Transforming Alpha Diversity Metrics for Mixed Effects Modeling

General Discussion
1

Hi Qiime 2 community!

I have a question regarding transforming alpha diversity metrics for mixed effects modeling and how to present the transformed data.

My study is a repeated measures design. I have analyzed four alpha diversity metrics (Chao1, Faith's PD, observed features, and Shannon's Index) using mixed effects modeling in SAS with the mixed procedure. I assessed each model for normality and found that the distribution of the residuals for Faith's PD is not normal. The QQ plot, residual differences plot, and histogram are not visually too wonky in my opinion, but the Shapiro-Wilk test and other statistical tests for normality are significant.

I then tried log transforming the Faith's PD data and the normality test passed.

Was this the correct transformation for this type of data?

I currently have a table of the means of each metric by experimental group. Do I replace the Faith's PD means with the log transformed data?

When I'm presenting this data I want to make graphs for visualization. Should I use the log transformed data for this as well and just tell the audience that it was transformed?

Would appreciate any opinions on this!

Thanks in advance!

2

Hey @Mahasti,

A log transform is pretty mild and I don't think there is any "right" transform for this data. You can try a sqrt or Box-Cox transform if you want to be complete. But I imagine you are just dealing with some mild skew which log is pretty good at minimizing since you mention the histograms look pretty fine.

I would definitely report that you have log-transformed the response for Faith's PD then leave it log-transformed for any further discussion and visualization, since that's what your model will be based on. Then you get to contextualize the complicated model with the simpler summaries instead of having a disconnect on one index.

1 Like
3

@ebolyen Thank you. This helps a lot!

1 Like