Post date: Sep 25, 2013 7:31:12 PM
Four of the bgsr MCMC analyses finished (the others are still running). I quickly looked at the results from the full model with the prior sd for model coefficients of 0.05 (I also looked a little at the results for sd = 0.01). There is more to do and I want to wait until more chains are done, but here is a summary of what I learned based on an e-mail I sent to Patrik:
"I ran a few quick summaries of the full model with the more liberal prior on selection (basically we have one set of runs with the prior expected average total selection on a locus of 0.01, and another of 0.05... both are total selection direct + indirect, the results that I am talking about are for 0.05). We are also now using the t-prior for the selection coefficients, geographic distance instead of elevation, and a beta-binomial model (rather than normal, I think this could make a pretty big difference given the small number of individuals and possibility that many initial or final allele frequencies are close to 0 or 1) for allele frequency change. The results are rather cool. There is still greater evidence of distance associated selection than host-plant associated selection (in particular the largest betas are larger for distance than host-plant). However, the effects of both distance and host-plant are non-significant (i.e., 95% CI's on the betas always overlap 0). But, we now do find significant evidence of constant (across treatments) selection on three SNPs. The truly cool thing is that these three SNPs are tightly linked (all on the same scaffolds and within 50 bp of each other, which the program does not know), Here is the scaffold, position, and effect of each of these SNPs
4313, 10750, beta0 -0.937125
4313, 10784, beta0 -0.947682
4313, 10799, beta0 -0.942799
These are definitely worth annotating. I am also including the loci with the biggest distance and host-plant effects (probably worth looking at too)
biggest distance effect:
1683, 119435, beta = -0.216152
biggest plant effect:
8157, 2183, beta -0.09554
I haven't looked quite as much at the results with the more conservative prior (expected average total selection of 0.01), but this seems to be such a strong prior for no/weak selection that it basically pulls all of the betas (including intercepts) back to 0. With that said the inferred allele frequency change under both priors is very similar (over correlation ~ 0.95), and for the three significant SNPs it is even higher (r > 0.999). Thus, under either prior for those three SNPs allele frequency changes range from -0.26 to -0.49 (across the 10 populations and three SNPs), but with the more restrictive prior this still isn't sufficient evidence to rule out drift.
I probably will not do much more until all of the chains are done (I want to make sure we get consistent results from independent chains), but the fact that the three significant SNPs show up right next to each other suggest that mcmc mixing and convergence is in pretty good shape.
One final thing, it is clear that drift was still a big player during the experiment. We have high confidence in the effective population size for the individuals that contributed to the offspring generation and these range from 3.5 to 11.5. Thus, the bottleneck was big.
The biggest change locus (biggest in each population, and biggest mean change of -0.7) does not come out with significant s or regression coefficients (the CI's are simply too big.. there is also more uncertainty in the actual allele frequency change for this locus); here is the info
scaffold 1782, position 2962"
The relevant outfiles and some R commands are in the projects/timema_reproduction_experiment/bgsr_results/ directory.