how to set the parameter of strata in adonis2

[Ivy-ops]

Hi developer,
I have been searching the whole website regarding how to set the "strata" in adonis2(), but it seems there is no confirmed answer regarding this. Could you guide me on the below 2 designs?
(1) paired design: for example 30 samples from 15 subjects, each subject has pre and post-treatment(described by "treatment" variable). Is that accurate to set: adonis2(distance ~ treatment, strata = subject)?
(2) repeated design: for example 30 samples from 10 subjects, each subject has 3 samples, and 5 subjects(5x3 = 15 samples) are treatment, 5 subjects(5x3 = 15 samples) are control. I want to compare treatment vs. control. Under this condition, how to set the formula? (assume "group" variable describes treatment vs. control, "subject" describes subject)?
Much appreciated!

[gavinsimpson]

For 1, you can do that {with strata], but I think it is best to learn how to define restricted permutations using the permute package. For a design-based fit you would create:

library("permute")
h1 <- with(your_data, how(blocks = subject))

and then pass h1 to the permutations argument of adonis2() (if you must use that)

For 2, are your samples equally-spaced in time?

[Ivy-ops]

@gavinsimpson Thank you!
For 1, do you mean:

h1 <- with(data, how(blocks = subject)); 
adonis2(distance ~ treatment, data = data, permutations = h1, strata =data$subject)

Is this correct? Here treatment is time (pre-treatment and post-treatment).

For 2, it is a repeated design, with no time variable. 30 samples from 10 subjects(each subject has 3 samples), 5 subjects belong to control (which means 5*3 =15 samples belong to control), and the rest 15 samples from the other 5 subjects belong to treatment. How to do that in adonis2()?

Looking forward to your thoughts!

[gavinsimpson]

No, you don't use strata at all - you specified the blocking structure needed in the call to how()

[Ivy-ops]

Thank you @gavinsimpson .
So for 1:
h1 <- with(data, how(blocks = subject));
adonis2(distance ~ treatment, data = data, permutations = h1, strata =NULL)
For 2: Do you have any thoughts on the repeated design? Much appreciated!

[gavinsimpson]

You don't need to specify anything for strata — just don't even mention it:

h1 <- with(your_data, how(nperm = 999, blocks = subject))
adonis2(distance ~ treatment, data = your_data, permutations = h1)

However I would suggest you do this via rda() because there is no easy way to exclude the between subject variation:

h1 <- with(your_data, how(nperm = 999, blocks = subject))
ord <- rda(response_df ~ treatment + Condition(subject), data = your_data)
anova(ord, permutations = h1)

would stop any between subject variation getting counted in the treatment effect.

As for your second design, you need to look at where the thing you want to test varies. Clearly it doesn't vary at the within-subject level: each subject only ever receives one of the treatment levels. So those samples shouldn't ever be permuted. Where the treatment varies is at the between-subject level. As such, under the null hypothesis of no effect of treatment and conditional upon us not permuting the samples within a subject, the observations at the between-subject are exchangeable. This means you want to permute the subjects, which is achieved through the use of the plots argument to how():

h2 <- with(your_data, how(nperm = 999,
  plots = Plots(strata = subject, type = "free"),
  within = Within(type = "none")))

then you fit your model as

adonis2(distance ~ treatment, your_data = data, permutations = h2)

# or
ord <- rda(response_df ~ treatment + Condition(subject), data = your_data)
anova(ord, permutations = h2)

where again, conditioning on subject will remove the between-subject effect from the estimate of the treatment effect. In both cases I don't think it (the conditioning) is needed for the purposes of the permutation test, but if you ignore it, you might inflate the inertia ("variance", sums of squares, etc) attributable to the treatment effect when you report the inertia explained value.

[jarioksa]

Please not that adonis2 defaults to Bray-Curtis distance but rda is based on Euclidean distance. For equivalent models you have two choices:

1. adonis2(..., method="euclidean") equals anova(rda(...)) for Euclidean distance in both.
2. adonis2(...) equals anova(dbrda(..., method="bray")) for Bray-Curtis in both.

In general adonis2() and anova(dbrda()) give identical results, although adonis2 may be marginally faster. With same permutations (or how object) they should give identical permutation statistics. adonis2 folds together fitting and testing, and with ordination (dbrda) permutation arguments should be given to anova so that these should be equivalent:

adonis2(formula, data, permutations, method, sqrt.dist, add, by, parallel)
## or alternatively with dbrda
ORD <- dbrda(formula, data, method, sqrt.dist, add) 
anova(ORD, permutations, by, parallel)

[Ivy-ops]

Thank you both @gavinsimpson @jarioksa! Your answers have been the most detailed I've received!

[jarioksa]

Please note that strata is a legacy argument in vegan. It was copied from the boot package over 20 years ago, and provides only the simplest blocking of permutations. You really need to use permute package for anything more sophisticated. I was about to remove the strata argument completely in favour of permute package, but I was convinced that it is better to maintain backward-compatibility and strata is sufficient for many a need. Actually, it was first dropped from adonis2, but people still used it although it was not there and was non-functional – the strata were just silently ignored. I decided it is best to put it back so that it works if people use it. The reason that you didn't find information about implementing your design with strata was that you cannot: you must see how to do these things with permute that is internally used in vegan.

[Ivy-ops]

Thank you! @jarioksa

[D-rgz]

Hi developer I want to do permutational analysis of variance to calculate variation between nematode haplotypes of the same species in samples , i have unbalanced haplotypes in samples and in regions (108 samples, 3 regions, sharing 27 haplotypes). Each sample has a range of between 1 to 9 haplotypes, I have built a pairwise distance matrix (hamming) for all haplotypes. I want to build an R script for performing analysis of molecular variance within samples and between samples in a regions and between regions (I have 3 regions). I guess my challenge is to build the correct dataframe shape ..i tried the nested design making Region and SampleID for each Haplotype so the rows match the dist matrix but it wouldnt run, and also i realize that i have to build it using the permute function over samples??? I also want to do an ordination for the different sample to show how they vary between the 3 regions
Your help will be much appreciated.

[gavinsimpson]

Please don't jump onto an existing discussion; start a new discussion question instead. If you do, please note that you will need to explain what you want to do and your setting with more detail otherwise we can't possibly help.

[D-rgz]

Thanks I will Deo

…

________________________________ From: Gavin Simpson ***@***.***> Sent: Friday, August 2, 2024 11:32 AM To: vegandevs/vegan ***@***.***> Cc: Deodatus Ruganuza ***@***.***>; Comment ***@***.***> Subject: Re: [vegandevs/vegan] how to set the parameter of strata in adonis2 (Discussion #600) [△EXTERNAL] Please don't jump onto an existing discussion; start a new discussion question instead. If you do, please note that you will need to explain what you want to do and your setting with more detail otherwise we can't possibly help. — Reply to this email directly, view it on GitHub<#600 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AP56GHK7EDVLDZHR4XKLGCLZPM73HAVCNFSM6AAAAABL4CZDHWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMRSGEYTQNI>. You are receiving this because you commented.Message ID: ***@***.***>