> I was especially interested in the issue of "noise" raised by
> Travis as my understanding is that grids with different
> constructs(but common elements)can be analysed using MDS if they
> are transformed into proximities. ....Often data for MDS is collected
> by asking subjects to rate how (dis)similar phenomena(eg
> elements) are without reference to any labels (eg construct
> poles). I had considered the approach adopted as analogous
> to the (dis)similarity approach, with the ratings indicating
> (dis)similarity.
That's an interesting way of viewing things, but unfortunately, your
clusters of elements will not be pure. I've been thinking it through
and the clearest explanation I can give is that if you allow elements
to vary on idiosyncratic dimensions, then people will be seeing
different patterns of similarity, i.e., "for Mark, Ethel is very
similar to Anne because she's a redhead and intelligent, while for
Dave, Anne and Ethel are opposites because Anne is tall and
introverted while Ethel is a petite extravert." While they would
agree on, say, a common male/female construct, Dave's and Marks spaces
are otherwise so different that they cancel each other out. Thus, any
patterns that exist on the common dimensions get washed out at worst
or fuzzified at best by the multiplicity of patterns inherent in
multiple constructions of the elements.
>
> Also of interest was Travis' comment regarding the possible
> significance of a 3rd dimension. This is an interesting
> proposition and upon reflection suggests to me,
> 1)the 2 dimensional model is a reasonable representation for the
> majority of participants in the research,
.and that raises the question of "who are the ones in peculiar
spaces?"
> 2)while the addition of a 3rd dimension adds little to explained
> variance, it is of significance to several participants...
What kind of result do you get when you exclude them? And also, are
the idiosyncratic constructs of the rest of the subjects semantically
close to each other in most cases?
> perhaps this may well be the difference between statistical
> significance and personal meaning. Certainly a 3rd dimension can
> be interpreted and allows for greater case discrimination (among
> a cluster of cases with several similar features).
Are there enough of those odd cases to do a separate INDSCAL? Or are
they demonstrably deviant in the semantic content of their
idiosyncratic constructs?
> Given the process of fitting a model to
> subjective data is a less exact process than scaling physical
> phenomena, perhaps it is too much to expect an exact fit in such
> circumstances ?
Yes, but a *close* fit should be possible.
> ... I also tried to analyse the 4
> scales using MDU though didn't find the 40 plots too illuminating
> (perhaps because when entered into regression the supplied
> constructs don't predict much of the variation in the release
> construct).
Sorry, I think I missed something there...I *presume* (and hope that
I'm right) that you had everyone rate the elements on a common
"release" construct and wish to use the subject weights to discern who
gave which rating...my first hunch is to identify those individuals
that I suggested you toss out (above) and see if there's a significant
chi-square. Given that they are *so* deviant, that's my best guess as
to who disagrees about "release" (whatever that is). Again, hope this
helps!
((((((((((((((((((((((((((((((((((((((((((((((((((((((((((
Travis Gee () tgee@alfred.carleton.ca ()
() () ()()()()
() () ()
() ()()()()()()()()()
"In science, the more we know the more extensive the
contact with nescience." -Spencer