cui bunny posted on Sunday, December 03, 2006 - 7:26 am
If I use Lisrel to test a measurement model with CFA, and the fit indices were reasonable but not great, and the only suggestion made by the modification indices was to correlate the errors of two items who load on the same latent factor (I have two latent factors in the model). This simple change made a huge improvement in the model fit. Why MI suggests correlated error? Is there any theoretical bases for that?