Discontinuous Modeling
Message/Author
 Nicole Johnson posted on Tuesday, December 17, 2013 - 11:54 am
Below is the model syntax:
MODEL:
int by precap@1 pscap@1 ptcap@1 p3cap@1 p6cap@1;
Slope1 by precap@-1.5 pscap@-.5 ptcap@.5 p3cap@1.5 p6cap@2.5;
Slope2 by precap@0 pscap@0 ptcap@.5 p3cap@1.5 p6cap@2.5;
[precap@0 pscap@0 ptcap@0 p3cap@0 p6cap@0];
[Int* Slope1* Slope2*];
Int with Slope1 Slope2 PStxu18; Slope1 with Slope2;
Slope1 on therapy;
Slope2 on therapy;

We keep receiving the warning
WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE
DEFINITE.

We examined tech4 and read that this could be due to a small sample size (n = 60) or due to skewed data (data after shelter stay or Slope 2 (highlighted error), is skewed). Any help would be appreciated.
 Bengt O. Muthen posted on Tuesday, December 17, 2013 - 3:54 pm
What are the TECH4 covariance and correlation matrices?
 Nicole Johnson posted on Friday, December 20, 2013 - 8:58 am
ESTIMATED COVARIANCE MATRIX FOR THE LATENT VARIABLES
INT SLOPE1 SLOPE2 PSTXU18 THERAPY
________ ________ ________ ________ ________
INT 466.783
SLOPE1 169.202 133.903
SLOPE2 -139.536 -104.654 27.688
PSTXU18 -88.192 0.000 0.000 6321.515
THERAPY 0.000 12.359 -13.792 0.000 7.127

ESTIMATED CORRELATION MATRIX FOR THE LATENT VARIABLES
INT SLOPE1 SLOPE2 PSTXU18 THERAPY
________ ________ ________ ________ ________
INT 1.000
SLOPE1 0.677 1.000
SLOPE2 -1.227 -1.719 1.000
PSTXU18 -0.051 0.000 0.000 1.000
THERAPY 0.000 0.400 -0.982 0.000 1.000
 Bengt O. Muthen posted on Friday, December 20, 2013 - 4:03 pm
You have two Tech4 correlations greater than 1. This is a problem. Perhaps you don't need slope2, or perhaps you need to center at the first time point for both slope1 and slope 2.
 Nicole Johnson posted on Friday, December 20, 2013 - 5:06 pm
Thank you for the suggestions. We will try them.

Best,
Nicole
 Nicole Johnson posted on Monday, December 23, 2013 - 12:46 pm
Dr. Muthen,

I tried your suggestions; however, we still receive the same warning message. Any additional suggestions? Also, what does it mean that we have two correlations greater than 1? What could this be a result of?

Thank you again,
Nicki
 Nicole Johnson posted on Monday, December 23, 2013 - 2:14 pm
Also, I read in a previous post that this error may be due to a "preponderance of zeros"; which starting at our post-treatment time point we have quiet a few zeros (8 at post treatment, 16 at 3 months post treatment, and 12 at 6 months post treatment). If this is the issue, what can we do about this?

Thank you,
Nicki
 Bengt O. Muthen posted on Monday, December 23, 2013 - 5:06 pm
Latent correlations greater than 1 may be due to channeling too much of the observed variable correlations through the latent variables instead of directly, or due to specifying too many latent variables. You can try to add correlations among the residuals of the outcomes, or delete one of the latent variables.

Preponderance of zeros is when you have a large percentage of zeros - say at least 20%.
 Nicole Johnson posted on Tuesday, December 24, 2013 - 7:29 am
Great! Thank you for the quick response. We do have a preponderance of zeros in 2 of our time points. Would you suggest doing anything to rectify this? Perhaps recoding the zeros to .001?

Best,
Nicki
 Bengt O. Muthen posted on Tuesday, December 24, 2013 - 5:18 pm
Recoding doesn't help. You may have to ignore it since the solution of two-part modeling would lead you to quite complicated modeling given that you also have piece-wise.