[BioC] CyberT: Paired Bayes t-test (Bayes corrected p-values)

Gordon Smyth smyth at wehi.edu.au
Wed Jun 29 07:00:23 CEST 2005

My understanding is that CyberT doesn't handle paired t-tests, although I 
could be wrong. The closely related but more general method in the limma 
package does however. Here's a small example.

First simulate data. In your case, replace Y with the exprSet object you 
get from gcrma().

Y <- matrix(rnorm(8912*8),8912,8)

Sample <- factor(rep(1:4,each=2))
BeforeAfter <- factor(rep(1:2,length=8))
design <- model.matrix(~Sample+BeforeAfter)

fit <- lmFit(Y,design)
fit <- eBayes(fit)
topTable(fit, coef="BeforeAfter2", adjust="fdr")


>[BioC] CyberT: Paired Bayes t-test (Bayes corrected p-values)
>Groot, Philip de philip.degroot at wur.nl
>Tue Jun 28 09:49:03 CEST 2005
>I have been digging in Bayesreg.R for several days now, but cannot find
>the solution for my problem. I want to calculate a Bayes PAIRED t-test
>if this is possible (I have the impression that it is not). With a
>paired t-test I am refering to a biological sample before and after
>treatment from the same object. This test is performed for 4 different
>objects, so in total I have 8 samples: 4 before treatment and 4 after
>treatment. The samples are hybridized on a HG-U133 plus2 array and
>quality control, GC-RMA normalization and IQR-filtering have been
>applied. In total, 8912 genes remained for calculating the Bayes paired
>When looking at this site:
>http://visitor.ics.uci.edu/genex/cybert/help/index.html I can find an
>example on how a paired Bayes t-test can be performed (that is: if I
>understand it correctly): add to the original non-2log transformed data
>(so transform the GC-RMA values back to the non-log scale) an additional
>'Expr Est' column (my statistical background is not as such that this
>makes sense to me). Well, I did this (using numR=8 and doLog=TRUE) and
>calculated p-values between 0 and roughly 10^-6... Of course, this can't
>be true... Searching for additional help (which is hardly available for
>paired t-tests) I found a CyberT refence saying that the paired t-test
>refers to the 2-dye situation. It is paired because the same array is
>used. Well, in my situation (Affymetrix) this is not true, but I can
>calculate the ratio after / before treatment (on the same object) and
>put it in. This did make sense, but now I find p-values between 0 and
>1.5! It is weird that Bayesreg.R allows the generation of p-values
>larger than 1, so I am wondering whether this is an artifact of Bayesian
>statistics or that I am doing something wrong? Can someone please give
>me a hint on how to do the paired Bayesian t-test properly (e.g. with a
>small example dataset)?
>Kind regards,
>Dr. Philip de Groot
>Wageningen University

More information about the Bioconductor mailing list