Dear Dr. Koper, Dear Michael
Thank you for your insight.I feel I should have given more description
around the data, but I feared being too lenghty.
The data I'm using is a state-wide data and the model I have is the most
simplified version of the predictions I am trying to obtain.
There are truly 4 nesting levels that we decided to collapse in order to be
able to run the analyses without loosing too much information.
There are about 15,000 subjects with an average of 10 repeated observations
for each , so the data set comprises a little less than 160,000
observations.
The fixed effects explanatory variables were entered in the models after
inspecting simple analyses and then performing a backward stepwise
regression.
In light of this information, do you think that I should change the
integration method or the starting values? Or is this something else that
gives that message?
My understanding is that by using the estimates of the GEE, it would
computationally be more effective.
Again, any help would be greatly appreciated.
Thank you
Hind
----- Original Message -----
From: "Nicola Koper" <nkoper@yahoo.com>
To: "Ali & Hind Lazrak" <alazrak@telus.net>
Sent: Friday, October 19, 2007 7:43 AM
Subject: Re: [S] glme warning message
Hi, Hind.
Occasionally, complex models have difficulty
converging if their starting values are too close to
the real values... because they don't have anywhere to
go. You might not need to specify the starting values
at all.
However, more likely the problem is that the model you
are trying to fit is too complex for the data. Unless
your data set is massive... 1000s of points?... you
are trying to do an awful lot with it. doubly-nested
models are complicated. Further, with a binary
distribution for the response variable, your model
actually has very little information... there is
little detail in the response variable... in order to
converge. So, it is common for complex models with a
binary distribution to not converge.
It would probably be best for you to simplify your
model by figuring out some way to avoid the two random
variables, obviously without violating any
assumptions.
Nicola Koper, Ph.D.
--- Ali & Hind Lazrak <alazrak@telus.net> wrote:
Dear S-news users
This is my first posting on this user group mailing
list. I just started using Splus for my master's
thesis, and I hope finding some help.
I am running a 2-level mixed effects model with
subjects (level 1) nested within firms (level 2).
Both subjects and firms are fitted as random
intercept. In the model the dependent variable is
binary (use) and I'm trying to predict the
determinants for use of protective equipment in a
workplace.
Before fitting the mixed model, I ran a gee model
where the coefficients obtained are used as starting
parameters for the REPQL integration method.
This model ran smoothly. My problem arises when I
run the mixed model itself as I get the following
warning message:
FALSE CONVERGENCE. in: ms( ~ - logLik(glmeSt,
glmePars), start = list(glmePars =
c(coef(glmeSt))), control = ....
I could not figure out what this warning means , and
the iterations are stopped.
Can anyone offer some help?
Thank you very much
Hind
P.S.: this is the command I used for the mixed model
sratglme <- glme(use ~ ethn + gender + curr.exp +
time + DOB + wd.dp + wd.group +
jobduration, random = ~ 1 | mill/studyno1, data =
srat.1, family = binomial(
link = "logit"), start = sratgee$coefficients,
method = "REPQL")
----- Original Message -----
From: "Michael O'Connell" <moconnell@insightful.com>
To: "Ali & Hind Lazrak" <alazrak@telus.net>
Sent: Friday, October 19, 2007 8:35 AM
Subject: RE: [S] glme warning message
can you share any data?
________________________________
From: s-news-owner@lists.biostat.wustl.edu on behalf of Ali & Hind Lazrak
Sent: Thu 10/18/2007 9:22 PM
To: s-news@lists.biostat.wustl.edu
Subject: [S] glme warning message
Dear S-news users
This is my first posting on this user group mailing list. I just started
using Splus for my master's thesis, and I hope finding some help.
I am running a 2-level mixed effects model with subjects (level 1) nested
within firms (level 2). Both subjects and firms are fitted as random
intercept. In the model the dependent variable is binary (use) and I'm
trying
to predict the determinants for use of protective equipment in a workplace.
Before fitting the mixed model, I ran a gee model where the coefficients
obtained are used as starting parameters for the REPQL integration method.
This model ran smoothly. My problem arises when I run the mixed model itself
as I get the following warning message:
FALSE CONVERGENCE. in: ms( ~ - logLik(glmeSt, glmePars), start =
list(glmePars =
c(coef(glmeSt))), control = ....
I could not figure out what this warning means , and the iterations
are stopped.
Can anyone offer some help?
Thank you very much
Hind
P.S.: this is the command I used for the mixed model
sratglme <- glme(use ~ ethn + gender + curr.exp + time + DOB
+ wd.dp + wd.group +
jobduration, random = ~ 1 | mill/studyno1, data = srat.1,
family = binomial(
link = "logit"), start = sratgee$coefficients, method =
"REPQL")
|