Can lme allow for serial correlation and 'pure' measurement error?

Dear mixed models list
 
Could someone please confirm my belief that lme does not allow one to
fit models with separate serial correlation and measurement error
components? In SAS proc Mixed, one can use serial correlation with a
"repeated", and adding the option "local" to this state adds an
additional independent error term. As far as I can tell from Pinheiro
and Bates, lme only allows specification of a single level of residual
covariance structure. As far as I understand, a nugget effect does not
give the same residual covariance structure that I want.
 
Just to give the context, I'm analysing a dataset in which subjects are
measured repeatedly over time. Subjects are measured at a number of
examinations, with multiple measurements made at each examination
(though not always the same number-otherwise I would just take the
mean). Conditional on a set of random effects, I believe there is serial
correlation, but since I have multiple measurements at identical times
for each subject, I need an additional measurement error component,
since for a model with just serial correlation a subject's measurements
at the same time point must be identical.
 
My apologies if the answer is obviously no, but I just wanted to check
I wasn't missing something obvious.
 
Many thanks
Jonathan Bartlett

	[[alternative HTML version deleted]]

I don't think that lme can do that out of the box.

One hack to get around the problem would be to use the mean of the
multiple measurements, also record the standard error of measurements
within examination and feed the latter into a variance model using the
weights argument.

I hope that this helps,

Andrew

On Thu, May 29, 2008 at 03:29:29PM +0100, Jonathan.Bartlett at lshtm.ac.uk wrote:

Dear mixed models list
 
Could someone please confirm my belief that lme does not allow one to
fit models with separate serial correlation and measurement error
components? In SAS proc Mixed, one can use serial correlation with a
"repeated", and adding the option "local" to this state adds an
additional independent error term. As far as I can tell from Pinheiro
and Bates, lme only allows specification of a single level of residual
covariance structure. As far as I understand, a nugget effect does not
give the same residual covariance structure that I want.
 
Just to give the context, I'm analysing a dataset in which subjects are
measured repeatedly over time. Subjects are measured at a number of
examinations, with multiple measurements made at each examination
(though not always the same number-otherwise I would just take the
mean). Conditional on a set of random effects, I believe there is serial
correlation, but since I have multiple measurements at identical times
for each subject, I need an additional measurement error component,
since for a model with just serial correlation a subject's measurements
at the same time point must be identical.
 
My apologies if the answer is obviously no, but I just wanted to check
I wasn't missing something obvious.
 
Many thanks
Jonathan Bartlett

	[[alternative HTML version deleted]]

-- 
Andrew Robinson  
Department of Mathematics and Statistics            Tel: +61-3-8344-6410
University of Melbourne, VIC 3010 Australia         Fax: +61-3-8344-4599
http://www.ms.unimelb.edu.au/~andrewpr
http://blogs.mbs.edu/fishing-in-the-bay/