home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!UTXVM.BITNET!TBN
- Message-ID: <STAT-L%92083115524914@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Mon, 31 Aug 1992 15:52:48 -0400
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: TBN@UTXVM.BITNET
- Subject: communalities > 1
- In-Reply-To: The letter of Thursday, 27 August 1992 3:07pm CT
- Lines: 162
-
- Jaap Hartog < ABRAMSE@HROEUR1.BITNET > writes
-
- " Dear Networkers,
- I am doing a factor analysis on 4 variables with SPSS PC+.
- One loading proved to be greater than 1 (1.01 in fact). May be
- rounding difficulties. So I tried an other extraction method
- (PAF). Extraction stopped because of communalities greater than
- one. So I tried another method (ALPHA). Again a loading greater
- than one.
- Is there someone among you who can shed some light on this
- dark matter.
- Jaap Hartog "
-
- A question very similar to this one was sent to the SPSSX-L bulletin
- board, and I have included the text of my letter on the subject below.
- One quick thing you may want to try would be to use a / Diagonal =
- 50 .50 .50 .50 subcommand followed by / Extraction = PAF. This
- will give you a principal-axis factor analysis with .5 as the
- initial starting value used in computing the estimates of
- communality. As I've noted in the letter below, I've had some
- luck with using .5 as a starting value, particularly in the case of
- factor analysis models.
-
- In my quick glance through the SPSS Base System User's Guide I did
- not come across any references to Heywood cases or how to deal with
- them. As I note below, SAS allows a user to arbitrarily (for better
- or worse) enforce a communality less-than-or-equal-to-one rule, which
- allows a factor analysis program to continue running when it would
- have crashed otherwise. I am curious to hear what SPSS users do when
- they encounter Heywood cases.
-
- Here is the similar letter posted to SPSSX-L and my response to it.
- Doubtless there will be other responses to it, and you may wish
- to monitor that bulletin board as well as this one for the many
- useful suggestions which will no doubt follow your and Susan
- Kashubek's postings....
-
- ***************************************************************
-
-
- To: spssx-l, sk7811R@acad.drake.edu
- Subject: Heywood Cases
-
- Susan Kashubek (SK7811R@ACAD.DRAKE.EDU) writes the following:
-
-
- " HELP!
- For those of you who use Lisrel 7:
-
- I am using Lisrel 7 to do some structual equation modeling and am having
- trouble with a recurring error message. This message states:
- Warning: Theta Eps is not positive definite
-
- I have played around some with the way the model is specified (adding or
- deleting variables, fixing or freeing paths) and sometimes get the same
- warning, other times the Theta Delta matrix or the Psi matrix will not
- be positive definite. The result is that the modification indices, t-values,
- residuals, etc. can't be computed, and I assume that the final parameter
- estimates are somewhat arbitrary.
-
- My sample size is about 190 subjects, with 23 measured variables (though I
- have tried deleting 4 or 5 to see if that helps). I did get the model to
- "run" once without any errors or warnings at all, but when I attempted to
- modify the model based on the results in the modification indices, I ended
- up getting more warnings that the Theta Eps matrix wasn't positive definite.
-
- Any ideas on why I am getting this message? Do I have too many variables?
- Are there common strategies for dealing with this problem? Any help would
- be greatly appreciated, as there are few, if any, people here at Drake
- who are familiar with Lisrel.
-
- Many thanks!
-
- Susan Kashubeck
- Drake University
- SK7811R@ACAD.DRAKE.EDU "
-
- What Ms. Kashubeck has encountered is what is popularly called a
- Heywood Case in the factor analysis literature. Heywood Cases arise
- in two situations in structural modeling. The first occurs when
- the unique properties of the correlation matrix which is being
- analyzed by LISREL lead LISREL to estimate a path value (or factor
- loading in the case of a confirmatory factor analysis model) that is
- larger that 1.00 (If you are using a standarized solution, the
- presence of such an estimate may be checked on the LISREL output).
-
- The result of a parameter estimate that is larger than 1.00 means
- that "negative error variance" exists--there is no positive variance
- "left over" to fill one of the error matrices. You will recall that
- LISREL has a total of eight matrices. The Theta Epsilon (Theta EPS)
- is the error matrix associated with Y-residuals (i.e., downstream
- variable residuals). I would recommend that you check the LY (Lamba Y)
- matrix
- parameter estimates and try to determine if any of them exceed 1.00
- in a standarized solution. LISREL is expecting some left over error
- variance that it can use to fashion a "gramian" (i.e., positive
- definite matrix) Theta Epsilon
- matrix, and if it cannot do so then it announces that Theta Epsilon
- is "not positive definite".
-
- What can one do in this situation? One quick solution is to try
- setting different starting values other than LISREL's default.
- I've had moderate success with using / ST .5 ALL. Another option is
- to move from a Maximum-Likelihood-based solution to a ordinary least-
- squares or a generalized least-squares solution, since the former
- method seems to be particularly prone to Heywood cases and the latter
- two methods much less so. I believe that ordinary least-squares and
- generalized least-squares solutions are available by typing
- UL or GLS on the / OU line, though I am not certain of this. ML
- , which is the default, will give you the maximum-likelihood solution.
-
- A second way you can obtain a Heywood case is called "Empirical
- Underidentification". If you have a situation, say, where you have
- two measurement variables leading into one latent variable on
- a model with two latent variables joined by one correlation, with
- 3 manifest variables attached to the other latent variable by one path
- each, the situation looks fine for a good solution, but
- there may be an infinite number of solutions possible for the path
- values (parameter estimates). This may be particularly true when the
- correlation or covariance which links the latent variables is
- small. To determine if this type of situation is contributing to
- a non-positive definite matrix, I would suggest checking the
- standard error of the estimates in the output. If these values are
- large, then this situation may be the cause of your difficulty.
- You can try to get around this difficulty by setting your residuals
- equal to one another with / EQ statements.
-
- Finally, if worse comes to worse, you could try to figure out which
- of the individual residual matrix elements (i.e., estimates in the
- THETA EPS) matrix are less than 0 and fix them to some value greater
- than zero, but I would try to exercise all other options before
- undertaking such a radical venture. I'm sure other SPSS-L readers
- will have ready solutions for you to try because this is not an
- uncommon problem for structural equation modelers to contend with.
-
- Heywood cases are definitely a royal pain in the statistical
- posterior. SAS in its Proc Calis has even gone so far as to add an
- option that allows users to quickly and simply specify and option
- which tells SAS not to let any parameter estimates exceed 1.00, and
- therefore all of the error matrices are kept positive definite.
- LISREL contains no such luxury, however, but the LISREL user can do
- the same thing via a little more work and patience.
-
-
- Tor Neilands, Systems Analyst
- Statistical and Mathematical Support Services
- ______
- Internet: TBN@Utxvm.cc.utexas.edu /_|_|_ ¢
- Phone: (512)-471-3241, Extension 263 |_|_|_ | - -
- Address: Com 1 |_|_|_ | - - 0
- Mail Code 12700 |_|_|_ | - -
- UT Ausin ¢|_|_/
- Austin, TX 78712 ||
- ||
- " Two roads diverged in a yellow wood, and ||
- I...I took the road less traveled, and /||¢
- that has made all the difference ".
- --- Robert Frost
- Disclaimer: All views expressed above and herein are only those of
- the author and do not necessarily represent the views of the University
- of Texas, the UT Comp Center, and anyone else I've forgotten to include
- in this disclaimer.
-