Yes / No / Not Sure responses - ordinal or nominal? ologit or mlogit?

[Diane Geelon]

Good morning everyone,

I am relatively new to regressions using categorical dependent variables and am struggling to decide on the most appropriate specification for my estimation model.

I am analysing the data from four survey questions around job quit intentions (e.g. are you intending to quit your job in the next six months?), three of which have "No" "Yes" "Not Sure" as the potential responses and one of which has "Never" "Occasionally" "Frequently" as the potential responses. n=994. There are 16 IVs in the full specification, some of which are categorical also (e.g. education, salary) and some of which are continuous (e.g. age, average affect scores).

1) Dealing with the three Yes No Not Sure questions first: The Not Sure responses account for 22.6% and 24% of the responses for the first two variables and 5.6% for the third variables. My starting point was therefore that merging the Not Sures with the Nos probably wasn't an option, thus ruling out standard logit / probit. I therefore proceeded to run an ologit model, the results of which successfully met the paralellel regression assumption (Brant test) apart from one of the five subcategories of education which was on the border. Having consulted Long and Freese (2014) in the meantime however, I am now confused as to whether or not ologit is indeed the right model as they clearly state on p.385 that "the category "don't know" invalidates models for ordinal outcomes". My question is whether 'don't know' and 'not sure' responses are equivalent (my instinct says no but happy to be corrected!) and if so, presumably I should be looking to use mlogit instead of ologit?

2) My second question refers to the "Never" "Occasionally" "Frequently" responses. Am I correct in my assumption that these responses constitute ordinal categorical responses and as such should be fitted with an ologit model (given that it passes the Brant test apart from one sub-category of the category education) or should I be considering using gologit2 (or even mlogit as above) instead?

Thanks in advance for your help / advice which is much appreciated

Diane

[Joseph Coveney]

What does the author of the questionnaire (survey form) say? Else, maybe there's something in the pertinent literature to guide you on this.

If nothing is forthcoming on either of those fronts, then maybe do it both ways. Use mlogit. Draft the paper, including the Conclusions and Discussion sections. Then use ologit. Draft the paper, including the Conclusions and Discussion sections. Compare these sections of the two drafts. Does it make a substantive difference?

[Nick Cox]

The order in which the categories appeared in the questionnaire can be important detail.

On this information I would regard "Yes" "Not sure" "No" as ordered. However, a test is not just a matter of thinking about meanings, but also one of seeing how coefficient estimates behave if you fit an ordered response model.

On categorical data analysis I stand in relationship to Long and Freese as a mouse to elephants. But I disagree with their emphatic dismissal of Don't know -- although my copy of their book is locked away (good contemporary excuse). It depends on the question. Sometimes "Don't know" is close to "Not sure" and in between "Yes" and "No". It's true that having a view and having no view are different issues, but consider the question.

[Diane Geelon]

Thanks all. Unfortunately I am the author of the survey Joseph and so have noone to check with! When designing it I included not sure to capture the complexity of the decision making process around quitting a job but I think in my own head not sure is probably closer to no than yes here. I will do as you suggest and run both models and see if it makes a substantive difference
.