Best model for an ordinal dependent variable which has a left-skewed distribution

[Tom Prosser]

Hello, I have an ordinal dependent variable which has a left-skewed distribution. So far, I've just been using OLS. Might another model be better? Should I run any particular diagnostics? Any help would be greatly appreciated - I'm not great with statistics. Best and thanks,
Tom

[Daniel Schaefer]

An ordered logistic regression might be more efficient here, particularly if there is reason to think adjacent categories of your ordinal variable are not equidistant from one another. See:

help ologit

But the ordered logistic regression can be difficult to interpret without the correct training. You might want to pick up a copy of Long and Freese 2014 to learn more. As for regression diagnostics, I believe this is a largely open area of research for the family of logit regressions, but for your OLS regression, just look at all of the usual stuff. Best practice is to test all of your regression assumptions systematically. Take a look at this text book to learn more about regression diagnostics.

[Leonardo Guizzetti]

Can you say more about what you are trying to model specifically? What is your ordered outcome?

[Tom Prosser]

An ordered logistic regression might be more efficient here, particularly if there is reason to think adjacent categories of your ordinal variable are not equidistant from one another. See:

help ologit

But the ordered logistic regression can be difficult to interpret without the correct training. You might want to pick up a copy of Long and Freese 2014 to learn more. As for regression diagnostics, I believe this is a largely open area of research for the family of logit regressions, but for your OLS regression, just look at all of the usual stuff. Best practice is to test all of your regression assumptions systematically. Take a look at this text book to learn more about regression diagnostics.

Thanks so much for the reply - I really appreciate it :-)

[Tom Prosser]

Can you say more about what you are trying to model specifically? What is your ordered outcome?

I'm modelling responses to this survey item:

‘Some people think that the way people talk needs to change with the times to be more sensitive to people from different backgrounds. Others think that this has already gone too far and many people are just too easily offended. Which is closer to your opinion?’

1. The way people talk needs to change a lot 2. The way people talk needs to change a little 3. People are a little too easily offended 4. People are much too easily offended

[Leonardo Guizzetti]

Thanks. I would agree with using some form of ordered logit regression. -ologit- is built-in with Stata and there is also Rich Williams' -gologit2- (Stata Journal) which fits generalized ordered logit models.