3
$\begingroup$

A standard approach prior to conducting a predictive or inferential analysis is to report some basic univariate descriptive statistics on the study variables: mean, median, minimum, maximum, variance, etc.

Why not perform unsupervised learning (such as cluster analysis) to help identify patterns in the data in higher dimensions?

I think it'd be useful to know, for example, there's an unusual concentration of young males residing in a specific zip code in your data, which you would miss looking at univariate descriptive statistics.

Improve this question
$\endgroup$
3
  • 3
    $\begingroup$ Agreed with Peter Flom's answer (+1) in that we should be cautious about it. Now, a situation where I find that more extensive exploration can be justified is the analysis of secondary data. I've occasionally found that even some data from official statistical services may have some problems you'll only discover through exploration. It can unveil bias or other problems with the data (e.g. clusters that you know shouldn't be there, ludicrous time series breaks due to undocumented changes of definitions, etc.), so there are situations where it can be certainly useful as a sanity check. $\endgroup$
    – J-J-J
    Commented May 8 at 20:50
  • 1
    $\begingroup$ Erratum: according to this comment in another thread, it seems that what I'm talking about in my comment above may be more akin to data profiling than to data exploration. $\endgroup$
    – J-J-J
    Commented May 9 at 6:44
  • 2
    $\begingroup$ @J-J-J Perhaps a good example of unsupervised learning used in exploratory analysis is detecting fraudulent insurance claims. For my job with a pharmaceutical company, part of the clinical trial process prior to randomly selecting subjects for control & treatment cohorts is characterizing the baseline population data - both for EA and data profiling. $\endgroup$
    – RobertF
    Commented May 9 at 13:36

1 Answer 1

Reset to default
4
$\begingroup$

I don't think we should do this "routinely". We should do it when it will be useful to us and when we won't abuse it and when we have the time to do it correctly.

You mention cluster analysis. Well, you have to choose things to do CA. Hierarchical or k means? Within each, which algorithm? What linkage? What criterion for stopping in hierarchical? What criterion for k in k means?

And you often have to think about the results and compare different results to see which is useful.

You mention that there might be an unusual concentration of males in a ZIP code, but, to do that, you have to have a measure of "unusualness" because, if you have a lot of ZIP codes, some may have a high proportion of men. This is particularly so because some ZIP codes have almost no people (in fact, there are quite a few ZIP codes with no residents at all, e.g. the NY Stock Exchange has its own ZIP code).

So, maybe use counties instead? Well, there are some counties with very few people. Loving County, TX has only 64 people!

So, will you do cross validation? How?

I am not against cluster analysis. But I do like David Cox's statement:

There are no routine statistical questions, only questionable statistical routines

You also mention mean and sd, but even those aren't always right.

Improve this answer
$\endgroup$
1
  • 3
    $\begingroup$ Thanks Peter - agreed one should proceed with caution. I should have added that sometimes describing the data, rather than answering a specific research question, is a goal in itself. Understandably, you don't want to identify spurious patterns in the data that disappear when you look at another sample, so cross validation sounds useful. $\endgroup$
    – RobertF
    Commented May 8 at 18:37

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.