comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

classic Classic list List threaded Threaded
6 messages Options
Reply | Threaded
Open this post in threaded view
|

comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Kornbrot, Diana
Hi

One might expect (hope) these procedures would have same results when GLMM has normal with identity link and some settings.
NOT SO. 
Design. this is complex as want to test limits, between factors are very unbalanced subject variable has n = 87
Factor 1, between: betwennrepeat, 2 level
Factor 2, between: Nlevels, 4 levels
Factor 3, repeated: analysis, 5 levels

Nearest I can get is below with following syntax
GLM repeated
GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels
  /WSFACTOR=analysis 5 Repeated
  /CONTRAST(Nlevels)=Repeated
  /METHOD=SSTYPE(3)
  /EMMEANS=TABLES(OVERALL)
  /EMMEANS=TABLES(bewennrepeat)
  /EMMEANS=TABLES(Nlevels)
  /EMMEANS=TABLES(analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels)
  /EMMEANS=TABLES(bewennrepeat*analysis)
  /EMMEANS=TABLES(Nlevels*analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)
  /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY
  /CRITERIA=ALPHA(.05)
  /WSDESIGN=analysis
  /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

GLMM

*Generalized Linear Mixed Models.Key settings in bold
GENLINMIXED
  /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis COVARIANCE_TYPE=UNSTRUCTURED
  /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE
  /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY
  /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels USE_INTERCEPT=TRUE
  /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95 DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0 SINGULAR=0.000000000001
  /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat CONTRAST=NONE
   /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Results. 
For GLM repeated have only included Roy’s largest root,as this is nearest to GLMM
Note on df
For between factors there are 7 =(4-1 for Nlevels) + (2-1 for bewennrepeat) +1 for grand mean
For GGLMM: df2 = 80 for all F tests
For repeated Roy’s: between factors still have df2 = 80
Repeated factors have lower df,as the related measures df are also taken into account
Comparison
GLMM unnstructured satterthwaite model Multivariate Tests Roy's largest root
Source F df1 df2 Sig. Effect F df1 Error df Sig.
analysis 6.52 4 80 .000136 analysis 6.27 4 77 .000201
bewennrepeat 21.04 1 80 .000016 bewennrepeat 21.04 1 80 .000016
Nlevels .10 3 80 .962591 Nlevels .10 3 80 .962591
analysis * bewennrepeat 7.79 4 80 .000024 analysis * bewennrepeat 7.49 4 77 .000038
analysis * Nlevels .76 12 80 .688855 analysis * Nlevels 1.76 4 79 .145261
bewennrepeat * Nlevels .25 2 80 .783290 bewennrepeat * Nlevels .25 3 80 .864661

The results are identical for between factors, 
BUT slightly different for repeated factor
WHY?, WHICH RESULTS are to be recommended?
best
Diana


________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 +44 (0) 208 444 2081                                                   



===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Maguin, Eugene

Isn’t one difference that the genlinmixed model uses an unstructured cov matrix while the glm model uses the standard repeated measures assumption of compound symmetry? What happens if you impose a compound symmetry cov matrix on the genlinmixed model?

 

Gene Maguin

 

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Kornbrot, Diana
Sent: Wednesday, March 01, 2017 4:56 AM
To: [hidden email]
Subject: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

 

Hi

 

One might expect (hope) these procedures would have same results when GLMM has normal with identity link and some settings.

NOT SO. 

Design. this is complex as want to test limits, between factors are very unbalanced subject variable has n = 87

Factor 1, between: betwennrepeat, 2 level

Factor 2, between: Nlevels, 4 levels

Factor 3, repeated: analysis, 5 levels

 

Nearest I can get is below with following syntax

GLM repeated

GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels

  /WSFACTOR=analysis 5 Repeated

  /CONTRAST(Nlevels)=Repeated

  /METHOD=SSTYPE(3)

  /EMMEANS=TABLES(OVERALL)

  /EMMEANS=TABLES(bewennrepeat)

  /EMMEANS=TABLES(Nlevels)

  /EMMEANS=TABLES(analysis)

  /EMMEANS=TABLES(bewennrepeat*Nlevels)

  /EMMEANS=TABLES(bewennrepeat*analysis)

  /EMMEANS=TABLES(Nlevels*analysis)

  /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)

  /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY

  /CRITERIA=ALPHA(.05)

