Assessment of estimator performance by posterior predictive expected squared error for the MAQC data sets. Total posterior predictive squared error relative to that of the gold standard model (the local-FDR mean expression estimator calculated using t-statistics and the theoretical null, labeled "t stat. with locfdr" in the figure) for Applied Biosystems, Affymetrix 1, and Affymetrix 2 data sets of the rat toxicogenomics subset of the MAQC study. Error definitions are the same as those of Figure 2. Algorithm definitions are the same as those of Figure 1. Results for the Applied Biosystems data set with the "all nulls true" estimator and for the both Affymetrix data sets with the local FDR estimator with empirical null (based on both Wilcoxon statistics and t statistics), the fold change shrinkage estimator, all fold change hard threshold estimators, the "all nulls true" estimator, and the "all nulls false" estimator are greater than 1.023 and are not plotted.