Skip to content
Prev 1465 / 20628 Next

Suspicious output from lme4-mcmcsamp

I meant to cc: this reply to the list.


---------- Forwarded message ----------
From: Douglas Bates <bates at stat.wisc.edu>
Date: Wed, Oct 8, 2008 at 12:27 PM
Subject: Re: [R-sig-ME] FW: Suspicious output from lme4-mcmcsamp
To: "De Woody J.A." <j.dewoody at soton.ac.uk>
On Wed, Oct 8, 2008 at 9:41 AM, De Woody J.A. <j.dewoody at soton.ac.uk> wrote:
As Dieter Menne indicated, this is one of the error messages that is
meant more as a reminder to myself that I should write the code for
this case.  The rather crytic terminology relates to the particular
way that the covariance matrix of the random effects is parameterized,
as sigma^2 * T %*% S %*% S %*% t(T) where S is a diagonal (S)cale
matrix with non-negative diagonal elements and T is a unit,
lower-(T)riangular matrix.  If the random effects are uncorrelated
then T is the identity matrix and there is no need to sample from the
posterior distribution of the parameters determining T.  I haven't yet
written the code to sample from that posterior distribution when the
random effects are correlated.
Expected by whom?  I don't know of examples where summary of an
merMCMC object has had a form other than what is shown below.

Such objects are in an S4 class and, in the fullness of time, as Bill
Venables is fond of saying, there will be extractor methods for them
that will provide more meaningful output.  At present there is not a
lot of useful output.  Mostly I look at the plots produced by

xyplot(sm2)

to see how the chains evolved.  At present, even that has a few
problems for models with multiple, uncorrelated random effects per
level of the grouping factor.
I don't understand why the fixef slot would have log(sigma^2), etc.
The fixef slot is the values of the fixed-effects parameters in the
sample.

The sigma slot is the sample from the distribution of the common scale
parameter, which is written as sigma.  The ST slot is the sample from
the distribution of the relative standard deviation of the random
effects.  To get the standard deviation of the random effects you
multiply ST by sigma.

I am not confident of the results from mcmcsamp at present and
regrettably I won't have time to look at it in more detail for a
couple of months.  If it were better to avoid confusion for the time
being I could disable it.  If someone else wants to experiment with it
send me email off-list and I will outline what should be happending in
that code.
Message-ID: <40e66e0b0810081027q7c902599i350728c2d59aa9bb@mail.gmail.com>
In-Reply-To: <40e66e0b0810081027k232a0158v70f4bb2ece6d5ec4@mail.gmail.com>