Let me ask you something - Effect Coding/MIXED

I'm a seasoned HLM person who relied on Hedeker's MIXREG package until SPSS
introduced MIXED.  I just ran some repeated measures analyses with 3
predictors: (1) treatment (2) week in treatment and (3) treatment by week
in treatment.  I have 2 treatment groups and potential time points.  There
are 74 subjects and a total of 450 observations.  The original coding for
treatment was 1 for experimental and 2 for control.


Run 1:

Using effect coding, I changed the experimental group to -1 and the control
group to + 1.  I computed the estimated outcomes from the parameter
estimates and plotted them - they were not a good fit to the raw data.

Run 2
Again using effect coding, I changed the experimental group to +1 and the
control group to 1.  After computing the estimated outcomes from the
paratement estimates, I plotted them and found them to fit the raw data -
AND run counter to the findings from the first run.

Finally, I ran both run 1 and run 2 in Hedeker's MIXREG package, just as a
comparison.  The parameter estimates produced the EXACT SAME estimates
whether the coding was as in run 1 or run 2.  This is what I would have
expected from SPSS MIXED.

What is going on?  Has anyone encountered this?  Is it a glitch?

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[parisec]

I have age1 age2 age3  and race1 race2 race3 race4 dummy variables.

Age1=refcat and Race1=refcat in the main effects model of a logistic regression..

For interaction terms i have:
age1 x race1
age2 x race1
age3 x race1

age1 x race2
age2 x race2
age3 x race2

age1 x race3
age2 x race3
age3 x race3

age1 x race4
age2 x race4
age3 x race4

now, since age1 and race1 are both refcats, which of these interaction terms go in the model and which stay out and become the reference categories?

my head has been buzzing on this for a few days now...

anyone have any insight?

Thanks
Carol

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Swank, Paul R]

anything with age1 or race1 are deleted. This leaves you with 6
variables which is the # of df.

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
Parise, Carol A.
Sent: Monday, October 22, 2007 5:38 PM
To: [hidden email]
Subject: Interactions - reference categories

I have age1 age2 age3  and race1 race2 race3 race4 dummy variables.

Age1=refcat and Race1=refcat in the main effects model of a logistic
regression..

For interaction terms i have:
age1 x race1
age2 x race1
age3 x race1

age1 x race2
age2 x race2
age3 x race2

age1 x race3
age2 x race3
age3 x race3

age1 x race4
age2 x race4
age3 x race4

now, since age1 and race1 are both refcats, which of these interaction
terms go in the model and which stay out and become the reference
categories?

my head has been buzzing on this for a few days now...

anyone have any insight?

Thanks
Carol

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command SIGNOFF SPSSX-L For a list
of commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[parisec]

Thank You...one more question...


Age1=refcat =45-70
Race1=refcat= White

what would be the most appropriate way to interpret a significant interaction of Age71+ x African American ?

Is it door number 1?

The odds that African Americans under over age 71 were  *** times more likely than African Americans aged 46-70 to be low SES
OR
door number 2?
                The odds that African Americans under over age 71 were  *** times more likely than Whites aged 46-70 to be low SES

Or is is something else?

Thanks
Carol



-----Original Message-----
From: Swank, Paul R [mailto:[hidden email]]
Sent: Monday, October 22, 2007 3:48 PM
To: Parise, Carol A.; [hidden email]
Subject: RE: Interactions - reference categories



anything with age1 or race1 are deleted. This leaves you with 6
variables which is the # of df.

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
Parise, Carol A.
Sent: Monday, October 22, 2007 5:38 PM
To: [hidden email]
Subject: Interactions - reference categories

I have age1 age2 age3  and race1 race2 race3 race4 dummy variables.

Age1=refcat and Race1=refcat in the main effects model of a logistic
regression..

For interaction terms i have:
age1 x race1
age2 x race1
age3 x race1

age1 x race2
age2 x race2
age3 x race2

age1 x race3
age2 x race3
age3 x race3

age1 x race4
age2 x race4
age3 x race4

now, since age1 and race1 are both refcats, which of these interaction
terms go in the model and which stay out and become the reference
categories?

my head has been buzzing on this for a few days now...

anyone have any insight?

