If you haven't, you should go and inspect your likelihoods to see if they've stabilized for each run (and therefore if your burnin is sufficient). Beyond that, it looks like all that I can see from your Structure Harvester plot points to K=7. Whether or not this makes sense I can't say.

The temptation is to use structure as a plug-and-play black box, but you should step back and critically examine what you're asking it to do. Here are some questions you need ask yourself in interpreting whether your results are believable:

Thanks, for you answer.  I use 100000 Burnings and 100000 MCMCs for each k, In some papers use 20 reps for each k.

What do you mean with runs converged?

I use the tool proyect, and set the 25 runs for each k .. 1 to 10 all in the same running.

sorry for my english...

If you haven't, you should go and inspect your likelihoods to see if they've stabilized for each run (and therefore if your burnin is sufficient). Beyond that, it looks like all that I can see from your Structure Harvester plot points to K=7. Whether or not this makes sense I can't say.

The temptation is to use structure as a plug-and-play black box, but you should step back and critically examine what you're asking it to do. Here are some questions you need ask yourself in interpreting whether your results are believable:

Sometimes the Evanno G. Method underestimate or overestimate the "k", check that out: 

https://groups.google.com/forum/#!topic/structure-software/iwYLK_xjX-k

In Evanno et al (2005) original paper, they claim: "We found that in most cases the estimated 'log probability of data' does not provide a correct estimation of the number of clusters" . Check: http://www.ncbi.nlm.nih.gov/pubmed/15969739

===========

I think you should try the analysis with 5 to 7 groups.

The priori=5 geographical origins

The final k=7

I agree with all previous considerations: however, given the correctness of analysis, not necessarily k coincides with geographical samples. The SEs of lnP(d) values may give you an indication of run convergence.

Ok  Arthur,I understand your point, I need read more... I'm New  in this topic. Thanks for your time.

STRUCTURE is a bit too sophisticated method. It requires good genetic data, understanding of underlying models etc. (comments above).  From some point of view neither result is true because an ideal model is considered. Why not to try more simple and understandable methods like PCA or MDS to analyze population structuring first?  GenALEx http://biology-assets.anu.edu.au/GenAlEx/Welcome.html is a good soft to perform PCA based on genotypic (individual) data. Probably I am wrong but if there is true structure in your data, you will get congruent results by STRUCTURE & by ordination methods.

Another point is that species are organized like a Russian (nesting) doll: genetic diversity is hierarchically structured. K=7 is a low(er) level of hierarchy while you are interested in some higher level… That’s why some people present results of STRUCTURE for different k’s, aiming to illustrate the hierarchy and to check particular hypothesis…

I fully agree with the approach suggested by Petr, i.e. considering first more simple approaches, and then considering the different K value as different levels of structure, without looking to a "true" number of clusters.

Dear Emily

In spite of agreeing with much of the above answers, I would like to suggest you provide as much as possible background information when asking some question.

How many markers do you have? What type of markers are they? This info can help to decide (based on somebody experience) whether your burning and mcmc chains are long enough. Indeed, the true answer will be obtained by looking at the chains convergence, as already told). To do it, you can use some statistical program (such as the R library coda, or Tracer), although usually looking at the Structure graphs is enough to get an idea.

How do your individual's ancestry plots look like? Is the variance of Ln Likelihood due to different groups for the same K, or to different admixture levels? If you find that different populations group together sometimes but not other (for the same K) then you probably need to run longer chains. Alternatively, you data could be trapped in local minima with similar likelihoods, making difficult to decide which genetic structure represents your data best.

These are just examples to illustrate how background info helps to answer questions. Cheers

Pablo