  /WSDESIGN=analysis

  /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

 

GLMM

 

*Generalized Linear Mixed Models.Key settings in bold

GENLINMIXED

  /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis COVARIANCE_TYPE=UNSTRUCTURED

  /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE

  /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY

  /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels USE_INTERCEPT=TRUE

  /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95 DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0 SINGULAR=0.000000000001

  /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE

   /EMMEANS TABLES=bewennrepeat CONTRAST=NONE

   /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE

   /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis CONTRAST=PAIRWISE

   /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE

   /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat CONTRAST=PAIRWISE

   /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis CONTRAST=PAIRWISE

  /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.

Results. 

For GLM repeated have only included Roy’s largest root,as this is nearest to GLMM

Note on df

For between factors there are 7 =(4-1 for Nlevels) + (2-1 for bewennrepeat) +1 for grand mean

For GGLMM: df2 = 80 for all F tests

For repeated Roy’s: between factors still have df2 = 80

Repeated factors have lower df,as the related measures df are also taken into account

Comparison

GLMM unnstructured

satterthwaite

model

Multivariate Tests

Roy's largest root

Source

F

df1

df2

Sig.

Effect

F

df1

Error df

Sig.

analysis

6.52

4

80

.000136

analysis

6.27

4

77

.000201

bewennrepeat

21.04

1

80

.000016

bewennrepeat

21.04

1

80

.000016

Nlevels

.10

3

80

.962591

Nlevels

.10

3

80

.962591

analysis * bewennrepeat

7.79

4

80

.000024

analysis * bewennrepeat

7.49

4

77

.000038

analysis * Nlevels

.76

12

80

.688855

analysis * Nlevels

1.76

4

79

.145261

bewennrepeat * Nlevels

.25

2

80

.783290

bewennrepeat * Nlevels

.25

3

80

.864661

 

The results are identical for between factors, 

BUT slightly different for repeated factor

WHY?, WHICH RESULTS are to be recommended?

best

Diana

 

 

________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 +44 (0) 208 444 2081                                                   

 

===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Bruce Weaver
Administrator
I was wondering about that too, Gene.  Diana, if that change doesn't get you all the way there, I'd also try changing DF to RESIDUAL.  


Maguin, Eugene wrote
Isn’t one difference that the genlinmixed model uses an unstructured cov matrix while the glm model uses the standard repeated measures assumption of compound symmetry? What happens if you impose a compound symmetry cov matrix on the genlinmixed model?

Gene Maguin

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Kornbrot, Diana
Sent: Wednesday, March 01, 2017 4:56 AM
To: [hidden email]
Subject: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Hi

One might expect (hope) these procedures would have same results when GLMM has normal with identity link and some settings.
NOT SO.
Design. this is complex as want to test limits, between factors are very unbalanced subject variable has n = 87
Factor 1, between: betwennrepeat, 2 level
Factor 2, between: Nlevels, 4 levels
Factor 3, repeated: analysis, 5 levels

Nearest I can get is below with following syntax
GLM repeated
GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels
  /WSFACTOR=analysis 5 Repeated
  /CONTRAST(Nlevels)=Repeated
  /METHOD=SSTYPE(3)
  /EMMEANS=TABLES(OVERALL)
  /EMMEANS=TABLES(bewennrepeat)
  /EMMEANS=TABLES(Nlevels)
  /EMMEANS=TABLES(analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels)
  /EMMEANS=TABLES(bewennrepeat*analysis)
  /EMMEANS=TABLES(Nlevels*analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)
  /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY
  /CRITERIA=ALPHA(.05)
  /WSDESIGN=analysis
  /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

GLMM

*Generalized Linear Mixed Models.Key settings in bold
GENLINMIXED
  /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis COVARIANCE_TYPE=UNSTRUCTURED
  /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE
  /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY
  /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels USE_INTERCEPT=TRUE
  /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95 DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0 SINGULAR=0.000000000001
  /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat CONTRAST=NONE
   /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Results.
For GLM repeated have only included Roy’s largest root,as this is nearest to GLMM
Note on df
For between factors there are 7 =(4-1 for Nlevels) + (2-1 for bewennrepeat) +1 for grand mean
For GGLMM: df2 = 80 for all F tests
For repeated Roy’s: between factors still have df2 = 80
Repeated factors have lower df,as the related measures df are also taken into account
Comparison
GLMM unnstructured

satterthwaite

model

Multivariate Tests

Roy's largest root

Source

F

df1

df2

Sig.

