Permutation of whole plots

[qutkat]

Hello there,

I have longitudinal data with 3 samples per individual ID. I would like to treat each ID as a plot and then permute the whole plots so I am keeping all observations from each individual together when I permute.
I tried using the following code per the Details section in the documentation but I am getting an error message.

perm = how(within = Within(type = "none"), plots = Plots(type=variables$PID), nperm = 9999)

The error I am getting is Error in match.arg(type) : 'arg' must be NULL or a character vector . I believe this is because type is expecting type = c("none","free","series","grid"), however int he details section it says to use type = foo in order to permute whole blocks. Not sure what I am doing wrong?

I am fairly new to R - could someone please assist? It would be greatly appreciated!

Many thanks
Katherine

[gavinsimpson]

You want:

how(within = Within(type = "none"),
    plots = Plots(strata = factor(variables$PID), type = "free"),
    nperm = 9999)

changing "free" to one of the other options if you need a specific type of restricted permutation design (regular time series/transect ("series") or regular spatial grid ("grid")).

[qutkat]

Hi Gavin, Thank you so much for getting back to me. I have tried the code you suggested and the adonis results look reasonable, but I am still a little concerned as when I ran shuffle to see what was happening it looks like the individual PIDs are all being randomly permuted within each week, whereas I need all PIDs from the same individual to be held together in the permutation. This is the code I am using - perhaps I am doing it incorrectly?
Prior to this I set variables$PID as a factor.

perm = how(within = Within(type = "none"), plots = Plots(strata=factor(variables$PID), type = "free"), nperm = 9999)

adonis2(arcsin ~ variables$d_rating + variables$Weight_kg + variables$BMI + variables$Age , permutations = perm, parallel = 7, na.rm=TRUE, by="margin")

variables[shuffle(nrow(variables), control = perm),c("PID", "week")]

[gavinsimpson]

I can't reproduce this. In fact I can demonstrate that things work correctly in a reproducible setting:

# setup
library("permute")

set.seed(1)
f1 <- gl(5, 10)            # in order
f2 <- f[sample(length(f))] # in random order

The easy setting, where everything is in order:

set.seed(21)
h1 <- how(within = Within(type = "none"), plots = Plots(strata = f1, type = "free"))
s1 <- shuffle(f, control = h1)
s1
f[s1]

r$> s1                                                                          
 [1]  1  2  3  4  5  6  7  8  9 10 21 22 23 24 25 26 27 28 29 30 41 42 43 44 45
[26] 46 47 48 49 50 11 12 13 14 15 16 17 18 19 20 31 32 33 34 35 36 37 38 39 40
r$> f[s1]                                                                       
 [1] 1 1 1 1 1 1 1 1 1 1 3 3 3 3 3 3 3 3 3 3 5 5 5 5 5 5 5 5 5 5 2 2 2 2 2 2 2 2
[39] 2 2 4 4 4 4 4 4 4 4 4 4
Levels: 1 2 3 4 5

Here the permutation is:

• 1 -> 1
• 2 -> 3
• 3 -> 5
• 4 -> 2
• 5 -> 4

And this continues to work with the randomly ordered f2

set.seed(21)
h2 <- how(within = Within(type = "none"), plots = Plots(strata = f2, type = "free"))
s2 <- shuffle(f2, control = h2)
s2
f2
f2[s2]

r$> s2                                                                          
 [1]  1  7  3  8  6  2  5 10 15 11  4 12 13 14 17 19 18  9 26 22 16 27 23 24 20
[26] 21 25 38 37 39 42 40 33 44 47 49 28 29 31 34 43 30 35 32 45 46 41 36 50 48
r$> f2                                                                          
 [1] 1 4 1 4 3 5 2 2 4 3 5 1 1 1 2 4 3 5 2 4 5 2 1 1 5 3 3 4 3 3 4 3 1 4 4 4 5 2
[39] 5 5 3 2 5 2 1 1 2 5 2 3
Levels: 1 2 3 4 5
r$> f2[s2]                                                                      
 [1] 1 2 1 2 5 4 3 3 2 5 4 1 1 1 3 2 5 4 3 2 4 3 1 1 4 5 5 2 5 5 2 5 1 2 2 2 4 3
[39] 4 4 5 3 4 3 1 1 3 4 3 5
Levels: 1 2 3 4 5

although it is much harder to see, but if you compare the last two outputs, all 4s became 2s etc.

As for the ordering within the PID issue, I can't reproduce that either. He's another reproducible example building on the elements above. Here I define a variable week that is in order 1, 2, ..., 10 alongside the correctly ordered f1 from earlier. Then I define a second data frame, df2, using the first, but scrambled in exactly the same way as f1 was scrambled to create f2 earlier.

df1 <- data.frame(f = f1, week = rep(1:10, times = 5))
# scramble df1 the same way we scrambled f1 to make f2
set.seed(1)
df2 <- df1[sample(length(f1)), ]
split(df2$week, df2$f)
split(df2$week[s2], df2$f)

The last two lines there are splitting the scrambled week according to f, and the same but for the shuffled (permuted) ordering of the rows in df2, the scrambled data frame

r$> split(df2$week, df2$f)                                                      
$`1`
 [1]  4  1 10  7  9  3  6  8  2  5
$`2`
 [1]  4  8  5  2 10  9  3  6  1  7
$`3`
 [1]  3  1  5  6  7  9  2  4  8 10
$`4`
 [1]  9  4  3 10  6  1  2  5  7  8
$`5`
 [1]  3  1  7  8  9  5  2  6  4 10

r$> split(df2$week, df2$f[s2])                                                  
$`1`
 [1]  4  1 10  7  9  3  6  8  2  5
$`2`
 [1]  9  4  3 10  6  1  2  5  7  8
$`3`
 [1]  4  8  5  2 10  9  3  6  1  7
$`4`
 [1]  3  1  7  8  9  5  2  6  4 10
$`5`
 [1]  3  1  5  6  7  9  2  4  8 10

If you compare $2 in the first output and $3 in the second, you'll notice that the order of the weeks is the same. This is because 2 got swapped with 3 in this particular permutation. Compare $5 in the first output with $4, again as per expectation as in this permutation 5 is swapped with 4. The $1 components are the same in both outputs because 1 was swapped with 1 in this permutation.

It's impossible to say what is wrong with your setting as your example is not reproducible by me. Check that you redefine your how() object after coercing it to be a factor; how() stores the factor defining the strata of the plots within the control object used to define permutations. So if you change the data or factor after defining the control object (h in your example) then things will get all out of whack. For example, I define h1 and h2 here because the factor indicating to which group each observation belongs changes when I scrambled the original factor f1. If you were to reuse h1 you'd get the wrong permutation because the stored strata don't match the data once we scramble f1 to make f2`.

[qutkat]

Thank you for taking the time to provide these examples. I have investigated further and it seems that the results of adonis are dependant on how the PIDs are sorted in the file when they are submitted. If I sort by PID so all the PIDs for an individual are together, then I am getting a significant result, whereas if I leave the file sorted by week so the PIDs are not grouped together, I get a very different result which is insignificant. The results do not vary for either of these settings when I change set.seed. Do you have any idea why this might be happening? I'm wondering if it is something to do with adonis?

[gavinsimpson]

You need to supply a reproducible example.

Also, don't build a model like this:

adonis2(arcsin ~ variables$d_rating + variables$Weight_kg + variables$BMI + variables$Age, ...)

use the data argument:

adonis2(arcsin ~ d_rating + Weight_kg + BMI + Age, data = variables, ...)