Rank ordered logit with reversed ranking

[Maya Lani]

Dear all,

I am using the rank ordered logistic regression (rologit). According to Stata manual, the rologit command treats the lower rank value as the least preferred and the higher rank one as the more preferred.

Take the following example of an employer evaluating the most preferred employee.

You can see a female, European employee with work experience is a lot more preferred than a female non-European with no work experience.

pref	female	European	Work experience
1	yes	no	no
6	yes	yes	yes

Now if the preferences are coded differently such that the lower rank indicates the most preferred option. For example:

pref	female	European	Work experience
6	yes	no	no
1	yes	yes	yes

Here a preference of 1 indicates that the candidate is ranked as the first (and hence more preferred option). In this kind of coding, Stata suggests to add "reverse" to the rologit regression. However, the Stata manual also indicates that it is not conclusive whether adding "reverse" is a better estimation than just using "rologit".

Any ideas on the better approach to use? should I simply use "rologit" and be careful on how I interpret the coefficients as I have a reversed ranking?

Thanks!

[Charlie Joyez]

Why not reverse the ranking before launching your command, this way you don't have to add the reverse option if you're not comfortable with it.

su pref
local max=r(max)
gen invpref=`max'-pref

rologit invpref ....

[Richard Williams]

Somewhat to my surprise, the manual points out that the reverse option does not simply reverse the signs; magnitudes of effects and their standard errors are also affected. Here is what the manual suggests:

Thus, although the weights of the attributes for reversed rankings are indeed mostly of opposite signs, the magnitudes of the weights and their standard errors differ. Which one is more appropriate? We have no advice to offer here. The specific science of the problem will determine what is appropriate, though we would be surprised indeed if this helps here. Formal testing does not help much either, as the models for the original and reversed rankings are not nested. The model-selection indices, such as the AIC and BIC, however, suggest that you stick to the rank-ordered logit model applied to the original ranking rather than to the reversed ranking [for the example given in the manual].

Read the manual entry for rologit for more detail.

As a minor point, if you want to use Charlie's suggestion, I would make it

gen invpref=`max'-pref + 1

so that the best option is coded 1 rather than 0.

[Maya Lani]

Richard Williams Dear Richard, I have read the whole manual very carefully, however, it seems that I was not clear enough in my question as I also noticed that the reverse option does not simply reverse the signs....

As you pointed out, the manual does not specify which method is more appropriate and that is why I was not sure how to proceed.

However, it seems like I simply need to re-code the ranking as you and Charlie Joyez suggest.

Thank you both for your feedback. I much appreciate it.