Discussion:
Advanced Question - Direct Oblimin Rotation
(too old to reply)
Daniel Soper
2004-10-06 19:25:00 UTC
Permalink
Greetings everyone...

I have a rather advanced SPSS question related to rotations in a factor
analysis, and I sincerely hope that one of the gurus out there will know the
answer.

Here's a little background on the problem:

SPSS has an implementation of the direct oblimin rotation strategy that can
be used when an oblique set of rotated factors is desired. The purpose of
direct oblimin rotation is to minimize the covariance of the squared
loadings in distinct columns. The loss function for direct oblimin rotation
as defined by McDonald (1985, p.86) contains a parameter known as "gamma"
that can be set to a value between zero and one. As gamma increases from
zero to one, the factors become less and less correlated. Unfortunately,
SPSS does not seem to directly adhere to the definitional formula for direct
oblimin rotation. Rather than providing the ability to adjust the gamma
parameter, SPSS instead provides access to a parameter known as "delta",
which is not a part of the definitional formula. The delta parameter seems
to operate in the opposite direction of the formally defined gamma
parameter, in that high values of delta yield higher correlations among
factors. According to the SPSS manual, the highest correlations among
factors are achieved when delta is left at its default value of zero,
however the maximum value of delta (which actually yields the highest
correlations) seems to be 0.80. As delta decreases into the negative range,
the factors become more and more orthogonal.

Given all of the background information listed above, here's my question:

Mathematically, what is this"delta" parameter that SPSS provides, and how
does it relate to the definitional formula? Does it have a basis in
literature, or did the good people at SPSS invent it? If anyone has any
information on this topic, please reply to this post. I've scoured the web
for a solution, and have found absolutely nothing.

Thank you in advance for your help!

-Dan
Bruce Weaver
2004-10-06 20:32:15 UTC
Permalink
Post by Daniel Soper
Greetings everyone...
I have a rather advanced SPSS question related to rotations in a factor
analysis, and I sincerely hope that one of the gurus out there will know the
answer.
--- background snipped ------
Post by Daniel Soper
Mathematically, what is this"delta" parameter that SPSS provides, and how
does it relate to the definitional formula? Does it have a basis in
literature, or did the good people at SPSS invent it? If anyone has any
information on this topic, please reply to this post. I've scoured the web
for a solution, and have found absolutely nothing.
Thank you in advance for your help!
-Dan
I can't tell if the manual you referred to is from the SPSS Algorithms
site. If not, go to http://support.spss.com. Log in as guest with
password guest. Select Statistics then Algorithms from the left hand
list. There is a PDF file with the algorithms used for Factor Analysis
that includes info on Rotation of Factors. You might find what you need
there.
--
Bruce Weaver
***@lakeheadu.ca
www.angelfire.com/wv/bwhomedir
Daniel Soper
2004-10-07 02:42:42 UTC
Permalink
This is exactly what I needed. Thank you so much for your advice!

-Dan
Post by Bruce Weaver
Post by Daniel Soper
Greetings everyone...
I have a rather advanced SPSS question related to rotations in a factor
analysis, and I sincerely hope that one of the gurus out there will know
the answer.
--- background snipped ------
Post by Daniel Soper
Mathematically, what is this"delta" parameter that SPSS provides, and how
does it relate to the definitional formula? Does it have a basis in
literature, or did the good people at SPSS invent it? If anyone has any
information on this topic, please reply to this post. I've scoured the
web for a solution, and have found absolutely nothing.
Thank you in advance for your help!
-Dan
I can't tell if the manual you referred to is from the SPSS Algorithms
site. If not, go to http://support.spss.com. Log in as guest with
password guest. Select Statistics then Algorithms from the left hand list.
There is a PDF file with the algorithms used for Factor Analysis that
includes info on Rotation of Factors. You might find what you need there.
--
Bruce Weaver
www.angelfire.com/wv/bwhomedir
optionstraderjeff
2005-01-16 08:15:50 UTC
Permalink
Bruce,

You might want to try Harris-Kaiser Orthoblique, or Primary Product
Functionplane. The former yields exceptional results when the data
have strong, independent clusters. The latter is the best available
method when a high hyperplane count is desired from a robust analytic
solution. Both methods yield consistently good solutions that have
just the right amount of obliquity to get a good fit, and no more.

I do not know if SPSS has implemented either of these methods, but they
are far superior to oblimin--and not highly sensitive to parameter
values.

If you are interested, I can post Fortran code for both of these
rotational methods on my anonymous ftp site or in one of these groups.
You should be able to write the unrotated factor matrix to a text file,
run it through one of these rotations, and then load the factor pattern
matrix and factor correlations back into SPSS.

Question: Can SPSS directly access external Fortran subroutines or
even complete programmes? Is it "extensible" like, for example,
Matlab?

