View topic - Random coefficient model again | Forum

Author

Message

< GenStat ~ Random coefficient model again

Nicholas Galwey

Posted: Wed Mar 14, 2012 9:48 am

Guest

Dear GenStatters,

A follow-up to my question last week about random coefficient models, answered by Murray Hannah.

I'm trying to fit one of these models to data with factors study_id (which identifies the individual subject) and treatment, explanatory variate Visit_Date, response variate total. The code is as follows:

VCOMPONENTS [FIXED = treatment*Visit_Date] RANDOM = study_id/Visit_Date
VSTRUCTURE [study_id/Visit_Date; CORR=unrest; FORM=whole; CINITIAL=_cinit]
REML [PRINT = model, components, Wald, means, effects; \
PTERMS = treatment*Visit_Date] total

When I add an autoregression model to this (as discussed in the previous query) it works fine, but as it stands this code gives me the following diagnostic:

Warning 38, code VC 31, statement 1 on line 8

Command: REML [PRINT = model, components, Wald, means, effects; PTERMS = treatm
Unsuccessful update for variance parameters - try smaller step.

Questions:

1. How do you change the step size?
3. If I change from CORR=unrest to CORR=positivedefinite or CORR=none, or even omit the VSTRUCTURE statement, the model converges. I guess that these simplifications are permissible, if the algorithm doesn't find enough information in the data to fit the fullest model. Do others agree? In this context, is setting CORR=none equivalent to omitting the VSTRUCTURE statement?

Best wishes,

Nick Galwey

________________________________

This e-mail was sent by GlaxoSmithKline Services Unlimited
(registered in England and Wales No. 1047315), which is a
member of the GlaxoSmithKline group of companies. The
registered address of GlaxoSmithKline Services Unlimited
is 980 Great West Road, Brentford, Middlesex TW8 9GS.

Post generated using Mail2Forum (http://www.mail2forum.com)

Murray Hannah

Posted: Thu Mar 15, 2012 12:20 am

Joined: 08 Oct 2008 Posts: 16

Hi Nick,

1. See VCYCLE for step length
2. I think CORR=positivedefinite should be okay. Also think setting CORR=none is equivalent to omitting the VSTRUCTURE statement, which model the subject slopes and intercepts as independent. With centred variate (Visit_Date), however, I would expect a positive correlation between slope and intercept (higher slope, higher centred intercept). Maybe you could get a sense of this with a trellis plot of total vs Visit_Date conditioned on study_id and treatment. It could turn out, with centred data, that slopes and intercepts are pretty much independent, but I'd still prefer to estimate their covariance. Maybe you could tweak _cinit to improve convergence.

Also a good idea to turn monitoring on (include "mon" in the PRINT list of REML), so you can see how convergence progressed.

Hope this helps
Murray.

Notice:
This email and any attachments may contain information that is personal, confidential,
legally privileged and/or copyright. No part of it should be reproduced, adapted or communicated without the prior written consent of the copyright owner.
It is the responsibility of the recipient to check for and remove viruses.
If you have received this email in error, please notify the sender by return email, delete it from your system and destroy any copies. You are not authorised to use, communicate or rely on the information contained in this email.
Please consider the environment before printing this email.

Post generated using Mail2Forum (http://www.mail2forum.com)

All times are GMT

View next topic View previous topic Page 1 of 1 Forum \| VSN International Forum Index ~ GenStat

You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
You can attach files in this forum
You can download files in this forum

Template by Mike Lothar // Protected by Anti-Spam ACP

| | | | | | |