You are here

Power estimation for the detection of rG and rE

2 posts / 0 new
Last post
Marianna's picture
Joined: 10/20/2016 - 19:09
Power estimation for the detection of rG and rE

How can one estimate the power for the detection of significant rG and rE in multivariate Cholesky models?

Specifically, the analysis employed a trivariate Cholesky (AE providing the best fit, with all significant rGs, and one significant rE). The CIs have been calculated for all estimates. The sample is on the small side: 200 same-sex pairs (half MZ, half DZ) and may have been underpowered for the other smaller rEs but I’m not sure what size effect I had enough power to detect.

I can estimate power for variance components in univariate decomposition models but cannot find anything specific to multivariate models or correlations. Any advice or script to get a power estimate specific to correlations would be super helpful!

Thanks in advance,

AdminNeale's picture
Joined: 03/01/2013 - 14:09
Brad Verhulst is putting

Brad Verhulst is putting finishing touches on a paper on this topic. For the time being, with the Cholesky approach, it isn't very simple to test the power to detect rG or rE in a trivariate model. What you can do, however, is to see how much the fit deteriorates if you add a non-linear constraint to force the rG or rE to equal a particular value.

# Example code to add a constraint that rG is zero
modelWithConstraint <- mxModel(modelWithoutConstraint, mxConstraint(0.0 == cov2cor(A)[1,2]) )

The difference in fit between the two models can be used as a non-centrality parameter for the chi-squared distribution.