You are here

Latent Interactions?

11 posts / 0 new
Last post
pdeboeck's picture
Joined: 08/04/2009 - 15:55
Latent Interactions?

I think this can be done, but I'm having problems seeing a clean solution. Is there a way with OpenMx to predict an outcome using the interaction of two latent variables?

Ryne's picture
Joined: 07/31/2009 - 15:12
Hey Pascal, The standard

Hey Pascal,

The standard citation on latent variable interactions is a Psycometrika article by Klein & Moosbrugger (2000). Mplus implements some form of this, though I recall it does some variation on the EM algorithm. Could be wrong, though. I'd imagine implementation is possible in OpenMx, but no one has done it yet to my knowledge.

neale's picture
Joined: 07/31/2009 - 15:14
There are some general

There are some general approaches to latent variable interactions which make the specification of the non-linear constraints much simpler (basically two matrix algebra expressions are required). I wrote about them here:

Neale, M.C. (1998) Modeling interaction and nonlinear effects with Mx: A general approach. In: G. Marcoulides & R. Schumacker Interaction and Non-linear Effects in Structural Equation Modeling pp. 43-61. New York: Lawrence Erlbaum Associates


pdeboeck's picture
Joined: 08/04/2009 - 15:55
Thanks for the pointers! I'll

Thanks for the pointers! I'll look into both articles. Thanks!

jam86hk's picture
Joined: 05/21/2010 - 03:22
Dear all, Among many latent

Dear all,

Among many latent interaction methods have been and being developed, the Kenny-Judd (1984) method (also see Dr Neale's (1998) chapter) is available for OpenMX. Marsh, Wen, and Hau (2004) unconstrained approach helps us to have less non-linear constraint commands.

For the LMS approach, if OpenMX provides numerical integration, Monte Carlo integration, or Gauss-Hermite quadrature integration, it is possible to be implemented by OpenMX.

Heining Cham

mooijaarta's picture
Joined: 01/30/2011 - 11:47
I think the methods with

I think the methods with product indicators (methods mentioned by Heining Cham, and others) are outdated now. There are other methods available in other packages, like Mplus and EQS.

However, for me it is still unclear how the ML method (with all kinds of its variations) can be implemented in openMx. Is there someone who has done this? And is the dev. team planning to implement this method?

Furthermore, in "Structural Equation Modeling" 2010, pages 357- 373, Mooijaart and Bentler propose a method using higher order moments in stead of using only means and covariances. This method is compared to the method of Mplus and seems to be promising also in cases with nonnormally distributed predictors.

I'm not sure if analyzing higher order moments can be done easily in openMx.

Ab Mooijaart

neale's picture
Joined: 07/31/2009 - 15:14
A couple of points: In

A couple of points:

In classic Mx, I had some success with implementing Gauss-Hermite quadrature using a mixture distribution (weighted with fixed weights). That should be possible with OpenMx also.

It is possible to specify user-defined fit functions using OpenMx. mxAlgebraObjective() would be the way to go. I recall Peter Bentler publishing a paper on higher moments early in the 80's (BJMSP if I recall correctly) and the matrix algebra was a bit heavy. However, I doubt whether it's anything that OpenMx can't handle.


neale's picture
Joined: 07/31/2009 - 15:14
Evaluate on Grid?

I think it should be possible using mxEvaluateOnGrid, but have not developed a script for it yet. Joshua Pritikin may do that sometime in the near future, however.

pehkawn's picture
Joined: 05/24/2020 - 19:45
**Recent developments?**

Since this is an old thread, has there been more recent developments on this topic? I'm fairly new to OpenMx and SEM in general, so any articles, scripts, tutorials etc. that demonstrates how to implement latent interactions in OpenMx would be highly appreciated.

jpritikin's picture
Joined: 05/24/2012 - 00:35
no progress so far

We have some experimental code, but nothing I can recommend at this time.

AdminNeale's picture
Joined: 03/01/2013 - 14:09
Manual specification of numerical integration

Hi Pål

I have a script which goes as far as specifying all 2-way interactions among latent variables, so F1 and F2 yield F1^2, F1F2 and F2^2 factors. I based it on a script for estimating non-normal latent densities (Schmitt J.E., Aggen S.H., Mehta P.D. Kubarych T.S., Neale, M.C. (2006) Semi-Nonparametric Methods for Detecting Latent Non-normality: A Fusion of Latent Trait and Ordered Latent Class Modeling. Multivariate Behavioral Research 2006; 41(4): 427-443)

So far it is not tested beyond seeing if it looks like it can estimate a path from the latent product variable F1^2. It's not terribly fast, especially if you crank up the number of integration points.

I don't consider this an optimal approach, but it may be functional for some purposes. Note also that the normal theory likelihood function doesn't have any non-normality bits in it, so it is likely inefficient compared to full-information estimators.

File attachments: