The simple repeated measurements design of the MAQC data set allowed us to study the trade-off between less degrees of freedom and variance reduction caused by inclusion of the array effect from a power perspective.
This indeed suggests that intensities do not seem to be influenced by the co-hybridized sample or array used, as the array effect appears to be smaller than the treatment effect.
We therefore, for the cell line dataset, compared the ranking of genes by p-values generated with two different ANOVA models: one including the array effect (the ratio analysis), and one without the array effect (the intensity analysis).
Since not all treatments are observed within the same array, the spot-specific array effect is corrected as a random effect to the model (1) and estimated by REML [ 36].
For dual-color arrays specifically, an array effect is included in the model to determine the technical noise introduced by any between-array differences.
Meanwhile the standardized estimates of differential expression (y i ) are adjusted for all other sources of variation fitted in the within-study model (i.e. dye and array effects), and standardized by the standard error of the estimate.
Finally, we apply SNM to a scenario based on two-color experiments where the signal is influenced by intensity-dependent dye and array effects.
To further investigate the magnitude of array-specific effects we compared the relative effect sizes of the array and treatment effects as defined by the ANOVA model.
Furthermore, we estimated the effects of a potential unmeasured or incompletely measured confounder on the main outcome, by using the array approach described by Schneeweiss.
This greatly reduces the possible confounding effects of inter-array variability and local array effects.
