Author Topic: Covariance matrix is not positive definite  (Read 618 times)

castleca

  • Newbie
  • *
  • Posts: 8
    • View Profile
    • Email
Covariance matrix is not positive definite
« on: April 06, 2017, 06:06:22 PM »
Hello,

I am trying to run a between-item multidimensional partial credit model, with 60 items, 3 dimensions, and 369 cases. The model appears to converge, and it gets 98% through computing empirical errors when it crashes and produces this error message: "Covariance matrix is not positive definite. This error is fatal." I have reexamined my command file and data files several times with no luck - I simply cannot figure out what is causing this problem. Can anyone point me in a helpful direction?

Thank you,

Courtney

castleca

  • Newbie
  • *
  • Posts: 8
    • View Profile
    • Email
Re: Covariance matrix is not positive definite
« Reply #1 on: April 06, 2017, 08:27:25 PM »
Hi again - when I try to run the same dataset as a within-item model, I get the following message (at roughly the same point, 98% of the way through computing errors): Element 75 in variance/covariance matrix is negative. Check model identification: -18521.1132. The same message appears for Element 77 of the variance/covariance matrix. Could this be related to the previous issue with a between-item model? Any assistance is greatly appreciated.

Thank you,
Courtney

dan_c

  • Administrator
  • Jr. Member
  • *****
  • Posts: 84
    • View Profile
    • Email
Re: Covariance matrix is not positive definite
« Reply #2 on: April 06, 2017, 09:19:58 PM »
Can you create a small working example to replicate the problem?

https://conquest-forums.acer.edu.au/index.php?topic=1239.msg1475#msg1475

castleca

  • Newbie
  • *
  • Posts: 8
    • View Profile
    • Email
Re: Covariance matrix is not positive definite
« Reply #3 on: April 07, 2017, 10:25:11 PM »
I am not sure it is possible for me to easily recreate the problem with a generic example, as I don't understand what is causing it in the first place. Could I send along my syntax and data file, to have you take a look?

Thank you for your attention,
Courtney

dan_c

  • Administrator
  • Jr. Member
  • *****
  • Posts: 84
    • View Profile
    • Email
Re: Covariance matrix is not positive definite
« Reply #4 on: April 11, 2017, 08:14:52 AM »
The benefit of producing a minimum working example is twofold - the community gets to see a example it can replicate and potentially solve, and producing the example often weeds out errors related to messy real-world data and syntax. Could you edit tutorial 7 to add partial credit and cross loadings and see if you can replicate your error?

Note, the other common issue related to identification of within-item models:

Quote
If ConQuest is being used to estimate a model that has within-item multidimensionality, then the set command argument constraints=cases must be provided. ConQuest can be used to estimate a within-item multidimensional model without constraints=cases. This will, however, require the user to define and import a design matrix. The comprehensive description of how to construct design matrices for multidimensional models is beyond the scope of this manual.
ConQuest Command Reference p88.