SPSSX Discussion

Confirmatory vs Exploratory Factor Analysis

Classic

List

Threaded

15 messages Options

Humphrey Paulie

Confirmatory vs Exploratory Factor Analysis

Dear list members,
Though, I understand the difference between CFA and EFA in principal, I have problems understanding the logic of CFA as I think EFA is enough.
In CFA we have an idea of the number of factors that the test is made up of and the loadings of the tests on each factor. Then we test this hypothesis by means of CFA to find out whether our hypothesized factor structure fits or not (is that right?).
What’s wrong with EFA under such a situation?
If we perform EFA and we get the hypothesized factor structure that we expected, then we have some evidence for our theory. If the EFA doesn’t give our expected factor structure then even the CFA cannot support our factor structure either (i.e., it won’t fit).
And if CFA supports our factor structure then the hypothesized factor structure should emerge when EFA is used afterwards.
This is the way I understand CFA & EFA.
It must be wrong since this argument makes CFA inconsequential.
Could you please make it clear for me?

Cheers
Humphrey

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

Swank, Paul R

Re: Confirmatory vs Exploratory Factor Analysis

You should try this question on the SEMNET listserve but I will take a stab at it. EFA will produce a set of factors but it is relatively easy, after the fact to see how he factors support you hypotheses. However, you have no measure of how well your theory is modeled. With CFA, you are forced a priori to specify a structure and then have a clear test (more or less) of whether or not the data supports the theoretical structure. EFA is also very loose in that often, if you apply a CFA to a model indicated by an EFA, even on the same data, it will not fit. This is because EFA allows loadings of indicators on factors that they don't belong to whereas CFA does not. If CFA supports the model then you have a much stronger argument for the validity of that model.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Humphrey Paulie
Sent: Friday, October 03, 2008 5:33 AM
To: [hidden email]
Subject: Confirmatory vs Exploratory Factor Analysis

Dear list members,
Though, I understand the difference between CFA and EFA in principal, I have problems understanding the logic of CFA as I think EFA is enough.
In CFA we have an idea of the number of factors that the test is made up of and the loadings of the tests on each factor. Then we test this hypothesis by means of CFA to find out whether our hypothesized factor structure fits or not (is that right?).
What?s wrong with EFA under such a situation?
If we perform EFA and we get the hypothesized factor structure that we expected, then we have some evidence for our theory. If the EFA doesn?t give our expected factor structure then even the CFA cannot support our factor structure either (i.e., it won?t fit).
And if CFA supports our factor structure then the hypothesized factor structure should emerge when EFA is used afterwards.
This is the way I understand CFA & EFA.
It must be wrong since this argument makes CFA inconsequential.
Could you please make it clear for me?

Cheers
Humphrey

=======
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

Garry Gelade

Re: Confirmatory vs Exploratory Factor Analysis

Hello Humphrey,

Paul is right, I'd add however that CFA is much more powerful than EFA.

Just a few examples of the possibilities: You can fix or free correlations
between any subset of factors; you can include additional correlations
between indicators; you can specify 2nd order or higher-order factor models,
and you can introduce cross-loadings for specific indicators. With multiple
groups of respondents, you can check for equivalence of factor stuctures or
factor loadings across groups.

However, even at at the basic level, you have much more control in CFA. Thus
you can omit an 'offending' item, and see what effect it has on model fit
statistics, while keeping the factor structure the same. You can't do that
in EFA, as the factor structure you recover may well be altered in
unpredictable ways when you change the item set. Testing competing facor
models can be done with CFA on a statistical basis, and that is outside the
scope of EFA.

Garry Gelade
Business Analytic Ltd.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Swank, Paul R
Sent: 03 October 2008 15:17
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

You should try this question on the SEMNET listserve but I will take a stab
at it. EFA will produce a set of factors but it is relatively easy, after
the fact to see how he factors support you hypotheses. However, you have no
measure of how well your theory is modeled. With CFA, you are forced a
priori to specify a structure and then have a clear test (more or less) of
whether or not the data supports the theoretical structure. EFA is also very
loose in that often, if you apply a CFA to a model indicated by an EFA, even
on the same data, it will not fit. This is because EFA allows loadings of
indicators on factors that they don't belong to whereas CFA does not. If CFA
supports the model then you have a much stronger argument for the validity
of that model.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Humphrey Paulie
Sent: Friday, October 03, 2008 5:33 AM
To: [hidden email]
Subject: Confirmatory vs Exploratory Factor Analysis

