Let's say I have multiple categorical predictors (X) and I want to see the influence on outcome (Y) by building a logistic regression model.

Some of my predictors have many levels

  • i.e. exercise frequency (with levels of 1,2,3,4,5,6,7,8 +) then collapsing to 1 - 3, 4 - 6, 8 +
  • or happiness level (excellent, very good, good, okay, bad, very bad, terrible) but then collapsing to good, fair or bad

For my outcome I am not interested in the details i.e the exact level of happiness that affects my outcome but how does happiness affect my outcome (Y). Therefore my form of collapsing is based on:

  • my own knowledge
  • results that are more interpretable
  • I have read that I would get a more precise estimate and higher degrees of freedom, is this correct?

My question is, is there anything technically wrong with this approach? Anything I should be cautious of? I would be grateful for any useful references too.

  • $\begingroup$ Is exercise frequency continuous, or can it be considered to be? Ordinal variables with many levels can sometimes be reasonably approximated as continuous. $\endgroup$
    – mkt
    Aug 2 at 14:03
  • $\begingroup$ It is categorical data from a survey (can not be considered continuous as some results are simply 7+) and most of the variables are not in that format (I just gave that as an example, most of my variables are similar to happiness) $\endgroup$
    – test1234
    Aug 2 at 14:05
  • $\begingroup$ I think collapsing can be fine, especially when you have many levels, are data-limited, and they cannot be assumed to be continuous. But be aware that you should treat them as ordinal variables in your analysis. Collapsing isn't to be done lightly though, because it is throwing away information and that is usually not a good idea. $\endgroup$
    – mkt
    Aug 2 at 14:11


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.