[adegenet-forum] Fwd: Significance of allelic contribution to discriminant functions

[Caitlin Collins]

Hello again Andrea,

Glad you found what you were looking for!

Incidentally, and in case anyone else on the forum is looking to visualise
the variable contributions to discriminant axes > 1, here is some code to
do so for a toy example. (The last chunk will be the relevant bit for
creating loading plots):

# make a simulated dataset with 5 "groups"
simpop <- glSim(200, 1000, 40, k=5, sort.pop=TRUE)
snps <- as.matrix(simpop)
phen <- simpop at other$ancestral.pops

# for fun/ as a check, quickly visualise the clusters
dapc1 <- dapc(snps, phen, n.pca=50, n.da=4)
scatter(dapc1)

# create an object called foo that contains the results of running snpzip
on your dataset
foo <- snpzip(snps, phen, xval.plot=TRUE, loading.plot=TRUE,
method="centroid")

# isolate the DAPC component of the snpzip results, calling it "dapc1"
dapc1 <- foo$DAPC
# specify that you want to run the following lines for all DA (ie. from
DA=1 to DA=(k-1), where K is the number of groups in your dataset)
DA <- c(1:dapc1$n.da)
par(ask=TRUE)
# generate separate loading plots for each DA
for(i in DA){
  title <- paste("Loading Plot for DA", i, sep=" ")
  maximus <- foo$FS[[i]][[2]]
  cutoff <-
abs(dapc1$var.contr[maximus,i][(which.min(dapc1$var.contr[maximus,i]))])-0.000001
  loadingplot(dapc1$var.contr[, i], threshold=cutoff, main=title)
}


Hope that helps!
And thanks for your input: I'll try and implement the above code within
snpzip to generate loadinplots for all DA automatically in the next release
of adegenet.

Cheers,
Caitlin.

On Mon, Oct 6, 2014 at 6:09 PM, Andrea Garavito <neagef at gmail.com> wrote:

> Hello again!
>
> I took a closer look into the object created by the snpzip tool, and I
> found the contributions for all the different axes.
> I didn't noticed them before as I was looking only at the plot obtained.
>
> Thanks anyway!
> Andrea
>
>
> 2014-10-06 12:30 GMT-03:00 Andrea Garavito <neagef at gmail.com>:
>
> Hello Caitlin,
>> I was taking a look to the adegenet forum and I found this previous
>> answer about a statistical threshold for marker contributions.
>>
>> Originally I was planing to retain for each one of my discriminant
>> functions, around the 0.3% of markers with the highest contributions by
>> establishing a threshold of  3-sigma. I'm not sure if these data are
>> distributed normally, but as I have almost 5000 markers I was assuming so.
>> Then I saw your post about the snpzip analysis and decided to give it a try.
>> I tested the function with all the methods available, and I think I'll
>> use the "median" method as with the others I'm getting to many markers
>> retained (and only one with the "single" method).
>> I see that the snpzip test make the analysis for the first discriminant
>> function, but is there a way to make it also for the other discriminant
>> functions found with DAPC?
>>
>> Thanks for your answer
>> Andrea
>>
>>
>> 2014-08-26 12:58 GMT-03:00 Caitlin Collins <caitiecollins at gmail.com>:
>>
>>> Yeah, it's new!
>>>
>>> I might as well note, in case you decide only to try a subset of the
>>> methods available:
>>> - Ward's method is most likely to select a very large number of
>>> variables to get the most complete picture
>>> - Single linkage hierarchical clustering will probably select the fewest
>>> - Centroid clustering will probably select a useful middle-ground.
>>>
>>> You can always check to see what proportion of the variance is contained
>>> in the subset of variables retained, or you could even try running a DAPC/
>>> PCA with just those variables to compare the discriminatory power of the
>>> entire set with that of the subset selected.
>>>
>>> Good luck.
>>>
>>> Cheers,
>>> Caitlin.
>>>
>>>
>>> On Tue, Aug 26, 2014 at 4:31 PM, Charlie Waters <cwaters8 at uw.edu> wrote:
>>>
>>>> Thanks Caitlin! I've never come across the snpzip function so I'll give
>>>> those clustering methods a try.
>>>>
>>>> Thanks,
>>>> Charlie
>>>>
>>>>
>>>> On Tue, Aug 26, 2014 at 3:49 AM, Caitlin Collins <
>>>> caitiecollins at gmail.com> wrote:
>>>>
>>>>> Hi Charlie,
>>>>>
>>>>> Good question. Technically, there is no one "correct" statistical
>>>>> solution to your problem. But, there *are *a number of ways of
>>>>> approaching the problem with more statistical rigour than simply using an
>>>>> arbitrary threshold as you have done.
>>>>>
>>>>> Have you taken a look at the snpzip function in the adegenet packge?
>>>>> If not, just type "?snpzip" into R with the adegenet package loaded. With
>>>>> this function, you can apply one of seven different hierarchical clustering
>>>>> formulas to the allelic contributions generated by dapc. Essentially, each
>>>>> hierarchical clustering method uses a unique approach to determine where
>>>>> the threshold should be drawn. I should note, however, that this
>>>>> descriptive approach will not have an associated p-value. You may want to
>>>>> try out a few different methods before deciding which variables you want
>>>>> to consider "most significant".
>>>>>
>>>>> I hope that helps!
>>>>>
>>>>> Best,
>>>>> Caitlin
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Charlie Waters
>>>> Box 355020
>>>> School of Aquatic and Fishery Sciences
>>>> University of Washington
>>>> Seattle, WA 98105
>>>>
>>>>
>>>
>>> _______________________________________________
>>> adegenet-forum mailing list
>>> adegenet-forum at lists.r-forge.r-project.org
>>>
>>> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/adegenet-forum
>>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/adegenet-forum/attachments/20141016/e1c7fc52/attachment.html>