Effect

F

df1

Error df

Sig.

analysis

6.52

4

80

.000136

analysis

6.27

4

77

.000201

bewennrepeat

21.04

1

80

.000016

bewennrepeat

21.04

1

80

.000016

Nlevels

.10

3

80

.962591

Nlevels

.10

3

80

.962591

analysis * bewennrepeat

7.79

4

80

.000024

analysis * bewennrepeat

7.49

4

77

.000038

analysis * Nlevels

.76

12

80

.688855

analysis * Nlevels

1.76

4

79

.145261

bewennrepeat * Nlevels

.25

2

80

.783290

bewennrepeat * Nlevels

.25

3

80

.864661


The results are identical for between factors,
BUT slightly different for repeated factor
WHY?, WHICH RESULTS are to be recommended?
best
Diana


________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626
[hidden email]<mailto:[hidden email]>
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 +44 (0) 208 444 2081


===================== To manage your subscription to SPSSX-L, send a message to [hidden email]<mailto:[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
--
Bruce Weaver
bweaver@lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/

"When all else fails, RTFM."

NOTE: My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.
Reply | Threaded
Open this post in threaded view
|

Re: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Kornbrot, Diana
Compound symmetry is much less like multivariate than unstructured as it gives very large df for analysis (repeated) and its interactions
Residuals is definitely wrong it gives df as total N -all dos which is MUCH too large.

Here is my problem
I want to recommend a SIMPLE CONSISTENT approach to psychologists that applies to both normal, identify AND binomial, logit analyses.
For normal,  the repeated option is very well known and  used in zillions of studies. As Thom says GLMM unstructured may well be slightly better, and is what I intend recommending.
BUT I would dearly love to understand why I am getting different answers

Another problem  with GLMM is that it often gives up immediately as matrices are problematic - unlike General Estimating Equations, GEE that gives a prompt answer whatever.

GEE is different again and gives chi-square inferential test statistic. 
Again I would dearly love to know why GEE gives chi-square not F
NB GEE also gives too large df2 for repeated predictor

When analyses give too large df2 the p value will be lower thus increasing chance of wrongly rejecting the nuke
All very puzzling - but definitely not going for overlarge df2
thanks for help
best
Diana


On 1 Mar 2017, at 14:46, Bruce Weaver <[hidden email]> wrote:

I was wondering about that too, Gene.  Diana, if that change doesn't get you
all the way there, I'd also try changing DF to RESIDUAL.   



Maguin, Eugene wrote
Isn’t one difference that the genlinmixed model uses an unstructured cov
matrix while the glm model uses the standard repeated measures assumption
of compound symmetry? What happens if you impose a compound symmetry cov
matrix on the genlinmixed model?

Gene Maguin

From: SPSSX(r) Discussion [mailto:

SPSSX-L@.UGA

] On Behalf Of Kornbrot, Diana
Sent: Wednesday, March 01, 2017 4:56 AM
To:

SPSSX-L@.UGA

Subject: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Hi

One might expect (hope) these procedures would have same results when GLMM
has normal with identity link and some settings.
NOT SO.
Design. this is complex as want to test limits, between factors are very
unbalanced subject variable has n = 87
Factor 1, between: betwennrepeat, 2 level
Factor 2, between: Nlevels, 4 levels
Factor 3, repeated: analysis, 5 levels

Nearest I can get is below with following syntax
GLM repeated
GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels
 /WSFACTOR=analysis 5 Repeated
 /CONTRAST(Nlevels)=Repeated
 /METHOD=SSTYPE(3)
 /EMMEANS=TABLES(OVERALL)
 /EMMEANS=TABLES(bewennrepeat)
 /EMMEANS=TABLES(Nlevels)
 /EMMEANS=TABLES(analysis)
 /EMMEANS=TABLES(bewennrepeat*Nlevels)
 /EMMEANS=TABLES(bewennrepeat*analysis)
 /EMMEANS=TABLES(Nlevels*analysis)
 /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)
 /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY
 /CRITERIA=ALPHA(.05)
 /WSDESIGN=analysis
 /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

