Reliability/Power of STRUCTURE results given lack of power in BAPS analysis

[Brock Geary]

Hello all, 

I am running a microsatellite data set through Structure and got results that suggest reasonably high gene flow/weak structure, with an optimal K of 2 based on Structure Harvester results and visual inspection of individual membership assignment plots. I've also run this data through the software BAPS and, while the clustering of populations also suggests K = 2 (with very high probability, >0.99), the somewhat low number of loci used (7) does not give BAPS enough power to detect any admixture in individuals (all are 100% assigned to one cluster or another). I spoke with an author of BAPS about this, and we agree that the # of loci combined with relatively low pairwise differentiation between pops is the cause of this problem. So as it stands, our results are at odds: highly admixed individuals in Structure vs. 'no' admixture in BAPS.

My question is, should the low power of BAPS to detect admixture in individuals given our data impact our confidence in the results we see in Structure? Our runs converged just fine, and we would not have suspected anything odd had we not run this through multiple programs. If anyone has experience running data sets through multiple pop gen programs, please let me know your thoughts on comparing results when things are hinkier in one program than another. I can provide plots etc. if they will help.

Thanks!
Brock

[Alex Putman]

Hi Brock,

Like many methods, Structure also has problems inferring admixture when the information content of the dataset is low.  I've listed a few papers that look at this issue or for the related issue of hybrid identification.  A few things:

1. Do exploratory methods such as a distance tree or PCA support K = 2? 

2. Have you looked at the 90% probability intervals for the membership assignments (Q)?  These should be in the output files for each run.  The width of these intervals could help determine how confident you should be.

3. In Structure, have you tried the correlated allele frequency or the prior population (LOCPRIOR) models?  The prior population assignment could be made using sampling location, some phenotype, or assignment as inferred with a generic clustering method (distance tree or multivariate method such as discriminant analysis of principal components)

4. Does the sample include any non-admixed individuals to your knowledge?  The problem of weak structure with too few loci is exacerbated when there aren't enough non-admixed individuals in the sample. See Pritchard et al. (2000) pg. 950 last paragraph

5. Try FLOCK. I don't have experience with it and it is new, but the paper updating it (Duchesne and Turgeon, 2012) claims that it outperforms STRUCTURE when differentiation is low.

I'm a fan of analysis using multiple methods.  In general, disagreement among different methods suggests some combination of: i) an assumption of one method has been violated, which may or may not provide biologically relevant information; or ii) the method(s) are being employed at or beyond the edge of their theoretical envelopes, and your confidence in the results may need to be reduced.  Good luck and let us know how it goes.

Best wishes,

Alex

Anderson, E. C., Thompson, E. A., 2002. Genetics. 160, 1217-1229.
Anderson, E. C., Dunham, K. K., 2008. Molecular Ecology Resources. 8, 1219-1229. doi:10.1111/j.1755-0998.2008.02355.x.
Duchesne, P., Turgeon, J., 2012. The Journal of Heredity. 103, 734-743. doi:10.1093/jhered/ess038.
Latch, E. K., Dharmarajan, G., Glaubitz, J. C., Rhodes, O. E., 2006. Conservation Genetics. 7, 295-302. doi:10.1007/s10592-005-9098-1.
Vähä, J.-P., Primmer, C. R., 2006. Molecular Ecology. 15, 63-72. doi:10.1111/j.1365-294X.2005.02773.x.
Vaughan, L. K. et al. 2009. Computational Statistics and Data Analysis. 53, 1755-1766. doi:10.1016/j.csda.2008.02.032.