Bayesian, MCMCglmm and multiple testing
Hi Fiona! Most of the FDR corrections using p-values are based on the assumptions that p-values are uniform between 0 and 1 under the null hypothesis, which would probably not be the case for the "pMCMC" output of MCMCglmm. Bayesian controls of the FDR do exist, but they are based either on posterior probabilities or Bayes Factors, no of which are yielded by MCMCglmm (to my knowledge). If you could achieve to derive a Bayes Factor from the output, then you could compute a q-value with (Bayesian) FDR control. Some information about that can be found on section 5 of the following document: http://onlinelibrary.wiley.com/store/10.1111/mec.12705/asset/supinfo/mec12705-sup-0001-Suppinfo.pdf?v=1&s=87375f6d520a98dc1a69594771db07065946b41f Hope this helps. Best Regards, Pierre. Le jeudi 23 juillet 2015, 07:39:06 Fiona Ingleby a ?crit :
Hi everyone, I?m working with gene expression data and am planning on running a mixed model with MCMCglmm for each gene in the dataset individually (>15000 models). With previous non-Bayesian approaches to this data, I have corrected results for multiple testing with the false discovery rate, and I?m wondering if there is a generally accepted way of correcting Bayesian results for multiple tests. I?ve had a look through some publications but I?m drawing a blank so would anyone be able to point me in the direction of some useful information? Either methods, or discussion about the consequences of multiple testing for Bayesian model results, would be really helpful. It has been suggested to me to simply use the pseudo-p-values in the MCMCglmm output to adjust p-values, but to be honest I?ve always ignored the pMCMC values as I?ve found the intervals much more useful, so I?m not sure how good a solution this would be. Thanks in advance for any help, Fiona
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models