We first drew 100 frequency-matched samples (see earlier Section on "Matched Sampling") from the data set, S1, S2, …, S100, each of size n = 10, 000.
Apart from high technical variance, FFPE samples typically exhibit low gene expression intensities and a compression of fold change across experimental groups relative to matched FF samples (see, for example [ 11]), thereby compromising the ability to detect DEGs in samples preserved in this manner.
To compare the magnitude of variability observed in the four types of quality control samples, we calculated the relative percent difference (RPD) between matched samples [for details regarding RPD calculations, see Supplemental Material, "Quality control samples" (http://dx.doi.org/10.1289/ehp.1205821)].
To identify tumour-specific genomic alterations and exclude regions of potential germ line copy number variation, we normalised the gastric cancer profiles against the matched gastric normal samples (see the Methods section and supplementary figure S1 (available online only) for representative profiles).
No failures occurred: three trisomy 18, two trisomy 13, and all 15 euploid matched samples were sequenced and correctly classified by both laboratories (see Supplementary Figure S2 online).
A significant enrichment of PECAM1 was observed in microdissected vasculature when compared with the epithelial cell marker cytokeratin 8 (KRT8); conversely, the latter marker was enriched in matched microdissected epithelial RNA samples (see Additional File 2, Figure S1C).
Furthermore, inverse probability of treatment weight adjustments showed that cancer outcomes favoured surgery over radiotherapy for patients without metastatic prostate cancer, with no significant differences for the metastatic group, but as with the propensity score matched sample, the point estimates were attenuated (see appendix 2 on bmj.com).
The disparity in overall mortality rates was not resolved by matching the CPRD and DIN samples (see online supplementary table S2: 55% vs 42%), nor was it resolved after further matching the samples on smoking status (51% vs 42%); hence, disparity cannot be attributed to cohort differences in cancer rates, year of diagnosis or patient demographic factors.
However when new Alzheimer's disease samples were also included, they did not cluster together with matching samples from the previous experiment, see Figure 2C.
In our case, the flank regions were often imprecise, and only up to 24% of the reads would pass such perfect matching criteria in the different samples (see Figure 1).
