AMOS ERROR MEASSAGE

[rks]

I am trying to test my model and I keep on getting this message "The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint." when I view the text output.Also I get "An error occurred while attempting to fit the model.The sample moment is not positive definite.It could fail to be positive definite for any of the following reasons..."then it give couple of reasons which I fail to understand.
Please someone help.

[David Marso]

You really need to be more specific about the model you are attempting to fit.
Is it a simple factor model?  
A more general SEM model?  
How many variables?  
How many cases?
What equations are being estimated?  
How many free parameters?  
What parameters are already constrained?  
What is the covariance structure of the latent variables?

"The sample moment is not positive definite....."
Perhaps quote the 'reasons which you fail to understand'.  
Someone here WILL understand.
Are any variables linear combinations of others (that is the most basic reason)?
Can you post the path diagram somewhere so we can see it?
Can you post the covariance matrix?
---
These are fundamental questions which we need to know the answers in order to help!!!
If any of this is Greek to you then you may want to review the theoretical foundations of SEM, Factor analysis and regression analysis so we are using an agreed upon vocabulary.
HTH, David

rks wrote

I am trying to test my model and I keep on getting this message "The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint." when I view the text output.Also I get "An error occurred while attempting to fit the model.The sample moment is not positive definite.It could fail to be positive definite for any of the following reasons..."then it give couple of reasons which I fail to understand.
Please someone help.

[rks]

Thanks David
I have attached the word file which contains the model I am testing.It has five constructs or factors twenty observed variables.249 cases. From estimation I have selected Maximum likelihood and Estimation means and intercepts. From output I select Minimisation history, Standardised estimates, Squared Multiple correlation and Correlation of estimates.The data file contains raw data for the observed variables for 249 cases, in numeric format since the questionnaire used 5 point Likert scale (1= Strongly disagree 5= Strongly agree).The covariance structure is in the Model.

The error message was from similar Model which was exactly the same but did not have the OBE construct and its resulting Path lines.An error occurred while checking for missing data in the group, Group number 1.
The sample moment matrix is not positive definite.It could fail to be positive definite for any of the following reasons:
1. The sample covariance matrix or the sample correlation matrix contains a data entry error.
2.The observed variables are linearly dependent (perhaps because the sample size is too small).
3. The sample covariance matrix or sample correlation matrix was computed from incomplete data using method of 'Pairwise deletion'.
4.The sample correlation matrix contains correlation coefficients other than product moment correlations (such as tetrachoric correlations).

For maximum likelihood estimation only,it may be appropriate to check "Allow non-positive definite sample covariance matrices" in the "Analysis Properties" window, or to use the NonPositive method.

I would also want to provide the data file,that would give a more clear picture.

ThanksTEST_DATA_2.savNew_Microsoft_Word_Document.doc

[rks]

This is the data File.TEST_DATA_2.sav

[David Marso]

Diagnostics:
Your 20x20 Matrix is of rank 9 (singular).
You have several exact copies of various variables!!
Why?  This will certainly create serious problems!

[kammel]

Hi

I'm trying to assess the fit of an established model to my data. It is a 6 factor model with 3 variables appearing twice in two different factors. I understand this may be a problem as I get this error message after selecting 'calculate estimates':

"An error occurred while attempting to fit the model.

The sample moment matrix is no positive definite. It could be for the following reasons:

1) The sample covariance matrix or the sample correlation matrix ocntains a data entry error.

2) The observed variables are linearly dependent (perhaps because the sample size is too small).

3) The sample covariance matrix or sample correlation amtrix was computed from incomplete data using the method of 'pairwise deletion'

4) The sample coorelation matrix contains correlation coefficients other than product moment correlations (such as tetrachoric correlations)".

I understand that the 3 variables appearing twice in two different factors are the causes of 'linear dependency', however as I am testing an established model, I see no way around this.

Can anyone suggets alternatives?

Kindest

[David Marso]

Please elaborate what this means: "3 variables appearing twice in two different factors".
You are SURELY NOT entering the same variables twice into the covariance matrix!
That would be truly unfortunate and doomed to fail miserably.
--

kammel wrote

Hi

I'm trying to assess the fit of an established model to my data. It is a 6 factor model with 3 variables appearing twice in two different factors. I understand this may be a problem as I get this error message after selecting 'calculate estimates':

"An error occurred while attempting to fit the model.

The sample moment matrix is no positive definite. It could be for the following reasons:

1) The sample covariance matrix or the sample correlation matrix ocntains a data entry error.

2) The observed variables are linearly dependent (perhaps because the sample size is too small).

3) The sample covariance matrix or sample correlation amtrix was computed from incomplete data using the method of 'pairwise deletion'

4) The sample coorelation matrix contains correlation coefficients other than product moment correlations (such as tetrachoric correlations)".

I understand that the 3 variables appearing twice in two different factors are the causes of 'linear dependency', however as I am testing an established model, I see no way around this.

Can anyone suggets alternatives?

Kindest

[Ryan]

Not enough information was provided to help....

1. type of SEM model (in detail)
2. covariance/correlation matrix from raw data
3. amos syntax
4. clear explanation about the variables, etc.
5. Very specific question(s)

It sounds like the OP may be speaking about cross-loadings, but that is a guess.

Ryan

On Apr 22, 2013, at 12:07 PM, David Marso <[hidden email]> wrote:

> Please elaborate what this means: "3 variables appearing twice in two
> different factors".
> You are SURELY NOT entering the same variables twice into the covariance
> matrix!
> That would be truly unfortunate and doomed to fail miserably.
> --
>
> kammel wrote
>> Hi
>>
>> I'm trying to assess the fit of an established model to my data. It is a 6
>> factor model with 3 variables appearing twice in two different factors. I
>> understand this may be a problem as I get this error message after
>> selecting 'calculate estimates':
>>
>> "An error occurred while attempting to fit the model.
>>
>> The sample moment matrix is no positive definite. It could be for the
>> following reasons:
>>
>> 1) The sample covariance matrix or the sample correlation matrix ocntains
>> a data entry error.
>>
>> 2) The observed variables are linearly dependent (perhaps because the
>> sample size is too small).
>>
>> 3) The sample covariance matrix or sample correlation amtrix was computed
>> from incomplete data using the method of 'pairwise deletion'
>>
>> 4) The sample coorelation matrix contains correlation coefficients other
>> than product moment correlations (such as tetrachoric correlations)".
>>
>> I understand that the 3 variables appearing twice in two different factors
>> are the causes of 'linear dependency', however as I am testing an
>> established model, I see no way around this.
>>
>> Can anyone suggets alternatives?
>>
>> Kindest
>
>
>
>
>
> -----
> Please reply to the list and not to my personal email.
> Those desiring my consulting or training services please feel free to email me.
> ---
> "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis."
> Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?"
> --
> View this message in context: http://spssx-discussion.1045642.n5.nabble.com/AMOS-ERROR-MEASSAGE-tp4667749p5719626.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]

To clarify David's comment --

Your error message complains about a matrix is not "positive definite."

If you actually did put the same variable twice into the list that results
in one covariance matrix, well, that is an error message that would show it. 

I don't use Amos, but I presume that you have mis-applied the syntax for
presenting the model that you want.  It seems less likely that Amos simply
can't do the model that you want.

