For this purpose one has to construct sets of cancers that are comparable; this is achieved by considering incident cancers in the two arms of the trial, ie removing length bias by ignoring cancers detected at the initial screening test for prevalence.
This plot is meant to be more heuristic in nature rather than any rigorous proof that dividing by length doesn't remove the length bias.
By removing the length bias from the EST database estimation, we achieve agreement between the two sets of data, thus explaining the apparent contradiction.
Using FET with only peaks near gene promoters removes nearly all of the length bias, but also ignores a large portion of the data.
The resulting permutations remove true biological association and the locus length bias inherent in the GO terms.
Arrays were then pre-processed using methods available in the R package "aroma.affymetrix" including techniques to remove systematic bias introduced due to PCR fragment length bias, differences in GC content and allelic cross talk.
All SNP Mapping arrays were first normalized using methods available in the R package "aroma.affymetrix" [19], including techniques to remove systematic biases introduced due to allelic cross talk, PCR fragment length bias and differences in GC content.
Dividing the mean by the length of the transcript removes the bias from that measure but subsequently introduces a length bias into the variance and the problem still persists.
The resulting permutations remove true biological association in the gene sets while maintaining any locus length bias.
Each experiment was normalized to mean 0 and std of 1, and the smoothed length bias (number of reads as a function of fragment length) was removed.
