Abstract:
|
There are two statistical formulations to comparison of gene expression levels: tests of equality and confidence intervals. The tests of equality formulation has been more popular, but the confidence interval formulation introduced by Kerr, Martin, and Churchill (2000) has the advantage of providing biologically meaningful information on the magnitude of differential expression levels. Their confidence intervals are individual, not simultaneous, confidence intervals however. In this paper, we recommend using generalized Latin Square designs to avoid confounding of dye and array effects with the treatment effect. We show that if, as we recommend, each gene appears an equal number of times with each dye and array combination in the design, then simultaneous confidence intervals based on least squares estimates are readily computable. This happens because the comparisons turn out to be a form of Multiple Comparison with the Mean in this case. The computations of both normal error and bootstrap-t simultaneous confidence intervals are illustrated using a published data set of Kerr and Churchill (2000).
|