Allow individual dzCr settings in multivariate umxACE()

[marco-deppe]

Might be more of a question than an issue - excuse me if this is the wrong place for this, not entirely sure how this works!

In a multivariate umxACE(), one might consider some variables that, according to their MZ/DZ correlations, would require D, and some variables that would require C to be estimated in order to show the 'full' picture (and improve fit). Is it possible to set individual dzCr (e.g., 0, .25, or 1) for different variables?

If not, would it be possible to implement this? E.g., allow dzCr = c(.25, .25, 1, 1, .25, .25), corresponding to the variable list provided? If I am not mistaken, dzCr currently only allows a specification for the entire model, not for single variables.

Any pointers to existing options to achieve this are much appreciated.

[tbates]

The trouble with this, though, is that in multivariate models, allowing some traits to have D but not C creates a problem: with only one or the other, trait covariance can be explained. But if one has D and the other C, that can't happen. So in practice, if there really is both D and C, people let A soak up the D, and model C using C.

FYI, umxACE re-uses the C matrix for D in D models, so coping with ACDE would be additional work.

[mcneale]

"in multivariate models, allowing some traits to have D but not C creates a problem: with only one or the other, trait covariance can be explained". Assuming that can means cannot hear, I don't actually think that this is a problem. Trait covariance can still occur with rA or rE, just not rD or rC. In the event that these latter two components are large for the two traits, and A and E small, the model prediction would be low correlation between the traits (the upper limit of the covariance being a1a2 + c1c2 if these are the path coefficients for the two traits' A and C effects). If the correlation turned out to be large, it would reject the C for this trait and D for that trait model. This in turn would add support for D and C both being in action, but their effects somewhat canceled out due to their confounding. In general, pure D no C, and pure C no D, seem unlikely models and a more nuanced take on what these estimates are is needed. IMO.