Clarification of strata argument for adoinis2

[cazzlewazzle89]

Hi all,

As the title suggests, I'm hoping for some clarification on the strata argument in adonis2.
I have paired samples from the faecal microbiome of mice, one is a 'total' faecal sample, the other is an extract of bacterial extracellular vesicles from the same sample.
I also have two treatment groups.

I am trying to see whether Treatment impacts the extracellular vesicle-producing component of the microbiome, which using the paired 'total' faecal sample to control for the fact that the microbiome will differ between individual mice.
Should I use MouseID as the strata argument for PERMANOVA?

A standard two way PERMANOVA gives me no significant effect of Treatment

adonis2(formula = bray_dist_paired ~ metadata_paired$Source * metadata_paired$Treatment)
                                                 Df SumOfSqs      R2       F Pr(>F)    
metadata_paired$Source                            1   4.7516 0.43150 26.7180  0.001 ***
metadata_paired$Treatment                         1   0.2809 0.02551  1.5795  0.142    
metadata_paired$Source:metadata_paired$Treatment  1   0.2882 0.02618  1.6208  0.134    
Residual                                         32   5.6910 0.51681                   
Total                                            35  11.0117 1.00000                   
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ‘ ’ 1

But adding the strata argument makes it significant

adonis2(formula = bray_dist_paired ~ metadata_paired$Source * metadata_paired$Treatment, strata = metadata_paired$MouseID)
                                                 Df SumOfSqs      R2       F Pr(>F)    
metadata_paired$Source                            1   4.7516 0.43150 26.7180  0.001 ***
metadata_paired$Treatment                         1   0.2809 0.02551  1.5795  0.001 ***
metadata_paired$Source:metadata_paired$Treatment  1   0.2882 0.02618  1.6208  0.275    
Residual                                         32   5.6910 0.51681                   
Total                                            35  11.0117 1.00000                   
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ‘ ’ 1

I am unsure which method is correct. The R2 value doesn't change, so I think I'm just unsure of what strata actually does.

Thanks in advance for any advice,
Calum

[gavinsimpson]

As per ?adonis2:

strata
Groups within which to constrain permutations. The traditional non-movable strata are set as Blocks in the permute package, but some more flexible alternatives may be more appropriate.

Hence strata doesn't change the model (which is why the estimated effects don't change), but it does affect the permutations. As per the documentation, permutations are restricted to be within the levels of strata. In your case this means we shuffle at random each pair of samples from an individual mouse. As I doubt (as your description is not complete — "I have two treatment groups" is, almost entirely, uninformative) the treatment varies at the Source level (i.e. in your setup, each mouse only experiences one of the two treatment levels), it seems to me that neither permutation test is correct.

If I surmise correctly, you want to permute the groups of paired samples, while keeping the samples within the pair unpermuted.

I also don't understand what you were hoping to do with the Source variable — all that can be explaining is the differences in "average" composition of the samples belonging to each Source, which doesn't sound like it will "control for the fact that the microbiome will differ between individual mice". If you want to to do that, you would need to include MouseID as a parametric factor term in the model effectively as a nuisance term.

You're going to have to explain more (generically - I don't want to know about the scientific details of your study, just the experimental design elements) if you want more specific help. Absent that, all I can suggest right now is that neither of your permutation designs is correct and your model is incorrectly specified for the effects you wish to estimate.

[jarioksa]

Does the Discussion #600 help you? Have you read the vignette of the permute package?

If your problem is "paired", you cannot use strata, but you must define your permutation scheme in the permute package. See help in the permute package.