CLR vs LOG transform

Downstream analysis and statistics MaAsLin
1

Hello there,

I am comparing Maaslin results for my dataset using various transformation methods. Per this previous post my understanding is that using CLR and Log transform should produce similar results Questions about normalization method CLR vs TSS - #3 by sma.

However, although I have some taxa with shared significance, many do not match. Do you have any idea why this may be occurring in my case?

I am uploading a relative abundance worksheet and running the following scenarios (all other parameters are the same):

  1. transform = NONE and normalisation = CLR
  2. transform = LOG and normalisation = NONE.

I should also mention that the relative abundances for samples do not add up to 1 as pre-filtering was applied, so most samples have total abundance of 0.75 - 0.85. Not sure if this will affect anything?

Many thanks for your assistance :slight_smile:

2

Hi there @Katb,

In general I wouldn’t be surprised to see differences in results when using CLR vs. TSS. The key thing to remember is that they are measuring the abundances in different ways. TSS is measuring proportions as a whole while CLR is measuring abundances in comparison to the geometric mean. So I wouldn’t be to concerned if you end up with different results but you should keep in mind the type of normalization you are using as it does impact the hypothesis that you are testing (and thus interpretation).

Thanks,
Jacob

3

Thanks Jacob. Just to confirm - is it correct to upload a relative abundance worksheet and select the CLR transformation? I wasn’t sure if I should be uploading count data? I thought for CLR the results would be the same whether I loaded count of relative abundance, but in this case I am getting different results. Is this because the count data has already been pre-filtered and only includes the counts for 0.75 - 0.85 relative abundance? Hope this makes sense…

Thanks,
Katherine

4

Hello,

In general when working with CLR transformations you should supply the raw count data to Maaslin2. It’s possible that filtering may be playing a part in the differences that you are seeing. To check that you could set the filtering to “NONE” and do your filtering outside of Maaslin2.

Thanks,
Jacob Nearing