Background: We provide a re-analysis of the Golden Spike dataset, a first generation "spike-in" control microarray dataset. The original analysis of the Golden Spike dataset was presented in a manuscript by Choe et al. and raised questions concerning the performance of several statistical methods for the control of the false discovery rate (across a set of tests for differential expression). These original findings are now in question as it has been reported that the p-values associated with the tests of differential expression for null probesets (i.e., probesets designed to be fold change 1 across the two arms of the experiment) are not uniformly distributed. Two recent publications have speculated as to the reasons the null distributions are non-uniform. A publication by Dabney and Storey concludes that the non-uniform distributions of null p-values are the direct consequence of an experimental design which requires technical replicates to approximate biological replicates. Irizarry et al. identify four characteristics of the feature level data (three related to experimental design and one artifact). Irizarry et al. argue that the four observed characteristics imply that the assumptions common to most pre-processing algorithms are not satisfied and hence the expression measure methodologies considered by Choe et al. are likely to be flawed.
Results: We replicate and extend the analyses of Dabney and Storey and present our results in the context of a two stage analysis. We provide evidence that the Stage I pre-processing algorithms considered in Dabney and Storey fail to provide expression values that are adequately centered or scaled. Furthermore, we demonstrate that the distributions of the p-values, test statistics, and probabilities associated with the relative locations and variabilities of the Stage II expression values vary with signal intensity. We provide diagnostic plots and a simple logistic regression based test statistic to detect these intensity related defects in the processed data.
Conclusion: We agree with Dabney and Storey that the null p-values considered in Choe et al. are indeed non-uniform. We also agree with the conclusion that, given current pre-processing technologies, the Golden Spike dataset should not serve as a reference dataset to evaluate false discovery rate controlling methodologies. However, we disagree with the assessment that the non-uniform p-values are merely the byproduct of testing for differential expression under the incorrect assumption that chip data are approximate to biological replicates. Whereas Dabney and Storey attribute the non-uniform p-values to violations of the Stage II model assumptions, we provide evidence that the non-uniformity can be attributed to the failure of the Stage I analyses to correct for systematic biases in the raw data matrix. Although we do not speculate as to the root cause of these systematic biases, the observations made in Irizarry et al. appear to be consistent with our findings. Whereas Irizarry et al. describe the effect of the experimental design on the feature level data, we consider the effect on the underlying multivariate distribution of putative null p-values. We demonstrate that the putative null distributions corresponding to the pre-processing algorithms considered in Choe et al. are all intensity dependent. This dependence serves to invalidate statistical inference based upon standard two sample test statistics. We identify a flaw in the characterization of the appropriate "null" probesets described in Choe et al. and we provide a corrected analysis which reduces (but does not eliminate) the intensity dependent effects.