Dear list members,
Though, I understand the difference between CFA and EFA in principal, I have
problems understanding the logic of CFA as I think EFA is enough.
In CFA we have an idea of the number of factors that the test is made up of
and the loadings of the tests on each factor. Then we test this hypothesis
by means of CFA to find out whether our hypothesized factor structure fits
or not (is that right?).
What?s wrong with EFA under such a situation?
If we perform EFA and we get the hypothesized factor structure that we
expected, then we have some evidence for our theory. If the EFA doesn?t give
our expected factor structure then even the CFA cannot support our factor
structure either (i.e., it won?t fit).
And if CFA supports our factor structure then the hypothesized factor
structure should emerge when EFA is used afterwards.
This is the way I understand CFA & EFA.
It must be wrong since this argument makes CFA inconsequential.
Could you please make it clear for me?

Cheers
Humphrey

=======
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

__________ NOD32 3492 (20081003) Information __________

This message was checked by NOD32 antivirus system.
http://www.eset.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

Mark A Davenport MADAVENP

Re: Confirmatory vs Exploratory Factor Analysis

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

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

Humphrey Paulie

Re: Confirmatory vs Exploratory Factor Analysis

Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same factor structure should emerg in EFA. Does this always happen? And if the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]> wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

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

Sara House

Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model than the CFA solution, but remember that since EFA is data-driven, it will provide what fits best for these particular data. Chance variations in your data, measurement error, etc., could affect a particular sample, and the EFA will take those into consideration to find the best model. However, this leaves you with a model that may only fit this sample and may not generalize back to the population. You would have to collect another sample and run a CFA using your EFA solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA, and get acceptable model fit, you have support for your theory. You may never get a model with absolute perfect fit because chance is lumpy. But you'll know that you have the ability to generalize your results back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same factor structure should emerg in EFA. Does this always happen? And if the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]> wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

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

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

Mark A Davenport MADAVENP

Re: Confirmatory vs Exploratory Factor Analysis

In reply to this post by Humphrey Paulie

If you are talking about why SPSS may give one solution and AMOS or EQS
may give another:

This would depend, in part, on what initial extraction method and rotation
method you tell SPSS to use. The various extraction methods (PC, PAF,
Alpha, etc.) distribute the variance a bit differently so even these
extraction methods could lead to different results for the same data.
Therein lies an important point. In EFA (using SPSS, SAS, whatever), the
data are telling you what factors it sees in the given sample; even if you
force a limit on the number of factors, there is so much more that you
can't control. In CFA, you are telling the data what factors are there
and you have much greater control over the solution. For these reasons,
and others, I long ago gave up expecting to get identical results from EFA
and CFA.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

Humphrey Paulie <[hidden email]>
Sent by: "SPSSX(r) Discussion" <[hidden email]>
10/03/2008 01:30 PM
Please respond to
[hidden email]

To
[hidden email]
cc

Subject
Re: Confirmatory vs Exploratory Factor Analysis

Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same factor
structure should emerg in EFA. Does this always happen? And if the answer
is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

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

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

Mark A Davenport MADAVENP

Re: Confirmatory vs Exploratory Factor Analysis

In reply to this post by Humphrey Paulie

Here is something you might want to try. Find some data for which you
have a good fitting CFA model with no correlated factors and no correlated
erros (AMOS, EQS, LISREL, whatever). Use that same data to set it up an
equivalent EFA solution in SPSS. Add the line

/EXTRACTION=PC
/ROTATION=VARIMAX

and you should get the same thing as you would get with no extraction
method (PCis the default). Print the results.

Now, run the same thing, but change the extraction method to PAF
(Principle Axis Factoring). Now see if the results are not closer to the
CFA solution.

/EXTRACTION=PAF
/ROTATION=VARIMAX

One thing PAF does is it accounts for error variance whereas principle
components ignores it.

I haven't run an example for myself before sending this, but I suspect
you'll find that the PAF solution is much closer to the CFA solution.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

Humphrey Paulie <[hidden email]>
Sent by: "SPSSX(r) Discussion" <[hidden email]>
10/03/2008 01:30 PM
Please respond to
[hidden email]

To
[hidden email]
cc

Subject
Re: Confirmatory vs Exploratory Factor Analysis

Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same factor
structure should emerg in EFA. Does this always happen? And if the answer
is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

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

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

Swank, Paul R

Re: Confirmatory vs Exploratory Factor Analysis

In reply to this post by Sara House

I don't see how an EFA can generate a better fitting solution when there
is not measure of fit in CFA. How do you compare the models except by
saying this one is closer to what I think?

