GEE problems with SPSS16

[Lee Woods]

I am having a few problems with running some GEE tests using SPSS16.

The data I am modelling consists of measures of car ownership for a number
of respondents over two time periods along with a list of socio-economic and
other explanatory variables.  The response (car ownership) is modelled as an
ordinal response.

Firstly, it is not possible to generate a measure of the overall fit of the
model such as a pseudo r-squared value.  Does anyone know why this might be
or if there is a good reason why this is not available?

Seccondly, when modelling the response as four categories, the model runs
perfectly well.  When re-running the exact same model for three categories
(a simpler model), SPSS states that the model can't be run due to
computational errors?  Any suggestions?  Maybe I should just try some
different software?

Any help would be much appreciated.

Lee Woods
University of Strathclyde, Glasgow, Scotland.

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

[Reutter, Alex]

Lee,

For the first problem, note that GEE is not a likelihood-based method of estimation, so measures based on likelihoods, such as the pseudo R squares that you see in Ordinal Regression (PLUM), are not possible for ordinal multinomial GEE models.

In _Generalized Estimating Equations_, Hardin and Hilbe suggest an extension of the R square statistic on pg 166.  I am not sure whether this measure has been generally adopted by the statistical community or whether it is useful as an absolute measure of model fit, but it was not included in SPSS.

If you simply need a relative measure of model fit for the purpose of comparing models, then the QIC and QICC are sufficient.


For the second problem, what's the exact wording of the error?

In terms of trying other software, my understanding is that SAS (GENMOD procedure) supports ordinal multinomial GEE, but only the independent working correlation structure is available (which is essentially a GZLM); Stata didn't include ordinal multinomial models in its GLM (for GZLMs) or XTGEE (for GEE models).  Perhaps there is an R package that does this.

Cheers,
Alex


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Lee Woods
Sent: Tuesday, February 12, 2008 9:16 AM
To: [hidden email]
Subject: GEE problems with SPSS16

I am having a few problems with running some GEE tests using SPSS16.

The data I am modelling consists of measures of car ownership for a number
of respondents over two time periods along with a list of socio-economic and
other explanatory variables.  The response (car ownership) is modelled as an
ordinal response.

Firstly, it is not possible to generate a measure of the overall fit of the
model such as a pseudo r-squared value.  Does anyone know why this might be
or if there is a good reason why this is not available?

Seccondly, when modelling the response as four categories, the model runs
perfectly well.  When re-running the exact same model for three categories
(a simpler model), SPSS states that the model can't be run due to
computational errors?  Any suggestions?  Maybe I should just try some
different software?

Any help would be much appreciated.

Lee Woods
University of Strathclyde, Glasgow, Scotland.

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

[Reutter, Alex]

I wrote:
> If you simply need a relative measure of model fit for the purpose of
> comparing models, then the QIC and QICC are sufficient.

...completely forgetting that the QIC and QICC aren't available in GEE models when the response is ordinal.  My apologies, Lee.

If you save predicted values, you can use Crosstabs to look at the observed vs. predicted values, and the measures in the Ordinal group on the Statistics subdialog could be useful, too.

Alex


-----Original Message-----
From: Reutter, Alex
Sent: Thursday, February 14, 2008 2:37 PM
To: [hidden email]
Subject: RE: GEE problems with SPSS16

Lee,

For the first problem, note that GEE is not a likelihood-based method of estimation, so measures based on likelihoods, such as the pseudo R squares that you see in Ordinal Regression (PLUM), are not possible for ordinal multinomial GEE models.

In _Generalized Estimating Equations_, Hardin and Hilbe suggest an extension of the R square statistic on pg 166.  I am not sure whether this measure has been generally adopted by the statistical community or whether it is useful as an absolute measure of model fit, but it was not included in SPSS.

If you simply need a relative measure of model fit for the purpose of comparing models, then the QIC and QICC are sufficient.


For the second problem, what's the exact wording of the error?

In terms of trying other software, my understanding is that SAS (GENMOD procedure) supports ordinal multinomial GEE, but only the independent working correlation structure is available (which is essentially a GZLM); Stata didn't include ordinal multinomial models in its GLM (for GZLMs) or XTGEE (for GEE models).  Perhaps there is an R package that does this.

Cheers,
Alex


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Lee Woods
Sent: Tuesday, February 12, 2008 9:16 AM
To: [hidden email]
Subject: GEE problems with SPSS16

I am having a few problems with running some GEE tests using SPSS16.

The data I am modelling consists of measures of car ownership for a number
of respondents over two time periods along with a list of socio-economic and
other explanatory variables.  The response (car ownership) is modelled as an
ordinal response.

Firstly, it is not possible to generate a measure of the overall fit of the
model such as a pseudo r-squared value.  Does anyone know why this might be
or if there is a good reason why this is not available?

Seccondly, when modelling the response as four categories, the model runs
perfectly well.  When re-running the exact same model for three categories
(a simpler model), SPSS states that the model can't be run due to
computational errors?  Any suggestions?  Maybe I should just try some
different software?

Any help would be much appreciated.

Lee Woods
University of Strathclyde, Glasgow, Scotland.

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