🚀 go-pugleaf

Organization: Purdue University
Date: Monday, 26 Jul 1993 09:11:05 EST
From: michael j. leiblein <LEIBLEIN@PURCCVM.BITNET>
Message-ID: <93207.091105LEIBLEIN@PURCCVM.BITNET>
Newsgroups: bit.listserv.spssx-l
Subject: ML vs. ULS estimation

   I am observing some odd results in a confirmatory model.  In comparing
a set of nested models I have obtained some (25% of models run) chi square
fit values that improve with the inclusion of additional constraints (ie
chsq, 19df= 87; chsq, 21df= 84).  Can anyone explain this?
   The data I am using consists of 12 highly correlated indicator variables
that have been theoretically linked to three subdimensions of a construct.
Because of the high correlation amongst the indicators (excessive commonality
is a term used by Fryxell (1985) to describe a similar situation) I originally
had identification problems with the model.
   Because the variances of the indicators vary by as much as two orders of
magnitude, I analyzed the data in the form of a correlation matrix.  All
attempts at using a covariance matrix have failed (even with the ridge option).
   Finally, I have had difficulty obtaining solutions with ML estimation (sigma
not pos def).  I have resorted to using ULS estimation- but I am observing
the odd fit values in about 25% of my models.  Could ULS estimation be the
cause of this problem?  I would greatly appreciate any insights into this
problem.  Responses may be sent to LEIBLEIN@PURCCVM.BITNET