Chris Evans responds to my ramblings about personal judgement (is that
the same as "subjectivity"?!) in factor analysis in an interesting item
which gives a v.handy reference to Horn (1967): thanks v. much, will get;
and goes on to say:
<snip>
> The Horn paper is actually a rather silly way of
>making Devi's point, particularly as the diagnostics I was wondering
>about really ought to have gone some way to diagnosing that the
>"data" were random. However, I thought it would amuse you all and
>send it as a bizarre seasonal greeting.
<snip>
This prompts the thought that knowing the "random-data" characteristics
of any computational procedure can be fruitful at times. My brain is too
fuddled with the shock of returning to work that I can't think through
the similarities between this and all the stuff I know (not a lot) on the
way we use sampling distributions to make inferences- I suspect there's a
great deal of similarity, as Chris' reference to "diagnostics" hints- but
his reference prompts me to ask about another.
When doing a Principal Components analysis ofan n x m grid, and making
judgements about variance proportions accounted for, it is useful to
know what the PC structure of any n x m sized randomly generated rating
grid might look like. Patrick Slater once provided a procedure for doing
just that, but the reference is very obscure.
Slater, P. (non-dated) "The reliability and significance of a grid"
internal mimeo, St. George's Hospital, London.
I wonder if anyone who was in contact with him at the time, (Chris, were
you at St. George's then? Mildred Shaw, are you there?), early 1970s I
should think, would have a copy?
Searching through my box-files to try and find my copy (no luck) I came
across an item highly relevant to this discussion.
Keen T.R. "Rotating factors in repertory grid analysis" A paper given at
the 4th International Congress on PCP, Toronto, Canada, May 1981; mimeo
Barbican Research group, Garnett College.
In it, he asserts, with some compelling empirical dat, that the type of
factor rotation one uses makes a much bigger difference to the factor
structure as interpreted, than does the type of principal components
algorithm used in the first place.
If anyone is interested in the latter, please e-mail me your snailmail
address and I'll send you a photocopy.
Cheers,
Devi Jankowicz
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%