Re:measuring distances in grid space

Devi Jankowicz (
Sun, 30 May 1999 00:44:42 +0000

Tony Downing repleis to my earlier item,

>Devi's point imply that we should regard it as bad practise not to check
>out the analysis with the participant?

Well, I wouldn't put it as strongly as that: depends on the context and
what you're seeking to get out of it! Maybe, "yes, of course" if you're
doing counselling/guidance/personal growth wor; "it depends" if you;re
doing anything else. (My own example was just an instance of me mucking
around to see what I could learn about different methods of cluster
analysis. It was never published, done out of sheer personal interest.)

By and large: if _you_ think that an opinion from the respondent is
feasible and will be useful, get it!

The example I quoted suggests it might be worthwhile: remember, a high
matching score between any two elements, or constructs, is just a number
representing a rating. If you want to draw inferences about about a
person's structure in general, rather than simply as shown by the
ratings, it's often faster to ask the person. Particularly, a high
matching score between two constructs may _not_ mean that the constructs
are causally linked ("whenever you think a - not a, you also tend to
think b - not b; does that mean that b - not b is implicationally
dependent on a - not a?"), but simply associated without cause; or,
indeed, simply coincidentally present, a function of how adequately you'd
sampled the whole realm of discourse in choosing the elements you used.

For example: suppose 8 elements are rated such that the matching score
chocolatey - milky
nice - nasty
is 95%.

If the elements were different brands of chocolate the inference that I
dislike milky chocolates is actually quite true. I dislike white
chocolate, and prefer bitter!

If the elements were skin colours of people I know, you'd really have to
ask me to find out, wouldn't you? It could be true for my construing of
those particular 8 people but certainly not of my construing of people in

How large is a sample of elements before it's adequate? When it comes to
rep grids, it seems to depend very much on the context, much more so than
on statistical properties like population size, acceptable alpha, and so

"If you want to know what a person thinks, why not ask him? He might just
tell you!" Kelly, passim.
The numbers are just one information source which determine meaning; they
don't do so pre-emptively.



PS re your PS. I still have a programme I wrote to cluster-analyse a 10 x
10 grid... on a Sinclair ZX81! The machine with the "massive optional
add-on 16k RAMpack", remember? Took two passes through the machine, one
for elicitation and another for analysis... but if you can fit that into
32k, it ought to be possible on a Psion. If there's a BASIC compiler for
the Psion and you're willing to do some tidying up, you're welcome to the