[R-sig-ME] How to fix individual elements of the residual covariance matrix in MCMCglmm (r_1470)

[claudia.kasper at iee.unibe.ch]

Dear Richard,

to fix the variance for the residual (co)variance matrix you could group the traits in a way that first you have those that should not be fixed and then those that should be fixed. For example, in a trivariate model I used fix=2 so that in a 3x3 (co)variance matrix cells nr 5,6,8,9 (numbered row-wise) are fixed. Hence, the variances of the second and third trait are fixed and their covariances whereas the other parameters are estimated. It is explained in "MCMC Methods for Multi-response Generalized Linear Mixed Models: The MCMCglmm R Package" by Jarrod Hadfield on page 9.

Best wishes,

Claudia Kasper
Institute for Ecology and Evolution
University of Bern
Switzerland

----------------------------------------------------------------------

Message: 1
Date: Wed, 22 Jun 2016 11:34:29 +0000 (UTC)
From: r_1470 <r_1470 at yahoo.co.uk>
To: "r-sig-mixed-models at r-project.org"
        <r-sig-mixed-models at r-project.org>
Subject: [R-sig-ME] How to fix individual elements of the residual
        covariance matrix in MCMCglmm
Message-ID:
        <1513886953.12791292.1466595269804.JavaMail.yahoo at mail.yahoo.com>
Content-Type: text/plain; charset="UTF-8"

Hello list,
I am using the MCMCglmm package in R for mixed-model multi-response regression, and I am trying to fix individual elements of the residual covariance matrix (when using rcov = ~ us(trait):units). I thought it might be possible using the 'fix' argument in the prior, but it seems that this either fixes the whole residual matrix to the prior specification, or allows all elements to be estimated. This fixes the whole residual matrix:
prior <- list(R = list(V = diag(5), n = 6, fix = 1), G = list( G1 = list(V = diag(5), n = 6)))
I believe 'fix = 0' would have the same effect as not specifying 'fix' (i.e. the whole matrix is estimated).
Is there a way to fix individual elements (e.g. I may want to fix the covariance between two of my five response variables to a specific value)
Thanks,
Richard.