Weight loss study.Mixed linear model. Help appreciated

[Erik Langer Madsen]

I am currently investigating a dataset including 68 subjects originating
from 3 different hospitals, who's had measured various parameters
including bodyweight at 5 different timepoints during a three year weight
loss study. The subjects can be characterized by different between
subjects factors: Gender: m/f, Diabetes: yes/no, treatment: active/placebo
and covariates such as age. Given that the data are correlated (repeated
measurements) and that some of the parameters that I wish to analyze have
missing values for some individuals I'd prefer using the mixed linear
model but I'm troubled by various outputs with different p-values
depending on how many fixed factors I add and which covariance structure I
use.

At present I use the ar1 covariance structure and time, gender, diabetes
and treatment as fixed factors. Including interactions between time and
the other fixed factors.
1. The AR1 structure seems to be my best choice given that the correlation
should be stronger regarding timepoints closer to each other. But does
anyone have arguments for choosing unstructured or compound symmetry
instead?
2.I ought to  add hospital as a random factor but then I get a non
positive hessian matrix. Any explanation for this problem or suggestions
for a solution would be appreciated.
3. Should I add subject (individuals) as a random factor under variance
components?
4. When gradually increasing the number of fixed factors into the model
then the p-values for levels and changes seems to differ a lot.
How should I test the validity of the model (Wald ? Overlapping of
confidence intervals?

Best regards
Erik Langer Madsen
[hidden email]
Phd-student
Aarhus Denmark

[Swank, Paul R]

Unfortunately, I don't use SPSS mixed to do my mixed models analyses but
I can tell you that the AR(1) structure is probably okay here. You can
test it but running a model that doesn't include any predictors other
than time and then comparing the variance covariance structure's impact
on fit. Once you have settled on the best var-cov structure, then youcan
test your hypothese about the fixed effects. How many hosptitals do you
have in the sample? If the number is small, say less than 20-30, then
you may not have enough variance in the hospitals to include it as a
random effect.

Usually, you only need to specify one level of the model as random. The
residual will represent the other effect.

Paul R. Swank, Ph.D. Professor
Director of Reseach
Children's Learning Institute
University of Texas Health Science Center-Houston

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Erik Langer Madsen
Sent: Wednesday, July 11, 2007 6:59 AM
To: [hidden email]
Subject: Weight loss study.Mixed linear model. Help appreciated

I am currently investigating a dataset including 68 subjects originating
from 3 different hospitals, who's had measured various parameters
including bodyweight at 5 different timepoints during a three year
weight loss study. The subjects can be characterized by different
between subjects factors: Gender: m/f, Diabetes: yes/no, treatment:
active/placebo and covariates such as age. Given that the data are
correlated (repeated
measurements) and that some of the parameters that I wish to analyze
have missing values for some individuals I'd prefer using the mixed
linear model but I'm troubled by various outputs with different p-values
depending on how many fixed factors I add and which covariance structure
I use.

At present I use the ar1 covariance structure and time, gender, diabetes
and treatment as fixed factors. Including interactions between time and
the other fixed factors.
1. The AR1 structure seems to be my best choice given that the
correlation should be stronger regarding timepoints closer to each
other. But does anyone have arguments for choosing unstructured or
compound symmetry instead?
2.I ought to  add hospital as a random factor but then I get a non
positive hessian matrix. Any explanation for this problem or suggestions
for a solution would be appreciated.
3. Should I add subject (individuals) as a random factor under variance
components?
4. When gradually increasing the number of fixed factors into the model
then the p-values for levels and changes seems to differ a lot.
How should I test the validity of the model (Wald ? Overlapping of
confidence intervals?

Best regards
Erik Langer Madsen
[hidden email]
Phd-student
Aarhus Denmark

[Erik Langer Madsen]

On Wed, 11 Jul 2007 10:25:21 -0500, Swank, Paul R
<[hidden email]> wrote:

>Unfortunately, I don't use SPSS mixed to do my mixed models analyses but
>I can tell you that the AR(1) structure is probably okay here. You can
>test it but running a model that doesn't include any predictors other
>than time and then comparing the variance covariance structure's impact
>on fit. Once you have settled on the best var-cov structure, then youcan
>test your hypothese about the fixed effects. How many hosptitals do you
>have in the sample? If the number is small, say less than 20-30, then
>you may not have enough variance in the hospitals to include it as a
>random effect.
>
>Usually, you only need to specify one level of the model as random. The
>residual will represent the other effect.
>
>Paul R. Swank, Ph.D. Professor
>Director of Reseach
>Children's Learning Institute
>University of Texas Health Science Center-Houston
>
>-----Original Message-----
>From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
>Erik Langer Madsen
>Sent: Wednesday, July 11, 2007 6:59 AM
>To: [hidden email]
>Subject: Weight loss study.Mixed linear model. Help appreciated
>
>I am currently investigating a dataset including 68 subjects originating
>from 3 different hospitals, who's had measured various parameters
>including bodyweight at 5 different timepoints during a three year
>weight loss study. The subjects can be characterized by different
>between subjects factors: Gender: m/f, Diabetes: yes/no, treatment:
>active/placebo and covariates such as age. Given that the data are
>correlated (repeated
>measurements) and that some of the parameters that I wish to analyze
>have missing values for some individuals I'd prefer using the mixed
>linear model but I'm troubled by various outputs with different p-values
>depending on how many fixed factors I add and which covariance structure
>I use.
>
>At present I use the ar1 covariance structure and time, gender, diabetes
>and treatment as fixed factors. Including interactions between time and
>the other fixed factors.
>1. The AR1 structure seems to be my best choice given that the
>correlation should be stronger regarding timepoints closer to each
>other. But does anyone have arguments for choosing unstructured or
>compound symmetry instead?
>2.I ought to  add hospital as a random factor but then I get a non
>positive hessian matrix. Any explanation for this problem or suggestions
>for a solution would be appreciated.
>3. Should I add subject (individuals) as a random factor under variance
>components?
>4. When gradually increasing the number of fixed factors into the model
>then the p-values for levels and changes seems to differ a lot.
>How should I test the validity of the model (Wald ? Overlapping of
>confidence intervals?
>
>Best regards
>Erik Langer Madsen
>[hidden email]
>Phd-student
>Aarhus Denmark

Dear Paul.R.Swank

Thank you for considering my problems.
We only have 3 hospitals included in this study so you may be right about
to little variance. How ever these hospitals do seem to differ somewhat
both regarding levels of weight and changes in weight.

 best regards
Erik