed for the simple effect of MAF at a given LD ρ, and for the simple effect of LD at a given MAF m. With Bonferroni corrections for family-wise α= 0.05, P-values for the effect of MAF <1.7 × 10−2, and P-values for the effect of LD <1 × 10−2 indicate significant effects.