0
$\begingroup$

I have a small dataset of 134 observations. The dataset consists of answers to a questionnaire. All the variables are measured in Likert-scale, ranging from 1 to 5 (strongly agree).

I want to perform a hierarchical cluster analysis on the variables but I am unsure about which linkage method and distance measure to use.

I have tried running Ward's linkage with squared Euclidean distance measure but I am getting conflicting answers online reading some of the posts and comments on different statistics forums. I know that Ward may not be the most appropriate linkage to use with ordinal Likert scale data like mine, but when I run the analysis with Ward's and squared Euclidean I get clusters that conceptually make a lot of sense and are well separated from each other.

So my questions:

  1. Should I use this approach with my data?

  2. Should I use some different measure?

Improve this question
$\endgroup$
6
  • $\begingroup$ Both your questions are really "should", not "can", so I've taken the liberty of editing accordingly. Euclid, Likert and Ward are all people's names. $\endgroup$
    – Nick Cox
    Commented Jan 15 at 12:52
  • $\begingroup$ How many such variables? $\endgroup$
    – Nick Cox
    Commented Jan 15 at 12:53
  • $\begingroup$ Three variables measured using the Likert-scale $\endgroup$
    – Lasse H
    Commented Jan 15 at 13:19
  • $\begingroup$ So, from one point of view there are $5^3 = 125$ distinct clusters -- from $111$ to $555$ --- that define themselves. Too many for some purposes, perhaps, but still worth looking at the pattern of frequencies. With any luck they won't all occur in the data. $\endgroup$
    – Nick Cox
    Commented Jan 15 at 13:22
  • $\begingroup$ Yes. I would like to minimize the number of clusters as much as possible. Any thought on whether I should use this approach with my data or whether I should use some different measure? $\endgroup$
    – Lasse H
    Commented Jan 15 at 14:04

1 Answer 1

Reset to default
0
$\begingroup$

My view (and others may differ) is that when it comes to cluster analysis, if a method gives you something useful, you can use it. You say you get clusters that are well separated and that make sense. Good.

More generally, ordinal data is tricky. Strictly speaking, the fact that a variable is ordinal means that you can't assume that the distance between (say) 1 and 2 is the same as between (say) 2 and 3. But ..... usually it's fairly close for some abstract definition of "close". Yes, you could recode 1, 2, 3, 4, 5 as 1, 2, 181,100291, 239918188 (if you are strict about what ordinal means) but, that just doesn't make any intuitive sense.

A long time ago, in grad school, I saw research trying to figure out how close some Likert scales are to being equally spaced, but I don't remember the details much less a citation.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thank you, Peter! $\endgroup$
    – Lasse H
    Commented Jan 15 at 14:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.