LISTSERV at the University of Georgia
Menubar Imagemap
Home Browse Manage Request Manuals Register
Previous messageNext messagePrevious in topicNext in topicPrevious by same authorNext by same authorPrevious page (February 2005)Back to main SPSSX-L pageJoin or leave SPSSX-L (or change settings)ReplyPost a new messageSearchProportional fontNon-proportional font
Date:         Mon, 14 Feb 2005 14:55:14 -0800
Reply-To:     Dale Glaser <glaserconsult@sbcglobal.net>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Dale Glaser <glaserconsult@sbcglobal.net>
Subject:      Re: help factor analysis
Comments: To: David Hitchin <d.h.hitchin@sussex.ac.uk>
In-Reply-To:  <1108419495.421123a730080@webmail.sussex.ac.uk>
Content-Type: text/plain; charset=us-ascii

And to add to this, even though I am reluctant to relax the convergence (or # of iterations) criteiron, if I do (e.g., IT = 50), even if I then obtain convergence, I like to be able to view the iteration history (as one can obtain from most SEM softwares) to see if achieving a global maximum was problematic..........did there seem to be many oscillations? (I believe part of this is revealed in the output for each step of the derivatives in LISREL). If so, and even if I did obtain convergence with 48 iterations, I'm going to be very circumspect about my evaluation/embracing of the factor model....................Dale

David Hitchin <d.h.hitchin@sussex.ac.uk> wrote:Hector,

I posted before seeing your reply, but I'm not really in disagreement with you. Both the number of iterations and the convergence margin can be adjusted. In my experience, increasing the number of iterations usually works, and doesn't require the syntax that some people dread, so I would suggest that option first, but if it failed I would relax the convergence criterion.

After many decades of performing factor analyses, as well as writing a lot of code to provide options excluded by SPSS, I have been struck by the fact that factor analyses either seem to work brilliantly the first time or not at all, and if they fail first time, then tinkering rarely produces much improvement.

David Hitchin

Quoting Hector Maletta :

> David, > I advised before rather to increase the convergence margin (set by > default > at 0.001) than to increase the number of iterations. In my > experience, > more > iterations often fail to achieve convergence if the convergence > criterion is > too tight, but then you don't normally need it to be so tight in the > > first > place, so it is simpler to increase it to, say, 0.01 or 0.005. Now if > > THAT > fails, it is a signal that the system would not easily converge. It > may > be > oscillating or the matrix may not have the right properties. In that > > cases > it is best to forget about rotation altogether. > > > Hector > > > > > > > -----Original Message----- > > From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] > > On Behalf Of David Hitchin > > Sent: Monday, February 14, 2005 5:27 PM > > To: SPSSX-L@LISTSERV.UGA.EDU > > Subject: Re: help factor analysis > > > > > > It often happens. The default setting just does not allow > > SPSS to do enough work to get the solution. Just increase the > > default maximum number of iterations (which is specified at > > the bottom of the "Factor Rotation" panel to something more > > realistic, such as 100. It doesn't take much longer to run, > > and this is generally enough to get you the solution. > > > > Many people like to create "summated rating scales", i.e. > > they get a rating score by adding together variables which > > have similar factor loadings, in other words rather than > > using a formula for factor scores which uses fractional > > weights, the weights are either 0 (the item does not belong > > to the subscale) or 1(it does). Its quite a lot of work > > building such scales, and rather than using factor analysis > > it may be better to use the "reliability" procedure to > > repeatedly chip out the subsets that you need. > > > > Of course for people scoring tests by hand, summated rating > > scales are easy, but they can't create proper factor scores > > by hand as there is too much complicated work. The surprise > > is, that if this is well done, the rating scales correlate > > pretty highly with the factor scores. In other words, you > > don't lose much by using 0/1 weightings rather than > > coefficients to several decimal places. > > > > > I am told that the correlation coefficient matrix that > > rotations are > > > based on in SPSS are not ideal for ordinal data? if not what > > > alternatives are available? > > > > > If there is a problem, it is in the factoring and the > > calculation of statistical tests, and not in the rotations. > > > > You have to face the fact that ordinal data (in a sense) > > contains less information than equal-interval data, and once > > you have lost (or failed to collect) information, then you > > can't recover it by doing more analysis. > > > > Perhaps your advisers are thinking of rank correlations as > > input to the factor analysis rather than the usual Pearson > > product-moment correlations. However, the Spearman method is > > equivalent to ranking your data and then proceeding in the > > usual way, while the Kendall method (if I remember rightly) > > is not guaranteed to produce a positive- definite matrix. > > > > You wrote: > > > I > > > want to use promax results but what I would like to do first is > to > > > compare it against Varimax and Oblimin rotations. I want to try > to > > > extract different factors and try different rotations and > > and report > > > the results that makes the most sense but I am getting this > > error in > > > most cases. > > > > Bear in mind that Varimax tries to give you the clearest > > orthogonal solution to your problem, i.e. the factor scores > > are uncorrelated. However, people who think about your data > > may think in terms of correlated constructs, and it may be > > easier to use factor scores that are aligned with the way > > that people think - even if these are not independent of each > other. > > > > If you are to use promax or oblimin solutions, that implies > > that you are already expecting a certain kind of factor > > structure - you know where you want the coefficients to be > > large and where you want them to be small. In this case you > > have a hypothesis to test, that a certain pattern will be > > present in the factor matrix. You won't find this structure > > in the most effective way by using ordinary factor analysis > > followed by rotations, and nor will this test such a hypothesis. > > > > For this you need a structural equation modelling routine, > > such as Lisrel, EQS,Amos or something similar. In this you > > specify the expected structure of your data, and the computer > > performs a version of factor analysis which seeks the optimum > > fit to your hypothesis, and tests the significance of the fit. > > > > David Hitchin > > > > > > Quoting Mpundu Mukanga : > > > > > I have likert type data...trying to perfom factor analysis. When > I > > > specify the number of factors I am getting the error as > > shown below. > > > Any syntax and pointers would greatly help. What is > > happening is here? > > > what am I doing wrong? what does that say about the data? > > Any reasons > > > why rotation is not converging? I maybe wrong here but I > > want to use > > > promax results but what I would like to do first is to compare > it > > > against Varimax and Oblimin rotations. I want to try to > extract > > > different factors and try different rotations and and report > the > > > results that makes the most sense but I am getting this > > error in most > > > cases. > > > > > > > > > Secondly is it a recommended practice to sum up or average > > the scales > > > scores? what i would like to do is create scales for each > factor. > > > > > > I am told that the correlation coefficient matrix that > > rotations are > > > based on in SPSS are not ideal for ordinal data? if not what > > > alternatives are available? > > > > > > Thanks for help > > > > > > > > > Rotated Component Matrix(a) > > > > > > > > > > > > a Rotation failed to converge in 25 iterations. > > (Convergence = .000). > > > > > > "Rotated Component Matrix(a) > > > > > > > > > > > > > > > --------------------------------- > > > Do you Yahoo!? > > > Yahoo! Mail - 250MB free storage. Do more. Manage less. > > > > > > > >


Back to: Top of message | Previous page | Main SPSSX-L page