02-18-2014 10:06 AM
My name is Madelaine, I am doing research on the party list rankings for open list elections.
Using a range of explanatory variable (age, education, etc.) I try to estimate the rank a candidate has on a party list.
As my dependent variabele (rank) is ordinal, I wanted to used a ordered logistic panel data regression. However, since the rank can take up values ranging from 1 to 80, I do not know how to interpret the odd ratio's which come out of the regression. I am also in doubt wheter the ordered logistic regression model is the correct model to use when the dependent rank variable has 80 categories.
I would very much appreciate it if someone could help me out on this,
Many thanks in advance, and kind regards,
02-18-2014 11:13 AM
My gut instinct is to say if you have 80 categories for rank, usually 1 to 80, you may want to treat it as continuous regression instead.
02-18-2014 11:19 AM
I agree with Reeza, especially if you primary interest in in the qualitative marginal response, as a linear panel model opens up the possibility of controlling for fixed effects (unobserved heterogeneity). Check out this link in the documentation to a two-way fixed effect linear model. At minimum, it is a good place to start.
02-18-2014 01:30 PM
First of all thanks for the quick reply, really great.
The thing i dont get, is if i have a ordinal dependent variable (the rank of someone on a party list), i can only use ordinal regression models right (ordered logit or ordered probit.)
02-18-2014 01:35 PM
No, because in linear regression, 3 is considered higher than 2, so the 'ordinal' portion of the analysis is maintained.
And if you have a large range of values, say 1 to 80 then it approximates a linear function, so linear regression is a valid option. I hesitate when people have 7 or 10 and use linear, but with 80 I think the you'd have to.
02-19-2014 12:50 PM
And if you are interested in models that predict values out near the edges of the range, you might want to look at quantile regression (see PROC QUANTREG).
One of the advantages to 'linear regression' is that it handles ties much better than nonparametric regression (so long as they don't overwhelm the number of records), and with a range of 80, I am willing to wager that there are a large number of ties in the data.