Figure 1.

When to use constant-mean normalization. The constant-mean assumption adds little noise for array designs with sufficiently large numbers of randomly selected genes. Assuming that the mean expression on arrays in a dataset would indeed be constant for an array monitoring the entire transcriptome, we chose random subsets of genes of each possible size and computed the CV of the mean expression level for hypothetical arrays monitoring just those subsets of genes. For arrays measuring more than about 10% of the genes, the level of variability introduced is not significantly larger than other sources of array variability, so normalization using the constant-mean assumption is reasonable. With fewer genes, the noise introduced by making this assumption grows dramatically, so other normalization methods may be desirable. Note that if there is bias in the selection of genes on the array, this effect may be much stronger.