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Sara House
Sent: Friday, October 03, 2008 12:45 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model
than the CFA solution, but remember that since EFA is data-driven, it
will provide what fits best for these particular data. Chance
variations in your data, measurement error, etc., could affect a
particular sample, and the EFA will take those into consideration to
find the best model. However, this leaves you with a model that may
only fit this sample and may not generalize back to the population. You
would have to collect another sample and run a CFA using your EFA
solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA,
and get acceptable model fit, you have support for your theory. You may
never get a model with absolute perfect fit because chance is lumpy.
But you'll know that you have the ability to generalize your results
back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same
factor structure should emerg in EFA. Does this always happen? And if
the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

************************************************************************
************************************************************************
***************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J.
W.
Tukey (1962)

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

==========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

Brandon Paris

Re: Confirmatory vs Exploratory Factor Analysis

I'm guessing a two-stage process, where for a set of data, you run an EFA. With that EFA in hand, you then run CFA of the resulting relationships. The CFA results can then be compared to other models. As has been previously mentioned, the fit should be high since the EFA and resulting CFA scores is based on the data gathered.

Of course, additional datasets to split the confounding would be helpful(i.e. build EFA using one dataset, test in CFA with second dataset -- but who ever has that option).

Thanks,
Brandon Paris
CI Manager, General Mills

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Swank, Paul R
Sent: Friday, October 03, 2008 2:33 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

I don't see how an EFA can generate a better fitting solution when there
is not measure of fit in CFA. How do you compare the models except by
saying this one is closer to what I think?

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Sara House
Sent: Friday, October 03, 2008 12:45 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model
than the CFA solution, but remember that since EFA is data-driven, it
will provide what fits best for these particular data. Chance
variations in your data, measurement error, etc., could affect a
particular sample, and the EFA will take those into consideration to
find the best model. However, this leaves you with a model that may
only fit this sample and may not generalize back to the population. You
would have to collect another sample and run a CFA using your EFA
solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA,
and get acceptable model fit, you have support for your theory. You may
never get a model with absolute perfect fit because chance is lumpy.
But you'll know that you have the ability to generalize your results
back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same
factor structure should emerg in EFA. Does this always happen? And if
the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

************************************************************************
************************************************************************
***************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J.
W.
Tukey (1962)

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

==========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

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

Swank, Paul R

Re: Confirmatory vs Exploratory Factor Analysis

However, the CFA solution of the EFA model is not the same as the EFA
model. In fact the CFA model is much more rigorous and often finds
significant lack of fit for EFA generated models since the EFA allows
small loadings on other factors besides the one each indicator loads on.
Thus, I still do not believe there is any way to compare the fit of a
CFA and EFA model.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: Brandon Paris [mailto:[hidden email]]
Sent: Monday, October 06, 2008 10:58 AM
To: Swank, Paul R; [hidden email]
Subject: RE: Confirmatory vs Exploratory Factor Analysis

I'm guessing a two-stage process, where for a set of data, you run an
EFA. With that EFA in hand, you then run CFA of the resulting
relationships. The CFA results can then be compared to other models.
As has been previously mentioned, the fit should be high since the EFA
and resulting CFA scores is based on the data gathered.

Of course, additional datasets to split the confounding would be
helpful(i.e. build EFA using one dataset, test in CFA with second
dataset -- but who ever has that option).

Thanks,
Brandon Paris
CI Manager, General Mills

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Swank, Paul R
Sent: Friday, October 03, 2008 2:33 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

I don't see how an EFA can generate a better fitting solution when there
is not measure of fit in CFA. How do you compare the models except by
saying this one is closer to what I think?

