Hi all,
I have a question about bivariate cholesky ACE model. For model fitting, I compared the reduced models (selectively dropping a21, c21 and e21) to the full bivariate ACE model and the best-fitting model was selected based on lowest AIC and nonsignificant likelihood ratio chi-square test.
However, I noticed that the best-fitting model’s ACE estimates do not seem to match with what I obtained from fitting each of the two variables in univariate ACE models. For example, in univariate model fitting I got:
First trait: C = 60% (CI: 0.46-0.71) and E = 40% (CI: 0.29-0.54).
(Full ACE model: A = 5% (CI: NA-0.60), C = 55% (CI: 0.04-0.71), E = 39% (CI: 0.27-0.54))
Second trait: A = 82.6% (CI: 0.74-0.88) and E = 17.4% (CI: 0.12-0.26)
(Full ACE model: A = 72.4% (CI: 0.31-0.88), C = 10% (CI: 0-0.51), E = 17.4% (CI: 0.12-0.26))
The best fitting bivariate model for the two traits is ACE (drop c21 and e21) and the estimates are as follows:
confidence intervals:
lbound estimate ubound note
ACEnce21.SA[1,1] 1.261549e-01 0.2769508 0.6573700
ACEnce21.SA[2,2] 3.757661e-01 0.7011000 0.8840258
ACEnce21.SC[1,1] 3.156919e-02 0.3914759 0.5619521
ACEnce21.SC[2,2] 1.939172e-42 0.1306183 0.4428392
ACEnce21.SE[1,1] 2.354506e-01 0.3315733 0.4584370
ACEnce21.SE[2,2] 1.133622e-01 0.1682817 0.2513258
ACEnce21.rA[2,1] NA -1.0000000 -0.6036494 !!!
ACEnce21.corrA[1,1] -5.549520e-01 -0.4406475 -0.3113180
(Full BivACE model)
confidence intervals:
lbound estimate ubound note
BivACE.SA[1,1] 1.669727e-03 0.17771597 0.68945684
BivACE.SA[2,2] 2.799714e-01 0.66712995 0.88409716
BivACE.SC[1,1] 2.803414e-03 0.48075051 0.69603875
BivACE.SC[2,2] 2.393405e-43 0.16523408 0.54170390
BivACE.SE[1,1] 2.416883e-01 0.34153352 0.47200335
BivACE.SE[2,2] 1.122780e-01 0.16763597 0.25194368
BivACE.rA[2,1] NA -1.00000000 -0.99750231 !!!
BivACE.rC[2,1] -1.000000e+00 -0.24981942 1.00000000
BivACE.rE[2,1] -4.071271e-01 -0.17633429 0.06402325
BivACE.corrA[1,1] -6.319900e-01 -0.34432491 -0.02917323
BivACE.corrC[1,1] -3.865143e-01 -0.07041025 0.20042770
BivACE.corrE[1,1] -1.118121e-01 -0.04219267 0.01531362
As can be seen, the bivariate ACE estimates are rather different for trait 1 especially. In view of this, should I trust the ACE estimates yielded by the univariate ACE modelling more than the ones obtained from bivariate ACE modelling?
Besides that, the rA (genetic correlation) between the two traits appear to be -100%, which seems quite drastic. Can I actually put much weight on the results and interpret the findings accordingly?
Will greatly appreciate the input from the various OpenMx experts on this forum.
Printer-friendly version
I would recommend using the bivariate estimates. The rationale is that trait 2 is providing new, if indirect, information about the genetic variance component. Essentially, it's saying that the pattern of cross-trait cross-twin correlations is inconsistent with such a low heritability for trait 1. This is more likely to happen when the measurement precision differs between the variables. An imprecisely measured trait is more likely to have its univariate estimates 'swayed' by multivariate data than one that is accurately measured.
On the whole, I'd rather see results of full models - ACE or ADE than certain submodels. They are easier to incorporate into meta-analyses, apart from anything else.
Hi Michael,
Thank you very much for your advice. As I did bivariate modelling to compare trait 1 with 10 task measures, I noticed that the bivariate ACE estimates for trait 1 varied quite considerably when modeled with different task measures, with trait 1's A estimates varying from 11%-34%. I'm not quite sure how should I decide which of these estimates is more "accurate"? Could it be the one with the narrowest CI?
Another question I have is this: when I did bivariate full modelling for trait 1 and trait 2, trait 1's A estimate is 14% whereas trait 2's is 41%. When I looked at the squared standardized path a matrix, a22 squared is 0, which means that 0% of the genetic variance of trait 2 is due to specific gene action unrelated to trait 1's genetic variance. I find it quite hard to make sense of this, because given that trait 2's A estimate is considerably greater than trait 1, how is it possible that all its A estimate is shared with trait 1? If my genetic or environmental correlations' CI include 0, does it mean that these correlations are unreliable?
Thanks and best regards,
Yi Ting
The best estimate is probably that from the truly multivariate model, rather than any of the bivariate ones. In principle it should have the narrowest CI's, and would avoid having to choose one.
It is entirely possible for a factor to have greater effect on one trait than another. This is regularly observed in a factor analysis, in which loadings vary in size. Your data seem to be telling you that the same genetic factors affect both traits but do so with different effect sizes.