ADONIS interpretation sanity check

General Discussion
1

Hello QIIMEers,

I'm staring at P-values and starting to question my own sanity. If someone out there can double-check the following interpretations I'd be very grateful!

The Experiment: I have microbiome data from an Icelandic glacial foreland. I've collected samples across three years (one collection per year, yearCollected). Those samples come from 5 age ranges in the foreland (how long that part of the land has been free of glacial ice, iceFreeYears).

I'm currently focusing on teasing out any differences based on the year the data was collected, since that seems to be a large driver of the dataset. To that end, I've done PERMANOVA, PERMDISP and ADONIS with both weighted and unweighted unifraq. Here are the p-value results:

Weighted Unifraq:

  • Significance results for yearCollected: PERMANOVA 0.001 PERMDISP 0.003

  • Signigicance results for iceFreeYears: PERMANOVA 0.042 PERMDISP 0.874

  • ADONIS formula yearCollected+iceFreeYears: yearCollected 0.001; iceFreeYears 0.16

  • ADONIS formula yearCollected*iceFreeYears: yearCollected: 0.001; iceFreeYears 0.187; yearCollected:iceFreeYears 0.755

Unweighted Unifraq:

  • Significance results for yearCollected: PERMANOVA 0.001 PERMDISP 0.064

  • Signigicance results for iceFreeYears: PERMANOVA 0.007 PERMDISP 0.912

  • ADONIS formula yearCollected+iceFreeYears: yearCollected 0.001; iceFreeYears 0.032

  • ADONIS formula yearCollected*iceFreeYears: yearCollected: 0.001; iceFreeYears 0.038; yearCollected:iceFreeYears 0.381

My interpretations:

  1. When looking at total ASVs (unweighted Unifraq), iceFreeYears and yearCollected have different centroids (based on significant p-vale of the PERMANOVA), but neither are different enough to be statistically significant (based on non-significant p-value from PERMDISP).

  2. When looking at most abundant organisms (weighted Unifraq), iceFreeYears and yearCollected have different centroids (based on significant p-vale of the PERMANOVA), but yearCollected is statistically significant (based on significant p-value from PERMDISP).

  3. ADONIS of total ASVs (unweighted Unifraq) shows that after partitioning out yearCollected, iceFreeYears is significant, but there's no significant interaction between yearCollected and iceFreeYears ?? (the yearCollected*iceFreeYears values). (this is where I'm getting turned around the most, hence the ??)

  4. ADONIS of the common ASVs (weighted Unifraq) shows that after partitioning out yearCollected, iceFreeYears is NOT significant, and there's no significant interaction between yearCollected and iceFreeYears ?? (the yearCollected*iceFreeYears values).

I've gone through many forum posts to get this far, thanks to everyone who has worked through similar questions with their own datasets. I am going through this data all alone and simply need someone else to tell me if my interpretations make sense (or where I went wrong!) given the test results.

Thank you!

1 Like
2

Hello Emmett!

That sounds like a cool project! (pun intended :sunglasses: :ice_cube:)

Would you be willing to summarize all those stat tests in a table? This lets us compare the different results from different tests.

Here's how I would summarize these stat tests, to get you started.

  • PERMANOVA and Adonis: look for overlap between samples in an ordination. While the math is different under the hood, they are trying to measure the same thing, so results should be similar.

  • PERMDISP measures dispersion / size / range / variance within each group, then compares groups to see if one group is more 'spread out / variable' than another.

  • abundant organisms (weighted Unifraq)

    Well said!

    total ASVs (unweighted Unifraq)

Well yes, but also no. Because it does not weight by the most common, you could also say unweighted UniFrac gives more preference towards less common microbes; The small fraction of uncommon microbes has a proportionally larger effect on unweighted UniFrac. :balance_scale:

1 Like
3

Thanks Colin.

Here's the Google Sheet summary of everything I have. I really appreciate having another set of eyes on this. I've typed some notes to myself in that doc, please critique any of my interpretations. :slight_smile:

4

I've collected samples across three years (one collection per year, yearCollected). Those samples come from 5 age ranges in the foreland (how long that part of the land has been free of glacial ice, iceFreeYears).

I see era with 4 degrees of freedom. This must be iceFreeYears.
I see yearcollected with 1 degrees of freedom, which implies 2 years. Where's your other year?

I notice you have tested all 4 ways of setting up this ANOVA. An ANOVA of ANOVAs, if you will.

var1+var2
var1*var2
var2+var1
var2*var1

ADONIS: IncludeInteraction+VarOrder (when we look at which variable to test first, controlling for if we include interaction)

Tell me what question is most important to you by removing the extra stat tests.

  • Do you want to include the interaction effect or not?
  • Are era and yearcollected independent?
    • If so, which one is most important biologically?
    • If not, which is dependent?
5

I haven't read the full thread here so I could be off, but just as a heads up, PERMDISP shouldn't be used with continuous variables, as it converts them to factors under the hood without any warning. This may explain the 1 degrees of freedom as every number is its own factor now. You can convert your variable into quantiles or something and use that instead.

1 Like