[BioC] Welch approximation: Degrees of freedom impact on pvalues

Benjamin Otto b.otto at uke.uni-hamburg.de
Thu Jun 15 10:57:47 CEST 2006


maybe this is rather a question for statisticians. Suppose I have two groups
of gene expression values, say group A and B with 10 samples each (20 in
total). Performing a t-test I would choose the one with unequal variance,
thus welch approximation. Usually the number of degrees of freedom (DF) in
this case seems a little bit more complicated than with the equal variance
version where I would substract 2 from the 20 (DF=18). The t.test package
performs this calculation automatically.

Now, digging in literature I still have the feeling, that I don't really
understand what HAPPENS when I just combine the easy DF version with the
welch approximation. Of course I get different pvalues and the DF will be
equal for each gene, although the variances might be different. But what is
the REAL consequence? Will the pvalues be unreliable by definition and why?



Benjamin Otto
Universitaetsklinikum Eppendorf Hamburg
Institut fuer Klinische Chemie
Martinistrasse 52
20246 Hamburg

More information about the Bioconductor mailing list