home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!paladin.american.edu!auvm!UNCVM1.BITNET!WDAVIS
- Message-ID: <STAT-L%92111311315027@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Fri, 13 Nov 1992 11:28:25 EST
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: Walter Davis <WDAVIS@UNCVM1.BITNET>
- Subject: reply to LISREL questions
- Lines: 102
-
- hi all,
-
- Over the last few days, Heiyi Xie has aske a couple of questions
- regarding latent variable models and LISREL in particular.
-
- First, regarding the 'assumption of multinormality' and LISREL, let
- me make a few clarifying points first.
-
- The ML estimator in LISREL was originally derived under the assumption
- of multinormality for all variables. However, Browne (1974) shows that
- only the assumption of no excessive multivariate kurtosis is necessary.
- Moreover, regardless of the distribution of the variables, all the
- LISREL estimators are consistent. The distribution only affects
- significance tests.
-
- That said, there are some other cases in which the distribution of the
- variables need not be multinormal nor have no excessive kurtosis.
- If you have exogenous X's (observed variables), e.g. dummy variables,
- and these are either fixed in repeated sampling or distributed
- independently of the error terms, then you need only assume that
- the error terms are multinormal. Or if all variables have the
- same kurtosis, then everything is fine.
-
- Bollen (1989) and D'Agostino (1986) provide tests for
- multinormality and multivariate kurtosis. Bollen (1989) also
- provides some results and citations to simulations investigating
- how well the ML and GLS estimators perform under non-normality
- (generally pretty well).
-
- When all else fails, Browne's (1984) Arbitrary Distribution Function
- (ADF, also known as WLS in LISREL and CALIS) provides consistent,
- asymptotically efficient and unbiased estimation with standard
- errors regardless of the distribution of the variables. This
- is a rather computationally intensive estimation and it requires
- the asymptotic covariances of the covariances (which can be
- calculated in PRELIS for LISREL, CALIS handles this automatically,
- and I'm not sure how EQS performs this). The question about
- this estimator is how large does the sample need to be before
- its nice properties hold. From my experience, ML, GLS and WLS
- will give very similar estimates and lead to the same substantive
- conclusions.
-
- I have written a SAS/IML macro for testing multivariate normality.
- If anyone would like a copy, please send me a note and I'll
- get it out to you as soon as possible.
-
- He then asked whether he could simply add up the indicators to
- form an index (adding also that he had a small sample and that
- the CFA had given him unexpected results). How well the additive
- index would reflect the latent variable depends on how reliable
- the various indicators are and whether their factor loadings
- are similar. Cronbach's alpha, a popular measure of reliability,
- assumes that the indicators have equal loadings and equal error
- variances. If your indicators closely approximate this, then
- Cronbach's alpha will be a fairly good estimate of how reliable
- the index is. If that reliability is high, you may be in good
- shape.
-
- For these questions, I strongly recommend that you read Ken
- Bollen's _Structural Equations with Latent Variables_ (Wiley: 1989)
- (disclaimer: he is my adviser). This answers all of the questions
- asked and a good deal more, including summaries of sample size
- effects and distribution requirements.
-
- Yesterday, Xie added a question about different packages.
- Structural equation software (LISREL, EQS, CALIS, AMOS, etc.)
- all do basically the same thing. They differ slightly in their
- options, but I believe they all provide the same estimators
- (ML, GLS, WLS/ADF), give very similar if not identical results
- (differences in my experience are usually due to different
- technical defaults such as convergence criteria, etc). Bollen
- and Ting did a review of LISREL vs. EQS for American Statistician
- within the last couple of years which highlights the similarities
- and differences. With a couple of exceptions (e.g. CALIS does
- not do multi-group analysis), the problem being analyzed will
- have no effect on your choice of software. More specifically they
- have the same small sample properties, etc.
-
- LISREL7 and LISCOMP can also deal with categorical endogenous
- variables. I have no experience with LISCOMP, but the LISREL
- package is relatively easy to use. Unfortunately, the asymptotic
- covariance matrix needed to estimate these models in LISREL is
- mis-calculated. See an article in _Journal of Marketing Research_
- (1991 or 92) by Rigdon and Ferguson. I understand that the problem
- will be corrected in the new version and that you may be able to
- get a corrected version for current stand-alone versions of
- LISREL from Scientific Software.
-
- My last point concerns Xie's question about the package LVPLS.
- I have not heard of this before, but assume from its name that
- it uses the Partial Least Squares (PLS) estimator. This
- estimator is an inconsistent estimator. That is, even with an
- infinite sample size, the estimator does not converge on the population
- value (see an article and dissertation by Dijkstra (cite not handy)).
- Given that, I know of no situations in which PLS can be recommended.
-
- hope this helps,
-
- Walter Davis <WDAVIS@UNCVM1>
- Department of Sociology
- Institute for Research in Social Science
- UNC - Chapel Hill
-