In wet-lab biology, gene expression is often measured as the difference between the expression levels of the gene of interest and a reference gene. An ideal reference gene exhibits minimal variance in expression level across a large variety of biologically interesting conditions. Identifying genes with lowest variation from high throughput gene expression data will yield predictions as to which are the best reference genes. To solve the problem, we will use a novel non-parametric probe-level gene expression analysis method DEMI (Differential Expression from Multiple Indicators). DEMI estimates differential gene expression by first establishing differential expression on probe level and then on the target (e.g. gene) level. Probe-level analysis will be conducted by plugging in a statistical test that calculates an estimate of variance and the associated lower-tail probability based on the empirical distribution or a theoretical H0 distribution. It is worth noting that highly expressed genes can exhibit artifactually low variance due to the limited dynamic range of microarrays (ceiling effect). Therefore it might also be necessary to estimate the inter-sample differential expression of the 10% of genes with lowest variation.