FastStructure and Harvester results interpretation

[Guo Yuanwen]

Hello,

 

I have a basic question about the Structure Harvester and fastStructure results. My dataset includes ~10,000 loci and 797 individuals, and our assumption for this population structure is not exceeding 6 clusters based on some prior knowledge. At first, I analyzed the whole dataset with fastStructure, and it showed the optimal K should between 1 to 5 (Model complexity that maximizes marginal likelihood = 5; Model components used to explain structure in data = 1), and the marginal likelihood for each K is showing as below:

K=1, Marginal likelihood =-1.17374

K=2, Marginal likelihood =-1.15117

K=3, Marginal likelihood =-1.14755

K=4, Marginal likelihood =-1.14636

K=5, Marginal likelihood =-1.14622

K=6, Marginal likelihood =-1.14682

K=7, Marginal likelihood =-1.14702

 

Then, I randomly selected 500 loci using Strauto as suggested by Vikram (Thanks!), and run the new dataset using Structure. From my understanding, because the uppermost point for Delta K is K=2, probably it is the optimal value I am looking for? It seems like commonly other Delta K figures will have a peak, but since mine starts from 2, so I am not sure whether or not my conclusion is reasonable.

 

 

Thank you!

 

Best,

Yuanwen

[Vikram Chhatre]

Given that you have used different size data sets with FastStructure and Structure, the results are not really comparable.  Before you run any further subsets with Structure, you may want to put together bar plots for K2 through K5 for both analyzes, which will shed more light on this.

V

--
You received this message because you are subscribed to the Google Groups "structure-software" group.
To unsubscribe from this group and stop receiving emails from it, send an email to structure-software+unsub...@googlegroups.com.
To post to this group, send email to structure-software@googlegroups.com.
Visit this group at https://groups.google.com/group/structure-software.
For more options, visit https://groups.google.com/d/optout.

[Guo Yuanwen]

Thank you very much for your reply, Vikram. I will take a look at those bar plots.

Best,

Yuanwen