--
Rich Ulrich

> Date: Mon, 22 Apr 2013 09:07:29 -0700

> From: [hidden email]
> Subject: Re: AMOS ERROR MEASSAGE
> To: [hidden email]
>
> Please elaborate what this means: "3 variables appearing twice in two
> different factors".
> You are SURELY NOT entering the same variables twice into the covariance
> matrix!
> That would be truly unfortunate and doomed to fail miserably.
> --
>
> kammel wrote
> > Hi
> >
> > I'm trying to assess the fit of an established model to my data. It is a 6
> > factor model with 3 variables appearing twice in two different factors. I
> > understand this may be a problem as I get this error message after
> > selecting 'calculate estimates':
> >
> > "An error occurred while attempting to fit the model.
> >
> > The sample moment matrix is no positive definite. It could be for the
> > following reasons:
> >
> > 1) The sample covariance matrix or the sample correlation matrix ocntains
> > a data entry error.
> >
> > 2) The observed variables are linearly dependent (perhaps because the
> > sample size is too small).
> >
> > 3) The sample covariance matrix or sample correlation amtrix was computed
> > from incomplete data using the method of 'pairwise deletion'
> >
> > 4) The sample coorelation matrix contains correlation coefficients other
> > than product moment correlations (such as tetrachoric correlations)".
> >
> > I understand that the 3 variables appearing twice in two different factors
> > are the causes of 'linear dependency', however as I am testing an
> > established model, I see no way around this.
> >
> > Can anyone suggets alternatives?
> >
> > Kindest
>
> ...

[kammel]

<quote author="David Marso">
Please elaborate what this means: "3 variables appearing twice in two different factors".
You are SURELY NOT entering the same variables twice into the covariance matrix!
That would be truly unfortunate and doomed to fail miserably.

David, thank you for your response.

To elaborate, I have six hypothesised latent factors: Discomfort, Succumbing, Information, Vulnerability 1, Coping and Vulnerability 2. According to the authors, item 3 of the questionnaire loads highly on Discomfort and Succumbing. Item 9 and 12 both load highly on Discomfort and Information.

Now, the problem is clear to me and should have been addressed by the authors in preliminary exploratory factor analyses. However, this model has subsequently been factor analysed (confirmatory) by many others since. In their articles, they do not outline whether they encountered such a problem, nor how they resolved it.

I am new to working at this level with SPSS and AMOS, and am tempted to delete the replicated items from the factors they least load on to. However, I want to be sure that this is my only option before I commit to it.

Kindest

[Ryan]

There are many possibilities, which is why providing syntax is so important...I can't tell if the OP has provided the variance-covariance matrix (which is usually in embedded in an SPSS data file) being fed into AMOS or the actual AMOS syntax (simultaneous equations) to fit the model. Redundant variables in the variance-covariance matrix as well as problematic models could yield a not p.d result.

 

If the OP provided the syntax, then, with little effort, someone (who had the time and access to AMOS) could fit the model to help the OP determine where the problem lies.

 

Ryan

On Mon, Apr 22, 2013 at 1:53 PM, Rich Ulrich <[hidden email]> wrote:

To clarify David's comment --

Your error message complains about a matrix is not "positive definite."

If you actually did put the same variable twice into the list that results
in one covariance matrix, well, that is an error message that would show it. 

I don't use Amos, but I presume that you have mis-applied the syntax for
presenting the model that you want.  It seems less likely that Amos simply
can't do the model that you want.

--
Rich Ulrich

> Date: Mon, 22 Apr 2013 09:07:29 -0700
> From: [hidden email]
> Subject: Re: AMOS ERROR MEASSAGE
> To: [hidden email]

>
> Please elaborate what this means: "3 variables appearing twice in two
> different factors".
> You are SURELY NOT entering the same variables twice into the covariance
> matrix!
> That would be truly unfortunate and doomed to fail miserably.
> --
>
> kammel wrote
> > Hi
> >
> > I'm trying to assess the fit of an established model to my data. It is a 6
> > factor model with 3 variables appearing twice in two different factors. I
> > understand this may be a problem as I get this error message after
> > selecting 'calculate estimates':
> >
> > "An error occurred while attempting to fit the model.
> >
> > The sample moment matrix is no positive definite. It could be for the
> > following reasons:
> >
> > 1) The sample covariance matrix or the sample correlation matrix ocntains
> > a data entry error.
> >
> > 2) The observed variables are linearly dependent (perhaps because the
> > sample size is too small).
> >
> > 3) The sample covariance matrix or sample correlation amtrix was computed
> > from incomplete data using the method of 'pairwise deletion'
> >
> > 4) The sample coorelation matrix contains correlation coefficients other
> > than product moment correlations (such as tetrachoric correlations)".
> >
> > I understand that the 3 variables appearing twice in two different factors
> > are the causes of 'linear dependency', however as I am testing an
> > established model, I see no way around this.
> >
> > Can anyone suggets alternatives?
> >
> > Kindest
>

> ...

