LISREL - Model does not converge, Matrix not positive definite

tanj

New Member
#1
Hello,

I have imported a large amount of data (250 observations, 15 variables) and defined a path diagram to develop SEM. But after building simplix syntax and running lisrel syntax, it says that the model does not converge and in the output file, following are the errors:

Matrix to be analyzed is not positive definite,
ridge option taken with ridge constant = 0.001


Serious problems were encountered during minimization.
Unable to continue iterations. Check your model and data.

Note: My variables are very strongly correlated (r squared > 0.9). Four of my latent variables have only one indicator each. And all the latent variables are connected with each other with a covariance curve.

And I found some solutions regarding multiplying the diagonals of my data matrix with a constant until the model is positive definite. I am not sure exactly how to do this. I am in urgent need of the solution.

Thank you.
 

Lazar

Phineas Packard
#2
Could you post your LISREL syntax. With one item indicators you need to constrain both the loading term and the error term. Have you done this?
 

tanj

New Member
#3
I didn't quite understand which file you wanted. I used 'build simplis syntax' and then 'run lisrel'. I think you need the spj file. No I haven't tried that for the single indicator latent variables. Can you please explain to me how it is done? I am at an amateur level in lisrel.
Is high correlation among the variables can create this non-convergence of model?

Thank you.
 

Lazar

Phineas Packard
#4
I didn't quite understand which file you wanted. I used 'build simplis syntax' and then 'run lisrel'. I think you need the spj file. No I haven't tried that for the single indicator latent variables. Can you please explain to me how it is done? I am at an amateur level in lisrel.
Is high correlation among the variables can create this non-convergence of model?

Thank you.
Tbh I have only scripted directly min LISREL never in simplex. In any case your error is mot likely due to model misspecification and not from a need to do anything strange to your covariance matrix. My guess is that you have misspecification the single item latent variables. The most common way to do this is to constrain the item loading to be 1 and the error to be zero. Barbara Byrne's LISREL book probably gives an example. It is a bit hard to fix the problem without looking at the model directly however.
 

tanj

New Member
#5
Thanks for your reply. I did as you said. Fixed the factor loading to 1 (value on the uni-directional arrow) and fixed the variance of the latent variable to 0. But it generates more errors as you can see in the attached file containing the output. The steps I following to develop the model is: Imported data as tab delimited txt file and saved as psf. Created a pth file as specified the observed and latent variable names and the psf data source. I specified the matrix to be analyzed as correlation. Then I drew the path diagram and fixed the factor loadings and variances for the latent variables with one indicator. Then, from setup, first 'Build Simplis Syntax' and then 'Run LISREL'. Then it says the model does not converge. I've also attached the path diagram I specified.

Previously I modified the values to make variances less different. But the same thing happened.
 

tanj

New Member
#6
I also tried with the LISREL syntax but it gives and error saying that the number of variables specified in the syntax do not match with the psf data file.
 

tanj

New Member
#7
Lazar, can you tell me what actually the values on the single headed arrows and double headed arrows mean when I have one observed variable for each latent variables which are connected by double headed or covariance arrows. The loadings on the arrows are fixed to 1.00 and the error variance on the latent variables are fixed to 0.00. If I select standardized coefficients from the menu, what does these values represents? Are they now correlations? Also, running this model provides

Degrees of Freedom = 0
Minimum Fit Function Chi-Square = 0.00 (P = 1.00)
Normal Theory Weighted Least Squares Chi-Square = 0.00 (P = 1.00)

What does all these values mean? Can degrees of freedom and chi-square can both be zero?

Thanks.
 
#8
I have a serious problem, and I can't solve it... The Model don't converge at CFA LISREL and at structural Model LISREL and I can't solve it... could any one help me necessarily?
 
#10
W_A_R_N_I_N_G: LAMBDA-X does not have full column rank


W_A_R_N_I_N_G: The solution was found non-admissible after 50 iterations.
The following solution is preliminary and is provided only
for the purpose of tracing the source of the problem.
Setting AD> 50 or AD=OFF may solve the problem

Did you Like to send you the output file of LISREL CFA ? or .pth files? could you kindly send me your E-mail to contact you more easier?
 
#11
We cannot help with only that limited amount of information.
Re: LISREL - Model does not converge, Matrix not positive definite
W_A_R_N_I_N_G: LAMBDA-X does not have full column rank


W_A_R_N_I_N_G: The solution was found non-admissible after 50 iterations.
The following solution is preliminary and is provided only
for the purpose of tracing the source of the problem.
Setting AD> 50 or AD=OFF may solve the problem

Did you Like to send you the output file of LISREL CFA ? or .pth files? could you kindly send me your E-mail to contact you more easier?
 

Lazar

Phineas Packard
#12
Turning AD=OFF will help but my guess is that you have miss-specified your model. In the initial output is a set of matrices relating to difference components of the model. Check that they match what you intended. Also check that you have an identified model (do you have degrees of freedom).

If you can not spot the issue I would guess that you might find it helpful to work with a package that specifies models using more natural language rather than specification of matrices. A good choice would be the lavaan package in R which is free.
 
#13
I would like to express my appreciation and respect for your valuable advises, time and efforts.
I tried your remarkable notes. I think that the model is not Identified and the indices of Good fit don't match the required cut off points at The CFA level. so Could you kindly help me to resolve this serious problem? Frankly, I couldn't use another program because This software was recommended by my supervisor to finish analyzing my Master Thesis data.... I am in a serious bottleneck problem
 
#14
Re: LISREL - Model does not converge, Matrix not positive definite
I would like to express my appreciation and respect for your valuable advises, time and efforts.
I tried your remarkable notes. I think that the model is not Identified and the indices of Good fit don't match the required cut off points at The CFA level. so Could you kindly help me to resolve this serious problem? Frankly, I couldn't use another program because This software was recommended by my supervisor to finish analyzing my Master Thesis data.... I am in a serious bottleneck problem
 

Lazar

Phineas Packard
#15
If you are getting fit for the your model then the model is identified (at least globally). Are the residuals all positive? Likewise, are the latent correlations all <=1? Again check the output closely. The most likely cause of your problem is that you have estimated a path you did not mean to.
 
#16
If you are getting fit for the your model then the model is identified (at least globally). Are the residuals all positive? Likewise, are the latent correlations all <=1? Again check the output closely. The most likely cause of your problem is that you have estimated a path you did not mean to.
thanks for your valuable note. I am trying to gain an insight into the output... R2 is very low --LESS THAN 20% in some cases>> is it better to remove paths associated with its observed variables; with low R2.>> could you kindly send me papers or any thing that could help me regarding this point? or any notes that could help me to determine which paths that must be removed at CFA Level of LISREL... I would like to ask you also about the composite reliability and VE variance extracted ? did we compute it? or is it given in the output file? why didn't it appear to me? because, the model don't converge!!! or what?
 
#17
thanks for your valuable note. I am trying to gain an insight into the output... R2 is very low --LESS THAN 20% in some cases>> is it better to remove paths associated with its observed variables; with low R2.>> could you kindly send me papers or any thing that could help me regarding this point? or any notes that could help me to determine which paths that must be removed at CFA Level of LISREL... I would like to ask you also about the composite reliability and VE variance extracted ? did we compute it? or is it given in the output file? why didn't it appear to me? because, the model don't converge!!! or what?