Anonymous posted on Wednesday, May 05, 2004 - 12:27 pm
I am using mPlus to analyze a two-level model consisting of nine latent variables (three of them are latent indicators of a second order variable), measured with a total of 32 items. All six latent variables are allowed to correlate. It is the same model in the within (n=460) and between part (n=105).
I regress one of the latent variables (A) in the between-part on an observed between-level variable (B). The correlation between the observed between-level variable (B) and one of the other latent variables (C) in the between-part of the model is estimated to be .35, t=2.36. However, when I simplify the model by excluding four latent variables (and their indicators) from the model (both within and between), the correlation between the same two original variables (B and C) are reduced to a non-significant .25, t=1.63. (It is exactly the same data that are analyzed, because I use the output of the big model as input for the smaller model.)
What might be the explanation that the correlation estimate varies depending on the size of the model, when the whole model is a correlation model? Might it be that I have too many observed variables in relation to the sample size in the between level, or might it be other explanations? I would appreciate your comments!