[kammel]

<quote author="R B">

It sounds like the OP may be speaking about cross-loadings, but that is a guess.

Ryan

Ryan,

Yes, I am referring to cross-loadings. I understand that this has serious implications for the proposed model and will discuss this in my thesis. However, I am most concerned with my options to continue with a CFA using this model.

Kindest.

[Ryan]

Okay. I actually found the schematic representation of the model attached to one of the posts. I wasn't able to view the covariance/correlation matrix. At a quick glance, the model seems to be constructed properly. Still, there could be a host of reasons a non p.d. matrix can happen, even with a properly constructed model. I've now also read the thread since the beginning, and from what I understand, it sounds like there are duplicates of variables. That's a mistake that can be easily remedied. If that is NOT the case (which sounds unlikely given David's diagnostics), then an examination of the output is important.

 

Best,

 

Ryan

On Mon, Apr 22, 2013 at 2:06 PM, R B <[hidden email]> wrote:

There are many possibilities, which is why providing syntax is so important...I can't tell if the OP has provided the variance-covariance matrix (which is usually in embedded in an SPSS data file) being fed into AMOS or the actual AMOS syntax (simultaneous equations) to fit the model. Redundant variables in the variance-covariance matrix as well as problematic models could yield a not p.d result.

 

If the OP provided the syntax, then, with little effort, someone (who had the time and access to AMOS) could fit the model to help the OP determine where the problem lies.

 

Ryan

On Mon, Apr 22, 2013 at 1:53 PM, Rich Ulrich <[hidden email]> wrote:

To clarify David's comment --

Your error message complains about a matrix is not "positive definite."

If you actually did put the same variable twice into the list that results
in one covariance matrix, well, that is an error message that would show it. 

I don't use Amos, but I presume that you have mis-applied the syntax for
presenting the model that you want.  It seems less likely that Amos simply
can't do the model that you want.

--
Rich Ulrich

> Date: Mon, 22 Apr 2013 09:07:29 -0700
> From: [hidden email]
> Subject: Re: AMOS ERROR MEASSAGE
> To: [hidden email]

>
> Please elaborate what this means: "3 variables appearing twice in two
> different factors".
> You are SURELY NOT entering the same variables twice into the covariance
> matrix!
> That would be truly unfortunate and doomed to fail miserably.
> --
>
> kammel wrote
> > Hi
> >
> > I'm trying to assess the fit of an established model to my data. It is a 6
> > factor model with 3 variables appearing twice in two different factors. I
> > understand this may be a problem as I get this error message after
> > selecting 'calculate estimates':
> >
> > "An error occurred while attempting to fit the model.
> >
> > The sample moment matrix is no positive definite. It could be for the
> > following reasons:
> >
> > 1) The sample covariance matrix or the sample correlation matrix ocntains
> > a data entry error.
> >
> > 2) The observed variables are linearly dependent (perhaps because the
> > sample size is too small).
> >
> > 3) The sample covariance matrix or sample correlation amtrix was computed
> > from incomplete data using the method of 'pairwise deletion'
> >
> > 4) The sample coorelation matrix contains correlation coefficients other
> > than product moment correlations (such as tetrachoric correlations)".
> >
> > I understand that the 3 variables appearing twice in two different factors
> > are the causes of 'linear dependency', however as I am testing an
> > established model, I see no way around this.
> >
> > Can anyone suggets alternatives?
> >
> > Kindest
>

> ...

