Attachment | Size |
---|---|

bivariate-BMI-hypertension.R | 22.46 KB |

Hello, everyone:

I'm running a bivariate twin model to estimate the correlation between a continuous variable (BMI) and a binary variable (CHD).

I have two questions:

First, i fund the means between MZ and DZ twins are different, do i need to constrain them with the same value in the ACE model;

Second, the p value between the saturated model and ACE model was＜0.05, what does this mean. Can i use ACE model to estimate the genetic and environmental correlation.

Thank you so much for your help.

Is there anyone could help me to explain the meaning of comparison between saturated model and ACE model.

I used univariate model to estimate the heritability of BMI, which was treated as a continuous variable. Unfortunately, the P value (comparison between saturated and ACE model) is close to 0.

When I treated BMI as a binary variable, the P value between these two models was not significant. But I fund the heritability between the continuous and binary model were not vary different.

Therefore, I’m confused by the comparison between saturated and ACE model.

What does P value＜0.05 mean, the ACE model decreased the fitness of the saturated model? The ACE model was not suitable for my variable? Is it necessary that the fit of ACE model is not different from the saturated model?

I just don’t know if I can fit ACE model for the further analysis.

Will the difference between saturated and ACE model, have influence on my estimation. And what I can do to minimize the influence?

I’ll appreciate with the help!

Hi

The results you have, significantly poor fit with BMI treated as a continuous variable, but ok fit when BMI is dichotomized might simply be a result of throwing away information in the BMI variable - less information with which to detect a departure from the saturated model. Better, I think, is to try to track down the source of the failure of the continuous BMI modeling.

The difference between the saturated model and the fitted model is that in the saturated case every statistic gets its own parameter - T1 and T2 means and (co)variances giving 14 parameters for each of the MZ and DZ groups. The fitted model constrains many of these to be equal - MZ/DZ and T1/T2 means and variances are all expected to be equal, but they aren't. You already alluded to different means for MZ and DZ twins, which in my experience is unusual. Also possible are differences in variances. Problems of this sort can emerge when the data are not bivariate normal. We can ignore the binary variable in this respect, because it isn't possible to distinguish means and variances. But BMI is known to be skewed in many samples, so it is possible that a log or a log-log transformation could help approximate multivariate normality and yield a better fit to the model.

I think it would help to know a bit about how the sample was ascertained - the MZ DZ mean differences are a bit of a worry. Absent the data, I cannot investigate further. Take a look at mxGetExpected(model$MZ,'covariance') and compare it with cov(mzdata, use='complete.obs') to get some clues about where the misfit is.

HTH!