Jeffrey Owen Katz, Ph.D.
v***@gmail.com
2020-05-09 21:30:31 UTC
Permalink
Post by Daniel Soper
Greetings everyone...
I have a rather advanced SPSS question related to rotations in a factor
analysis, and I sincerely hope that one of the gurus out there will know the
answer.
SPSS has an implementation of the direct oblimin rotation strategy that can
be used when an oblique set of rotated factors is desired. The purpose of
direct oblimin rotation is to minimize the covariance of the squared
loadings in distinct columns. The loss function for direct oblimin rotation
as defined by McDonald (1985, p.86) contains a parameter known as "gamma"
that can be set to a value between zero and one. As gamma increases from
zero to one, the factors become less and less correlated. Unfortunately,
SPSS does not seem to directly adhere to the definitional formula for direct
oblimin rotation. Rather than providing the ability to adjust the gamma
parameter, SPSS instead provides access to a parameter known as "delta",
which is not a part of the definitional formula. The delta parameter seems
to operate in the opposite direction of the formally defined gamma
parameter, in that high values of delta yield higher correlations among
factors. According to the SPSS manual, the highest correlations among
factors are achieved when delta is left at its default value of zero,
however the maximum value of delta (which actually yields the highest
correlations) seems to be 0.80. As delta decreases into the negative range,
the factors become more and more orthogonal.
Mathematically, what is this"delta" parameter that SPSS provides, and how
does it relate to the definitional formula? Does it have a basis in
literature, or did the good people at SPSS invent it? If anyone has any
information on this topic, please reply to this post. I've scoured the web
for a solution, and have found absolutely nothing.
Thank you in advance for your help!
-Dan
Hello Dan,

I am dealing with the exact same problem that you mentioned here and trying to understand why the R and SPSS outputs do not match. Could you please share your suggestions on this. I am unable to access the spss support website mentioned above.

-Vinita
Rich Ulrich
2020-05-10 17:14:17 UTC
Permalink
Post by v***@gmail.com
Post by Daniel Soper
Greetings everyone...
I have a rather advanced SPSS question related to rotations in a factor
analysis, and I sincerely hope that one of the gurus out there will know the
answer.
SPSS has an implementation of the direct oblimin rotation strategy that can
be used when an oblique set of rotated factors is desired. The purpose of
direct oblimin rotation is to minimize the covariance of the squared
loadings in distinct columns. The loss function for direct oblimin rotation
as defined by McDonald (1985, p.86) contains a parameter known as "gamma"
that can be set to a value between zero and one. As gamma increases from
zero to one, the factors become less and less correlated. Unfortunately,
SPSS does not seem to directly adhere to the definitional formula for direct
oblimin rotation. Rather than providing the ability to adjust the gamma
parameter, SPSS instead provides access to a parameter known as "delta",
which is not a part of the definitional formula. The delta parameter seems
to operate in the opposite direction of the formally defined gamma
parameter, in that high values of delta yield higher correlations among
factors. According to the SPSS manual, the highest correlations among
factors are achieved when delta is left at its default value of zero,
however the maximum value of delta (which actually yields the highest
correlations) seems to be 0.80. As delta decreases into the negative range,
the factors become more and more orthogonal.
Mathematically, what is this"delta" parameter that SPSS provides, and how
does it relate to the definitional formula? Does it have a basis in
literature, or did the good people at SPSS invent it? If anyone has any
information on this topic, please reply to this post. I've scoured the web
for a solution, and have found absolutely nothing.
Thank you in advance for your help!
-Dan
Hello Dan,
I am dealing with the exact same problem that you mentioned here and trying to understand why the R and SPSS outputs do not match. Could you please share your suggestions on this. I am unable to access the spss support website mentioned above.
I found the original thread, and the original answers.

SPSS was bought by IBM since Bruce gave 2004's support URL,
which no longer exists.

Looking for the SPSS algorithms manual, I found
ftp://public.dhe.ibm.com/software/analytics/spss/documentation/statistics/24.0/en/client/Manuals/

- which includes a listing for the algorithms manual.

Notice, the manual has an FTP prefix, not http (where you
see the Link translated, probably at the bottom of your screen.)
I clicked on the manual multiple times in Firefox without having
it open, before I noticed FTP and realized that "FTP" isn't going
to open. FTP is an old file-transfer protocol which I haven't
used in years. But clicking the link in Firefox did download the
pdf file every time I had clicked.

Anyway, I discovered multiple versions of the manual now
exist in my DOWNLOADs directory. The one that is 61 MB
is version 24. Earlier clicking around had gotten me a
larger document which is version 25. The section you want
will not have changed, I expect.

Hope this helps. - BTW, the fact that I did not find a more
common, HTTP, on-line file does not prove that one does
not exist.
--
Rich Ulrich
Loading...