edgeR with uneven groups

[sindrle]

Hi!
I have two groups:

Healthy=26
Diabetic=22

> colnames(design)
[1] "(Intercept)" "DiseaseDiabetic"
[3] "DiseaseHealthy:Patients2" "DiseaseDiabetic:Patients2"
[5] "DiseaseHealthy:Patients3" "DiseaseDiabetic:Patients3"
[7] "DiseaseHealthy:Patients4" "DiseaseDiabetic:Patients4"
[9] "DiseaseHealthy:Patients5" "DiseaseDiabetic:Patients5"
[11] "DiseaseHealthy:Patients6" "DiseaseDiabetic:Patients6"
[13] "DiseaseHealthy:Patients7" "DiseaseDiabetic:Patients7"
[15] "DiseaseHealthy:Patients8" "DiseaseDiabetic:Patients8"
[17] "DiseaseHealthy:Patients9" "DiseaseDiabetic:Patients9"
[19] "DiseaseHealthy:Patients10" "DiseaseDiabetic:Patients10"
[21] "DiseaseHealthy:Patients11" "DiseaseDiabetic:Patients11"
[23] "DiseaseHealthy:Patients12" "DiseaseDiabetic:Patients12"
[25] "DiseaseHealthy:Patients13" "DiseaseDiabetic:Patients13"
[27] "DiseaseHealthy:TreatmentAfterTreatment" "DiseaseDiabetic:TreatmentAfterTreatment"

> fit <- glmFit(d,design)
Error in glmFit.default(y = y$counts, design = design, dispersion = dispersion, :
Design matrix not of full rank. The following coefficients not estimable:
DiseaseDiabetic:Patients12 DiseaseDiabetic:Patients13

How can I adjust for uneven groups?

Thanks!

[dpryan]

You're trying to test with patients that don't exist (since there are only 11 diabetic patients, you can't look for a DiseaseDiabetic:Patients12 interaction). Presumably you mistyped the design formula (it looks like you used ~Disease:Patients + Disease + Disease:Treatment, which doesn't really make sense given your experiment).

[sindrle]

Thats exactly what I did, because I dont understand the command...

design <- model.matrix(~Disease+Disease:Patients+Disease:Treatment)

I will google some. I dont understand the difference between really basic aspects. i.e. "*" vs "+" vs ":" vs "~".

I have fiddled around a bit, Ill try some more..


For future people struggling with the same read this:

https://stat.ethz.ch/pipermail/bioconductor/2013-July/053686.html

[dpryan]

I will google some. I dont understand the difference between really basic aspects. i.e. "*" vs "+" vs ":" vs "~".

Ah, yeah, that'll cause some issues then. Have a look at a few tutorials like this one (or even just the "details" section of "help(lm)" in R) to get a better feel about what these things are actually doing.

[sindrle]

Thanks!
I dont even understand why I should read up on linear regression.. Wow, I have alot to learn!

To sum up the fix for uneven groups:

Delete alle columns with only zeros i.e.

design <- design[,-24]

[dpryan]

FYI, once you've read up on linear models, you'll know why "design <- design[,-c(24,26)]" will fix the error but still not answer the actual biological question you want to ask

[sindrle]

Oh man, I look forward to that!

But hold on.. Is this incorrect then?

fit <- glmFit(d,design)

#Genes responding to Treatment in healthy patients
ExHealthy <- glmLRT(fit, coef="DiseaseHealthy:TreatmentTreatment")

#Genes responding to Treatment in diabetic patients
ExDiabetics <- glmLRT(fit, coef="DiseaseDiabetic:TreatmentTreatment")

#Genes responding to Treatment in any group
ExAny <- glmLRT(fit, coef=25:26)

#Genes responding differently to Treatment in diabetic vs normals
ExAny <- glmLRT(fit, contrast=c)

I dont know what "contrast" does.. Googling that aswell.

My brain is boiling, must sleep some hours..