You are here

Result for Univariate and Multivariate Genetic Models

2 posts / 0 new
Last post
JudyP's picture
Offline
Joined: 04/29/2019 - 14:05
Result for Univariate and Multivariate Genetic Models

Hello,

I have a general question about univariate and multivariate genetic models. Do the ACE estimations roughly align with each other when using univariate and multivariate models (i.e., Cholesky Decomposition, Independent Pathway Model, Common Pathway Model) for the same variables? For example, if the univariate model for variable 1 shows no additive genetic factor but significant C and E, should the ACE estimation for variable 1 from the multivariate models show no A when compared with other variables? Sometimes, the best fitting model contains a latent genetic variable in an independent/common pathway model even for variable 1 with no A tested in the univariate model. Is it acceptable? Or is there something wrong with the analysis?

I get positive AICs for my univariate analysis, but negative AICs for my multivariate analysis on the same variables. Are the negative AICs acceptable?

Thanks,
Judy

AdminNeale's picture
Offline
Joined: 03/01/2013 - 14:09
It happens. Trust multivariate

The multivariate model is using information from cross-twin cross-variable covariances. Sometimes these show more of a pattern of rMZ_AB > rDZ_AB than the univariate rMZ_AA vs. rDZ_AA comparison. See https://vipbg.vcu.edu/vipbg/Articles/psychological-bulimia-1992.pdf for an example. In that case, there was more information on one of the two measures than the other, due to different thresholds for affected status. The more informative measure, depression, had an AE architecture, whereas bulimia was close to CE. The cross-twin bulimia-depression correlations echoed the pattern of the depression correlations, so the model "decided" the reason the two traits correlated was more likely due to A.

If the level of measurement is the same, and if there are thresholds they are also equivalently spaced in the two variables, then this source of difference in precision will not be the likely source. However, sample size for different traits may differ with FIML, which allows for missingness on a case-by-case basis.