Generalized Ordered Logit: Robust erros by defualt and Psedo R

[Greta Petersen]

Hi,
I'm running gologit2 as the proportional odds assumption is violated in ordered logit. However, without me adding robust in the command, both models are run with robust standard errors. Is this normal? I can't find any information that the models have robust errors by default.

I also have a very low pseudo R squared, 0.06. Is this a serious issue? This is for a bachelor thesis, so I'm not expected to have extremely robust results, but I do feel I have to mention it. Past theses from my university using ordered logit do not report pseudo R, but I do feel like I should do it despite the low value.

Thanks!

[Carlo Lazzaro]

Greta:
welcome to this forum.
Richard Williams ' -gologit2- is a community-contributed module (as FAQ kindly request Stata users to mention it).
Have you alredy taken at look at gologit2 manual?
In addition, as per FAQ again, please share what you typed and what Stata gave you back. Thanks.

[Richard Williams]

gologit2 does NOT default to robust standard errors. Are you perhaps using other options, such as cluster, that force the use of robust standard errors with any command? If you show commands and output we may be able to better advise you.

With any model, the goal is to get the correct model, not to get the highest R^2. A low R^2 may just indicate that there is a lot of random variability in outcomes.

Having said that, a very low R^2 might also make you wonder if the model is correct. Are important variables excluded from the model, perhaps because they aren't included in the data set? Is the measurement of variables poor? Is the theory just not that good? Your discussion may note ways in which the data were not entirely adequate for your intended task. Or, you may be able to think of better ways to use the data you do have, e.g. include additional relevant variables.

I often warn my own students that they are at the mercy of the data. If they've done a good job but the results don't come out the way they hoped, well, that's life. They may have trouble getting the thing published but that doesn't mean I will give them a poor grade.

I also often suggest they try to come up with plausible counter-hypotheses explaining why things may not go like they think they will. That way you may be able to argue that the results are interesting no matter how they come out.

You can also see the gologit2 support page and troubleshooting FAQ:

https://www3.nd.edu/~rwilliam/gologit2/index.html

https://www3.nd.edu/~rwilliam/gologit2/tsfaq.html

[Mike Lacy]

In addition to Richard's useful points:

In general, I'd advise against focusing on the absolute size of the pseudo-R^2 because it tends to be pretty low relative to the conventional R^2 that would be obtained for the continuous variable that (theoretically) underlies the observed ordinal response. That pseudo-R^2 is more useful if used in relative terms, for example, in examining the proportional increase some predictor added to the model would make. For this and related issues regarding R^2 measures for ordinal response models, my old-ish article might be useful: Lacy, M.G., 2006. "An Explained Variation Measure for Ordinal Response Models With Comparisons to Other Ordinal R² Measures." Sociological Methods & Research, 34(4), pp.469-520.

I would not recommend that Greta get into that article in any depth for the current project, but it does give perspective on the performance of different R^2 measures for ordinal responses.

[Greta Petersen]

Thanks everyone!
Figured out stata adds robust errors when using pweights!

Thanks also for the advice regarding pseudo R!