0

I have 9 categories of response variable, and facing interpretation problems.

  • Could I use this ordinal data as continuous data?
  • If not then please refer me to some example with more than 5 categories.

Thanks!

mdewey
  • 16,541
  • 22
  • 30
  • 57
love-stats
  • 1,119
  • 3
  • 11
  • 16
  • 1
    Do you think it is fair to say that your ordinal data is also *interval* data? That is, is the difference between level 4 and level 3 the same as the difference between level 8 and level 7, etc..? If not, you can't treat it as continuous. – Macro Aug 17 '11 at 13:45
  • They are not equal and we cannot treat them as interval/ratio scale. Could anybody provide me some examples of more than 6 categories of response for ordinal logistic regression? – love-stats Aug 17 '11 at 15:10
  • 2
    What do you mean by an example? Logistically (no pun intended) it is the same as less than six categories. – Macro Aug 17 '11 at 15:24

2 Answers2

1

As Macro has pointed out correctly, the interpretation of an ordered probit or logit model does not depend on the number of categories of the dependent variable. However, there might be some issues with the reporting of the results. If you have a dependent variable with 9 categories and you want to present outcome-specific marginal effect for each of your explanatory variables in detail, the whole thing becomes very cumbersome. In that case, you have to be a bit inventive. You could for instance rely on the graphical methods that have been advised elsewhere.

Another possibility would be to reduce the number of categories. Does it make sense to keep nine categories, or could some categories be pooled?

Community
  • 1
1

Technically, ordinal is ordinal and continuous is continuous, and you shouldn't treat one as the other. However, in practical terms, it gets a bit blurry somewhere. Is your ordinal variable really continuous "underneath"? That is, are the nine levels really just points on a continuous scale? If they are, then one thing you could do is run both ordinal and OLS regression and see if the results are very different. If they are not, you can go with the OLS results and note what you did. Perhaps not exactly technically correct, but "all models are wrong and some models are useful".

Peter Flom
  • 94,055
  • 35
  • 143
  • 276