At an early stage of genomic investigations, a small sample of microarrays is used in gene expression experiments to identify small subsets of candidate genes for a further accurate investigation. Unlike the statistical analysis methods for a large sample of microarrays, an appropriate statistical method for identifying small subsets is a randomization test that provides exact P values. These exact P values from a randomization test for a small sample of microarrays are discrete. The possible existence of differentially expressed genes in the sample of a full set of genes can be tested for the null hypothesis of a uniform distribution. Subsets of smaller P values are of prime interest for a further accurate investigation and identifying these outlier cells from a multinomial distribution of P values is possible by M test of Fuchs et al. (1980). Above all, the genome-wide gene expressions in microarrays are correlated, but the majority of statistical analysis methods in the microarray analysis are based on an independence assumption of genes and ignore the possibly correlated expression levels. We investigated with simulation studies the effect that correlated gene expression levels could have on the randomization test results and M test results, and found that the effects are often not ignorable.