Mon, 04/19/2010 - 06:14

Question from a novice in OpenMx:

I want to convert an ACE script into an ADE script.

Am I correct in that, basically, the only thing I need to change is the mxAlgebra of the DZ group? Like this:

from:

mxAlgebra( expression= rbind ( cbind(A+C+E , 0.5%x%A+C),

cbind(0.5%x%A+C , A+C+E)), name="expCovDZ" )

into:

mxAlgebra( expression= rbind ( cbind(A+C+E , 0.5%x%A+0.25%x%C),

cbind(0.5%x%A+0.25%x%C, A+C+E)), name="expCovDZ" )

Yes, that would do it. However, I strongly recommend renaming all the C matrices to D, and the c path to d, just to avoid confusion when interpreting the output. It's not such a problem when you first do it, but can be a real head-scratcher when you look at it again some months/years down the road.

Exactly how should you decide when to use ACE or ADE? I the literature I found some calculate r(mz) and r(dz) values, and if r(mz) is at least twice as large as r(dz) then they only calculate ADE models. Based on r(mz) and r(dz) values should I conclude which model to use, or should I run ACE and ADE also and comapre them using the AIC and BIC?

I have some data where r(mz) values are more than twice as big as r(dz) values, and running both ACE and ADE models with no covariates, and also with sex and age as moderators, the ACE model with no covariates has lower AIC and BIC values then the ADE model with covariates. How should I conclude this? Should I only compare those ACE ADE models to each other which have the same covariates? Should I even run ACE ADE models without sex and age as covariates?

I don't think it is good statistical practice to inspect the data and decide which model to fit ahead of time. In part this is because this procedure capitalizes on chance and makes whichever model is chosen seem to fit better than it should on average. In addition, there are times when the correlations can be misleading as to which model is better, particularly if there are total variance differences between MZ and DZ pairs.

It's probably ok to compare the full set of models - with and without covariates and find whichever has the lowest AIC & (hopefully) BIC.

I see the point. Thanks!