Hi @jwdebelius,
Thank you SO MUCH for your detailed response. Incredibly kind of you.
I've read the DAG-related post you linked here--this is helpful for understanding how to think about where features belong in models (i.e., should the taxa be our outcomes or our predictors? And does this even matter a whole lot?).
I've worked with interaction models before, and I have a statistician I can bug if needed as well. Thanks for the lead there.
For my benefit and the benefit of others who see this later (if I'm understanding correctly), it sounds like the tldr of all this is that I should
- Flip the way I'm thinking about modeling this and use existing tools that have the predicted outcome be the features (taxa), rather than using features (taxa) as predictors.
- For a specific modeling approach, I should use an interaction model where I interact time with my covariate. I'm aware of multiple existing tools that can do that, so I should choose one that performs well based on benchmark papers, seems to be appropriate for microbiome data (e.g., acknowledges compositionally, etc.), and run with that.
Is this an accurate summary?
Thank you again!