SPSS Statistics

Hello folks.  I'm working on some new teaching slides for the fall, and have a question about the REPEATED sub-command for MIXED.  The documentation says:

• The default covariance type for repeated effects is DIAG. 

A bit further down, it says this:

COVTYPE(type)

Covariance structure. Specify the covariance structure of the identical blocks [emphasis added] for the residual covariance matrix (see Covariance Structure List (MIXED command)). The default structure for repeated effects is DIAG.

Meanwhile, I've been looking at these lecture slides.  Notice in particular slides 9-12, and notice that the author describes the matrix on slide 10 as BLOCK diagonal, not just diagonal. 

All this leaves me thinking that when I include a REPEATED sub-command, a BLOCK DIAGONAL matrix like the one on slide 10, is generated automatically (i.e., by default), and that COVTYPE(type) allows me to specify the structure of the n k×k SIGMA matrices for individual subjects (i.e., the identical blocks mentioned above).  But AFAICT, there is no way (and maybe no good reason) to change that BLOCK DIAGONAL matrix to some other structure.  

So, what do you think?  Am I (close to) understanding things correctly?  If not, can you set me straight?  ;-)  

Thanks,
Bruce

PS- If I am right, I don't understand why the description of COVTYPE(type) says again that "The default structure for repeated effects is DIAG."  Why would it say that if COVTYPE(type) can change the structure of the n SIGMA matrices only, but not the structure of the overall, BLOCK DIAG matrix?  This makes me think I may still not be understanding things correctly.  OTOH, if I am correct, maybe the documentation just needs some tweaking? 

I think I have now resolved my confusion.*  If I include a REPEATED sub-command without the COVTYPE(type) option, the Residual Covariance (R) Matrix is indeed a diagonal k x k matrix, where k is the number of measurement occasions.  So, this is one of the SIGMA matrices shown in the larger BLOCK DIAGONAL matrix on slide 10 here.  In other words, the documentation for REPEATED simply does not make any explicit reference to that larger BLOCK DIAG matrix. 

I've also compared examples on the website for Lesa Hoffman's LDA book to compare R matrices from SPSS, SAS, and Stata.  The output from SAS and Stata indicate that the matrix is for "personid = 1", or words to that effect.  It might be nice if the SPSS output also included that. 

* It is uncanny how frequently crafting a question to put to others helps one think more clearly about a problem themself.  And I daresay much of that "thinking" is done by the tortoise mind rather than the hare brain. 

Hello folks.  I'm working on some new teaching slides for the fall, and have a question about the REPEATED sub-command for MIXED.  The documentation says:

• The default covariance type for repeated effects is DIAG. 

A bit further down, it says this:

COVTYPE(type)

Covariance structure. Specify the covariance structure of the identical blocks [emphasis added] for the residual covariance matrix (see Covariance Structure List (MIXED command)). The default structure for repeated effects is DIAG.

Meanwhile, I've been looking at these lecture slides.  Notice in particular slides 9-12, and notice that the author describes the matrix on slide 10 as BLOCK diagonal, not just diagonal. 

All this leaves me thinking that when I include a REPEATED sub-command, a BLOCK DIAGONAL matrix like the one on slide 10, is generated automatically (i.e., by default), and that COVTYPE(type) allows me to specify the structure of the n k×k SIGMA matrices for individual subjects (i.e., the identical blocks mentioned above).  But AFAICT, there is no way (and maybe no good reason) to change that BLOCK DIAGONAL matrix to some other structure.  

So, what do you think?  Am I (close to) understanding things correctly?  If not, can you set me straight?  ;-)  

Thanks,
Bruce

PS- If I am right, I don't understand why the description of COVTYPE(type) says again that "The default structure for repeated effects is DIAG."  Why would it say that if COVTYPE(type) can change the structure of the n SIGMA matrices only, but not the structure of the overall, BLOCK DIAG matrix?  This makes me think I may still not be understanding things correctly.  OTOH, if I am correct, maybe the documentation just needs some tweaking? 

Hi Bruce

We would definitely consider your suggestion for the documentation of REPEATED subcommand.  Thanks for pointing out that.  In addition, I request more clarity on the change in SPSS output that you suggested above. It would be highly appreciated if you could share exactly what you need extra in the current SPSS output.

Regards

I think I have now resolved my confusion.*  If I include a REPEATED sub-command without the COVTYPE(type) option, the Residual Covariance (R) Matrix is indeed a diagonal k x k matrix, where k is the number of measurement occasions.  So, this is one of the SIGMA matrices shown in the larger BLOCK DIAGONAL matrix on slide 10 here.  In other words, the documentation for REPEATED simply does not make any explicit reference to that larger BLOCK DIAG matrix. 

