Cross Loading Removal in Factor Analysis

[tonid...@gmail.com]

I am running Factor Analysis in my university thesis that have Cross loading in its "Rotated Component Matrix" I need to remove cross loading in such a way by which I can have at least 2 questions from the questionnaire on which factor analysis is run.

Need help.

Toni D'Souza

[David Marso]

Perhaps you would care to take the time to explain your specific situation in greater detail? Or do you expect some stranger to lead you through the great unknown blindfolded?

[Rich Ulrich]

On Wed, 4 May 2016 21:04:50 -0700 (PDT), tonid...@gmail.com wrote:

>I am running Factor Analysis in my university thesis that have Cross loading in its "Rotated Component Matrix" I need to remove cross loading in such a way by which I can have at least 2 questions from the questionnaire on which factor analysis is run.
>
>Need help.
>

As a blindfolded stranger, I wonder what your N is, the number
of variables, and the general size of the r's. Also, are these
from a specific and well-chosen universe of psychological scale items
(where I have most experience) or are they some questionnaire
variables that have better (or worse) reliability than that? For
scale items, I usually dropped all loadings less than a cutoff of
(absolute) 0.40, plus or minus 0.05 (which often worked to
produce unique loadings for all items).

For scale items, I don't remember accepting a 'factor' with
only two items, partly because of limits on its expected reliability.
That would not be a problem for a factor analysis using various,
previously-derived scaled scores (or other robust measures).
For something called a questionnaire, I would be careful about
factors with only 2 -- I have seen them arise when one question
was, in effect, asked twice, resulting in an atypically high r
(over 0.90) and an artifactual factor. For that, I would drop one
of them or replace them with an average of the two; and re-analyze.


I remember being annoyed by the labels provided by some FA
programs -- it has been a long time since I ran anything. I
don't remember that "Rotated Component Matrix" is what I
was looking at, so be careful that you are looking at the right
set of numbers.

--
Rich Ulrich

[sobot...@gmail.com]

Hello, I am running a factor analysis for my MA thesis and I am facing with cross loading factored problems. I don't either how to interpret or how to delete the overlapping factors. I do need your help to explain about it , recommend any document to read or give me any helpful link to check, Thanks !!

[Rich Ulrich]

On Tue, 4 Jul 2017 08:20:10 -0700 (PDT), sobot...@gmail.com wrote:

>Hello, I am running a factor analysis for my MA thesis and I am facing with cross loading factored problems. I don't either how to interpret or how to delete the overlapping factors. I do need your help to explain about it , recommend any document to read or give me any helpful link to check, Thanks !!

Roderick P Mcdonald wrote a fine book on factor analysis, and it
covers the basics. Check your llibrary.

"Principal factors with varimax rotation" does a pretty good job
of providing distinct factors; which is one reason why it is a
standard. - I'm thinking especially about rating scales, but it
works well elsewhere, too.

What you do with crossloadings depends on what you are trying
to achieve, and how many variables you have to work with.

If you have surplus items, the simplest answer is to drop anything
with more than one "high" loading. (Over 0.45? over 0.35?).
I have never done that, but it serves a purpose.

For my purposes, it has usually been sufficient to use an item
where it loads highest. This is also simple to describe and defend.
Where two loadings are very close, I could consider using the
item in the factor with fewer defining items. If there is no need
to do otherwise, of course, the simplest rule is to load an item
in more than one factor if it crosses the cutoff for more than one.

If /I/ had more than a couple of items in question on a rating
scale handed to me, I might wonder about whether the questions
were well-asked or the terms well-defined ... because that does
not happen very often. But I don't know anything about what
sort of data you are factoring.

--
Rich Ulrich