Number of Iterations

[mm.j...@gmail.com]

Hello, 

I am running dataset consisting of 340 samples and 28000 SNPs, at 10000 burnin and 10000MCMC reps, K=2 to K=7. Initially I chose 20 iterations as this was the recommended minimum in the Structure tutorial, but it took a very a long time to run and I revised the iterations to 5. Does it really matter the number of iterations I choose in the long run? Kindly advise. 

Thanks.

[Janice Boyd]

From personal experience I can say that will depend on how stable the results are for each run.  Sometimes you get pretty much the same results for each run.  But if your data is violating the many assumptions made in deriving STRUCTURE (eg, uneven group sizes or lots of relatedness) you can get quite different results and many iterations are needed.  Just look at the results for each of your 5 runs and see if they are pretty much the same to verify if 5 is enough.

Janice Boyd

Texas A&M

--
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-softw...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/structure-software/d0747909-0564-453d-bd72-f79555b6a5b9n%40googlegroups.com.

[Ana Nikolic]

Hello,

My question is what if results from,for example, 20 iterations for one K are highly different from each other? Which one to chose? Or how to change number of iterations? Or what else to do?

Thanks!

Ana

To view this discussion on the web visit https://groups.google.com/d/msgid/structure-software/CAAfJRa7vcNGr_Pu0K_4fVOadADc%2BySJTCdrivOvvBSAY%3Dn2a_Q%40mail.gmail.com.

[Vikram Chhatre]

You need to perform cluster matching on the 20 sets using CLUMPP.

https://rosenberglab.stanford.edu/clumpp.html

However, if your iterations are truly divergent from each other, you may have multimodality in your data.  

To view this discussion on the web visit https://groups.google.com/d/msgid/structure-software/CA%2B5wYtW7uw%3Dn9yp%2B85y2Vwshj%2BcoDrubV5jx875WQHP%3DqHpoJA%40mail.gmail.com.

[Heidy M. Villalobos Barrantes]

Also in R with the package pophelper (Roy Francis) you can process all the Results of Structure.

Best regads

Heidy

To view this discussion on the web visit https://groups.google.com/d/msgid/structure-software/CAJZnH0keAEDW%2BexG_X1%2B2s%3DNBzNQwwX2e0Jo1d2PVOSPZPGWbw%40mail.gmail.com.

--

Heidy M Villalobos B., M.Sc.
Doctora (c) en Sistemática y Biodiversidad 

Laboratorio BIOMAS

Facultad de Ciencias Naturales y Oceanográficas

Universidad de Concepción

Concepción, Chile

Tel cel Chile: +56 9 6725 8005

Número internacional: +506 4001 2483

hema...@gmail.com

heidy.villal...@ucr.ac.cr
skype: hemavb0509

"Existen dos formas de ver la vida. Una es pensar que no existen los milagros y la otra es pensar que todo es un milagro" A.E.

[banta....@gmail.com]

I'd be remiss not to mention my how-to video for using pophelper. I wrote it for Linux, but it works just as well for any R users.

https://www.youtube.com/watch?v=HJgJ4fVJq2s&t=50s

Josh Banta

[B. Heinze (gmail)]

Hi Josh:

I watched the tutorial, great work and very nicely explained!

One remark for your sample data set though, in a situation where there seems to be exceptionally high variability of ln(K) at one set of Ks (graph panel A), I look at the individual runs at that K. In your case, that's K=5: very low variability at K=3, 4, 6, but a huge error bar at K=5. Often it happens in such cases that one of the runs was "going wild", i.e. shows a very large deviation from the others (in ln(K) values). It may be due to the run not "converging on alpha" (correct? I am not the most knowledgeable person in this regard). What I usually do (if I have enough replicates of runs at each K), I remove this deviating run from the data set, and re-do the post-STRUCTURE analysis. I have more confidence in the data if the pattern of the ln(K) graph is more regular - with a clearer trend in ln(K) values, and increasing standard deviation from some point onwards.

Maybe the same can be achieved when increasing the burn-in run period (as you mention in the tutorial), but the way outlined above is a much quicker fix.

best regards, Berthold

Berthold Heinze

BFW - Austrian Federal Research Centre for Forests

Vienna, Austria (EU)

--

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-softw...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/structure-software/e1a4a3e2-150a-4b4f-aed1-a575500ce756n%40googlegroups.com.

[Eva T.]

Hi all,

this thread has already been very helpful, thank you! I have a small sample of 94 individuals and approx. 3000 loci. I was advised to run 5 iterations of Structure for 50.000 generations with 5.000 burnin for K=1-10, check for best K and then run Structure for that K only for 500.000 generations and use that as the final result.

The results of the first part, analysed with Harvester, are attached. It seems K=2 is the best supported result judging from delta K values, but Mean LnP(K) is the least negative at K=3. Could such discrepancy be due to the small number of iterations?

Thank you very much in advance,

Eva