Should replicate analyses of ADMIXTURE all have the same cross-validation values

Question

[devenvyas]

This is a pretty simple question. I have noticed when I run my ADMIXTURE values for K=2 to 20 with ten replicates, my output CV values for each value of K are always identical for all 10 replicates. I've seen literature where there is some variation in replicate CVs. They do have different loglikelihoods, but the CVs fall out the same.

I was wondering, is this normal?

Answer 1

[biomonte]

I have also noticed identical CV-error values for every replicate ADMIXTURE run. This is due to the initial seed number that is always 43. There is an option of generating a pseudo-random seed using the system clock (flag -s). Here is how I ran it:

admixture -s time --cv file.ped K -jN

Answer 2

[Brice Sarver]

I've seen this with my ADMIXTURE runs in the past. For what it's worth, I've noticed more variation in the likelihoods of STRUCTURE runs, meaning that there can be a bit of a difference among sets of replicates when selecting an 'optimal' K (whatever that means). That's not using cross validation, though.