GLMM

*Generalized Linear Mixed Models.Key settings in bold
GENLINMIXED
 /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis
COVARIANCE_TYPE=UNSTRUCTURED
 /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE
 /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY
 /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat
analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels
USE_INTERCEPT=TRUE
 /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING
INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95
DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0
SINGULAR=0.000000000001
 /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS TABLES=bewennrepeat CONTRAST=NONE
  /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE
  /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis
CONTRAST=PAIRWISE
  /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat
CONTRAST=PAIRWISE
  /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis
CONTRAST=PAIRWISE
 /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Results.
For GLM repeated have only included Roy’s largest root,as this is nearest
to GLMM
Note on df
For between factors there are 7 =(4-1 for Nlevels) + (2-1 for
bewennrepeat) +1 for grand mean
For GGLMM: df2 = 80 for all F tests
For repeated Roy’s: between factors still have df2 = 80
Repeated factors have lower df,as the related measures df are also taken
into account
Comparison
GLMM unnstructured

satterthwaite

model

Multivariate Tests

Roy's largest root

Source

F

df1

df2

Sig.

Effect

F

df1

Error df

Sig.

analysis

6.52

4

80

.000136

analysis

6.27

4

77

.000201

bewennrepeat

21.04

1

80

.000016

bewennrepeat

21.04

1

80

.000016

Nlevels

.10

3

80

.962591

Nlevels

.10

3

80

.962591

analysis * bewennrepeat

7.79

4

80

.000024

analysis * bewennrepeat

7.49

4

77

.000038

analysis * Nlevels

.76

12

80

.688855

analysis * Nlevels

1.76

4

79

.145261

bewennrepeat * Nlevels

.25

2

80

.783290

bewennrepeat * Nlevels

.25

3

80

.864661


The results are identical for between factors,
BUT slightly different for repeated factor
WHY?, WHICH RESULTS are to be recommended?
best
Diana


________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626

d.e.kornbrot@.ac

&lt;mailto:

d.e.kornbrot@.ac

&gt;
http://dianakornbrot.wordpress.com/
http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
+44 (0) 208 444 2081


===================== To manage your subscription to SPSSX-L, send a
message to

LISTSERV@.UGA

&lt;mailto:

LISTSERV@.UGA

&gt; (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

LISTSERV@.UGA

(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





-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/

"When all else fails, RTFM."

NOTE: My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/comparing-GeneraliZed-linear-MIXED-GLMM-with-GLM-REPEATED-tp5733923p5733926.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
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

________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 +44 (0) 208 444 2081                                                   



===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Ryan Black
In reply to this post by Kornbrot, Diana
Diana,

I haven't had time to read your post carefully, but a quick glance at your syntax looks like you employed a mixed (within subjects and between subjects factors) ANOVA. As others mentioned, the tests generated that include the repeated measures factor assume Sphericity from an ANOVA. For a true comparison between a MANOVA and a linear mixed model, you should employ a MANOVA, which assumes an unstructured Sigma matrix. I would also suggest you use the MIXED procedure to start since you are assuming the DVs are conditionally MVN.

What you are trying to accomplish has been done before. The linear mixed modeling procedure in SPSS, and other software for that matter, can accomplish just about everything a general linear modeling procedure can do and more (e.g., better handle unbalanced designs, utilize more restrictive covariance structures).

Ryan

On Wed, Mar 1, 2017 at 4:56 AM, Kornbrot, Diana <[hidden email]> wrote:
Hi

One might expect (hope) these procedures would have same results when GLMM has normal with identity link and some settings.
NOT SO. 
Design. this is complex as want to test limits, between factors are very unbalanced subject variable has n = 87
Factor 1, between: betwennrepeat, 2 level
Factor 2, between: Nlevels, 4 levels
Factor 3, repeated: analysis, 5 levels

Nearest I can get is below with following syntax
GLM repeated
GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels
  /WSFACTOR=analysis 5 Repeated
  /CONTRAST(Nlevels)=Repeated
  /METHOD=SSTYPE(3)
  /EMMEANS=TABLES(OVERALL)
  /EMMEANS=TABLES(bewennrepeat)
  /EMMEANS=TABLES(Nlevels)
  /EMMEANS=TABLES(analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels)
  /EMMEANS=TABLES(bewennrepeat*analysis)
  /EMMEANS=TABLES(Nlevels*analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)
  /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY
  /CRITERIA=ALPHA(.05)
  /WSDESIGN=analysis
  /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