[David Marso]

Note in Ryan's first post that he asked 5 or 6 very specific questions including COV matrix and Amos Syntax (supplying a pic of the path diagram would be useful as well).  Barring that, you are not likely to receive further assistance.  SEM models are a PITA to diagnose in the first place even WITH complete info.
Expecting a group of strangers to poke around in the dark when you are in a position to help them help you by providing answers to specific questions is not a good use of this resource.
--
<quote author="kammel">

R B wrote

It sounds like the OP may be speaking about cross-loadings, but that is a guess.

Ryan

Ryan,

Yes, I am referring to cross-loadings. I understand that this has serious implications for the proposed model and will discuss this in my thesis. However, I am most concerned with my options to continue with a CFA using this model.

Kindest.

[David Marso]

Ryan, NOTE that that path model representation is probably from a much earlier thread from a different OP (I don't know about 3 months ago).  EDIT: Actually is is over a year old (how time flies).
ANOTHER GOOD REASON WHY PEOPLE SHOULD START NEW THREADS RATHER THAN HIJACKING OLD ONES!  It is confusing for all concerned.

R B wrote

Okay. I actually found the schematic representation of the model attached
to one of the posts. I wasn't able to view the covariance/correlation
matrix. At a quick glance, the model seems to be constructed properly.
Still, there could be a host of reasons a non p.d. matrix can happen, even
with a properly constructed model. I've now also read the thread since the
beginning, and from what I understand, it sounds like there are duplicates
of variables. That's a mistake that can be easily remedied. If that is NOT
the case (which sounds unlikely given David's diagnostics), then an
examination of the output is important.

Best,

Ryan


On Mon, Apr 22, 2013 at 2:06 PM, R B <[hidden email]> wrote:

> There are many possibilities, which is why providing syntax is so
> important...I can't tell if the OP has provided the variance-covariance
> matrix (which is usually in embedded in an SPSS data file) being fed into
> AMOS or the actual AMOS syntax (simultaneous equations) to fit the model.
> Redundant variables in the variance-covariance matrix as well
> as problematic models could yield a not p.d result.
>
> If the OP provided the syntax, then, with little effort, someone (who had
> the time and access to AMOS) could fit the model to help the OP determine
> where the problem lies.
>
> Ryan
> On Mon, Apr 22, 2013 at 1:53 PM, Rich Ulrich <[hidden email]> wrote:
>
>> To clarify David's comment --
>>
>> Your error message complains about a matrix is not "positive definite."
>>
>> If you actually did put the same variable twice into the list that results
>> in one covariance matrix, well, that is an error message that would show
>> it.
>>
>> I don't use Amos, but I presume that you have mis-applied the syntax for
>> presenting the model that you want.  It seems less likely that Amos simply
>> can't do the model that you want.
>>
>> --
>> Rich Ulrich
>>
>>
>> > Date: Mon, 22 Apr 2013 09:07:29 -0700
>> > From: [hidden email]
>> > Subject: Re: AMOS ERROR MEASSAGE
>> > To: [hidden email]
>>
>> >
>> > Please elaborate what this means: "3 variables appearing twice in two
>> > different factors".
>> > You are SURELY NOT entering the same variables twice into the covariance
>> > matrix!
>> > That would be truly unfortunate and doomed to fail miserably.
>> > --
>> >
>> > kammel wrote
>> > > Hi
>> > >
>> > > I'm trying to assess the fit of an established model to my data. It
>> is a 6
>> > > factor model with 3 variables appearing twice in two different
>> factors. I
>> > > understand this may be a problem as I get this error message after
>> > > selecting 'calculate estimates':
>> > >
>> > > "An error occurred while attempting to fit the model.
>> > >
>> > > The sample moment matrix is no positive definite. It could be for the
>> > > following reasons:
>> > >
>> > > 1) The sample covariance matrix or the sample correlation matrix
>> ocntains
>> > > a data entry error.
>> > >
>> > > 2) The observed variables are linearly dependent (perhaps because the
>> > > sample size is too small).
>> > >
>> > > 3) The sample covariance matrix or sample correlation amtrix was
>> computed
>> > > from incomplete data using the method of 'pairwise deletion'
>> > >
>> > > 4) The sample coorelation matrix contains correlation coefficients
>> other
>> > > than product moment correlations (such as tetrachoric correlations)".
>> > >
>> > > I understand that the 3 variables appearing twice in two different
>> factors
>> > > are the causes of 'linear dependency', however as I am testing an
>> > > established model, I see no way around this.
>> > >
>> > > Can anyone suggets alternatives?
>> > >
>> > > Kindest
>> >
>> > ...
>>
>
>

