Do little interactions get lost in dark random forests?

BMC Bioinformatics

Table 2 Simulation parameters

Parameter	Description	Values
β _M	Effect of marginal-only SNPs	0.4, 0.8
β _I	Interaction effect (see Table 1)	0.4, 0.8
M A F _M	Minor allele frequencies for marginal-only SNPs	0.2, 0.4
M A F _I	Minor allele frequencies for interacting SNPs	0.2, 0.4
mtry	Number of SNPs randomly selected as splitting candidates	10, 50

All combinations of these parameters were simulated. The interaction models Interaction only, Modifier SNP, No interaction, Redundant and Synergistic (see Fig. 2) were considered. As variable importance measures, we determined the Gini importance, permutation importance, pairwise permutation importance, joint importance by maximal subtrees and joint variable importance, resulting in a total of 800 simulation scenarios. In addition, one simulation with only 2 marginal-only SNPs and one simulation with 2493 noise SNPs was performed. In both cases, m t r y=50, β _M=β _I=0.4 and M A F _M=M A F _I=0.2 was set. Finally, a simulation with simulated linkage disequilibrium was performed, see the Methods section for a description. All simulation scenarios were replicated 100 times

ISSN: 1471-2105