[R] Use of betatree {betareg} to estimate true positives from multiple tests
r_1470
r_1470 at yahoo.co.uk
Sat Jan 30 15:43:52 CET 2016
Hi, Doesanyone use the 'betatree' function in the betareg package to do a kind of falsediscovery rate (FDR) test for a set of many p values? I wasthinking to compare the beta parameters of the true distribution of about 1000p values with p values of permuted data, and test whether the two distributionswere significantly different, with a greater mass of low p values in the truedata. I know this is possible with betatree, I'm just not sure if it's a 'normal'method. I knowuniform-beta mixture models are commonly used to examine FDR, and there are R packagesavailable, but my tests are non-independent so the null distribution may not beuniform and has to be found empirically. The mixture model approach assumes there are two groups: truepositives (beta distribution) and true negatives (uniform distribution). I think my approach would assume a continuous distributionof effect sizes in the true data, rather than 2 distinct groups, with a point mass at 0 effect size for thepermuted data.
I also looked at using the 'betamix' function, but this is much less accurate and takes longer to run. Best Richard.
[[alternative HTML version deleted]]
More information about the R-help
mailing list