[BioC] unbalanced design

Arne.Muller at aventis.com Arne.Muller at aventis.com
Mon Dec 6 18:11:31 CET 2004

Dear All,

I'm wondering which method to use to analyse a rather unbalanced design (Affy). I've three studies (S1..S3) in which several doses of a drug have been tested. The table blow gives the number of observations (the replicates, cell cultures) per study/dose combination:

		S1	S2	S3
0.00mM	4	3	3
0.01mM	0	0	3
0.10mM	3	3	0
0.25mM	2	3	3	
0.50mM	3	3	0
1.00mM	3	3	3

For the moment the entire experiment is normalized all together (RMA+Quantile).

I expect a very strong study effect (different laboratory protocols have been used!), and the dose effect should be mainly independent (small interaction). I think lme would be the right choice, since the study effect is an (unwanted) random effect, but I'm not sure using it for an unbalanced design. I'd use:

value ~ dose, random = ~ 1|study

In addition I'd like to test the interaction (I don't expect much interaction),  with a fixed effects model:

value ~ study + dose + study:dose

Again, I'm not sure whether this model would be meaningful because of the unbalanced design (especially with repsect to testing interactions).

I'm interested whether there's a general dose effect, i.e. genes consistently altered across  studies.

Would limma handle such a design, or are there other packages that could be used for this (e.g. testing for a trend or dose dependence across studies)?

I'd be happy for any discussion and comments.

	kind regards,


More information about the Bioconductor mailing list