Jon Heron posted on Tuesday, February 06, 2018 - 12:19 am
we've estimated a bifactor model with one general and four specific factors. we now want to consider a series of risk factors for these traits + have been having much discussion about how to maintain orthogonality for the five factors. This turned into a conversation about whether orthogonality or RESIDUAL-orthogonality was needed. (i.e. should the factors be conditionally or unconditionally independent).
It feels to me that residual orthgonality is what is needed, since if X predicts both y1 and y2 then these y's much surely be correlated in a model that excludes X. As you say in an earlier post, this residual orthogonality is most easily introduced using "model = nocovariances".
The problem is that its customary, at least in epidemiology, to build up ones models by introducing confounders. If I'm not mistaken, "model = nocovariances" is going to rescale the model at each step rendering any observed change in the effects of X hard to interpret.
Can you see an obvious solution to this? Have been wondering about using plausible values but that feels a little excessive and might be hard to get past the reviewers.
Yes, I agree that residual orthogonality is all that's needed. The Model=nocovariances should affect only Psi (or Theta) - which would refer to residual covariances - so I don't think there is a problem of rescaling. In any case, you can set up the zero covariances you want "by hand" instead.
Jon Heron posted on Tuesday, February 13, 2018 - 10:50 am