I have a more statistical question regarding all the different significance calculations done in qiime2 and especially their interpretations.
For my data, I tried to follow the PD mouse tutorial beta diversity section as much as possible.
I performed a PERMANOVA (beta-group-significance command in qiime2) using metadata column "treatment" and metadata column "room"
for room: unweighted unifrac (p= 0.011) and weighted unifrac (p= 0.818)
for treatment: unweighted unifrac (p=0.043) and weighted unifrac (p=0.319)
What is the difference between unweighted and weighted and what does this say about my data?
I continued with the unweighted treatment data and also performed a permdisp on this data. This was not significant (p=0.349). From this, I conclude that the significant result in unweighted unifrac distances for treatment is not caused by a high dispersion. Is this correct?
then I continued with an adonis using the formula: treatment+room
There I found both room and treatment to be significant (room - p = 0.002, treatment p = 0.008)
I also tried running the adonis with formula = room ( p = 0.002) and formula = treatment (p = 0.041). These were also both significant.
I am a bit lost in the statistical maze of all kinds of different analysis methods
What have I exactly calculated and what do these results mean?
I hope someone can give me a bit of clarity on the different methods!
Great questions! I'm more of an numerical ecologist than a 'real' statisticien, but hopefully I can help.
Here, the unifrac distance is being 'weighted' (or not!) by the abundance of the ASVs you are comparing between samples. Unweighted compares all differences in branch lengths equally, while weighted gives more weight to ASV based on their abundance in the data set.
This means that when you look at all ASVs there's a significant difference (p=0.043), but the most common ASVs that make up most of the community are not very different (p=0.319).
It's not caused by a difference in dispersion. But yeah.
Check out these posts! (probably easier to read then a stats paper, and with more emoji )
And of course, let us know if you have any questions!
First of all thank you for you elaborate explanation! This really helped me in understanding my research! However, I have one more question.
In my research I analyse the microbiome of "unchallenged" and "challenged" chickens. In my experimental set-up these chickens were divided over 2 rooms, so half of the challenged birds were groups together with half of the unchallenged birds in one room, the other half of both groups were put in the other room.
I am still having trouble in understanding the difference between PERMANOVA (the beta-group-significance command in qiime2) and ADONIS (qiime diversity adonis) and which one I should use in my analysis.
So with the permanova I test whether ONLY the effect of treatment or ONLY the effect of room is significant. and with adonis I test the effect of treatment and room together.
But since my treatment groups were equally divided between the rooms, does it make sense to perform this adonis?
I hope you can give me some clarity on this topic once again!
Without getting into the underlying math, these are very similar tests and let you measure the same thing: how different are groups from each other.
Yes, you can do this with both the basic permanova and the distance permanova (adonis) test.
Yep, and sequentially in the order of your choosing!
This is the true power of the adonis test, and what differentiates it from a basic permanova. When the distance matrix is partitioned, to compare vs or Room1 vs Room2, you can first 'pre-partition' it to remove variance (statistically control) factors you are not interested in!
So you can partition out the chicken variance so you can measure the effect of the room adonis(~chicken + room)
or partition out the room variance so you can measure the effect of challenging your chickens! adonis(~room + chicken)
Even better, you control for one and test the other
It can even do interaction effects! adonis(~room * chicken)
To make use of the full power of this test, you need enough blocking in your study design so the variables in question do not confound each other. Based on what you have told me so far, I think the ADONIS test is a great fit for your question, as it should help you remove variance due to room and let you focus on variance due to the chicken challenge!
By putting genotype first in the formula (genotype + donor, not donor+genotype) I adjust for donor to see if genotype will still have an effect.
So if I want to analyse the effect of treatment on my chickens I should write the formula line like this: --p-formula treatment + room.
And if I want to analyse the effect of room I do it the other way around.