I'm aware that this makes sense algorithmically. The underlying rationale of my question is if the tool (here k-means) is applied to provide a solution to a trivial problem (two-dimensional, at best) which could be easily treated otherwise (e.g. create clusters based on quartiles or +/-1SD above/below the mean on the two variables).
K-means is a good option only when clusters are linearly separalbe.
Dear Angelos,
If you have only two continuous variables I simply would do a scatterplot to see the relation between the variables. If there are any meaningful clusters you will see them and you could run k-means or whatever, but this is not data reduction as it is the object of scaling or cluster techniques.
Best wishes,
Jörg
Angelos,
looking at a XY-scatterplot will give the most information.
my advice is to show this scatterplot to people working in the clinical/practical world and let them decide and or draw borders.
however,
a clusteranalysis does give extra information: the distance from each specific case to each cluster center. this gives a typicality score.
you can select cases with a distance under a criterion value and start (or repeat) analyses on external variables as validation.
succes,
Eddy
I would look at density plot instead a scatter plot and do and two-dimensional gaussian mixture model. Write me, if you want to know how.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA