Fold change

Fold change is used in one of two cases (in general). First, you only
have one chip for experiment and control, so that is all you can do. In
this case you have to assume that the genes with the big fold changes
are actually differentially expressed. There are two problems with this.
First, you are not accounting for the variability in the measurement of
that gene, so the big fold change may be due completely by chance.
Second, there is no inherent control for the absolute magnitude of the
intensities. For instance, if the intensity for the experimental gene is
100, and the control is 25, then the fold change will be four-fold and
you will assume differential expression. However, this is very close to
the noise level, so you have no idea if you can trust either value.
Contrast this with an intensity for the experiment of 10,000, and 2500
for the control. Here you likely have a difference, but just looking at
fold changes you cannot distinguish between the two.

If you are using fold change when you have replicated data (instead of
a t-test), you are ignoring the estimate of variability. The t-test has
been shown to be the universally most powerful test for comparing
(normally distributed, equal variance) data, so you *cannot* devise a
test that will do a better job. You can argue that the data are not
normal, and don't have constant variance, but you still will have a hard
time beating a t-test, and things like SAM are simply designed to
increase the power of the t-test for specific violations of the
underlying assumptions.

Jim
10:28:11

Newton, MA et al, J. Comp Biol, vol8, no1, pp37-52, 2001
Baggerly, KA, et al, J. Comp Biol, vol8, no6, pp639-659, 2001
Chen, Y, et al, J. Biomed Optics, 2, 364-374, 1997
Dudoit, et al,
http://www.stat.berkeley.edu/users/sandrine/Docs/Papers/sinica.final.pdf
Lee, MLT, et al, PNAS, 97, 9834-9839, 2000