[R-sig-ME] G vs R posterior correlation in bivariate MCMCglmm
xav.harrison at gmail.com
Wed Oct 26 14:16:04 CEST 2016
Thanks for the swift and helpful reply. That makes a lot of sense. I should
have said in the original email that I have taken this approach rather than
fitting a Disease ~ Phenotype model because we have predictors that
influence both simultaneously. For example we know both mean disease
intensity and mean phenotype were lower in 2014. I was hoping to use the
bivariate model approach to estimate the posterior correlation between the
two traits once 'controlling' for the predictors and seeing what's left.
Does this sound sensible?
If you wouldn't mind elaborating further, I'm trying to work out what it
means when there is a non-zero correlation at the site level but not at the
units level, which some models have recovered. Is this a problem suggestive
of insufficient data to estimate both matrices accurately, or something
that could be biologically plausible? I'm afraid of thinking I've found the
golden egg but in fact have built myself a nice random number generator by
asking too much of the models.
On 26 October 2016 at 10:34, Jarrod Hadfield <j.hadfield at ed.ac.uk> wrote:
> If the relationship between the two is causal - lets say phenotype
> affects disease - then the regression at the two levels will be identical:
> COV(Disease, phenotype)/VAR(phenotype) is the same at the site level and
> the units level. However, they can easily differ. Imagine that the amount
> of resource varies between sites, but within sites all individuals have
> access to the same amount of resource. If there is a trade-off then, for a
> given amount of resource, there will be a positive relationship (if high
> values of the phenotype are 'good') between the two variables observed at
> the units level. However, imagine that as the amount of resource increases
> individuals can increase their phenotype but also reduce the amount of
> disease. As a consequence the between site correlation may well be
> negative. So, correlations at both levels tell you something interesting.
> However, it should be noted that if you can assume causality you are better
> just fitting a univariate model with phenotype in as a predictor: you get
> an increase in precision for your assumption.
> On 26/10/16 08:46, Xav Harrison wrote:
>> Hi Folks
>> A potentially rookie question here, but here goes.
>> I'm using MCMCglmm to fit a bivariate response model where one response is
>> a Poisson count of pathogen load (Disease) and one is a Gaussian
>> phentotypic measure of the host (Phenotype).
>> I've fit a model with fixed effects that one would expect to influence
>> of these predictors, and including a random effect of site, of the form:
>> prior1<-list(R=list(V=diag(2), nu=3),G=list(G1=list(V=diag(2),nu=3,
>> MCMCglmm(cbind(Disease,Phenotype) ~ trait-1 +
>> The golden egg here, and my hypothesis, is that after accounting for some
>> predictors there will be a negative posterior correlation between the
>> response traits, where the posterior correlation is calculated as:
>> Cov(Disease,Phenotype) / sqrt(Var(Disease)*Var(Phenotype))
>> My question is whether this correlation is relevant at the G structure
>> level (Site random effect) or at the R structure level, which I take to be
>> the residual variance at the observation (individual) level?
>> I suspect the answer is the latter, but I'm struggling to interpret what
>> means if there is a negative correlation at the Site level? Does it mean
>> that the variances of the two traits at the site level are not
>> in that higher values in one trait for a site tend to produce lower values
>> for the other?
>> Any help greatly appreciated.
> The University of Edinburgh is a charitable body, registered in
> Scotland, with registration number SC005336.
Dr Xavier Harrison
Institute of Zoology
London NW1 4RY
+44 (0) 207 449 6621
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models