SB Chi square value

[Ravindra Nath]

Dear All,

I am facing a peculiar problem. I have estimated a SEM using asymptotic
covariances through the ML method. I get a SB Chi square value of 0 with a P
value of 1. Furthermore all my standardized residuals are 0. However, the
standard errors of my estimates a very large. Does this suggest that the
covariances among variables that I am trying to model are statistically
insignificant, and there is no point in modelling them?

I have attached LISREL's Goodness of Fit Statistics below. I will
appreciate any help.
Degrees of Freedom = 228
Minimum Fit Function Chi-Square = 3459.80 (P = 0.0)
Normal Theory Weighted Least Squares Chi-Square = 3494.14 (P = 0.0)
Satorra-Bentler Scaled Chi-Square = 0.00 (P = 1.00)
Chi-Square Corrected for Non-Normality = 2023252861.51 (P = 0.0)
Estimated Non-centrality Parameter (NCP) = 0.0
90 Percent Confidence Interval for NCP = (0.0 ; 0.0)

Minimum Fit Function Value = 2.43
Population Discrepancy Function Value (F0) = 0.0
90 Percent Confidence Interval for F0 = (0.0 ; 0.0)
Root Mean Square Error of Approximation (RMSEA) = 0.0
90 Percent Confidence Interval for RMSEA = (0.0 ; 0.0)
P-Value for Test of Close Fit (RMSEA < 0.05) = 0.00

Expected Cross-Validation Index (ECVI) = 0.26
90 Percent Confidence Interval for ECVI = (0.26 ; 0.26)
ECVI for Saturated Model = 0.42
ECVI for Independence Model = 86.34

Chi-Square for Independence Model with 276 Degrees of Freedom =
122904.00
Independence AIC = 122952.00
Model AIC = 3638.14
Saturated AIC = 600.00
Independence CAIC = 123102.29
Model CAIC = 4089.00
Saturated CAIC = 2478.58

Normed Fit Index (NFI) = 1.00
Non-Normed Fit Index (NNFI) = 1.00
Parsimony Normed Fit Index (PNFI) = 0.83
Comparative Fit Index (CFI) = 1.00
Incremental Fit Index (IFI) = 1.00
Relative Fit Index (RFI) = 1.00

Critical N (CN) = ********************


Root Mean Square Residual (RMR) = 1.09
Standardized RMR = 0.13
Goodness of Fit Index (GFI) = 0.83
Adjusted Goodness of Fit Index (AGFI) = 0.78
Parsimony Goodness of Fit Index (PGFI) = 0.63


Regards,
Ravi

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Ed Rigdon]

Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
Dear All,

I am facing a peculiar problem. I have estimated a SEM using
asymptotic
covariances through the ML method. I get a SB Chi square value of 0
with a P
value of 1. Furthermore all my standardized residuals are 0. However,
the
standard errors of my estimates a very large. Does this suggest that
the
covariances among variables that I am trying to model are
statistically
insignificant, and there is no point in modelling them?

I have attached LISREL's Goodness of Fit Statistics below. I will
appreciate any help.
Degrees of Freedom = 228
Minimum Fit Function Chi-Square = 3459.80 (P = 0.0)
Normal Theory Weighted Least Squares Chi-Square = 3494.14 (P =
0.0)
Satorra-Bentler Scaled Chi-Square = 0.00 (P = 1.00)
Chi-Square Corrected for Non-Normality = 2023252861.51 (P =
0.0)
Estimated Non-centrality Parameter (NCP) = 0.0
90 Percent Confidence Interval for NCP = (0.0 ; 0.0)

Minimum Fit Function Value = 2.43
Population Discrepancy Function Value (F0) = 0.0
90 Percent Confidence Interval for F0 = (0.0 ; 0.0)
Root Mean Square Error of Approximation (RMSEA) = 0.0
90 Percent Confidence Interval for RMSEA = (0.0 ; 0.0)
P-Value for Test of Close Fit (RMSEA < 0.05) = 0.00