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Sara House
Sent: Friday, October 03, 2008 12:45 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model
than the CFA solution, but remember that since EFA is data-driven, it
will provide what fits best for these particular data. Chance
variations in your data, measurement error, etc., could affect a
particular sample, and the EFA will take those into consideration to
find the best model. However, this leaves you with a model that may
only fit this sample and may not generalize back to the population. You
would have to collect another sample and run a CFA using your EFA
solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA,
and get acceptable model fit, you have support for your theory. You may
never get a model with absolute perfect fit because chance is lumpy.
But you'll know that you have the ability to generalize your results
back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same
factor structure should emerg in EFA. Does this always happen? And if
the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

************************************************************************
************************************************************************
***************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J.
W.
Tukey (1962)

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

==========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

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

Brandon Paris

Re: Confirmatory vs Exploratory Factor Analysis

I agree, although I have to admit that I don't have enough experience to say CFA often finds sig lack of fit for EFA models (makes sense, though).

But when I think about implications coming from a typical EFA, I think most people ignore the small loadings as trivial and equal to zero. So net-net, they are hypothesizing a simpler model -- the EFA minus the trivial connections.

IMHO, taking that belief and testing it in CFA (again with new data, not the same data), reflects the scientific process. If poor fit results at the CFA, then the hypothesis developed needs to be revised and retested.

Thanks,
Brandon
[hidden email]

-----Original Message-----
From: Swank, Paul R [mailto:[hidden email]]
Sent: Monday, October 06, 2008 11:16 AM
To: Brandon Paris; [hidden email]
Subject: RE: Confirmatory vs Exploratory Factor Analysis

However, the CFA solution of the EFA model is not the same as the EFA
model. In fact the CFA model is much more rigorous and often finds
significant lack of fit for EFA generated models since the EFA allows
small loadings on other factors besides the one each indicator loads on.
Thus, I still do not believe there is any way to compare the fit of a
CFA and EFA model.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: Brandon Paris [mailto:[hidden email]]
Sent: Monday, October 06, 2008 10:58 AM
To: Swank, Paul R; [hidden email]
Subject: RE: Confirmatory vs Exploratory Factor Analysis

I'm guessing a two-stage process, where for a set of data, you run an
EFA. With that EFA in hand, you then run CFA of the resulting
relationships. The CFA results can then be compared to other models.
As has been previously mentioned, the fit should be high since the EFA
and resulting CFA scores is based on the data gathered.

Of course, additional datasets to split the confounding would be
helpful(i.e. build EFA using one dataset, test in CFA with second
dataset -- but who ever has that option).

Thanks,
Brandon Paris
CI Manager, General Mills

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Swank, Paul R
Sent: Friday, October 03, 2008 2:33 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

I don't see how an EFA can generate a better fitting solution when there
is not measure of fit in CFA. How do you compare the models except by
saying this one is closer to what I think?

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Sara House
Sent: Friday, October 03, 2008 12:45 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model
than the CFA solution, but remember that since EFA is data-driven, it
will provide what fits best for these particular data. Chance
variations in your data, measurement error, etc., could affect a
particular sample, and the EFA will take those into consideration to
find the best model. However, this leaves you with a model that may
only fit this sample and may not generalize back to the population. You
would have to collect another sample and run a CFA using your EFA
solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA,
and get acceptable model fit, you have support for your theory. You may
never get a model with absolute perfect fit because chance is lumpy.
But you'll know that you have the ability to generalize your results
back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same
factor structure should emerg in EFA. Does this always happen? And if
the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

************************************************************************
************************************************************************
***************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J.
W.
Tukey (1962)

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

==========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

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

Mark A Davenport MADAVENP

Re: Confirmatory vs Exploratory Factor Analysis

As some have alluded to, there are some deeper philosphical issues at work
here. The purpose of EFA is to find relationships between items in the
dataset, When these occur, one might argue that these relationships may
reflect underlying latent factors. Generally, most would not (or should
not) presume that these factors hold much water outside of the immediate
sample. The purpose of CFA, is to confirm that the pattern of latent
factors that are believed to exist in the population are also apparent in
your sample (assumed to represent that population). These really are two
different approaches to understanding the nature of the relationship
between items in a dataset. In EFA, you are not fitting data to a model,
you are teasing a model out of the data (and I use the term 'model' VERY
loosely here). This hole idea of fitting data in EFA just doesn't make
sense. This is also complicated by the fact that the default extraction
method in EFA is principle components, which lumps shared and unique
variance together. The structure of CFA is really much more consistent
with principal axis factoring or PAF, also available in SPSS, where shared
variance is treated apart from unique (error) variance.

