i'd like to repeat what was done as part of: 'Summaries of Affymetrix GeneChip probe level data' by Irizarry et al. Nucleic Acid Res., 2003, 31(4):e15 http://nar.oupjournals.org/cgi/content/full/31/4/e15 that is, compare the Affymetrix I/D/NC calls/p-values for a differential expression experiment involving just 2 arrays with those based on the RMA method. in the above paper, it says: 'Test statistics can be created for RMA [...] based on estimates of standard error obtained from probe level data.' at this point, the authors reference: 'Exploration, normalization, and summaries of high density oligonucleotide array probe level data' by Irizarry et al. Biostatistics 4:249-264 (2003) http://biostatistics.oupjournals.org/cgi/content/abstract/4/2/249 it spells out the linear model that is least-squares-fitted to the measured intensities for the probes in a probe set: Y = mu + alpha + epsilon, where mu is the expression estimate (log- scale expression level for the respective array), alpha a probe affinity effect, and epsilon an independent identically distributed error with mean 0. the rma function of the affy R package provides just mu for a probe set, but not epsilon. function fitPLM of the affyPLM package returns what is called 'chip level parameter estimates' and 'standard errors'. i reckon that the latter correspond to mu and epsilon, respectively. please correct me if i'm wrong here. given mu and its standard error (SE) for the same probe set, but on 2 different arrays, i'd naively use t = |mu1 - mu2| / (SE1 + SE2) as the test statistic where df = 2 * number of probes in the probe set - 2. that provides me w/ a p-value that i can compare to Affymetrix'. does this make sense, and is this what was done in the paper at the top? cheers, Reiner -- (*)->[]->()->[]->(**)->[]->()->[]->(*)->[]->()->[]->()->[]->()->[]->() ->[] (Humboldt University Berlin, Germany)->[]-> ... (University of Maryland, USA)->[]-> ... (King's College London, UK) http://www.cs.umd.edu/~rschulz