Expected Cross-Validation Index (ECVI) = 0.26
90 Percent Confidence Interval for ECVI = (0.26 ; 0.26)
ECVI for Saturated Model = 0.42
ECVI for Independence Model = 86.34

Chi-Square for Independence Model with 276 Degrees of Freedom =
122904.00
Independence AIC = 122952.00
Model AIC = 3638.14
Saturated AIC = 600.00
Independence CAIC = 123102.29
Model CAIC = 4089.00
Saturated CAIC = 2478.58

Normed Fit Index (NFI) = 1.00
Non-Normed Fit Index (NNFI) = 1.00
Parsimony Normed Fit Index (PNFI) = 0.83
Comparative Fit Index (CFI) = 1.00
Incremental Fit Index (IFI) = 1.00
Relative Fit Index (RFI) = 1.00

Critical N (CN) = ********************


Root Mean Square Residual (RMR) = 1.09
Standardized RMR = 0.13
Goodness of Fit Index (GFI) = 0.83
Adjusted Goodness of Fit Index (AGFI) = 0.78
Parsimony Goodness of Fit Index (PGFI) = 0.63


Regards,
Ravi

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Ravindra Nath]

Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
oneother problem I have faced in the past:
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO (continuous)
or OR (ordinal) depending on their nature.

Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.

Regards,

Ravi

Post by Ed Rigdon
Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon
Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Cameron McIntosh]

Hi Ed and Ravi,

Ed - I had thought that the fact that the ACM is not inverted with S-B scaling would guard against such a problem. It seems to work well at small sample sizes (i.e., less than 200), but then again the simulations I've read about usually don't involve huge numbers of observed variables. With too low N in such a case, you still won't get a stable enough estimate of the ACM to work properly even with the S-B procedure? Would that explain the crazy fit output that Ravi got?

Ravi - How many measured variables do you have and what is your sample size? By the way, it's fine if the program treats some variables as continuous and some as ordinal - this is not an error, actually, this is the way it should be done. It just means that the input matrix is a combination of different types of correlations that respect the different categorizations of the variables. Then the information from the ACM (the covariance matrix of the correlations) is incorporated into the estimation method to obtain correct fit statistics and standard errors. Of course, the correctness depends on meeting the conditions required for using the alternative estimation method.

Hope this helps,

Cam


Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney's Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]

-----Original Message-----
From: Structural Equation Modeling Discussion Group [mailto:***@BAMA.UA.EDU] On Behalf Of Ravindra Nath
Sent: September 25, 2006 10:24 AM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value


Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in some of my other analysis as well, I though I could share with the group oneother problem I have faced in the past:
While specifying an asymptotic covariance matrix through PRELIS, the asymptotic covariances might be misleading incase the program treats some of the variables as continuous and some others as ordinal. This is specially true if the analysis involves a likert like scale and the researcher has rescaled the variables. Its best to specify the variables as CO (continuous) or OR (ordinal) depending on their nature.

Another issue that I have experienced while attempting SB is that the fitted covariance matrix is not positive definite.

Regards,

Ravi


On 9/25/06, Ed Rigdon <***@langate.gsu.edu> wrote:

Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of the message as: SIGNOFF SEMNET Search the archives at http://bama.ua.edu/archives/semnet.html


--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Gabriel Dy-Liacco]

Hi Cam, Ed, Ravi,



I sometimes get output that looks similar to Ravi’s even with an adequate
sample size and also get the non-positive definite warning. Another person
on the listserv who uses LISREL suggested sometime ago that PRELIS sometimes
would create matrix files with nonsense characters in them and this would
result in weird output. The remedy was to delete all of the files associated
with the particular output and start over again. So far, this has worked for
me.



HTH,



Gabriel





_____

From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Cameron McIntosh
Sent: Monday, September 25, 2006 1:22 PM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value



Hi Ed and Ravi,



Ed - I had thought that the fact that the ACM is not inverted with S-B
scaling would guard against such a problem. It seems to work well at small
sample sizes (i.e., less than 200), but then again the simulations I've read
about usually don't involve huge numbers of observed variables. With too low
N in such a case, you still won't get a stable enough estimate of the ACM to
work properly even with the S-B procedure? Would that explain the crazy fit
output that Ravi got?



