Communality estimates >1: Warning #11382
Daniel L. Snyder
I ran my dissertation variables through a Maximum Likelihood factor analysis
procedure. I received the message: "Warning # 11382 One or more
communality estimates greater than 1.0 have been encountered during
iterations. The resulting improper solution should be interpreted with
caution."
Curiously, I looked at some printouts I did this summer which reproduced a
stats class assignment wherein we did a PCA of some Minnesota Multiphasic
Personality Inventory (MMPI) variables. My PCA conducted this summer was
exactly the same as what I obtained for the assignment (as it should).
While I was at it, I ran the same variables through all of the other FA
procedures on SPSS 6.1.1 (PowerMac). As I looked back on those printouts, I
discovered that the same message was found for Maximum Likelihood and for
GLS.
My deduction is that ML and GLS procedures operate under some tight
assumptions which make them "finicky" (for lack of a better word) for some
data sets.
Can someone confirm if this is the case? Otherwise, propose a possible
solution for using ML.
Could multicolinearity be the cultprit? I am yet in the process of trying
some data transformations and then examining the correlation matrix for
multicollinearity.
I would like to attempt a "true" FA procedure, so if ML and GLS are out of
the question, could someone suggest a good alternative? Otherwise, I will
have to go with PCA, which works, but I know is really not the same thing a
factor analysis.
Thank you very much for the assistance.
--
So I'll cherish the Old Rugged Cross,
Till my trophies at last I lay down;
I will cling to the Old Rugged Cross,
And exchange it someday for a crown.
Rich Ulrich
<sn...@tc.umn.edu> wrote:
> I still have not received an answer from my inquiry last weekend.
... I guess that we folks who read it did not find it exciting or
challenging or worth the bother -- maybe it is something that you
should find in the literature without too much hunting?
> I ran my dissertation variables through a Maximum Likelihood factor analysis
> procedure. I received the message: "Warning # 11382 One or more
> communality estimates greater than 1.0 have been encountered during
> iterations. The resulting improper solution should be interpreted with
> caution."
[ ... ]
I think I have seen it with "too many variables" rather than merely
too much intercorrelation between variables. (Do you have 10 times
the number of cases as variables, for your social-science-type data?)
You don't *have* to "iterate on the communalities" which is what was
happening when the message came about. For a safe solution to
contrast to whatever you see, you can plug in the maximum practical
Reliability along the diagonal, say, 0.75 for rating-scale data. If
you have too-many variables, you should not be trying the
iteration-solution, anyway.
--
Rich Ulrich, wpi...@pitt.edu
http://www.pitt.edu/~wpilib/index.html
Daniel L. Snyder
have been very helpful.
I have been deleting some variables - due to multicollinearity and some
which I do not think make sense (e.g. a ratio variable which is essentially
interpreted categorically).
As I have done so, my Measures of Statistical Adequacy have improved.
So far, I have only run PCA, but will try the reduced variable set again
with ML. Perhaps this will solve my problem.
-Dan
in article ufojgskg4g72rh4b1...@4ax.com, Rich Ulrich at
wpi...@pitt.edu wrote on 4/28/00 2:16 PM:
--
Code without integration of client data, but NEVER
interpret without it.
Barry
Herman Rubin
Daniel L. Snyder <sn...@tc.umn.edu> wrote:
>I still have not received an answer from my inquiry last weekend.
>I ran my dissertation variables through a Maximum Likelihood factor analysis
>procedure. I received the message: "Warning # 11382 One or more
>communality estimates greater than 1.0 have been encountered during
>iterations. The resulting improper solution should be interpreted with
>caution."
>Curiously, I looked at some printouts I did this summer which reproduced a
>stats class assignment wherein we did a PCA of some Minnesota Multiphasic
>Personality Inventory (MMPI) variables. My PCA conducted this summer was
>exactly the same as what I obtained for the assignment (as it should).
>While I was at it, I ran the same variables through all of the other FA
>procedures on SPSS 6.1.1 (PowerMac). As I looked back on those printouts, I
>discovered that the same message was found for Maximum Likelihood and for
>GLS.
>My deduction is that ML and GLS procedures operate under some tight
>assumptions which make them "finicky" (for lack of a better word) for some
>data sets.
I am not sure exactly which GLS assumptions are made.
Properly, ML should "correct" to communality 1, but it
makes the problem more difficult, as there may be multiple
corrections necessary. In fact, I would not be willing to
call it maximum likelihood if the parameter estimate is
outside the parameter space.
