Re: possible distortion of individual construing if MDS is used
to analyse grids with common elements but some different
constructs.
I agree 2 individuals could use different constructs but rate
cases similarly. If 10 grids were analysed I would assume there
would be some "commonality" of construing even if different
construct labels were used. MDS, aims to reduce data to its
minimum dimensionality (which must involve loss of the richness
of "personal construing") and Indscal particularly aims to
represent common dimensions. The issue must therefore be whether
a particular MDS model generates separation of elements which has
some "common" or useful/important meaning.
As grids would mostly consist of at least 10 elements,the
proximities which would be calculated would be based on the
overall interrelationships. Indscal would appear most useful in
research applications where careful attention should be given to
element sampling. As a summary measure an Indscal MDS must result
in some loss of "personal construing". Also, if there were a
small number of grids I wouldn't be inclined to use Indscal.
Re: are the constructs of people whom the model doesn't fit as
well, idiosyncratic?
I'll have to examine this more carefully once the content
analysis is completed. However, inspection of the grids for
people who fit the model poorest suggests use of constructs
relevant to the 3rd dimension(dim)interpreted. I did undertake
analysis of the 20 best fitting subjects, the 13 and 20 worst
fitting subjects. Of interest for all groups was a consistent
first dim. The 2 dim. model for the best fitting group was the
same as the 2 dim. model for the whole group. There were also
similarities between the 3rd dim of the 3 dim model for the whole
group and the 2nd dim for the 13 worst fitting subjects. (I hope
this isn't too confusing). Re what is poor fit, Young and Harris
(1990)in the SPSS Base System Users Guide (p.439), refer to a RSQ
range of .4-.967 as " indicating that there are no judges who
are being fit very poorly". I am now inclined to think the 3
dimensional model is appropriate; Kruskal and Wish 1978 provide
a rationale for accepting a dimension which is used by some
subjects although variance shows small increase.
Some confusion may have resulted from my reference to regression.
Two different regressions were performed, 1) predicting release
ratings from the other supplied constructs (elements were
common).... grids were very useful in explaining why
dangerousness didn't predict release, and 2)regression using
subject weights to predict values on supplied constructs,
including release. I'm interested in the comments re using the
chisquare .. I assume it is meant good and poor fitting subjects
are considered in terms of release. This is something I haven't
done.
Regards,
Bob Green.