Re: Factor analysis

I will try to explain;

I have a data set with N=2000, and n=20 of those are fullfilling a
sufficient amount of criteria for diagnosis A (they have 5 or more out
of 8 symptoms). The rest, 1980 persons, are having diagnosis B, C,
D.... and H.
Now, there is some comorbidity and overlapping, that is; some of the
ones with diagnosis A, also have diagnosis G, some have diagnosis, D, E

and F, and so on.
Also, someone with diagnosis A usually have some of the criteria from
other diagnoses, but often not enough to have that diagnosis as well.

What I want to test, is if there is a latent taxon, that is a special
group, which can be measured with the 8 criteria thought to
operationalized Diagnosis A.
To do this, I have reliability-tested the criteria (cronbach=0.65),
done a factor analysis out of the 20 people who have this diagnosis
(found one or two factors), and checked for
comorbidity with other diagnoses (relatively common).

Now, I need to check for the specificity, sensitivity, positive
predictive value and negative predictive value of the 8 criteria in
diagnosis A. (false positives, true positives etc). But this takes some

time, and I wonder if anyone here have a good syntax or procedure for
doing this?

Also, I want to do a MAMBAC and MAXCOV, to check if there truly is a
latent taxon or discrete group of the ones identified by these
criteria. But how to do this? I found some syntaxes for a statistic
program called R, but I feel more comfortable with SPSS.

Or maybe other procedures, if you have some good ideas?

Any help will be a tremendous relief for me!!



John Uebersax skrev:
You can do either or both.

A factor analysis of the 20 will show the factor/correlation structure
among patients within the diagnostic category. A factor analysis of the
total sample will show the structure in the total population.
Sometimes the two structures will differ significantly, sometimes not.

I'm assuming that all symptoms/criteria are assessable and assessed
relative to all patients in the sample.

A third possibility: you can calculate pooled within-category
correlations (considering all categories jointly) and then
factor-analyze these. Among other things, that can avoid limitations of
a small N within any one category. This is perhaps more common than
factoring within each category separately.

A case can be made for either factoring within-group or total sample
correlations. Analysis of the entire sample "confounds" the
correlation attributable to casual connection among the symptoms with
the merely statistical correlation attributable to two symptoms being
both associated with the same diagnosis. On the other hand, some
diagnostic systems aim to produce categories with minimal correlation
of symptoms within categories--then it would make less sense to examine
total factor structure.

(I neglected to mention yet another possibility--factoring
between-group correlation structure; i.e., the total correlation
structure can be decomposed into a within-category and a
between-category structure, and both can be factored independently;
however, factor analysis of between-group correlations is relatively
rare.)

It's probably most common to use the entire sample; but, as Ray
mentioned, it's a substantive question and depends on the study's
goals. Unless you have a definite reason for examining within-group
structure, the default would probably be to use the full sample.

You might possibly check the literature related to your research and
see what other's have done.

--
John Uebersax PhD
http://ourworld.compuserve.com/homepages/jsuebersax/

ErikLater wrote:
Do you know if i need to select data when performing a factor analysis?
That is, I have a large sample of people, and around 100 different
criteria for different diagnoses. 8 of these criteria make up one
diagnosis, and 20 people have this diagnosis. When performing a factor
analysis on these 8 criteria, do I only perform this on the 20 people
or on everybody?

.

Relevant Pages

  • Re: More on Klempners deceptiveness
    ... The CDC surveillance criteria provide as follows for for laboratory ... Second National Conference on Serologic Diagnosis of Lyme Disease. ... case definition criteria for laboratory confirmation. ...
    (sci.med.diseases.lyme)
  • Re: Factor analysis
    ... A case can be made for either factoring within-group or total sample ... the merely statistical correlation attributable to two symptoms being ... diagnosis, and 20 people have this diagnosis. ...
    (sci.stat.edu)
  • Re: Lyme Newsgroup Questions
    ... Can I just add that the government of the UK is misusing CDC criteria, ... body EUCALB on diagnosis, which are generally based on the CDC ... > patients with persistent Lyme disease through the Lyme Times ...
    (sci.med.diseases.lyme)
  • Re: Interesting Blogger POET
    ... Scape that Goat ... Some of the pamphlets seriously creep me out. ... Diagnosis and treatment ... Using the DSM IV criteria ...
    (alt.gathering.rainbow)
  • Re: Is it possible?
    ... factors for diabetes who had a FBG of 189 some years ago. ... "In the absence of unequivocal hyperglycemia, these criteria ... Criteria for the diagnosis of diabetes mellitus ... Symptoms of diabetes plus casual plasma glucose ...
    (alt.support.diabetes)