This is a series of diagnostic plots that were used to evaluate how well a particular statistical model fits the data and explains the sources of variation in an Illumina 450k dataset.

• mean-var-model.pdf shows the variance trend modeling performed by voom, a method originally designed for mean-variance modeling in RNA-seq data. In this case, it models the mean-variance dependency induced by the logistic transform used for converting beta values (i.e. percent methylation) to M-values (i.e. ratio of methylated to unmethylated signal) in methylation data. Page 2 shows the mean-variance trend after fitting the model with the voom weights to cancel out the trend.
• sample-weights.pdf Shows the results of limma's arrayWeights method, which detects and down-weights outlier samples, plotted against all known clinical covariates for those samples. Diabetes status had a significant association with the sample weights, indicating that the Type I diabetes samples were overall more consistent and had fewer outlier observations that Type II diabetes samples.
• pcoa.pdf shows a Principle Coordinate Plot (similar to a PCA plot) of all the samples after subtracting out the effects of known covariates. Points are sized by their sample weight, and a crosshair shows the center of mass of each group.
• pval-histograms.pdf and pval-cdf.pdf show the p-value distributions for each contrast of interest, presented as a histogram and as an empirical cumulative distribution function. Each is annotated with asymptotes indicating the estimated fraction of probes affected by that contrast.