Ravi - How many measured variables do you have and what is your sample size?
By the way, it's fine if the program treats some variables as continuous and
some as ordinal - this is not an error, actually, this is the way it should
be done. It just means that the input matrix is a combination of different
types of correlations that respect the different categorizations of the
variables. Then the information from the ACM (the covariance matrix of the
correlations) is incorporated into the estimation method to obtain correct
fit statistics and standard errors. Of course, the correctness depends on
meeting the conditions required for using the alternative estimation method.



Hope this helps,



Cam



Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la
santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney’s Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]

-----Original Message-----
From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Ravindra Nath
Sent: September 25, 2006 10:24 AM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value

Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
oneother problem I have faced in the past:
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO (continuous)
or OR (ordinal) depending on their nature.

Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.

Regards,

Ravi

On 9/25/06, Ed Rigdon <***@langate.gsu.edu> wrote:

Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Ed Rigdon]

Gabriel--

That's a very good point. LISREL is still a program where
rebooting LISREL or your computer, or cleaning out old versions
of files, sometimes eliminates odd results. If you think you have
fixed a problem, but it doesn't go away, this is something to try.

--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
Hi Cam, Ed, Ravi,



I sometimes get output that looks similar to Ravi’s even with an adequate
sample size and also get the non-positive definite warning. Another person
on the listserv who uses LISREL suggested sometime ago that PRELIS sometimes
would create matrix files with nonsense characters in them and this would
result in weird output. The remedy was to delete all of the files associated
with the particular output and start over again. So far, this has worked for
me.



HTH,



Gabriel





_____

From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Cameron McIntosh
Sent: Monday, September 25, 2006 1:22 PM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value



Hi Ed and Ravi,



Ed - I had thought that the fact that the ACM is not inverted with S-B
scaling would guard against such a problem. It seems to work well at small
sample sizes (i.e., less than 200), but then again the simulations I've read
about usually don't involve huge numbers of observed variables. With too low
N in such a case, you still won't get a stable enough estimate of the ACM to
work properly even with the S-B procedure? Would that explain the crazy fit
output that Ravi got?



Ravi - How many measured variables do you have and what is your sample size?
By the way, it's fine if the program treats some variables as continuous and
some as ordinal - this is not an error, actually, this is the way it should
be done. It just means that the input matrix is a combination of different
types of correlations that respect the different categorizations of the
variables. Then the information from the ACM (the covariance matrix of the
correlations) is incorporated into the estimation method to obtain correct
fit statistics and standard errors. Of course, the correctness depends on
meeting the conditions required for using the alternative estimation method.



Hope this helps,



Cam



Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la
santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney’s Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]

-----Original Message-----
From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Ravindra Nath
Sent: September 25, 2006 10:24 AM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value

Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
oneother problem I have faced in the past:
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO (continuous)
or OR (ordinal) depending on their nature.

Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.

Regards,

Ravi

On 9/25/06, Ed Rigdon <***@langate.gsu.edu> wrote:

Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Gabriel Dy-Liacco]

Ed, Ravi,

You're welcome and glad it worked. Incidentally, 3 other people I know have
recently experienced similar problems with the LISREL 8.72 package. I'll let
SSI know.

Gabriel

Gabriel S. Dy-Liacco, Ph.D.

