Partial Least Squares in SPSS

Post by Johnny Sandaire
I was wondering if anyone has some insight on how I can run a PLS
analysis in SPSS. I was able to configure SPSS to run a PLS (that was
not easy), then tried to run the data. I have four variables that
formulate a single construct (such as, under the preference construct,
I would have: I love to eat desert in the morning; desert tastes
better with my morning coffee; breakfast is the best time to have
deserts; and I enjoy deserts better with my first meal of the day). I
am using a 7 point likert scale to record the users responses from
Strongly disagree to strongly agree. Now, I wanted to use these
variables as a dependent variable set to determine their influence on
the remainning 28 variables, which are a set of 4 per construct, such
as energy durign the day, appetite during lunch, etc...
Any assistance with this approach would be greatly appreciated.
Regards,
Johnny

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.

With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.

See SPSS Developer Central for details (www.spss.com/devcentral).

HTH,
Jon

Johnny Sandaire

2008-12-25 13:13:35 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

Hello Jon:
Thank you for your response. I have prepared a Web page that explains
my purpose. Since I am not able to publish images on this group, I
though I would provide a link so you can review what I have done and
provide your recommendations and future steps.

http://www.webzest.com/home/PLSSetup.htm

I hope you can assist.

Regards,

Johnny

JKPeck

2008-12-26 04:16:48 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

Thank you for your response. I have prepared a Web page that explains
my purpose. Since I am not able to publish images on this group, I
though I would provide a link so you can review what I have done and
provide your recommendations and future steps.
http://www.webzest.com/home/PLSSetup.htm
I hope you can assist.
Regards,
Johnny

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.

My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.

Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.

When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.

The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.

Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.

The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.

The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.

Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.

I hope that will get you started.

Regards,
Jon Peck

Johnny Sandaire

2008-12-26 12:51:36 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

Jon:

Thank you for your excellent response. Followed your guidelines and
produced the following:
http://www.webzest.com/PLSRegression.htm

To answer your question, I am attempting to repeat the following
experiment: http://www.webzest.com/utaut.pdf
The author used PLS to analyze his data and I am attempting to follow
his path to derive at a similar set of results. Additionally, I have
a small set of respondents; 37, to be exact. My sample size was small
due to a limited pilot study that I am conducting.

On your initial suggestion, I peaked at the multivariate GLM and I
like it. Given my data set, do you think I can use that approach as
well? How would I set-it-up? I tried it and my results were not all
there. Most of the responses were null.

Please advise...

Regards,

Johnny

JKPeck

2008-12-26 14:18:04 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

Thank you for your excellent response. Followed your guidelines and
produced the following:http://www.webzest.com/PLSRegression.htm
To answer your question, I am attempting to repeat the following
experiment:http://www.webzest.com/utaut.pdf
The author used PLS to analyze his data and I am attempting to follow
his path to derive at a similar set of results. Additionally, I have
a small set of respondents; 37, to be exact. My sample size was small
due to a limited pilot study that I am conducting.
On your initial suggestion, I peaked at the multivariate GLM and I
like it. Given my data set, do you think I can use that approach as
well? How would I set-it-up? I tried it and my results were not all
there. Most of the responses were null.
Please advise...
Regards,
Johnny

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.

So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.

HTH,
Jon

Johnny Sandaire

2008-12-26 16:53:37 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.
So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.
HTH,
Jon- Hide quoted text -
- Show quoted text -

Jon:

I reviewed the data and found one response that was all 1s out of a 7-
point likert scale. I held it back (removed it) and regenerated the
results. What do you think of this set?
http://www.webzest.com/PLSRegression2.htm

I do not think I can get any more results. The pilot was a small
group and I only invited 50 people.

Johnny

JKPeck

2008-12-27 02:48:05 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.
So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.
HTH,
Jon- Hide quoted text -
- Show quoted text -

I reviewed the data and found one response that was all 1s out of a 7-
point likert scale. I held it back (removed it) and regenerated the
results. What do you think of this set?http://www.webzest.com/PLSRegression2.htm
I do not think I can get any more results. The pilot was a small
group and I only invited 50 people.
Johnny

So, what is your ultimate goal for this analysis? It looks like the
incremental variance is falling off with the 5th factor, so you
probably are not going to improve the fit much more, but what do you
want to say about the results? The variance importance table might be
helpful in seeing which variables are driving what you have so far.

