Figure 2. Detection power comparison of forward stepwise penalized logistic regression (stepPLR), Random Forests (RF), tuned Relief-F (TuRF) and Evaporative Cooling (EC) filter for 100 simulated replicates of each of the models defined in Figure 1 using 500 cases and 500 controls and 1500 total SNPs.

Detection power is defined as the percentage of replicates for which both causal SNPs are found above the cutoff in the filter's rank list. The power is plotted as a function of the rank-list cutoff on a log scale, from the top 2 up to the top 100 final SNPs. EC, RF and TuRF results are sorted by importance score. Detection was counted for stepPLR if both causal SNPs had nonzero coefficients in the LR models.