[BioC] How to adjust the power.t.test R function to reflect different experimental designs?

[Davis, Sean (NIH/NCI) [E]]

Hi, Nadia.

Microarray studies are a bit different from typical statistical problems because the number of tests is typically quite large compared to the number of samples.  So, power.t.test will probably not get you directly to your desired result.  

There have been numerous papers that attempt to define exactly what you are trying to define (at least if I understand you correctly).  You might start by searching:

http://scholar.google.com/

for "microarray design".  

Sean


-----Original Message-----
From: Nadia Messerschmidt [mailto:nadia.messerschmidt at gmail.com]
Sent: Sat 10/21/2006 9:39 AM
To: bioconductor at stat.math.ethz.ch
Cc: Dave Berger
Subject: [BioC] How to adjust the power.t.test R function to reflect different experimental designs?
 
Hi,

I'm am currently a postgraduate bioinformatics student at the University 
of Pretoria, South Africa.

My background is from computer science, and I'm trying to create 
software that would enable biologists to get a better feel for what the 
statistics behind the experiments are. And how different experimental 
designs (loop or reference) would influence the statistics.

I thought I would start with time course experiments since they are used 
quite a lot at our lab. I will need a generic approach that enables all 
time point comparisons and the profiles of genes across all time points.

If you for instance do a small pilot study (would the design matter 
here?) and find that the population coefficient of variability is 30 % 
for argument's sake, you can enter that value into the power.t.test and 
get a feel for the different possibilities (changing sig level, power ect).

But now, for the large-scale study, the question is, would a loop or a 
reference design be better? Would it be possible to adapt the 
power.t.test parameters somehow so that is would reflect the different 
designs, or would you have to do two pilots studies, one loop and one 
reference, get the var of each and put those into the power.t.test? Or 
is there some way that you can take the variance from the pilot study, 
say using a loop design, and adjust that to reflect a ref design?

Is what I'm trying to achieve feasible at all? Because I read another 
paper saying that the effectiveness of the loop design becomes less if 
you have more than 10 time points - something that surely can't be 
accounted for in the power.t.test params?