Figure 9

Correlations of pairs of genes sampled from subsets with an increasing minimum percentage of present calls for probe sets. From the breast cancer data, 5000 pairs of genes were randomly sampled from the subset of probe sets that were present on at least one of the chips, the subset of probe sets that were present on at least 20% of the chips etc., finally from the subset of probe sets present on all chips. The three curves represent the three expression measures. Excluding the probe sets that are always absent already reduces the residual correlation considerably, excluding the probe sets that are more than 80% absent breaks the pattern of highest correlation at smallest variability for RMA and MBEI seen in Figures 2, 3, 6, and 7. Using only probe sets that are always present however leads to a notable increase in residual correlation, especially for MAS5.