Post by Cameron McIntosh
-----Original Message-----
From: Structural Equation Modeling Discussion Group
Sent: Monday, September 25, 2006 4:51 PM
Subject: Re: SB Chi square value
Gabriel--
That's a very good point. LISREL is still a program where
rebooting LISREL or your computer, or cleaning out old versions
of files, sometimes eliminates odd results. If you think you have
fixed a problem, but it doesn't go away, this is something to try.
--Ed Rigdon
Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
Hi Cam, Ed, Ravi,
I sometimes get output that looks similar to Ravi’s even with an adequate
sample size and also get the non-positive definite warning. Another person
on the listserv who uses LISREL suggested sometime ago that PRELIS sometimes
would create matrix files with nonsense characters in them and this would
result in weird output. The remedy was to delete all of the files associated
with the particular output and start over again. So far, this has worked for
me.
HTH,
Gabriel
_____
From: Structural Equation Modeling Discussion Group
Sent: Monday, September 25, 2006 1:22 PM
Subject: Re: SB Chi square value
Hi Ed and Ravi,
Ed - I had thought that the fact that the ACM is not inverted with S-B
scaling would guard against such a problem. It seems to work well at small
sample sizes (i.e., less than 200), but then again the simulations I've read
about usually don't involve huge numbers of observed variables. With too low
N in such a case, you still won't get a stable enough estimate of the ACM to
work properly even with the S-B procedure? Would that explain the crazy fit
output that Ravi got?
Ravi - How many measured variables do you have and what is your sample size?
By the way, it's fine if the program treats some variables as continuous and
some as ordinal - this is not an error, actually, this is the way it should
be done. It just means that the input matrix is a combination of different
types of correlations that respect the different categorizations of the
variables. Then the information from the ACM (the covariance matrix of the
correlations) is incorporated into the estimation method to obtain correct
fit statistics and standard errors. Of course, the correctness depends on
meeting the conditions required for using the alternative estimation method.
Hope this helps,
Cam
Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la
santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney’s Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]
-----Original Message-----
From: Structural Equation Modeling Discussion Group
Sent: September 25, 2006 10:24 AM
Subject: Re: SB Chi square value
Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO (continuous)
or OR (ordinal) depending on their nature.
Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.
Regards,
Ravi
Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon
Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html
-------------------------------------------------------------- To
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Ravindra Nath]

Thanks Ed, Gabriel, Cam,

Rebooting LISREL helped. I was able to get SB chi square after all. Here are
the results:
Degrees of Freedom = 228
Minimum Fit Function Chi-Square = 1585.89 (P = 0.0)
Normal Theory Weighted Least Squares Chi-Square = 1465.03 (P = 0.0)
Satorra-Bentler Scaled Chi-Square = 905.62 (P = 0.0)
Chi-Square Corrected for Non-Normality = 2203.88 (P = 0.0)
Estimated Non-centrality Parameter (NCP) = 683.62
90 Percent Confidence Interval for NCP = (594.62 ; 780.16)

Minimum Fit Function Value = 1.11
Population Discrepancy Function Value (F0) = 0.48
90 Percent Confidence Interval for F0 = (0.42 ; 0.55)
Root Mean Square Error of Approximation (RMSEA) = 0.047
90 Percent Confidence Interval for RMSEA = (0.043 ; 0.050)
P-Value for Test of Close Fit (RMSEA < 0.05) = 0.00

Expected Cross-Validation Index (ECVI) = 1.14
90 Percent Confidence Interval for ECVI = (0.68 ; 0.81)
ECVI for Saturated Model = 0.42
ECVI for Independence Model = 68.50

Chi-Square for Independence Model with 276 Degrees of Freedom =
97495.15
Independence AIC = 97543.15
Model AIC = 1621.03
Saturated AIC = 600.00
Independence CAIC = 97693.44
Model CAIC = 2109.46
Saturated CAIC = 2478.58

Normed Fit Index (NFI) = 0.99
Non-Normed Fit Index (NNFI) = 0.99
Parsimony Normed Fit Index (PNFI) = 0.80
Comparative Fit Index (CFI) = 0.99
Incremental Fit Index (IFI) = 0.99
Relative Fit Index (RFI) = 0.99

Critical N (CN) = 431.74


Root Mean Square Residual (RMR) = 75.38
Standardized RMR = 0.11
Goodness of Fit Index (GFI) = 0.92
Adjusted Goodness of Fit Index (AGFI) = 0.89
Parsimony Goodness of Fit Index (PGFI) = 0.68


Cam,

I am modeling 24 manifest variables. My sample size is 800.

On your point on variable scale, I thought it to be a good practise to
instruct PRELIS on the nature of my variables, or else it treats even
ordinal variables with more than 10 scale points as continuous.

