CNTRL BY u1 u2 c3; PLSR BY u7 u8 u9; SLFRLZTN BY u10 u11 u12 ; CASP BY CNTRL PLSR (1) SLFRLZTN (2); [CNTRL@0PLSR@0SLFRLZTN@0];
My indicators are categorical and I'm using the WLSMV estimator.
However, I'm getting a non positive definite covariance matrix.
I think that this is due to the fact that CASP, the second-order latent variable, is strongly related to SLFRLTZN, one of the first order latent variable. The estimated correlation is a bit higher than 1.
Moreover, the completely standardized factor loading for SLFRLZTN is a bit higher than 1 (STDXY loading=1.026) and it's residual variance is very small, negative and non-significant (residual=-.052, p=.999).
I read on the forum that I could fix the residual variance to 0. Is this what you would advice in my case?
Theoretically, would then this mean that CASP = SLFRLZTN ? In other words, that my second order latent variable is redundant?
When factors correlate one or greater, it means that they are not statistically distinguishable. You might try an EFA on your first-order factors or look at modification indices for cross-loadings to see if the problem is with the specification of the first-order factors.