MCMCglmm : Difference in additive genetic variance estimated in univariate vs bivariate models
Hi Stephane, If I am worried about prior influence and my model is multi-trait Gaussian I often run it through ASReml as a check. Generally I find parameter expanded priors to be weaker than the inverse-Wishart and usually use them. Unfortunately I haven't seen much work done regarding their properties for covariance matrices. However, having V=diag(n), nu=n, alpha.mu=c(0,0), alpha.V=diag(n)*s where n is the dimension of the matrix and s a scale parameter (assuming the traits are on the same scale) seems to work well under at least some circumstances. Even with low replication the posterior modes are close to their REML estimates, and if one of the variance is close to zero then the posterior for the correlation is close to being uniform on the -1/1 interval, as you would hope - approximate standard errors for this correlation obtained from REML analyses often seem to be anti-conservative. Cheers, Jarrod Quoting Stephane Chantepie <chantepie at mnhn.fr> on Thu, 19 Jul 2012 17:20:24 +0200:
Hi Jarrod and all other, You said "your posterior is just reflecting your prior" : I was thinking about this issue but how can I test it? The V I used is V=(Phenotypic variance/2)= 0.47 so it is bigger than the posterior mode. I have tried to used V=1 but the Va posterior results remains the same (Va=0.10 for the univariate model spz_9). For the bivariate model c(spz_9,spz_5) I have kept the same V=(Phenotypic variance/2)but nu=2. To answer your question "if nu=1 then V*nu/(nu-1) is ..." V*nu/0 so +? or -? ;-) Just to give you an idea of the my posteriors: The Va posterior mode resulting from the spz_9 univariate model hit the 0 (http://ubuntuone.com/2IGcmcdqkjcVQCdZgvMhnP) whereas the Va posterior mode resulting from the (spz_9,spz_5) bivariate model seems to be better shaped(at the top http://ubuntuone.com/3PaBOJ5dF6kDIhV6ahPnlM). To conclude, the estimation of Va with bivariate models is biologically consistent with senescence theories while the Va estimated with univariate model is not. So : Do do you think that I could use bivariate results or it is better to consider that I do not have enought information? Thank a lot for your help all the best stephane
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.