Mark

***************************************************************************************************************************************************************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J. W.
Tukey (1962)

Brandon Paris <[hidden email]>
Sent by: "SPSSX(r) Discussion" <[hidden email]>
10/06/2008 01:09 PM
Please respond to
Brandon Paris <[hidden email]>

To
[hidden email]
cc

Subject
Re: Confirmatory vs Exploratory Factor Analysis

I agree, although I have to admit that I don't have enough experience to
say CFA often finds sig lack of fit for EFA models (makes sense, though).

But when I think about implications coming from a typical EFA, I think
most people ignore the small loadings as trivial and equal to zero. So
net-net, they are hypothesizing a simpler model -- the EFA minus the
trivial connections.

IMHO, taking that belief and testing it in CFA (again with new data, not
the same data), reflects the scientific process. If poor fit results at
the CFA, then the hypothesis developed needs to be revised and retested.

Thanks,
Brandon
[hidden email]

-----Original Message-----
From: Swank, Paul R [mailto:[hidden email]]
Sent: Monday, October 06, 2008 11:16 AM
To: Brandon Paris; [hidden email]
Subject: RE: Confirmatory vs Exploratory Factor Analysis

However, the CFA solution of the EFA model is not the same as the EFA
model. In fact the CFA model is much more rigorous and often finds
significant lack of fit for EFA generated models since the EFA allows
small loadings on other factors besides the one each indicator loads on.
Thus, I still do not believe there is any way to compare the fit of a
CFA and EFA model.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: Brandon Paris [mailto:[hidden email]]
Sent: Monday, October 06, 2008 10:58 AM
To: Swank, Paul R; [hidden email]
Subject: RE: Confirmatory vs Exploratory Factor Analysis

I'm guessing a two-stage process, where for a set of data, you run an
EFA. With that EFA in hand, you then run CFA of the resulting
relationships. The CFA results can then be compared to other models.
As has been previously mentioned, the fit should be high since the EFA
and resulting CFA scores is based on the data gathered.

Of course, additional datasets to split the confounding would be
helpful(i.e. build EFA using one dataset, test in CFA with second
dataset -- but who ever has that option).

Thanks,
Brandon Paris
CI Manager, General Mills

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Swank, Paul R
Sent: Friday, October 03, 2008 2:33 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

I don't see how an EFA can generate a better fitting solution when there
is not measure of fit in CFA. How do you compare the models except by
saying this one is closer to what I think?

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Sara House
Sent: Friday, October 03, 2008 12:45 PM
To: [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

It is possible that an EFA solution would provide a better-fitting model
than the CFA solution, but remember that since EFA is data-driven, it
will provide what fits best for these particular data. Chance
variations in your data, measurement error, etc., could affect a
particular sample, and the EFA will take those into consideration to
find the best model. However, this leaves you with a model that may
only fit this sample and may not generalize back to the population. You
would have to collect another sample and run a CFA using your EFA
solution to see if your model still fits.

If, on the other hand, you create a theory-driven model, perform a CFA,
and get acceptable model fit, you have support for your theory. You may
never get a model with absolute perfect fit because chance is lumpy.
But you'll know that you have the ability to generalize your results
back to a population.

Sara

Sara M. House, M.A.
Adjunct Faculty
Loyola University Chicago, Psychology Department
Email: [hidden email]
Teaching: Research Methods, Psychology & Law

>>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
Thanks for your replies,
If a specified factor structure fits in CFA, then I expect the same
factor structure should emerg in EFA. Does this always happen? And if
the answer is no, doesnt this invalidate CFA?
Humphrey

--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
wrote:

From: Mark A Davenport MADAVENP <[hidden email]>
Subject: Re: Confirmatory vs Exploratory Factor Analysis
To: [hidden email]
Date: Friday, October 3, 2008, 11:04 AM

Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
the name). Any addition to the model (correlated errors, etc.) may make
practical sense, but in a technical sense, is evidence that the theory
does not accurately describe what the sample is showing you. In my
experience, I have rarely been able to get EFA to produce an 'expected'
factor structure. That's really not it's purpose. It's far more
general
in it's application. Yet, I can often force a CFA model to fit an
expected structure by tweaking the diagram. It seems like every year a
new way to assess model fit is released. If you are going to use CFA
properly, you REALLY need to keep up with the literature.