In another context, in market research, it is often a practise to rescale a
five point Likert measurement (1 = Strongly disagree, 5 = strongly agree),
by a factor of 20 (20 = Strongly disagree, 100 = strongly agree). Often,
values which would be non-existent in the original scale appear in the
rescaled dataset because of some kind of imputation (while I agree this is
not entirely scientific, but the practise has its own merits). In a dataset
of this kind, if some of my variables are treated as ordinal and others as
continuous my covariances could be misleading. In this situation, if I am
using a Psf datafile and simplis directly estimates a covariance matrix
(without me having to input one), it treats all variables as continuous. Now
if I was to use PRELIS computed ACM, which would have treated some of these
variables as ordinal, my estimnation could go wrong.

Thanks Again,
Ravi

Post by Ed Rigdon
Gabriel--
That's a very good point. LISREL is still a program where
rebooting LISREL or your computer, or cleaning out old versions
of files, sometimes eliminates odd results. If you think you have
fixed a problem, but it doesn't go away, this is something to try.
--Ed Rigdon
Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
Hi Cam, Ed, Ravi,
I sometimes get output that looks similar to Ravi's even with an adequate
sample size and also get the non-positive definite warning. Another person
on the listserv who uses LISREL suggested sometime ago that PRELIS
sometimes
would create matrix files with nonsense characters in them and this would
result in weird output. The remedy was to delete all of the files
associated
with the particular output and start over again. So far, this has worked
for
me.
HTH,
Gabriel
_____
From: Structural Equation Modeling Discussion Group
Sent: Monday, September 25, 2006 1:22 PM
Subject: Re: SB Chi square value
Hi Ed and Ravi,
Ed - I had thought that the fact that the ACM is not inverted with S-B
scaling would guard against such a problem. It seems to work well at small
sample sizes (i.e., less than 200), but then again the simulations I've
read
about usually don't involve huge numbers of observed variables. With too
low
N in such a case, you still won't get a stable enough estimate of the ACM
to
work properly even with the S-B procedure? Would that explain the crazy
fit
output that Ravi got?
Ravi - How many measured variables do you have and what is your sample
size?
By the way, it's fine if the program treats some variables as continuous
and
some as ordinal - this is not an error, actually, this is the way it
should
be done. It just means that the input matrix is a combination of different
types of correlations that respect the different categorizations of the
variables. Then the information from the ACM (the covariance matrix of the
correlations) is incorporated into the estimation method to obtain correct
fit statistics and standard errors. Of course, the correctness depends on
meeting the conditions required for using the alternative estimation
method.
Hope this helps,
Cam
Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de
la
santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney's Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]
-----Original Message-----
From: Structural Equation Modeling Discussion Group
Sent: September 25, 2006 10:24 AM
Subject: Re: SB Chi square value
Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some
of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO
(continuous)
or OR (ordinal) depending on their nature.
Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.
Regards,
Ravi
Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon
Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To
of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html
-------------------------------------------------------------- To
of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Deborah Bandalos]

Hi everyone - I was interested in the subject of the SB chi-square because I
have also obtained anomalous results. However, in my case these have been
obtained when using the SB chi-square in chi-square difference testing in a
multiple group situation. I am aware that SB difference tests must be
corrected and have applied the appropriate correction, but this has
sometimes resulted in negative values for the difference test, even though a
simple subtraction of either the usual ML chi-square values or of the SB
chi-square values results in a positive difference (as expected).

I am guessing that these results are due to a combination of relatively
small sample sizes (210 and 219 for the 2 groups) and extreme nonnormality
(kurtosis of one variable was over 70 for one group and values for several
other variables had values between 5 and 13 in both samples).

I was wondering if anyone else has had similar results under these
conditions.

Thanks in advance for any insights list members might have.

Debbi Bandalos

Deborah L. Bandalos
Professor, Department of Educational Psychology and Instructional Technology
325V Aderhold Hall
University of Georgia
Athens, GA 30606
706-542-4540 (phone)
706-542-4240 (fax)


There are three kinds of people; those who can do math and those who can't.
_____


--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Colin R. Harbke]

Hi Deborah,

I have encountered anomalous SB chi-square difference tests in multiple-group
comparisons from time to time. Some of them have been with large data sets so
I don't think the issue is simply a matter of sample size. The correction
factors have to be different between the competing models -- which may be more
likely in small sample situations.

I was curious regarding the appropriate correction to the SB difference test?
Is this the SB chi-square difference test itself, or an alternative version
for multiple-group comparisons?

Best wishes,

Colin

Post by Deborah Bandalos
Hi everyone - I was interested in the subject of the SB chi-square because I
have also obtained anomalous results. However, in my case these have been
obtained when using the SB chi-square in chi-square difference testing in a
multiple group situation. I am aware that SB difference tests must be
corrected and have applied the appropriate correction, but this has
sometimes resulted in negative values for the difference test, even though a
simple subtraction of either the usual ML chi-square values or of the SB
chi-square values results in a positive difference (as expected).
I am guessing that these results are due to a combination of relatively
small sample sizes (210 and 219 for the 2 groups) and extreme nonnormality
(kurtosis of one variable was over 70 for one group and values for several
other variables had values between 5 and 13 in both samples).
I was wondering if anyone else has had similar results under these
conditions.
Thanks in advance for any insights list members might have.
Debbi Bandalos
Deborah L. Bandalos
Professor, Department of Educational Psychology and Instructional Technology
325V Aderhold Hall
University of Georgia
Athens, GA 30606
706-542-4540 (phone)
706-542-4240 (fax)
There are three kinds of people; those who can do math and those who can't.
_____
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

-------------------------------------------------
This mail sent through IMP: http://horde.org/imp/

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Deborah Bandalos]

Hi Colin - Thanks for your info on this. I think the role of sample size may
be in the degree of stability to estimate the correction factor, but I could
be wrong about this.

Anyway, the correction factor I mentioned just refers to the fact that the
difference tests are not computed as simple differences in chi-square
values, as they are using the usual ML-based chi-square values.


Debbi

Deborah L. Bandalos
Professor, Department of Educational Psychology and Instructional Technology
325V Aderhold Hall
University of Georgia
Athens, GA 30606
706-542-4540 (phone)
706-542-4240 (fax)

There are three kinds of people; those who can do math and those who can't.


-----Original Message-----
From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Colin R. Harbke
Sent: Tuesday, September 26, 2006 4:13 PM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value

Hi Deborah,

I have encountered anomalous SB chi-square difference tests in
multiple-group
comparisons from time to time. Some of them have been with large data sets
so
I don't think the issue is simply a matter of sample size. The correction
factors have to be different between the competing models -- which may be
more
likely in small sample situations.

I was curious regarding the appropriate correction to the SB difference
test?
Is this the SB chi-square difference test itself, or an alternative version
for multiple-group comparisons?

Best wishes,

Colin

Post by Deborah Bandalos
Hi everyone - I was interested in the subject of the SB chi-square because I
have also obtained anomalous results. However, in my case these have been
obtained when using the SB chi-square in chi-square difference testing in a
multiple group situation. I am aware that SB difference tests must be
corrected and have applied the appropriate correction, but this has
sometimes resulted in negative values for the difference test, even though a
simple subtraction of either the usual ML chi-square values or of the SB
chi-square values results in a positive difference (as expected).
I am guessing that these results are due to a combination of relatively
small sample sizes (210 and 219 for the 2 groups) and extreme nonnormality
(kurtosis of one variable was over 70 for one group and values for several
other variables had values between 5 and 13 in both samples).
I was wondering if anyone else has had similar results under these
conditions.
Thanks in advance for any insights list members might have.
Debbi Bandalos
Deborah L. Bandalos
Professor, Department of Educational Psychology and Instructional Technology
325V Aderhold Hall
University of Georgia
Athens, GA 30606
706-542-4540 (phone)
706-542-4240 (fax)
There are three kinds of people; those who can do math and those who can't.
_____
--------------------------------------------------------------
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

-------------------------------------------------
This mail sent through IMP: http://horde.org/imp/