[kammel]

Gething.amwIDP_covariance.spv

Dear Ryan and David,

Thank you for your responses.

I apologise if my posting on here has caused confusion. I am new to this form of communication and using AMOS so please bare with me if I am vague or unable to fill in necessary details. I am hugely appreciative of any help you can give.

So Ryan, in response to your questions:

1) I am seeking to assess the model fit of an established questionnaire - the Interactions with disabled persons scale- to my data sample using confirmatory factor analysis.

2) The scale has 6 factors (Discomfort, Succumbing, Information, Vulnerability1, Coping and Vulnerability2). The authors have presented a model in which 3 of items have high loadings on two factors (cross-loading as you pointed out Ryan). I believe that these may be the cause of the P.D. problem.

3)The SEM model and the correlation matrix is attached to this post.

4) I'm afraid I am not sure how to get the Amos Syntax, I have not being using a script up till this point.

Kindest.

[kammel]

IDP_model.docx

Hi,

You should be able to see the model in this doc.

Thank you

[David Marso]

NOPE!  My old version of Word does NOT read docx!
Please post generic formats.  OTOH:  I did glance at your correlation matrix and the existence of off diagonal 1 values suggest you DID COPY VARIABLES AND USE THEM TWICE!
--
That is NOT the way to get Amos or any other SEM program to cooperate.
From out the the gate the covariance matrix is NPD.
Maybe start again and have ONE COPY of any given variable and direct paths from the single instance to multiple outcomes.   I nave no idea how much stats background you have or how much reading you have done in the SEM literature, but that might be something should consider exploring further.
If you are copying someone else's model specification and they are using multiple copies of variables then it is a classic case of the blind leading the blind.
-----

kammel wrote

IDP_model.docx

Hi,

You should be able to see the model in this doc.

Thank you

[kammel]

Doc8.doc

David,

I have changed the format of the document.

Thank you for answering my question. I have read a sufficient amount to identify why the covariance matrix is NPD. My question to the group, really is what are my other options for assessing the fit of the model. Not, why has this occurred.

As mentioned before, there have been subsequent published analyses of the construct validity of the IDP. These have failed to mention how they dealt with the cross loading items. In addition, there is no mention of this issue in text books or statistical journal articles.

I will take your advice to 'have one copy of any given variable'. I was tempted to do so (through common sense and no guiding theory) but was seeking the opinion of those more experienced in statistical analyses at such a high level.

Thank you for your time and patience.

[David Marso]

1. Nuke the dups.
2. Draw paths from the appropriate latents to the observed.
AFAIK: There is nothing forbidden about having a variable load on more than one latent variable.
OTOH: The nature of that observed variable violates basic concepts of simple structure favored in the factor analytic world.  Not much more to say on that.
--

kammel wrote

Doc8.doc

David,

I have changed the format of the document.

Thank you for answering my question. I have read a sufficient amount to identify why the covariance matrix is NPD. My question to the group, really is what are my other options for assessing the fit of the model. Not, why has this occurred.

As mentioned before, there have been subsequent published analyses of the construct validity of the IDP. These have failed to mention how they dealt with the cross loading items. In addition, there is no mention of this issue in text books or statistical journal articles.

I will take your advice to 'have one copy of any given variable'. I was tempted to do so (through common sense and no guiding theory) but was seeking the opinion of those more experienced in statistical analyses at such a high level.

Thank you for your time and patience.