I've also compared examples on the website for Lesa Hoffman's LDA book to compare R matrices from SPSS, SAS, and Stata.  The output from SAS and Stata indicate that the matrix is for "personid = 1", or words to that effect.  It might be nice if the SPSS output also included that. 

* It is uncanny how frequently crafting a question to put to others helps one think more clearly about a problem themself.  And I daresay much of that "thinking" is done by the tortoise mind rather than the hare brain. 

Hello folks.  I'm working on some new teaching slides for the fall, and have a question about the REPEATED sub-command for MIXED.  The documentation says:

• The default covariance type for repeated effects is DIAG. 

A bit further down, it says this:

COVTYPE(type)

Covariance structure. Specify the covariance structure of the identical blocks [emphasis added] for the residual covariance matrix (see Covariance Structure List (MIXED command)). The default structure for repeated effects is DIAG.

Meanwhile, I've been looking at these lecture slides.  Notice in particular slides 9-12, and notice that the author describes the matrix on slide 10 as BLOCK diagonal, not just diagonal. 

All this leaves me thinking that when I include a REPEATED sub-command, a BLOCK DIAGONAL matrix like the one on slide 10, is generated automatically (i.e., by default), and that COVTYPE(type) allows me to specify the structure of the n k×k SIGMA matrices for individual subjects (i.e., the identical blocks mentioned above).  But AFAICT, there is no way (and maybe no good reason) to change that BLOCK DIAGONAL matrix to some other structure.  

So, what do you think?  Am I (close to) understanding things correctly?  If not, can you set me straight?  ;-)  

Thanks,
Bruce

PS- If I am right, I don't understand why the description of COVTYPE(type) says again that "The default structure for repeated effects is DIAG."  Why would it say that if COVTYPE(type) can change the structure of the n SIGMA matrices only, but not the structure of the overall, BLOCK DIAG matrix?  This makes me think I may still not be understanding things correctly.  OTOH, if I am correct, maybe the documentation just needs some tweaking? 

Hello Bindu.  Thanks for your reply.  Now that I have studied the output more carefully, I see that when I use COVTYPE(UNR) or COVTYPE(CSR), the correlations are displayed, but they are in the Estimates of Covariance Parameters table, not in the Residual Covariance (R) Matrix table.  The latter invariable shows a covariance matrix.  I was expecting to see the latter table replaced by a correlation matrix, or to have a (full) residual correlation matrix tacked on after the R covariance matrix. 

As I write this, and think about the way things are presented in some books on linear mixed modeling, I think it would be great if there was an option to display variances on the diagonal, covariances above the diagonal, and correlations below the diagonal.  (See Table 4.3 in Lesa Hoffman's Longitudinal Analysis book, for example.)

I also think it would be great if there was a general option to display residual correlations, regardless of which covariance structure one selects.  At the moment (in v28, at least), the only options allowing me to display correlations are UNR and CSR.

Thanks,
Bruce

Hi Bruce

We would definitely consider your suggestion for the documentation of REPEATED subcommand.  Thanks for pointing out that.  In addition, I request more clarity on the change in SPSS output that you suggested above. It would be highly appreciated if you could share exactly what you need extra in the current SPSS output.

Regards

I think I have now resolved my confusion.*  If I include a REPEATED sub-command without the COVTYPE(type) option, the Residual Covariance (R) Matrix is indeed a diagonal k x k matrix, where k is the number of measurement occasions.  So, this is one of the SIGMA matrices shown in the larger BLOCK DIAGONAL matrix on slide 10 here.  In other words, the documentation for REPEATED simply does not make any explicit reference to that larger BLOCK DIAG matrix. 

I've also compared examples on the website for Lesa Hoffman's LDA book to compare R matrices from SPSS, SAS, and Stata.  The output from SAS and Stata indicate that the matrix is for "personid = 1", or words to that effect.  It might be nice if the SPSS output also included that. 

* It is uncanny how frequently crafting a question to put to others helps one think more clearly about a problem themself.  And I daresay much of that "thinking" is done by the tortoise mind rather than the hare brain. 

Hello folks.  I'm working on some new teaching slides for the fall, and have a question about the REPEATED sub-command for MIXED.  The documentation says:

• The default covariance type for repeated effects is DIAG. 

A bit further down, it says this:

COVTYPE(type)

Covariance structure. Specify the covariance structure of the identical blocks [emphasis added] for the residual covariance matrix (see Covariance Structure List (MIXED command)). The default structure for repeated effects is DIAG.

Meanwhile, I've been looking at these lecture slides.  Notice in particular slides 9-12, and notice that the author describes the matrix on slide 10 as BLOCK diagonal, not just diagonal. 

