Using the full potential of microarray data as an experimental tool will require appropriate statistical analyses. By its nature, it is quantitative data: data which exhibit continuous, rather than discrete, variation. The data may be fluorescence intensity, or ratios of fluorescence intensity at different wavelengths representing two- or four-color systems. Investigators may impose thresholds on the data to make them appear discrete, but information may be lost or is at best underutilised in the process. Further, some if not all of the genetic variation being measured is itself quantitative (e.g. regulatory variation). An appropriate statistical analysis which takes this into account is desirable. Further, microarray data has the additional complication that hundreds of tests are assessed simultaneously. Thus it is likely that some of the tests will appear significant purely by chance (type I error). However, traditional approaches to multiple testing, such as the Bonferroni approach, are unsatisfactory because they are too conservative, that is, too likely to reject effects of biological interest. This is particularly true if the tests are not independent, as would be the case where genes are part of a single developmental cascade or network. Two tests are proposed that address these issues, an analysis of variance (ANOVA) approach and permutation. The ANOVA allows the examination of all sources of variation contributing to continuous data on a spot by spot basis, evaluating not only the statistical significance but the relative contributions of these effects. The ANOVA includes the main effects of genotype and treatment, as well as assessing variation between independent RNA preparations and between chips. This test functions to identify all loci of potential interest, but ignores the multiple testing problem. In contrast, the permutation test focuses on multiple testing. This test reshuffles, or permutes, the identities of the spots and their fluorescence values to generate an empirical null distribution appropriate to each unique dataset. In contrast to Bonferroni corrections, however, permutation takes the correlation structure within the dataset into account, thereby generating a biologically appropriate response to non-independence. The tests will be compared using a sample microarray dataset comparing the well-studied response of Drosophila melanogaster to heat shock across five genetically distinct stocks. By using the heat shock system, we will be able to assess how well the tests perform in a defined context.