www.cmm.bristol.ac.uk/forum

Hi all
I am running an unordered random effects model on a survey item about types of civic actions against corruption. This variable has six categories ranging from report corruption (1) to initiate/ join a protest (6). The residuals of categories 3, 4, 5 and 6 have an average correlation coefficient of r= 0.55. They are negatively correlated with the residuals of response categories 1 and 2. I want to run a binary logit instead of multinomial model, since the former is relatively easier to present in a table and interpret. Is it statistically reasonable to collapse categories (and run a simpler model) on the basis that random effects are highly correlated?

Thank you.

Moletsane Monyake

Hi Moletsane Monyake,
It is perfectly acceptable to collapse ordered categories into smaller numbers of categories on practical grounds and as you say on ease of interpretation grounds. I wouldn't worry too much about things like residuals etc. which are often hard to interpret for ordered category models but instead think more about what your collapsing of categories means from a practical stand point.
Best wishes,
Bill.

Hi Bill
Thank you for your answer. Just to clarify, the categories are 'UNORDERED'. E.g, some respondents said that ordinary people cannot contribute anything in the fight against corruption; others said "ordinary people can report corruption", "ordinary people can vote for clean candidates" "ordinary people can write letters to media and raise awareness" etc. So, the depedent variable is definitely unordered categorical. I ran an empty model and inspected the correlations between the residuals of these alternatives. The intercept variances are greater for vote clean canidates and protest. But as I say, four of these responses are positively correlated with an average correlation coefficient of 0.55. They are negatively correlated with the remaining two alternatives.