SEM in AMOS - questions on error terms, covariance and negative chi square value

[Nick]

Dear all,

I am pretty new to SEM in AMOS and would be very grateful for your help with a few questions regarding my model (attached with standardized estimates), specified as follows:
- four predictors (P1 to P4) of one dependent variable (AV1)
- AV 1 in turn predicts a second dependent variable (AV2), AV2 being a latent variable with three indicator variables (outcome1 to outcome 3)
- four variables, that I want to control for, (CV1 to CV4) also influence AV2.

I have a sample of n=170 with some missing data. Data are mean scores from questionnaires and are not skewed.

Here are my questions:
- is it necessary to use the double-headed covariance arrows between all predictor variables (P1 to P4; and CV1 to CV4)

- are the error terms correctly placed and specified with regression weight = 1 ?

- why do I get a chi square value that is likely to be nonsensical:
Number of distinct sample moments: 90
Number of distinct parameters to be estimated: 48
Degrees of freedom (90 - 48): 42
Result (Default model)
Minimum was achieved
Chi-square = -14504178770189,500


Many thanks for a reply in advance,
all the best,
Nick

[Swank, Paul R]

You cannot do attachments to this list, I believe, so we can't see your model. Obviously, something is wrong. I can tell you that it is usually possible to set the covariance between the independent variables and the covariates to 0 but I bet it would mean a considerably worse fit unless somehow they are all independent of one another, which I doubt. Error terms are usually estimated.

Paul R. Swank, Ph.D.
Professor, Department of Pediatrics
Medical School
Adjunct Professor, Health Promotions and Behavioral Sciences
School of Public Health
University of Texas Health Science Center at Houston

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Nick
Sent: Monday, June 25, 2012 4:59 AM
To: [hidden email]
Subject: SEM in AMOS - questions on error terms, covariance and negative chi square value

Dear all,

I am pretty new to SEM in AMOS and would be very grateful for your help with a few questions regarding my model (attached with standardized estimates), specified as follows:
- four predictors (P1 to P4) of one dependent variable (AV1)
- AV 1 in turn predicts a second dependent variable (AV2), AV2 being a latent variable with three indicator variables (outcome1 to outcome 3)
- four variables, that I want to control for, (CV1 to CV4) also influence AV2.

I have a sample of n=170 with some missing data. Data are mean scores from questionnaires and are not skewed.

Here are my questions:
- is it necessary to use the double-headed covariance arrows between all predictor variables (P1 to P4; and CV1 to CV4)

- are the error terms correctly placed and specified with regression weight = 1 ?

- why do I get a chi square value that is likely to be nonsensical:
Number of distinct sample moments: 90
Number of distinct parameters to be estimated: 48 Degrees of freedom (90 - 48): 42 Result (Default model) Minimum was achieved Chi-square = -14504178770189,500


Many thanks for a reply in advance,
all the best,
Nick





http://spssx-discussion.1045642.n5.nabble.com/file/n5713770/Model_standardizedEstimates.jpg


--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/SEM-in-AMOS-questions-on-error-terms-covariance-and-negative-chi-square-value-tp5713770.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

[Bruce Weaver]

Hi Paul.  The attachment can be seen via the Nabble archive:

http://spssx-discussion.1045642.n5.nabble.com/file/n5713770/Model_standardizedEstimates.jpg

Cheers,
Bruce

Swank, Paul R wrote

You cannot do attachments to this list, I believe, so we can't see your model. Obviously, something is wrong. I can tell you that it is usually possible to set the covariance between the independent variables and the covariates to 0 but I bet it would mean a considerably worse fit unless somehow they are all independent of one another, which I doubt. Error terms are usually estimated.

Paul R. Swank, Ph.D.
Professor, Department of Pediatrics
Medical School
Adjunct Professor, Health Promotions and Behavioral Sciences
School of Public Health
University of Texas Health Science Center at Houston

[Nick]

Hi Bruce and Paul,
thank you for your replies and for clarifying how the attached figure with the model and estimates can be seen.

Looking at the figure, do you or others have any suggestions why my model does not work (see my last posting, e.g. the chi square value)?

Any advice would be greatly appreciated,
all the best,
Nick

[David Marso]

Chi square looks *WONKY* certainly some mispecification!
Why are you setting the error(s) to 1?
In this case since you are using observed vars rather than a latent variable you might consider setting them to 0 (as a start -treating them as observed without error) or if you have an idea of the reliability use ??1-reliability?? not sure of this exactly.  Google is your friend in that regard.

Nick wrote

Hi Bruce and Paul,
thank you for your replies and for clarifying how the attached figure with the model and estimates can be seen.

Looking at the figure, do you or others have any suggestions why my model does not work (see my last posting, e.g. the chi square value)?

