I am trying to test if four variables ( different reading test) can be grouped into two constructs or just one. My hypothesis is that is should be two and not one. For my dissertation proposal I plan to use Principal Component Analysis and Confirmatory Factor Analysis. Once I have collected the data, if I run Exploratory Factor Analysis either by principal component analysis or by maximum likelihood, I obtain only one factor. If I run CFA, the two factor model fits better than the one factor model. The problem is that the four variables correlate more than .7 each other. I think that the correlation can be explained because one resulting construct could be dependent on the other, and that is theoretically possible ( One factor is decoding words aloud, the other is decoding and identifying the meaning). I have been told that if there is a dependence relationship between the resulting constructs, EFA is useless and I should use Confirmatory Factor Analysis, but I canít find any reference to document that. Do you know of any reference that talks about the inappropriate use of EFA when there is a relationship of dependence among the resulting constructs?
bmuthen posted on Saturday, January 21, 2006 - 7:40 am
If a 2-factor CFA model fits better than a 1-factor model, it seems like EFA would be able to show those same 2 factors. If one factor influences the other, the factors are correlated. I don't see how that would invalidate the EFA. Even in an EFA with uncorrelated factors (orthogonal solution such as with Varimax rotation), we know that an equivalent representation of the observed-variable correlation matrix can use correlated factors (oblique solution, such as with Promax rotation).