I'm new to datascience (from a medical/medical science background).

My supervisor (social sciences background) asked me to assist in rewriting a paper where we do a cluster analysis for a questionnaire (unvalidated) with 63 questions relating to attitudes and behaviours and then to give insights about this data using an ANOVA on outcome variables (also related to attitudes and behaviors, basically an average of the responses to questions that seem related to a specific theme) based on questions from within this dataset. A previous paper used the questionnaire and stopped at defining the clusters ("we found these clusters, this is how each of the clusters answered each of the questions"). My supervisor wanted to gain additional insights to find determinants of the outcome variables of interest.

I ended up doing a PCA on the 63 questions and using those for a cluster analysis. I spend quite a bit of time in the results section defining the clusters based on their similarities/differences, but then in the next results section rather than doing an anova (or in my case multivariate kruskal wallis tests), on the exact same data i used to make up the clusters (which seems super not valid), I decided to do a hierarchical multiple regression on various demographic factors that were not included in the cluster analysis, and then in the final mode I added the clusters. Obviously, the clusters are most predictive of the outcome variables (because they use the same data).

I'm now trying to write my discussion and i need to say that what i've done is not valid, but then suggest a more useful way to perform this, and the usefulness of cluster analysis for describing behaviour/attitudes of a population in general. Is it reasonable to suggest that clustering be done on either demographic variables or on the outcome variables, and then compare clusters to the variables not included in the analysis? Is that valid?

What analyses can be validly performed after clusters have been defined, and how would those analyses be useful?

I've spent months trying to teach myself responsible statistics, and any help would be hugely appreciated.

  • $\begingroup$ what is your actual resseach question? or what are you trying to find? $\endgroup$
    – rep_ho
    Commented May 12, 2021 at 12:34
  • $\begingroup$ @rep_ho thanks for your response - we're looking for determinants of the outcome variables (a specific behaviour/attitude), and in particular my supervisor wanted to look how participant clusters (based on a number of attitudes/behaviours) interact with the outcome variables, along with a number of demographic factors (age, gender, etc) (sorry i can't be more specific about my actual variables/research question) $\endgroup$
    – scvbelle
    Commented May 13, 2021 at 5:04
  • $\begingroup$ do you need clusters for that? isn't it enough to look at the relationship between variables and the outcome? $\endgroup$
    – rep_ho
    Commented May 13, 2021 at 8:37

1 Answer 1


I don't think that validity is an issue here. The clusters were defined on variables/questions that were attitudinal, behavioural etc. of nature. The demographic variables are complete separate (from a conceptual view). Although there may be (would likely be) some correlations between demographic and the questions, CONCEPTUALLY they are completely different. So you HAD to do the cluster analysis EXCLUDING the demographic variables since else you'd cluster demographic variables into attitudinal/behavioural/etc which are DEPENDENT variables but have to be kept completely SEPARATELY from a conceptual point of view. Hope my argument is clear (if not, let me know). My two cents. I haven't got references.

  • $\begingroup$ thanks! Unfortunately I do run the clusters (based on attitudinal/behavioural factors from the PCA) against the dependent variables (made up of summed scores for responses to questions relating to attitudes and behaviours)... I added the demographics as independent variables as separate elements in the hierarchical regression to try and get some valid data out of it. I now want to write a paragraph on where/how cluster analysis should reasonably be performed and what value it can give to inform policy makers trying to shift behaviours/attitudes, but also where it is not useful. $\endgroup$
    – scvbelle
    Commented May 12, 2021 at 12:04
  • $\begingroup$ i don't think it's valid, you are basically just checking if the variables used for clustering are related to those other variables, and if you want to check that, there are better ways to do it than clustering $\endgroup$
    – rep_ho
    Commented May 12, 2021 at 12:35

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.