[Sigia-l] RE: Cluster analysis is dead - was Jakob wrong? (Try MDS)

[nathan at nathancurtis.com]

Indeed, cluster analysis - with its assumptions and applications - falls
somewhat short in it's direct applicability to analyzing results of card
sort research.  In fact, multidimensional scaling may be a more precise
technique.

Multidimensional scaling (MDS) analyzes participant's judgments of the
similarities or differences between items (such as cards serving as pages
within an architecture) frequently used in market and psychological
research. Data is input for each participant as item pairs and along with
a judged similarity between the two items.  While MDS supports a more
robust range of data input (such as a scale of observed similarity from
0-9), in general a card sort's groups could lead one to derive simple
inputs of 0 if two cards were not included in the same group, 1 if they
were.

The output is displayed via a visualization commonly referred to as a
perceptual map.  While the results are not as rigid as the actual tree via
cluster analysis, the researcher can extract far more information by
interpreting item (card) similarity & groupings by perceiving the distance
of these mapped items visually across two or more dimensions.

An unfortunate aspect of using MDS for card sort analysis is that the
technique is not commonly available (as far as I know) beyond devoted
statistical analysis software packages like SAS, SPlus, or SPSS.  These
may or may not be within the researchers budget or reach, but if you have
access to the software necessary to analyze such data, I suggest learning
more about this statistical method.

Nathan Curtis
Information Architecture & Interaction Design
nextel.com User Experience
Nextel Communications, Inc.