Any advice would be greatly appreciated,
all the best,
Nick

[Andy W]

It is a very complicated model, so I can't say any specifics. The only advice I could give is to estimate the individual models separately for diagnostic checks at the onset. Perhaps say starting with the latent variable model (AV2 -> Indicators), then sequentially adding in the other components (say CV's first, then AV1, then predictors of AV1). It is possible misspecification in any of the sub-models could result in strange estimates for the full model. I don't know, do the multiple causes in MIMIC models cause misspecification if a causal indicator is not correlated with the latent variable?

The SEMNET mailing list may be a better alternative to direct your question to.

Andy

[David Matheson-3]

Hi Nick,
 You don't mention whether you encountered the negative chi-square on the
first run of the model file. However, I think that you've encountered a
bug in AMOS 20 that is triggered on subsequent runs of a model file where
there is missing data  and the computer's local settings use a comma as a
decimal point. There is a fix available at:

http://www-01.ibm.com/support/docview.wss?uid=swg24032174

You will need to register on the site to download the fix pack. In
the "Download description" area, you'll see a "Fix list" link that
provides descriptions of the issues corrected by the fix pack. This issue
is described in more detail in the third link in the fix list. (You'll
also need to be registered on the site to open the description.)

The negative chi-square and other strange values in subsequent
runs of the model are due to  the *.AmosP file, which is created
the first time you fit a model to a specific subset of variables
in a specific data file that has missing values. The *.AmosP
file contains the discrepancies for the independence and
saturated models. In the first analysis, the discrepancies are
calculated. On subsequent analyses of the same subset of
variables in the same data file, the discrepancies for the
independence and saturated models are read from the *.AmosP
file.
A temporary workaround is to delete any AmosP files in the
working directory, i.e. the directory with the AMOS program
(.amw) file, before running each analysis.


David Matheson
IBM SPSS Support


On Mon, 25 Jun 2012 02:58:59 -0700, Nick <[hidden email]> wrote:

>Dear all,
>
>I am pretty new to SEM in AMOS and would be very grateful for your help
with

>a few questions regarding my model (attached with standardized estimates),
>specified as follows:
>- four predictors (P1 to P4) of one dependent variable (AV1)
>- AV 1 in turn predicts a second dependent variable (AV2), AV2 being a
>latent variable with three indicator variables (outcome1 to outcome 3)
>- four variables, that I want to control for, (CV1 to CV4) also influence
>AV2.
>
>I have a sample of n=170 with some missing data. Data are mean scores from
>questionnaires and are not skewed.
>
>Here are my questions:
>- is it necessary to use the double-headed covariance arrows between all
>predictor variables (P1 to P4; and CV1 to CV4)
>
>- are the error terms correctly placed and specified with regression

weight

>= 1 ?
>
>- why do I get a chi square value that is likely to be nonsensical:
>Number of distinct sample moments: 90
>Number of distinct parameters to be estimated: 48
>Degrees of freedom (90 - 48): 42
>Result (Default model)
>Minimum was achieved
>Chi-square = -14504178770189,500
>
>
>Many thanks for a reply in advance,
>all the best,
>Nick
>
>
>
>
>
>http://spssx-

discussion.1045642.n5.nabble.com/file/n5713770/Model_standardizedEstimates.
jpg
>
>
>--
>View this message in context: http://spssx-
discussion.1045642.n5.nabble.com/SEM-in-AMOS-questions-on-error-terms-
covariance-and-negative-chi-square-value-tp5713770.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

[Nick]

Dear David and all,

thank you very much for your feedback.
Yes, it was the bug in AMOS 20 that caused the nonsensical chi square value. With the software update (and the same model), chi square is now about 130, RMSEA about 0.09.

Given the complexity of the model and my difficulty to know how it is best specified (error terms, covariances between predictor variables), I was wondering whether I should not better calculate two multiple linear regressions instead:
regression 1 with AV1 as dependent variable (DV) and P1 to P4 as independent variables (IVs)
and
regression 2 with AV2 as DV (average of standardized outcomes 1 to 3) and AV1 and CV1 to CV4 as IVs.

Two regressions may be a much more straightforward analytic approach than this complex SEM model.
Any thoughts?

Thanks a lot again for your help,
Nick

[Andy W]

Nick,

That certainly does make the model more simple. One thing to note is that constructing AV2 as the "(average of standardized outcomes 1 to 3)" forces a very explicit parallel measurement model for AV2. It would certainly be advised to confirm whether this measurement model is appropriate separately (i.e. you would just conduct a normal confirmatory factor analysis for the [AV2 -> manifest] part). It looks like this is perhaps reasonable given the original set of loadings in your attached image (if they are standardized loadings that is).

I think it is reasonable to estimate separate models in an exploratory phase to identify separately model fit of the individual components (it is also easier to interpret the model in small parts than it is as a whole). Whether or not you should subsequently build up to the full model in the end is a difficult question to answer though.

HTH,
Andy