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

[Cameron McIntosh]

Hi Ravi,
Looks to me like you got things straightened out with running the program, but remember that S-B scaling is not completely theoretically appropriate for ordinal variables, even though Joreskog suggests using it (in an online paper on Karl's corner). I would actually recommend DWLS estimation in your case, although it sometimes produces wacky output as well. DWLS just requires the asymptotic variance matrix (AV) of the elements of your correlation matrix.
Cam


Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney's Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]

-----Original Message-----
From: Structural Equation Modeling Discussion Group [mailto:***@BAMA.UA.EDU] On Behalf Of Ravindra Nath
Sent: September 26, 2006 12:44 AM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value


Thanks Ed, Gabriel, Cam,

Rebooting LISREL helped. I was able to get SB chi square after all. Here are the results:
Degrees of Freedom = 228
Minimum Fit Function Chi-Square = 1585.89 (P = 0.0)
Normal Theory Weighted Least Squares Chi-Square = 1465.03 (P = 0.0)
Satorra-Bentler Scaled Chi-Square = 905.62 (P = 0.0)
Chi-Square Corrected for Non-Normality = 2203.88 (P = 0.0)
Estimated Non-centrality Parameter (NCP) = 683.62
90 Percent Confidence Interval for NCP = (594.62 ; 780.16)

Minimum Fit Function Value = 1.11
Population Discrepancy Function Value (F0) = 0.48
90 Percent Confidence Interval for F0 = (0.42 ; 0.55)
Root Mean Square Error of Approximation (RMSEA) = 0.047
90 Percent Confidence Interval for RMSEA = (0.043 ; 0.050)
P-Value for Test of Close Fit (RMSEA < 0.05) = 0.00

Expected Cross-Validation Index (ECVI) = 1.14
90 Percent Confidence Interval for ECVI = (0.68 ; 0.81)
ECVI for Saturated Model = 0.42
ECVI for Independence Model = 68.50

Chi-Square for Independence Model with 276 Degrees of Freedom = 97495.15
Independence AIC = 97543.15
Model AIC = 1621.03
Saturated AIC = 600.00
Independence CAIC = 97693.44
Model CAIC = 2109.46
Saturated CAIC = 2478.58

Normed Fit Index (NFI) = 0.99
Non-Normed Fit Index (NNFI) = 0.99
Parsimony Normed Fit Index (PNFI) = 0.80
Comparative Fit Index (CFI) = 0.99
Incremental Fit Index (IFI) = 0.99
Relative Fit Index (RFI) = 0.99

Critical N (CN) = 431.74


Root Mean Square Residual (RMR) = 75.38
Standardized RMR = 0.11
Goodness of Fit Index (GFI) = 0.92
Adjusted Goodness of Fit Index (AGFI) = 0.89
Parsimony Goodness of Fit Index (PGFI) = 0.68


Cam,

I am modeling 24 manifest variables. My sample size is 800.

On your point on variable scale, I thought it to be a good practise to instruct PRELIS on the nature of my variables, or else it treats even ordinal variables with more than 10 scale points as continuous.

In another context, in market research, it is often a practise to rescale a five point Likert measurement (1 = Strongly disagree, 5 = strongly agree), by a factor of 20 (20 = Strongly disagree, 100 = strongly agree). Often, values which would be non-existent in the original scale appear in the rescaled dataset because of some kind of imputation (while I agree this is not entirely scientific, but the practise has its own merits). In a dataset of this kind, if some of my variables are treated as ordinal and others as continuous my covariances could be misleading. In this situation, if I am using a Psf datafile and simplis directly estimates a covariance matrix (without me having to input one), it treats all variables as continuous. Now if I was to use PRELIS computed ACM, which would have treated some of these variables as ordinal, my estimnation could go wrong.

Thanks Again,
Ravi




On 9/26/06, Ed Rigdon <***@langate.gsu.edu> wrote:

Gabriel--

That's a very good point. LISREL is still a program where
rebooting LISREL or your computer, or cleaning out old versions
of files, sometimes eliminates odd results. If you think you have
fixed a problem, but it doesn't go away, this is something to try.