************************************************************************
************************************************************************
***************
Mark A. Davenport Ph.D.
Senior Research Analyst
Office of Institutional Research
The University of North Carolina at Greensboro
336.256.0395
[hidden email]

'An approximate answer to the right question is worth a good deal more
than an exact answer to an approximate question.' --a paraphrase of J.
W.
Tukey (1962)

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

==========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

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

Bob Schacht-3

Re: Confirmatory vs Exploratory Factor Analysis

In reply to this post by Brandon Paris

At 07:07 AM 10/6/2008, Brandon Paris wrote:

>I agree, although I have to admit that I don't have enough experience to
>say CFA often finds sig lack of fit for EFA models (makes sense, though).
>
>But when I think about implications coming from a typical EFA, I think
>most people ignore the small loadings as trivial and equal to zero. So
>net-net, they are hypothesizing a simpler model -- the EFA minus the
>trivial connections.
>
>IMHO, taking that belief and testing it in CFA (again with new data, not
>the same data), reflects the scientific process. If poor fit results at
>the CFA, then the hypothesis developed needs to be revised and retested.
>
>Thanks,
>Brandon
>[hidden email]

Brandon,
Aren't you assuming that the EFA results in factors that are readily
explained in terms of an external theory, which is what CFA is supposed to
test?
If you just go from one to the other, without paying attention to any
theoretical context (and I am not referring to statistical theory here, I'm
referring to theory about the original variables and how they are seen to
relate to each other), doesn't it all just amount to statistical
mumbo-jumbo and pushing numbers around?

In other words, I guess that I am saying that yes, I suppose that you can
do as you propose, but what does it all mean, if anything?

P.S. Now that I've read Mark's response, I agree with his position, which
is much more elegantly stated than mine.

Bob Schacht

>-----Original Message-----
>From: Swank, Paul R [mailto:[hidden email]]
>Sent: Monday, October 06, 2008 11:16 AM
>To: Brandon Paris; [hidden email]
>Subject: RE: Confirmatory vs Exploratory Factor Analysis
>
>However, the CFA solution of the EFA model is not the same as the EFA
>model. In fact the CFA model is much more rigorous and often finds
>significant lack of fit for EFA generated models since the EFA allows
>small loadings on other factors besides the one each indicator loads on.
>Thus, I still do not believe there is any way to compare the fit of a
>CFA and EFA model.
>
>Paul R. Swank, Ph.D
>Professor and Director of Research
>Children's Learning Institute
>University of Texas Health Science Center
>Houston, TX 77038
>
>-----Original Message-----
>From: Brandon Paris [mailto:[hidden email]]
>Sent: Monday, October 06, 2008 10:58 AM
>To: Swank, Paul R; [hidden email]
>Subject: RE: Confirmatory vs Exploratory Factor Analysis
>
>I'm guessing a two-stage process, where for a set of data, you run an
>EFA. With that EFA in hand, you then run CFA of the resulting
>relationships. The CFA results can then be compared to other models.
>As has been previously mentioned, the fit should be high since the EFA
>and resulting CFA scores is based on the data gathered.
>
>Of course, additional datasets to split the confounding would be
>helpful(i.e. build EFA using one dataset, test in CFA with second
>dataset -- but who ever has that option).
>
>Thanks,
>Brandon Paris
>CI Manager, General Mills
>
>-----Original Message-----
>From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
>Swank, Paul R
>Sent: Friday, October 03, 2008 2:33 PM
>To: [hidden email]
>Subject: Re: Confirmatory vs Exploratory Factor Analysis
>
>I don't see how an EFA can generate a better fitting solution when there
>is not measure of fit in CFA. How do you compare the models except by
>saying this one is closer to what I think?
>
>Paul R. Swank, Ph.D
>Professor and Director of Research
>Children's Learning Institute
>University of Texas Health Science Center
>Houston, TX 77038
>
>
>-----Original Message-----
>From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
>Sara House
>Sent: Friday, October 03, 2008 12:45 PM
>To: [hidden email]
>Subject: Re: Confirmatory vs Exploratory Factor Analysis
>
>It is possible that an EFA solution would provide a better-fitting model
>than the CFA solution, but remember that since EFA is data-driven, it
>will provide what fits best for these particular data. Chance
>variations in your data, measurement error, etc., could affect a
>particular sample, and the EFA will take those into consideration to
>find the best model. However, this leaves you with a model that may
>only fit this sample and may not generalize back to the population. You
>would have to collect another sample and run a CFA using your EFA
>solution to see if your model still fits.
>
>If, on the other hand, you create a theory-driven model, perform a CFA,
>and get acceptable model fit, you have support for your theory. You may
>never get a model with absolute perfect fit because chance is lumpy.
>But you'll know that you have the ability to generalize your results
>back to a population.
>
>Sara
>
>Sara M. House, M.A.
>Adjunct Faculty
>Loyola University Chicago, Psychology Department
>Email: [hidden email]
>Teaching: Research Methods, Psychology & Law
>
> >>> Humphrey Paulie <[hidden email]> 10/3/2008 12:25 PM >>>
>Thanks for your replies,
>If a specified factor structure fits in CFA, then I expect the same
>factor structure should emerg in EFA. Does this always happen? And if
>the answer is no, doesnt this invalidate CFA?
>Humphrey
>
>--- On Fri, 10/3/08, Mark A Davenport MADAVENP <[hidden email]>
>wrote:
>
>From: Mark A Davenport MADAVENP <[hidden email]>
>Subject: Re: Confirmatory vs Exploratory Factor Analysis
>To: [hidden email]
>Date: Friday, October 3, 2008, 11:04 AM
>
>Keep in mind that CFA is theory driven, EFA is not necessarily so (hence
>the name). Any addition to the model (correlated errors, etc.) may make
>practical sense, but in a technical sense, is evidence that the theory
>does not accurately describe what the sample is showing you. In my
>experience, I have rarely been able to get EFA to produce an 'expected'
>factor structure. That's really not it's purpose. It's far more
>general
>in it's application. Yet, I can often force a CFA model to fit an
>expected structure by tweaking the diagram. It seems like every year a
>new way to assess model fit is released. If you are going to use CFA
>properly, you REALLY need to keep up with the literature.
>
>************************************************************************
>************************************************************************
>***************
>Mark A. Davenport Ph.D.
>Senior Research Analyst
>Office of Institutional Research
>The University of North Carolina at Greensboro
>336.256.0395
>[hidden email]
>
>'An approximate answer to the right question is worth a good deal more
>than an exact answer to an approximate question.' --a paraphrase of J.
>W.
>Tukey (1962)
>
>=====================
>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
>
>==========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
>
>=====================
>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

