Attachment | Size |
---|---|
mi.r [6] | 380 bytes |
sorbom.r [7] | 2.16 KB |
sorbom.txt [8] | 2.72 KB |
I am starting a new thread on this but see
https://openmx.ssri.psu.edu/thread/831
for background.
I am still attempting to replicate the results from Sorbom's paper using the code posted by Rine and some of the new options to mxRun outlined in the old thread. I am attaching three files.
mi.r contains my code for modification indexes
sorbom.r creates the correlation matrix from one of the examples in Sorbom's paper and runs it
sorbom.txt is the resultant output showing the value of mi I get,
the values of the gradient, and the principal diagonal of the Hessian. It also has my session info.
Since I am not quite sure what combination of Major iterations and useOptimizer to use I have done it for several choices. As you can see I am getting zero values for mi. A number of possibilities occur to me in non-increasing order of my prior probability (a) I have failed to implement Rine's code properly (b) I have made some other error in specification (c) the changes made to the optimizer do not support this sort of calculation.
Even more worrying is that if you re-run sorbom.r several times in the same session you do not get the same results. I can understand that it OpenMx often needs to start from random values but that does not seem appropriate here although i may have misunderstood.
And a final query about degrees of freedom. I get the same chi-sqaure for the model without additional paths as Sorbom but the degrees of freedom differ by 7 which is the size of the correlation matrix (OpenMx 6, Sorbom 13)