Hi Benjamin, Jim gives a good answer to your question. The MTP function uses a bootstrap null distribution to compute p-values for Cox regression coefficients (permuation null distribution is available for t- and F-tests, but not Cox or linear regression). If your arrays are iid and your sample size is large enough, these should be good p-values. But, they can be problematic in other settings. The nice thing is that you do not have to make assumptions about the distribution of the test statistics when you use the bootstrap estimated null distribution. FYI: there was a bug in multtest_1.5.2 so that rawp (and hence adjp) computed with the MTP function were too small. This has been fixed in multtest_1.5.4, which is available under developmental packages on the BioC website. You should install and use this package or multtest_1.6.0, which will be released next month. Cheers, Katie ***************************************************************** Date: Sun, 17 Apr 2005 18:50:09 +0200 From: Benjamin Haibe-Kains <bhaibeka@ulb.ac.be> Subject: [BioC] multtest and Cox regression To: Bioconductor Mailing List <bioconductor@stat.math.ethz.ch> Hi, I would want to perform a filtering of my genes according to the p-values of univariate Cox regression (so, if I have 100 genes, I compute 100 Cox regression with only one gene at a time). Because I perform multiple statistical test, I would want to use a multiple testing procedure to get adjusted p-values. I have seen that the MTP function from the multtest package does exactly what I need. When I used this function with the Cox parameter r <- MTP(X=data, Y=Surv(time, event), test="coxph.YvsXZ", B=100) and I look the unadjusted p-values (r@rawp), they do not correspond to the univariate p-values returns by the summary of the coxph function (p.value <- 1 - pchisq(z^2, df=1) where z is the z-statistic). Actually, when I look at the code, the computed statistic is identical but the distribution against this statistic is compared is not the same (in the MTP function, this function is estimated by bootstrap, it's not the chisq). I don't understand this fact. Can anyone give me further details about that ? Best, -- Benjamin Haibe-Kains [http://www.ulb.ac.be/di/map/bhaibeka/] ------------------------------ Date: Sun, 17 Apr 2005 17:33:27 -0400 From: "James W. MacDonald" <jmacdon@med.umich.edu> Subject: Re: [BioC] multtest and Cox regression To: Benjamin Haibe-Kains <bhaibeka@ulb.ac.be> Cc: Bioconductor Mailing List <bioconductor@stat.math.ethz.ch> With microarray data it is difficult if not impossible to determine the correct null distribution for each of the genes on a given chip. Given this fact, you really have two choices; either assume that the null is what you would expect (e.g., use a t-distribution for t-tests, an F-distribution for F-tests, a chisquare distribution for a Cox model), and take the chance that you are wrong for some (most, all?) of your genes, or don't assume anything about the null and estimate it using a bootstrap or permutation distribution. There are pluses and minuses to each approach, and you have to decide which method is applicable to your situation. The multtest package is designed to use bootstrap or permutation null distributions. If you simply want to rely on the chi-square distribution and adjust for multiplicity, you might want to look at ?p.adjust. Best, Jim