My inclination would be to simplify the model dramatically, since
there really isn't a lot of information in the small dataset you
have. But it all depends on your purpose.

-Jon

Johnny Sandaire

2008-12-29 14:23:23 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.
So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.
HTH,
Jon- Hide quoted text -
- Show quoted text -

So, what is your ultimate goal for this analysis? It looks like the
incremental variance is falling off with the 5th factor, so you
probably are not going to improve the fit much more, but what do you
want to say about the results? The variance importance table might be
helpful in seeing which variables are driving what you have so far.
My inclination would be to simplify the model dramatically, since
there really isn't a lot of information in the small dataset you
have. But it all depends on your purpose.
-Jon- Hide quoted text -
- Show quoted text -

Jon...

Thank you for your response and guidance. I sent you a response via
the Reply to Author and wanted to get your opinion on my data efforts
so far. I did not think it was helpful to share my data sets with the
Group; however, if that is not possible, then I understand.
Nevertheless, I would be interested in hearing your point-of-view on
the variance subject. My original research reference stated that
their process explained 70% of the variance in a specific area. I
wanted to understand their approach and repeat the process with my own
data and analysis. How do you suggest I proceed?

Regards,

-Johnny

JKPeck

2008-12-29 17:35:09 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.
So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.
HTH,
Jon- Hide quoted text -
- Show quoted text -

So, what is your ultimate goal for this analysis? It looks like the
incremental variance is falling off with the 5th factor, so you
probably are not going to improve the fit much more, but what do you
want to say about the results? The variance importance table might be
helpful in seeing which variables are driving what you have so far.
My inclination would be to simplify the model dramatically, since
there really isn't a lot of information in the small dataset you
have. But it all depends on your purpose.
-Jon- Hide quoted text -
- Show quoted text -

Jon...
Thank you for your response and guidance. I sent you a response via
the Reply to Author and wanted to get your opinion on my data efforts
so far. I did not think it was helpful to share my data sets with the
Group; however, if that is not possible, then I understand.
Nevertheless, I would be interested in hearing your point-of-view on
the variance subject. My original research reference stated that
their process explained 70% of the variance in a specific area. I
wanted to understand their approach and repeat the process with my own
data and analysis. How do you suggest I proceed?
Regards,
-Johnny

I think that this is a question you should discuss with your thesis
advisor. I am not familiar with the particular model you are applying
or the part this plays in your thesis. I would think that replicating
someone else's study on your data, while it might be useful, would not
typically be thesis material. The question is what your original
contribution is and how this analysis supports it.

Good luck,
Jon

Johnny Sandaire

2008-12-30 12:58:32 UTC

With SPSS versions 16 or 17 you can use the Python-based PLS module
from Developer Central. It supports multiple dependent variables, and
it has a dialog box interface.
With 14 or 15, there is an older Python module that you can use, but
it supports only a single dependent variable and does not have a
dialog box.
See SPSS Developer Central for details (www.spss.com/devcentral).
HTH,
Jon- Hide quoted text -
- Show quoted text -

The web page is an interesting way to present your questions. I wish
all questions were so clearly posed.
My first question back is why you want to use PLS rather than,
multivariate GLM. PLS is very useful in certain situations, but it
doesn't come with all the inferential apparatus of more traditional
procedures.
Getting to your specific issues, first, measurement level. The
measurement levels of the dependent and independent variables in the
analysis affect how the variables are treated, so it is important to
set these correctly. In the Define Variable Properties dialog you
chose, you would move all the variables you plan to use, or at least
those whose measurement levels might not be set correctly, to the
Variables to Scan list and press continue. SPSS will scan the data
and present you with a dialog where you can set the measurement level
(including getting hints at what the correct level might be) and set
other metadata. Alternatively, you can set the MLs in the Data Editor
Variable View from a simple drop-down menu.
When you do the PLS regression, you would include all the desired
predictors and all the dependent variables. PLS will extract the most
important combinations of predictors for those dependent variables.
It will not give you significance tests, however.
The Case ID variable is optional, but if you have such a variable, it
will make the plots easier to read. Otherwise it will just use the
case number.
Choosing the maximum number of latent factors is not an entirely
scientific process, but you will see in the output the proportion of
variance explained by each factor. Often it will happen that you will
see rapidly diminishing returns to adding more factors, which gives
you some idea of where to stop. There are some ways to test the
number of factors, but they are not part of this procedure.
The other tables include the Parameters table, which shows you
regression coefficients for each dependent variable, the Variable
Importance table, which shows you how variables impact each latent
factor.
The Model tab lets you set up interaction effects among your
predictors. Leave it alone for now.
Finally, the Options tab includes the ability to get a variety of
useful plots. Check all the boxes and enter a dataset name for each
and see what happens.
I hope that will get you started.
Regards,
Jon Peck- Hide quoted text -
- Show quoted text -