Robert M. Schacht, Ph.D. <[hidden email]>
Pacific Basin Rehabilitation Research & Training Center
1268 Young Street, Suite #204
Research Center, University of Hawaii
Honolulu, HI 96814

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

Brandon Paris

Re: Confirmatory vs Exploratory Factor Analysis

Hi --

Let me start by saying I agree with all of the comments here; I was just adding my two cents to a two dollar conversation (so in the end, my thoughts probably aren't worth that much).

Also, I always try to be brief in my emails, but I want to clarify myself at this time. I apologize for the length.

Bob Schacht stated:
If you just go from one to the other, without paying attention to any
theoretical context (and I am not referring to statistical theory here, I'm
referring to theory about the original variables and how they are seen to
relate to each other), doesn't it all just amount to statistical
mumbo-jumbo and pushing numbers around?
----*snip*----

Yes. I naively assumed that the "grounding in theory" step is always being considered. In market research practitioner speak, the first litmus test for EFA analyses (other approaches as well) is often a "face validity" judgement -- an attempt to ground the results in some idea of what I "should expect" (more on this shortly).

But how are theories themselves developed? By observing a pattern in the world, proposing a means for why it occurs, and then testing to see if that "answer" consistently holds across various streams of new observations.

So if I am starting from the beginning, either with no underlying theory to utilize (or, more likely, a theory exists and I am unaware of it), then why can't we gather some data and analyze it in some way? EFA facilitates that in the factor analytic realm. And based on those results, I can build a hypothesis, then test it, evaluate it, refine it, retest, etc.

To conclude, let me say that I work in Market Research, and I feel that, unlike fields such as Psychology, the Marketing discipline in general has considerable gaps in established social scientific theory. Often we are tackling business questions without available theory to rely upon; as much as I hate to say it, we rely on "face validity" too much.

Thanks,
Brandon Paris

-----Original Message-----
From: Bob Schacht [mailto:[hidden email]]
Sent: Monday, October 06, 2008 2:23 PM
To: Brandon Paris; [hidden email]
Subject: Re: Confirmatory vs Exploratory Factor Analysis

At 07:07 AM 10/6/2008, Brandon Paris wrote: