Re: cluster vs principal components analysis

bob green (
Sat, 22 May 1999 07:50:18 +1000

Hello Sharon,

I pulled my hair out when I faced a similar problem. As I had common
elements, my unwieldy solution was to do an individual differences
multidimensional scaling on the whole data set.

Then I examined the factor analyses of people with high subject weights on
the respective dimensions. For example, I interpreted the dimensions as:
seriosuness of offence, mental health, history of antisocial behaviour.
Particulalry for those with the highest weight on seriousness of offence, I
checked if their factor analysis (done via spss ) had constructs which
loaded highly on offending vs mental health variables. For those with high
weights on dimension two, I checked their patterns of mental health
loadings. This worked out to confirm the interpretation I made of the MDS,
though no doubt there is a neater statistical way to do what I did. I
eventually chose a 3 dimensional solution, which 1 person had a really high
weight on. The factor loadings of his grid were helpful to confirm the mDS
interpretation I made.

The MDS model didn't fit some participants as well as others. As cluster
analysis can be based on distances, as is MDS, I used the individual cluster
analyses to check out why the elements clustered differently to the group
mDS solution (clearly being able to do individual MDS on each grid would
have been better).

Something I still find difficult is not only the choice between methods
(Slater and Rump debated this issue in 1974), but that different programs
generate different results.

If you have common elements, the other comment I would make is to also
consider methods which allow comparison between key elements. A post I will
soon send on measuring change, will elaborate on this issue.

If you want to discuss this further or show me some of your data let me
know. I'm in the city (32212511) at the moment and not Wacol.