GLMM

*Generalized Linear Mixed Models.Key settings in bold
GENLINMIXED
  /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis COVARIANCE_TYPE=UNSTRUCTURED
  /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE
  /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY
  /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels USE_INTERCEPT=TRUE
  /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95 DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0 SINGULAR=0.000000000001
  /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat CONTRAST=NONE
   /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Results. 
For GLM repeated have only included Roy’s largest root,as this is nearest to GLMM
Note on df
For between factors there are 7 =(4-1 for Nlevels) + (2-1 for bewennrepeat) +1 for grand mean
For GGLMM: df2 = 80 for all F tests
For repeated Roy’s: between factors still have df2 = 80
Repeated factors have lower df,as the related measures df are also taken into account
Comparison
GLMM unnstructured satterthwaite model Multivariate Tests Roy's largest root
Source F df1 df2 Sig. Effect F df1 Error df Sig.
analysis 6.52 4 80 .000136 analysis 6.27 4 77 .000201
bewennrepeat 21.04 1 80 .000016 bewennrepeat 21.04 1 80 .000016
Nlevels .10 3 80 .962591 Nlevels .10 3 80 .962591
analysis * bewennrepeat 7.79 4 80 .000024 analysis * bewennrepeat 7.49 4 77 .000038
analysis * Nlevels .76 12 80 .688855 analysis * Nlevels 1.76 4 79 .145261
bewennrepeat * Nlevels .25 2 80 .783290 bewennrepeat * Nlevels .25 3 80 .864661

The results are identical for between factors, 
BUT slightly different for repeated factor
WHY?, WHICH RESULTS are to be recommended?
best
Diana


________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
<a href="tel:+44%201707%20284626" target="_blank" value="+441707284626">+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 <a href="tel:+44%2020%208444%202081" target="_blank" value="+442084442081">+44 (0) 208 444 2081                                                   



===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: comparing GeneraliZed linear MIXED (GLMM) with GLM REPEATED

Ryan Black
Diana,

You are correct that the multivariate omnibus tests are in fact multivariate, but those are followed by univariate tests in the output. There is much more flexibility using the MANOVA command. 
 
Ryan 

Sent from my iPhone

On Mar 5, 2017, at 10:24 AM, Kornbrot, Diana <[hidden email]> wrote:

Ryan

repeated does NOT assume sphericity fro multivariate tests - it uses unstructured  covariance and  makes various assumption about df2. it does assume normality

further investigation suggests that REPEATED with Roy’s largest root is very close to mixed, generalised linear, with normal and identity covariance unstructured, build Sattertthwaite df, with model covariance matrix

here mixed gives df2 = N - number of estimated parameters from main and between factors for all comparisons
while repeated also subtracts the df appropriate to the relevant within comparison

not a lot of people know that!

in fact, the number of ways of choking the options is rather more than the number of contributors to the debate
best
Diana


On 3 Mar 2017, at 13:03, Ryan Black <[hidden email]> wrote:

Diana,

I haven't had time to read your post carefully, but a quick glance at your syntax looks like you employed a mixed (within subjects and between subjects factors) ANOVA. As others mentioned, the tests generated that include the repeated measures factor assume Sphericity from an ANOVA. For a true comparison between a MANOVA and a linear mixed model, you should employ a MANOVA, which assumes an unstructured Sigma matrix. I would also suggest you use the MIXED procedure to start since you are assuming the DVs are conditionally MVN.

What you are trying to accomplish has been done before. The linear mixed modeling procedure in SPSS, and other software for that matter, can accomplish just about everything a general linear modeling procedure can do and more (e.g., better handle unbalanced designs, utilize more restrictive covariance structures).

