COvaraince not Positive definite for CFA

[cicilia]

Hello All

Been through all the msgs for this NPD problem Tried the different soultions. I have N=200, 14 varaibles for a CFA. They go into 4 LV's. Everything seems ok but I get NPD msg. Any suggestions?

[David Marso]

So what have you *actually* tried relative to the various solutions.
I doubt that it is possible for anyone to assist you to sort this without access to the raw data or at least the covariance matrix and the CFA model description.

cicilia wrote

Hello All

Been through all the msgs for this NPD problem Tried the different soultions. I have N=200, 14 varaibles for a CFA. They go into 4 LV's. Everything seems ok but I get NPD msg. Any suggestions?

[Art Kendall]

I did not see the original message but have you checked out the archive of this list for discussion of ordinary exploratory factor analysis?
Some variable is highly predictable from some other variable(s). Common reasons for this are
-- a variable name is actually included twice.  Did an item accidentally get included in the scoring key for more than 1 scale and you cut-and-pasted those list from other syntax?
-- some pair of items have (close to) a perfect correlation.
-- some item has a very high squared multiple correlation with other items.
-- more items than cases [but you do not have this unless there is an extreme missing data situation.]

After  eyeballing the variable list for more than one occurrence of a variable name, a quick and dirty way to double check is to paste your variable list from the factor analysis into RELIABILITY. first make the 4 scales and then a scale from all 14 items.
Look at the statistics about the correlations.  Is what is the highest/lowest interitem correlation?
Look  4 scales.
If you still have not found where the redundancy came from look at the corrected item correlations and squared multiple correlations for the 14 item scale.

if you still have not tracked down the redundancy go to the archives and look for the articles that deal with regression diagnostics for multicollinearity.

On 5/30/2012 6:00 AM, David Marso wrote:

So what have you *actually* tried relative to the various solutions.
I doubt that it is possible for anyone to assist you to sort this without
access to the raw data or at least the covariance matrix and the CFA model
description.

cicilia wrote

Hello All

Been through all the msgs for this NPD problem Tried the different
soultions. I have N=200, 14 varaibles for a CFA. They go into 4 LV's.
Everything seems ok but I get NPD msg. Any suggestions?

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/COvaraince-not-Positive-definite-for-CFA-tp5713427p5713428.html
Sent from the SPSSX Discussion mailing list archive at Nabble.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

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

[Poes, Matthew Joseph]

The problem is you have a data issue that is causing your 4 latent variables to over-tax the 14 variables that make them up.  I don't know enough about these 14 variables, but you have some set of problems amongst one or all of them that is making it too difficult to extract the 4 latent variables.  I would be sure there isn't a lack of variance, a constant, anything of that sort.  I would also consider going back to the EFA and see if a simpler model worked as well, maybe a 2 or 3 LV model.  Then test that.

I've heard different arguments on sample size for CFA.  I've heard everything from 30 per LV with all variables meeting assumptions, 100 if they don't fully.  I've heard 10 per manifest variable.  I've heard 30 per manifest variable.  There doesn't seem to be clear consensus on this, but your situation may indicate a power issue.

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: [phone]
email: [hidden email]


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of cicilia
Sent: Wednesday, May 30, 2012 4:27 AM
To: [hidden email]
Subject: COvaraince not Positive definite for CFA

Hello All

Been through all the msgs for this NPD problem Tried the different soultions. I have N=200, 14 varaibles for a CFA. They go into 4 LV's.
Everything seems ok but I get NPD msg. Any suggestions?

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/COvaraince-not-Positive-definite-for-CFA-tp5713427.html
Sent from the SPSSX Discussion mailing list archive at Nabble.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

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

[Rich Ulrich]

Complete data: analyzing r or covariances --
If you have the same variable twice, or some other
redundancy (subtotals and total?), the matrix will
have zero as the determinant.  Not "negative", but
not positive, either.

So, is the determinant negative?  That  would indicate
that your correlations are inconsistent, impossible to
be achieved for a set of complete data.  You might get
that by analyzing incomplete data with the option of
"pairwise correlations", when the data (and missing)
exist such that that a multiple correlation that would
be near 1.0 for complete data becomes, by formula,
greater than 1.0.  (There are also limits shown by the
formulas for partial correlation, but that is less likely
to be relevant for PCA.)

--
Rich Ulrich

> Date: Wed, 30 May 2012 02:27:06 -0700
> From: [hidden email]
> Subject: COvaraince not Positive definite for CFA
> To: [hidden email]
>
> Hello All
>
> Been through all the msgs for this NPD problem Tried the different
> soultions. I have N=200, 14 varaibles for a CFA. They go into 4 LV's.
> Everything seems ok but I get NPD msg. Any suggestions?
 ...

[cicilia]

I am posting the data for the 14 variables here. These are the agg scores of the subscales and this is what I was trying to analyze. Am unable to get the original EFA as this is a scale that I havent developed. I need to do the CFA to establish validity and reliability and to check for equivalence across cultures.

Heres the correct list of 14 variables.....had uploaded the wrong file. Apologies to all.....pls do go through this one. There are 7 subscales here.GOALS.sav

Thanks in advance.

[David Marso]

Well the data are of full rank so it must be model specific.
--

cicilia wrote

I am posting the data for the 14 variables here. These are the agg scores of the subscales and this is what I was trying to analyze. Am unable to get the original EFA as this is a scale that I havent developed. I need to do the CFA to establish validity and reliability and to check for equivalence across cultures.variables.sav 

Thanks in advance.

[cicilia]

Errrmmmm...that would mean????

[Art Kendall]

I received a "file not found" when I clicked on the link.

What syntax are you using for the CFA?

What happened with the RELIABILITY run? 
Did you do an EFA?  Did that come up with the same problem? If not does the resulting scoring key correspond to the one from the earlier study?

On 5/31/2012 5:00 AM, cicilia wrote:

I am posting the data for the 14 variables here. These are the agg scores of
the subscales and this is what I was trying to analyze. Am unable to get the
original EFA as this is a scale that I havent developed. I need to do the
CFA to establish validity and reliability and to check for equivalence
across cultures.
http://spssx-discussion.1045642.n5.nabble.com/file/n5713444/variables.sav
variables.sav

Thanks in advance.

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/COvaraince-not-Positive-definite-for-CFA-tp5713427p5713444.html
Sent from the SPSSX Discussion mailing list archive at Nabble.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

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

[David Marso]

Google is your friend!
http://en.wikipedia.org/wiki/Rank_%28linear_algebra%29
Having a little matrix algebra background is essential to doing any sort of stats modeling ie SEM CFA etc?

cicilia wrote

Errrmmmm...that would mean????