small sample number

[Alina-Ba]

Hi! Thanks for developing the method! I have a very small number of samples (n=3 per condition). My samples are also paired - 3 donors, each control and stimulated. Is it realistic to get credible effects with this small number of samples? Is there any way to include the paired designe? Do you have any recommendations for me? Thanks!!

[mschilli87]

scCODA can even be used to find credible effects using a single sample per condition

[Alina-Ba]

In theory maybe. But I have a quite strong change in one of my cell types and it is not detected as credible (I assume due to the small sample size) - so I am wondering if there are any parameters I can adjust. I'm guessing there is no way to take the paired design into account, right?

[johannesostner]

Hi! Yes, you are correct in your assessment that scCODA can only detect very large effects if the sample size is small. This is the case in any kind of statistical model - fewer samples means higher uncertainty. Regarding the paired design, there is unfortunately no official support for such scenarios in scCODA atm. Adding sth like a subject-specific random intercept into the model formula might not throw any errors with the current implementation, but I am unsure about how the resulting parameters can be interpreted or if this will improve your results

[Alina-Ba]

Thanks a lot for your response @johannesostner Could you maybe specify how you would suggest including a subject-specific random intercept? Do you have any other suggestions for changes in variables, FDR cutoffs, or methods considering my small sample number?

[mschilli87]

Have you tried pooling your replicates (i.e. use the condition as 'sample')? This obviously takes away all advantages of doing replicates but as you get more cells per 'sample' you might find your expected change to be credible. Not sure how to interpret such a result though.

[Alina-Ba]

Thanks @mschilli87!! I tried it and indeed I get an effect - at least for a FDR of 0.2!

@johannesostner were you suggesting something like this: mod.CompositionalAnalysis(data_all, formula="Condition + Donor",reference_cell_type="Pericytes") Would this be valid to take into account the paired design (Donor 1-3, each control and stimulated)?

Thanks a lot for your help with this!