Optimizers

I have conducted a large scale numerical simulation study to see small
sample properties of SEM and then have the following question:
I would appreciate it if you could teach me why MLEs in a CFA model are
different between analyses of raw data and covariance matrix data in
OpenMx package of R. Are the optimization methods employed different?
The maximum difference I encountered is 0.084267.

When I compare the results, I adjusted their scales, i.e., multiplied by
sqrt(N/N-1) for factor loadings and by N/N-1 for error variances.

Hello,

I just found in a run that the calculated Hessian has a negative eigenvalue (which also result in NAs for some standard errors) and some of the gradients seems large. However, the convergence code is 0 so no error or warning is displayed. Should a warning be displayed in this case?

I tried another starting value and this problem goes away, with objective function decreases about 0.28, indicating the first run does not reach the minimum.

The code and data are in the attachment.

Thanks.

- Hao

Hello!

Today, I felt like fixing all my free parameters and call mxRun() on my model. I expected the optimizer to converge instantly or rather not to be invoked. Instead, it seems the backend is running into an infinite loop somehow. Maybe the backend needs to check the condition of no free parameters in the model.

I'm trying to spec out a least-squares based optimizer using the mxAlgebraObjective, but I've been unable to include data in an algebra. Is that possible/feasible, or do I need to move to mxRObjective?

I have run into an interesting problem. A model (which is admittedly of low stability) that I used in class today ran on the Mac (with warnings, but reasonably close estimates) but not on the PC (non-invertible). Both are Intel processors, both running R 2.9.2 and OpenMx 0.2.2. The model is of simulated data with boundary constraints on residuals because variance explained is very high.

I attach the script and data as a test case for the limits of similarities between architectures.

Just got this crash running a FIML version of the front page factor model without updating the object to FIML

> data(demoOneFactor)
> factorModel <- mxModel("One Factor",
+ mxMatrix("Full", 5, 1, values=0.2,
+ free=T, name="A"),
+ mxMatrix("Symm", 1, 1, values=1,
+ free=F, name="L"),
+ mxMatrix("Diag", 5, 5, values=1,
+ free=T, name="U"),
+ mxAlgebra(A %*% L %*% t(A) + U, name="R",
+ dimnames = list(names(demoOneFactor),
+ names(demoOneFactor))),
+ mxMLObjective("R"),

Here's an interesting one...

So I've recently been working on converting an mx script to OpenMx (file attached). I might be doing something I shouldn't, but the script caused R on my Mac to close w/out explanation or warning. Everything goes fine until I hit "mxRun." I ran this under Windows XP and got the following error:

> EstModel <- mxRun(CoupledModel)
Running CoupledModel
Error in mxRun(CoupledModel) :
BLAS/LAPACK routine 'DGEMM ' gave error code -13

i am following up here on the discussion previously started on the email list. i'm still unsure, after perusing the code and the npsol manual as to weather the raw likelihood or the -2 log likelihood is being returned from the optimizer. i'm wondering if anyone might know where i could find this info? or how i might reasonably search for it in the code (?)