Negative B-Statistic Values In Limma/Microarray Differential Expression

Question

[CrazyB]

A question concerning the interpretation of microarray expression data -

After running GEO2R, which uses limma-based package for value comparison, I got a table of genes with their

"ID"    "adj.P.Val"    "P.Value"    "t"    "B"    "logFC"    "Gene.symbol"    "Gene.title"

From what I read on google, B-value and the moderated t (and hence P or adj P value) should rank the genes in the same order, which is true for my output table. However, according to other postings on google, only probes with positive B-values are thought to have a "differential expression". In my output, among the 54000 probes, only 85 probes show a "SIGNIFICANT" p-value (<0.05) but with positive B-values (many with p-value <0.0006 also have negative B-values).

How do we interpret those probes with low p-value but negative B-values ? Any suggestion or direction for more information on this issue ? Great many thanks.

[Sean Davis]

A p-value of 0.0006 may not be significant with 54000 tests. Are the lowest p-values also the largest B-values?

[CrazyB]

Ah, so is it correct that we should or must take into account the total number of tests in limma to gauge the significance ? And yes, the lowest p is 0.00000265 and it does have the largest B (3.055469). But this would mean most of the genes in the array show NO differential expression, right?

[Neilfws]

The output from topTable includes the "adj.P.Val" column. This is the p-value adjusted for multiple testing. By default, the method used is "BH" (Benjamini & Hochberg), also known as "fdr" (false discovery rate). See ?p.adjust.

Answer 1

[Sean Davis]

The B-statistic is the posterior odds of differential expression. Gordon Smyth has written a bit about it, but the take-home message is that since we do not often know the prior probability of differential expression, the B-statistic is not very useful. Instead, the focus should be on biological significance as measured by fold change and statistical significance as measured by a multiple-test-adjusted significance measure, such as the false discovery rate.