Ryan

On Wed, Mar 1, 2017 at 4:56 AM, Kornbrot, Diana <[hidden email]> wrote:
Hi

One might expect (hope) these procedures would have same results when GLMM has normal with identity link and some settings.
NOT SO. 
Design. this is complex as want to test limits, between factors are very unbalanced subject variable has n = 87
Factor 1, between: betwennrepeat, 2 level
Factor 2, between: Nlevels, 4 levels
Factor 3, repeated: analysis, 5 levels

Nearest I can get is below with following syntax
GLM repeated
GLM raw_f lgt_f z_f FlogitVC_f FprobitVC_f BY bewennrepeat Nlevels
  /WSFACTOR=analysis 5 Repeated
  /CONTRAST(Nlevels)=Repeated
  /METHOD=SSTYPE(3)
  /EMMEANS=TABLES(OVERALL)
  /EMMEANS=TABLES(bewennrepeat)
  /EMMEANS=TABLES(Nlevels)
  /EMMEANS=TABLES(analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels)
  /EMMEANS=TABLES(bewennrepeat*analysis)
  /EMMEANS=TABLES(Nlevels*analysis)
  /EMMEANS=TABLES(bewennrepeat*Nlevels*analysis)
  /PRINT=DESCRIPTIVE ETASQ HOMOGENEITY
  /CRITERIA=ALPHA(.05)
  /WSDESIGN=analysis
  /DESIGN=bewennrepeat Nlevels bewennrepeat*Nlevels.

GLMM

*Generalized Linear Mixed Models.Key settings in bold
GENLINMIXED
  /DATA_STRUCTURE SUBJECTS=id REPEATED_MEASURES=analysis COVARIANCE_TYPE=UNSTRUCTURED
  /FIELDS TARGET=trans1 TRIALS=NONE OFFSET=NONE
  /TARGET_OPTIONS DISTRIBUTION=NORMAL LINK=IDENTITY
  /FIXED  EFFECTS=analysis bewennrepeat Nlevels analysis*bewennrepeat analysis*Nlevels bewennrepeat*Nlevels analysis*bewennrepeat*Nlevels USE_INTERCEPT=TRUE
  /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100 CONFIDENCE_LEVEL=95 DF_METHOD=SATTERTHWAITE COVB=MODEL PCONVERGE=0.000001(ABSOLUTE) SCORING=0 SINGULAR=0.000000000001
  /EMMEANS TABLES=analysis COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat CONTRAST=NONE
   /EMMEANS TABLES=Nlevels COMPARE=Nlevels CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
   /EMMEANS TABLES=bewennrepeat*Nlevels COMPARE=bewennrepeat CONTRAST=PAIRWISE
   /EMMEANS TABLES=analysis*bewennrepeat*Nlevels COMPARE=analysis CONTRAST=PAIRWISE
  /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Results. 
For GLM repeated have only included Roy’s largest root,as this is nearest to GLMM
Note on df
For between factors there are 7 =(4-1 for Nlevels) + (2-1 for bewennrepeat) +1 for grand mean
For GGLMM: df2 = 80 for all F tests
For repeated Roy’s: between factors still have df2 = 80
Repeated factors have lower df,as the related measures df are also taken into account
Comparison
GLMM unnstructured satterthwaite model Multivariate Tests Roy's largest root
Source F df1 df2 Sig. Effect F df1 Error df Sig.
analysis 6.52 4 80 .000136 analysis 6.27 4 77 .000201
bewennrepeat 21.04 1 80 .000016 bewennrepeat 21.04 1 80 .000016
Nlevels .10 3 80 .962591 Nlevels .10 3 80 .962591
analysis * bewennrepeat 7.79 4 80 .000024 analysis * bewennrepeat 7.49 4 77 .000038
analysis * Nlevels .76 12 80 .688855 analysis * Nlevels 1.76 4 79 .145261
bewennrepeat * Nlevels .25 2 80 .783290 bewennrepeat * Nlevels .25 3 80 .864661

The results are identical for between factors, 
BUT slightly different for repeated factor
WHY?, WHICH RESULTS are to be recommended?
best
Diana


________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
<a href="tel:+44%201707%20284626" target="_blank" value="+441707284626" class="">+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 <a href="tel:+44%2020%208444%202081" target="_blank" value="+442084442081" class="">+44 (0) 208 444 2081                                                   



===================== 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

________________________________________
Professor Diana Kornbrot
Work
University of Hertfordshire
College Lane, Hatfield, Hertfordshire AL10 9AB, UK
+44 (0) 170 728 4626
[hidden email]
http://dianakornbrot.wordpress.com/
 http://go.herts.ac.uk/Diana_Kornbrot
skype:  kornbrotme
Home
19 Elmhurst Avenue
London N2 0LT, UK
 +44 (0) 208 444 2081                                                   



===================== 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