Table 4. Effects of preprocessing methods on gene selection under the null hypothesis of no association between genes and the bacterial phenotype in the Candidemia dataset [11].

Preprocessing	Number of significant probes
	Mean	St. Dev.	95% Interval
No preprocessing	82.6	640.7	0.0	607.0
Center (subtract global mean)	82.6	640.7	0.0	607.0
Standardize (subtract global mean and divide by stdev)	82.6	640.7	0.0	607.0
Scale each probe to [0,1]	82.6	640.7	0.0	607.0
Batch correction from the supplementary software of Zaas et al. [9]	221.8	1098.0	0.0	3543.5
ComBat	253.2	1174.3	0.0	3991.5

The phenotype variable was randomly permuted 10,000 times. On each permutation, we applied a given preprocessing method and then performed gene selection using a two-sample t-test with the false discovery rate (FDR) correction at level 0.2 [32], [33].