[R-sig-ME] binary trait correlation across environments (experimental trials) using MCMCglmm?
j@h@dfield @ending from ed@@c@uk
Thu Jan 3 10:24:31 CET 2019
The prior is not fixed at one. Also, the priors you are using are quite informative. I would use
prior <- list(R = list(V = 1, fix=1), G = list(G1 = list(V = diag(2), nu = 2, alpha.mu=c(0,0), alpha.V=diag(2)*100)))
Your equation for the genetic correlation is correct.
You shouldn't expect the correlation in family means to equal the model based estimate of the genetic correlation for a number of reasons; most importantly they are on different scales (data versus latent) and the variances of the family means contains residual variation but the covariance doesn't so the correlation in family means is a (downwardly) biased estimator.
On 02/01/2019 22:18, Plough, Louis wrote:
Thanks Jarrod for the prompt response!
I think the residual variance was set at 1. My prior looks like this:
prior <- list(R = list(V = 1, nu = 3), G = list(G1 = list(V = diag(2), nu = 2)))
Not sure why the difference for nu in R vs G (3 vs 2)...might have been a typo. Is there guidance on setting the gamma parameter for this kind of binary trait, cross environment correlation model with the 'threshold' family? In the low iteration (toy) run I did for the R-Sig-ME post, the correlation was a bit lower than I would expect based on simply phenotypic correlation of family means, so I think I might need to tweak nu for both?
Can I also confirm that I am estimating the rG correctly with the following code:
Any reason to use 'TrialA:TrialB.animal' vs 'TrialB:TrialA.animal' in the numerator? Or are they basically equivalent? I wouldn't want add them, would I?
Thanks for your help!!!
On Wed, Jan 2, 2019 at 3:17 PM HADFIELD Jarrod <j.hadfield using ed.ac.uk<mailto:j.hadfield using ed.ac.uk>> wrote:
Your model bi_model_trial is the correct one. However you must fix the residual variance at one in the prior. You have estimated the genetic (co)variance matrix for the two trials, from which you can obtain the genetic correlation. Alternatively, you could fit animal+animal:Trial which assumes the genetic variances in the two trials are the same and the correlation is positive. The correlation in this latter model is obtained as VAR(animal)/(VAR(animal)+VAR(animal:Trial)). Also, there seems to be a problem with your Surv data as it has three levels rather than 2 and so a cutpoint is being estimated.
On 2 Jan 2019, at 16:33, Plough, Louis <lplough using umces.edu<mailto:lplough using umces.edu>> wrote:
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models