Hi, normally we search for differentially expressed genes in different observation or treatment groups. So, in a very basic way, one performs a t.test for each gene between the two groups and takes the p-value as measure for significance. Now, is it a) possible and b) reasonable to test whether the two treatments may lead to differentially high expression variances (not means) in the groups? To give a very simple biological example I could compare non-tumor to tumor cells. By intuition I would conclude that the non-tumor cells should have not only no differentially expressed genes but also nearly no variance in expression level per gene between the samples which are member of this group. However the tumor cells could have as one possibility higher/lower expressed genes (different means, the normal thing) or as second thought genes which are just kicked out of balance and thus exhibit an extraordinary high variance between the tumor samples. Now how do I test that? With a simple F-test between the two groups across each gene? And for a more global test with a hypothesis like "Tumor cells exhibit more variance in gene expression across samples than non-tumor cells", do I compute the variance across each gene for each group and perform a t.test afterwards between the tumor- and non-tumor-variances? If this approach seems reasonable, then what is the correct measure to use, variance or standard deviation? The funny thing is, that when I perform a t.test for two "variance" groups of mine I get a p-value of 0.3 while the test for "sqrt(variance)" returns one of 2.3e-16. That really surprises me. Regards Benjamin -- Benjamin Otto Universitaetsklinikum Eppendorf Hamburg Institut fuer Klinische Chemie Martinistrasse 52 20246 Hamburg

Hi Benjamin, the question is very reasonable. More general, you might not just look for differential variance, but for different distributions in the groups. Have a look at the EDD package. Also, the work by Margarete Pepe on the pAUC statistic might be relevant. You need a substantial number of samples in each group to meaningfully do these things (more than if you just test for difference of location) - that is probably one of the reasons why difference of the locations is what most people go for. Best, Wolfgang > Hi, > > normally we search for differentially expressed genes in different > observation or treatment groups. So, in a very basic way, one performs a > t.test for each gene between the two groups and takes the p-value as measure > for significance. Now, is it a) possible and b) reasonable to test whether > the two treatments may lead to differentially high expression variances (not > means) in the groups? > > To give a very simple biological example I could compare non-tumor to tumor > cells. By intuition I would conclude that the non-tumor cells should have > not only no differentially expressed genes but also nearly no variance in > expression level per gene between the samples which are member of this > group. However the tumor cells could have as one possibility higher/lower > expressed genes (different means, the normal thing) or as second thought > genes which are just kicked out of balance and thus exhibit an extraordinary > high variance between the tumor samples. Now how do I test that? With a > simple F-test between the two groups across each gene? > > And for a more global test with a hypothesis like "Tumor cells exhibit more > variance in gene expression across samples than non-tumor cells", do I > compute the variance across each gene for each group and perform a t.test > afterwards between the tumor- and non-tumor-variances? > > If this approach seems reasonable, then what is the correct measure to use, > variance or standard deviation? The funny thing is, that when I perform a > t.test for two "variance" groups of mine I get a p-value of 0.3 while the > test for "sqrt(variance)" returns one of 2.3e-16. That really surprises me. > > Regards > > Benjamin > > -- ------------------------------------------------------------------ Wolfgang Huber EBI/EMBL Cambridge UK http://www.ebi.ac.uk/huber