All this leaves me thinking that when I include a REPEATED sub-command, a BLOCK DIAGONAL matrix like the one on slide 10, is generated automatically (i.e., by default), and that COVTYPE(type) allows me to specify the structure of the n k×k SIGMA matrices for individual subjects (i.e., the identical blocks mentioned above).  But AFAICT, there is no way (and maybe no good reason) to change that BLOCK DIAGONAL matrix to some other structure.  

So, what do you think?  Am I (close to) understanding things correctly?  If not, can you set me straight?  ;-)  

Thanks,
Bruce

PS- If I am right, I don't understand why the description of COVTYPE(type) says again that "The default structure for repeated effects is DIAG."  Why would it say that if COVTYPE(type) can change the structure of the n SIGMA matrices only, but not the structure of the overall, BLOCK DIAG matrix?  This makes me think I may still not be understanding things correctly.  OTOH, if I am correct, maybe the documentation just needs some tweaking? 

Hello Bindu.  Thanks for your reply.  Now that I have studied the output more carefully, I see that when I use COVTYPE(UNR) or COVTYPE(CSR), the correlations are displayed, but they are in the Estimates of Covariance Parameters table, not in the Residual Covariance (R) Matrix table.  The latter invariable shows a covariance matrix.  I was expecting to see the latter table replaced by a correlation matrix, or to have a (full) residual correlation matrix tacked on after the R covariance matrix. 

As I write this, and think about the way things are presented in some books on linear mixed modeling, I think it would be great if there was an option to display variances on the diagonal, covariances above the diagonal, and correlations below the diagonal.  (See Table 4.3 in Lesa Hoffman's Longitudinal Analysis book, for example.)

I also think it would be great if there was a general option to display residual correlations, regardless of which covariance structure one selects.  At the moment (in v28, at least), the only options allowing me to display correlations are UNR and CSR.

Thanks,
Bruce

Hi Bruce

We would definitely consider your suggestion for the documentation of REPEATED subcommand.  Thanks for pointing out that.  In addition, I request more clarity on the change in SPSS output that you suggested above. It would be highly appreciated if you could share exactly what you need extra in the current SPSS output.

Regards

I think I have now resolved my confusion.*  If I include a REPEATED sub-command without the COVTYPE(type) option, the Residual Covariance (R) Matrix is indeed a diagonal k x k matrix, where k is the number of measurement occasions.  So, this is one of the SIGMA matrices shown in the larger BLOCK DIAGONAL matrix on slide 10 here.  In other words, the documentation for REPEATED simply does not make any explicit reference to that larger BLOCK DIAG matrix. 

I've also compared examples on the website for Lesa Hoffman's LDA book to compare R matrices from SPSS, SAS, and Stata.  The output from SAS and Stata indicate that the matrix is for "personid = 1", or words to that effect.  It might be nice if the SPSS output also included that. 

* It is uncanny how frequently crafting a question to put to others helps one think more clearly about a problem themself.  And I daresay much of that "thinking" is done by the tortoise mind rather than the hare brain. 

Hello folks.  I'm working on some new teaching slides for the fall, and have a question about the REPEATED sub-command for MIXED.  The documentation says:

• The default covariance type for repeated effects is DIAG. 

A bit further down, it says this:

COVTYPE(type)

Covariance structure. Specify the covariance structure of the identical blocks [emphasis added] for the residual covariance matrix (see Covariance Structure List (MIXED command)). The default structure for repeated effects is DIAG.

Meanwhile, I've been looking at these lecture slides.  Notice in particular slides 9-12, and notice that the author describes the matrix on slide 10 as BLOCK diagonal, not just diagonal. 

All this leaves me thinking that when I include a REPEATED sub-command, a BLOCK DIAGONAL matrix like the one on slide 10, is generated automatically (i.e., by default), and that COVTYPE(type) allows me to specify the structure of the n k×k SIGMA matrices for individual subjects (i.e., the identical blocks mentioned above).  But AFAICT, there is no way (and maybe no good reason) to change that BLOCK DIAGONAL matrix to some other structure.  

So, what do you think?  Am I (close to) understanding things correctly?  If not, can you set me straight?  ;-)  

Thanks,
Bruce

PS- If I am right, I don't understand why the description of COVTYPE(type) says again that "The default structure for repeated effects is DIAG."  Why would it say that if COVTYPE(type) can change the structure of the n SIGMA matrices only, but not the structure of the overall, BLOCK DIAG matrix?  This makes me think I may still not be understanding things correctly.  OTOH, if I am correct, maybe the documentation just needs some tweaking?