One of the situations where PLS is used is when there are too few
observations. With only 37 cases and such a large number of
variables, you are clearly in that situation. Perhaps when you get a
larger dataset you will be able to use the GLM approach. I saw only
one latent factor in the PLS results, and it didn't explain much of
the variance. Furthermore, the regression plot has all but one of the
points clustered very close together and one extremely different
point. That one point probably dominates all the results.
So it seems to me that as yet you don't have nearly enough data to
pursue such a complex model. Perhaps you can gain some insight if you
can really cut down the model. Otherwise, I'd wait for more data.
HTH,
Jon- Hide quoted text -
- Show quoted text -

So, what is your ultimate goal for this analysis? It looks like the
incremental variance is falling off with the 5th factor, so you
probably are not going to improve the fit much more, but what do you
want to say about the results? The variance importance table might be
helpful in seeing which variables are driving what you have so far.
My inclination would be to simplify the model dramatically, since
there really isn't a lot of information in the small dataset you
have. But it all depends on your purpose.
-Jon- Hide quoted text -
- Show quoted text -

Jon...
Thank you for your response and guidance. I sent you a response via
the Reply to Author and wanted to get your opinion on my data efforts
so far. I did not think it was helpful to share my data sets with the
Group; however, if that is not possible, then I understand.
Nevertheless, I would be interested in hearing your point-of-view on
the variance subject. My original research reference stated that
their process explained 70% of the variance in a specific area. I
wanted to understand their approach and repeat the process with my own
data and analysis. How do you suggest I proceed?
Regards,
-Johnny

Jon: That is good advice. Your SPSS guidance was great! Have a
wonderful New Year 2009.

Warmest regards,

-Johnny

morteza mofidi

2015-12-12 00:22:35 UTC

hi John
My thesis is in connection with indicators of sustainability. I have 95 index and 45 observations. My goal is to reduce the number of indicators. Is pca test is appropriate For my purpose? Or pls test is appropriate? In past software version 3.10 pls test is exist but I do not know how to run pls test and interpret the results ...
I'm grateful guide me

Zahra Hamze

2023-09-18 12:52:10 UTC

Post by morteza mofidi
hi John
My thesis is in connection with indicators of sustainability. I have 95 index and 45 observations. My goal is to reduce the number of indicators. Is pca test is appropriate For my purpose? Or pls test is appropriate? In past software version 3.10 pls test is exist but I do not know how to run pls test and interpret the results ...
I'm grateful guide me

greetings
I am trying to use PLS regression in SPSS but there is not any result in the end! I fixed all errors and installed the required libraries and no longer get any errors but still there is no result. I was wondering if you could help me fix this problem.

Rich Ulrich

2023-09-18 17:19:53 UTC

On Mon, 18 Sep 2023 05:52:10 -0700 (PDT), Zahra Hamze

Post by Zahra Hamze

You say,
"I ran something and it didn't work. Please help."

I'm not a user of PLS so I don't have a Usual Problem file.

It is probably not an problem with SPSS, but with your
syntax and/or data.

You might mention the N of cases, N of variables, and provide
the syntax that you ran.

Also: What generated the data? What do the correlations look like?

Citing the messages for the 'errors' that you fixed probably
would provide further clues.

--
Rich Ulrich

a***@gmail.com

2013-03-20 01:54:34 UTC