--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
Hi Cam, Ed, Ravi,



I sometimes get output that looks similar to Ravi's even with an adequate
sample size and also get the non-positive definite warning. Another person
on the listserv who uses LISREL suggested sometime ago that PRELIS sometimes
would create matrix files with nonsense characters in them and this would
result in weird output. The remedy was to delete all of the files associated
with the particular output and start over again. So far, this has worked for
me.



HTH,



Gabriel





_____

From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU] On Behalf Of Cameron McIntosh
Sent: Monday, September 25, 2006 1:22 PM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value



Hi Ed and Ravi,



Ed - I had thought that the fact that the ACM is not inverted with S-B
scaling would guard against such a problem. It seems to work well at small
sample sizes (i.e., less than 200), but then again the simulations I've read
about usually don't involve huge numbers of observed variables. With too low
N in such a case, you still won't get a stable enough estimate of the ACM to
work properly even with the S-B procedure? Would that explain the crazy fit
output that Ravi got?



Ravi - How many measured variables do you have and what is your sample size?
By the way, it's fine if the program treats some variables as continuous and
some as ordinal - this is not an error, actually, this is the way it should
be done. It just means that the input matrix is a combination of different
types of correlations that respect the different categorizations of the
variables. Then the information from the ACM (the covariance matrix of the
correlations) is incorporated into the estimation method to obtain correct
fit statistics and standard errors. Of course, the correctness depends on
meeting the conditions required for using the alternative estimation method.



Hope this helps,



Cam



Cameron N. McIntosh, MA
Analyst / Analyste
Health Analysis and Measurement Group / Groupe d'analyse et de mesure de la
santé
Statistics Canada / Statistique Canada
24-Q R.H. Coats Building
100 Tunney's Pasture Driveway
Ottawa, ON
K1A 0T6
Phone: [phone]
Fax: [phone]

-----Original Message-----
From: Structural Equation Modeling Discussion Group
[mailto:***@BAMA.UA.EDU ] On Behalf Of Ravindra Nath
Sent: September 25, 2006 10:24 AM
To: ***@BAMA.UA.EDU
Subject: Re: SB Chi square value

Thanks Ed,
It seems a SB correction did not hold here. Since I have been using SB in
some of my other analysis as well, I though I could share with the group
oneother problem I have faced in the past:
While specifying an asymptotic covariance matrix through PRELIS, the
asymptotic covariances might be misleading incase the program treats some of
the variables as continuous and some others as ordinal. This is specially
true if the analysis involves a likert like scale and the researcher has
rescaled the variables. Its best to specify the variables as CO (continuous)
or OR (ordinal) depending on their nature.

Another issue that I have experienced while attempting SB is that the
fitted covariance matrix is not positive definite.

Regards,

Ravi

On 9/25/06, Ed Rigdon < ***@langate.gsu.edu> wrote:

Dear Ravi--
I think that the SB correction has failed, in this instance,
perhaps
because your sample size is too low for the number of observed
variables. The low GFI also suggests that your sample size is low
relative to your 228 DF. Perhaps you can bootstrap the weight
matrix that you need for the SB correction--or perhaps you'll have
to live without it, in this case.
--Ed Rigdon

Edward E. Rigdon, Professor and Chair,
Department of Marketing
Georgia State University
P.O. Box 3991
Atlanta, GA 30302-3991
(express: 35 Broad St., Suite 1300, zip 30303)
phone [phone] fax [phone]
-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

-------------------------------------------------------------- To
unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of
the message as: SIGNOFF SEMNET Search the archives at
http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html

--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html



-------------------------------------------------------------- To unsubscribe from SEMNET, send email to ***@bama.ua.edu with the body of the message as: SIGNOFF SEMNET Search the archives at http://bama.ua.edu/archives/semnet.html


--------------------------------------------------------------
To unsubscribe from SEMNET, send email to ***@bama.ua.edu
with the body of the message as: SIGNOFF SEMNET
Search the archives at http://bama.ua.edu/archives/semnet.html