You are here

Univariate Sex Limitation Model Test of Rc

7 posts / 0 new
Last post
newsomjr's picture
Offline
Joined: 10/13/2009 - 01:27
Univariate Sex Limitation Model Test of Rc
AttachmentSize
Binary Data Sex Limitation to Test Rg and Rc.R7.95 KB

Hi all,

I am still learning OpenMx, and have tweaked a script from the 2012 Boulder workshop. I'd like to test whether the same genetic and shared environmental factors are operating in males and females. I believe I am doing it correctly, but was hoping someone might be able to verify? I have attached my script, and would greatly appreciate any assistance!

neale's picture
Offline
Joined: 07/31/2009 - 15:14
Looks good

Hi Jamie

Your code looks good to me, as long as nv=1. There are problems with multivariate models of this type, and the way the code is written such models would (thankfully) not run for nv > 1.

Do the results seem reasonable? Taking a look at the 5 observed rMZ, rDZ and rDZOS correlations, do the estimates make sense? This is always an important step in model-fitting.

Cheers
Mike

newsomjr's picture
Offline
Joined: 10/13/2009 - 01:27
Thanks and One More Question

Hi Mike,

Thank you for taking the time to look that over. In my script, nv=1. I did check the correlations prior to estimating the models, and my results seem to be consistent with those. All of this has led me to another question, though. My next step would be to test if the magnitude of the effects differ between males and females. If I find that fixing rg and rc to 0 does not reduce the fit of the models (when tested separately), is it appropriate to fix both of them to 0 before testing if the paths for males and females can be equated?

Thanks again for your help!
Jamie

neale's picture
Offline
Joined: 07/31/2009 - 15:14
Usually rg=rc=1

The typical model would be either rg=rc=1 or allow one of them to be freely estimated if this model fits better. I think both rg=0 or rc=0, or both is a bit strange. It says that completely different factors influence variation in males & females. More usually we'd be heading towards a model where Am=Af, Cm=Cf and Em=Ef, plus rg=rc=1, i.e., no sex differences in variance components at all. One in which Am=Af, Cm=Cf and Em=Ef but rg<1 or rc<1 would be a bit strange, as it would be some sort of coincidence that the sources of variance have the same effects in the two sexes, while actually being completely different factors.

Cheers
Mike

newsomjr's picture
Offline
Joined: 10/13/2009 - 01:27
I think I understand where I

I think I understand where I was going wrong now. Glad I posted my question! Again, thank you for your time and assistance!

Jamie

newsomjr's picture
Offline
Joined: 10/13/2009 - 01:27
I have re-estimated some of

I have re-estimated some of my models, which has led me to another peculiar finding. The estimates for rg are equal to the lower boundary. Fixing rg=1 did not reduce the fit of the model, but standard errors for af, am, and cm returned NaN. I believe the estimate for af would be negative (if not bounded to 0-1) and the estimate for am is positive. Indeed, when I relaxed the lower boundary of af it did produce a negative value and the standard errors were correctly estimated. I believe this is causing the issue with rg equaling the lower boundary, but I am not sure how this can be addressed. Based on the intraclass correlations, it looks as though af=0 and cm=0. Is this possible and/or reasonable?

neale's picture
Offline
Joined: 07/31/2009 - 15:14
Reasonable

Yes, indeed it is reasonable. Boundary conditions can be difficult to overcome, and when parameters are at bounds the Hessian can be difficult to estimate so the Standard Errors come out NaN.

In the old days, we would use software that didn't allow boundary constraints, and effect them by specifying two paths with a dummy variable in between:

... sqrt(d) ............. sqrt(d)
A ---------> Dummy ---------> B

This works (although the relationship with standard error changes, see [1]). I don't know whether it reduces the boundary problem with std errors, but I suspect not.

[1] Neale, M.C., Heath, A.C., Hewitt, J.K., Eaves, L.J. Fulker, D.W. (1989) Fitting genetic models with LISREL: hypothesis testing. Behavior Genetics 19: 37-49.