>Can someone confirm if this is the case? Otherwise, propose a possible
>solution for using ML.
The correct way to do this would be to take a variable
whose communality is one, make it a factor, and get the
residuals of all other scores from their regression on it.
Run FA on those (one can use matrix methods, and get the
moments directly), and convert the results back. Repeat on
the reduced problem if necessary.
There may be more than one result. To decide which one is
the true ML requires comparing the likelihood functions at
the various solutions.
>Could multicolinearity be the cultprit? I am yet in the process of trying
>some data transformations and then examining the correlation matrix for
>multicollinearity.
If you have a linear model, linear methods are appropriate.
Any non-linear transformations would destroy the linearity
of the model. Try to understand your assumptions, and
what the statistical procedures are doing, before running
any canned program. Statistics is not a collection of
incantations to be used blindly.
It could, but not necessarily. It is easy to give a
correlation matrix for 3 variables, without too heavy a
correlation, with this. All which is needed here to get a
perfect fit with a communality larger than 1 is for one
correlation to be larger than the product of the other two,
assuming all correlations are positive.
>I would like to attempt a "true" FA procedure, so if ML and GLS are out of
>the question, could someone suggest a good alternative? Otherwise, I will
>have to go with PCA, which works, but I know is really not the same thing a
>factor analysis.
PCA has nothing to do with factor analysis.
What I question from your questions is, do you know what
is meant by the answers you are getting?
--
This address is for information only. I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907-1399
hru...@stat.purdue.edu Phone: (765)494-6054 FAX: (765)494-0558
Daniel L. Snyder
in article 8eklrq$r...@odds.stat.purdue.edu, Herman Rubin at
hru...@odds.stat.purdue.edu wrote on 5/1/00 2:27 PM:
--
David Hitchin
: I ran my dissertation variables through a Maximum Likelihood factor analysis
: procedure. I received the message: "Warning # 11382 One or more
: communality estimates greater than 1.0 have been encountered during
: iterations. The resulting improper solution should be interpreted with
: caution."
: My deduction is that ML and GLS procedures operate under some tight
: assumptions which make them "finicky" (for lack of a better word) for some
: data sets.
: Can someone confirm if this is the case? Otherwise, propose a possible
: solution for using ML.
The procedures that you are using solve the problem by iterative
methods; an initial guess at the solution is followed by a series of
steps, each of which tries to maximise a likelihood or reduce an error
sum of squares. This iterative procedure may try searching in an
"impossible" region, in which case the computer may give up or may stick
on the boundary of the feasible region. This may happen because the
computer has started from an unfortunate estimate with no clear path to
a good solution which exists, or because there is no good solution. In
theory you could specify an alternative starting position and try
searching again, but modern programs usually generate such good starting
positions that it may not be worth trying alternative initial estimates.
The basic requirements for a feasible solution are (1) you are fitting
the correct model (2) that you have enough data and (3) that the
errors are sufficiently close to normally distributed.
One way of trying to fit an INcorrect model is to try more factors than
exist (according to the information in the data). There's no clear
definition of "enough data" but bear in mind that you need a good
estimate of the correlation or covariance matrix. Where the relationship
between the variables and the factors is clear you can get away with a
smaller sample - but you don't really know that until AFTER you have
done the analyis.
The normality assumption is less important, unless it is grossly
violated. Although the ML model is defined in terms of normality, it is
a fortunate (though not well known fact) that you will get ML estimate
of many of the parameters WHATEVER the error distribution - but the
standard errors and significance tests DO rely on normality.
David Hitchin
University of Sussex
Daniel L. Snyder
exploratory, I am not really entering with preconceived notions as to what
factors might/might not exist. Some of my variables were quite skewed
and/or kurtotic - and even with logarithmic transformations, may yet not be
close enough to normal. Hopefully I have enough data - about 155 cases, and
I am currently working with 14 variables.
I probably need to review the correlation matrix again. Although there are
some strong correlations, there are also many which are very weak. Perhaps
there is no good solution.
in article 8emp7a$3fd$1...@infa.central.susx.ac.uk, David Hitchin at
cc...@central.susx.ac.uk wrote on 5/2/00 9:37 AM:
--
All she really needs from me is more of You...
That's all she really needs from me - I know it is, I know its true
More understanding, more tenderness - A love that goes beyond my humanness
That's all she really needs from me ... Is more of You - more of You