Thanks
Carol

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command SIGNOFF SPSSX-L For a list
of commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Swank, Paul R]

You're dealing with an interaction which basically means a difference
between differences. Thus, the difference between 71+ and 46-70 is not
the same for African_Americans as it is for Whites. It may be that the
odds are greater for an older African-American compared to a younger
African-American to be at a lower SES than it would be for and older
White repondent comapred to a younger White respondent.


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


-----Original Message-----
From: Parise, Carol A. [mailto:[hidden email]]
Sent: Monday, October 22, 2007 6:01 PM
To: Swank, Paul R; [hidden email]
Subject: RE: Interactions - reference categories

Thank You...one more question...


Age1=refcat =45-70
Race1=refcat= White

what would be the most appropriate way to interpret a significant
interaction of Age71+ x African American ?

Is it door number 1?

The odds that African Americans under over age 71 were  *** times more
likely than African Americans aged 46-70 to be low SES OR door number 2?
                The odds that African Americans under over age 71 were
*** times more likely than Whites aged 46-70 to be low SES

Or is is something else?

Thanks
Carol



-----Original Message-----
From: Swank, Paul R [mailto:[hidden email]]
Sent: Monday, October 22, 2007 3:48 PM
To: Parise, Carol A.; [hidden email]
Subject: RE: Interactions - reference categories



anything with age1 or race1 are deleted. This leaves you with 6
variables which is the # of df.

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
Parise, Carol A.
Sent: Monday, October 22, 2007 5:38 PM
To: [hidden email]
Subject: Interactions - reference categories

I have age1 age2 age3  and race1 race2 race3 race4 dummy variables.

Age1=refcat and Race1=refcat in the main effects model of a logistic
regression..

For interaction terms i have:
age1 x race1
age2 x race1
age3 x race1

age1 x race2
age2 x race2
age3 x race2

age1 x race3
age2 x race3
age3 x race3

age1 x race4
age2 x race4
age3 x race4

now, since age1 and race1 are both refcats, which of these interaction
terms go in the model and which stay out and become the reference
categories?

my head has been buzzing on this for a few days now...

anyone have any insight?

Thanks
Carol

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command SIGNOFF SPSSX-L For a list
of commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[peter link]

Hi Charla -

I am only speculating here, but, if you enter the treatment group in the
'factor' box rather than the 'covariate' box, maybe that would get results
consistent with Don Hedeker's MIXREG program.

Keep in mind that changing the coding from +1 to -1 should change the sign
of the treatment group parameter estimate.  Maybe more than this is
happening, though?

Hope this helps some,

Peter Link
VA San Diego Healthcare System

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
Charla Lopez
Sent: Monday, October 22, 2007 12:57 PM
To: [hidden email]
Subject: Let me ask you something - Effect Coding/MIXED


I'm a seasoned HLM person who relied on Hedeker's MIXREG package until SPSS
introduced MIXED.  I just ran some repeated measures analyses with 3
predictors: (1) treatment (2) week in treatment and (3) treatment by week
in treatment.  I have 2 treatment groups and potential time points.  There
are 74 subjects and a total of 450 observations.  The original coding for
treatment was 1 for experimental and 2 for control.


Run 1:

Using effect coding, I changed the experimental group to -1 and the control
group to + 1.  I computed the estimated outcomes from the parameter
estimates and plotted them - they were not a good fit to the raw data.

Run 2
Again using effect coding, I changed the experimental group to +1 and the
control group to 1.  After computing the estimated outcomes from the
paratement estimates, I plotted them and found them to fit the raw data -
AND run counter to the findings from the first run.

Finally, I ran both run 1 and run 2 in Hedeker's MIXREG package, just as a
comparison.  The parameter estimates produced the EXACT SAME estimates
whether the coding was as in run 1 or run 2.  This is what I would have
expected from SPSS MIXED.

What is going on?  Has anyone encountered this?  Is it a glitch?

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD