gologit2

[Ania Aksan]

Hello Statalist,

I am running ordered probit regressions and have been using gologit2 rather than oprobit, because the proportional odds assumption fails for some of the variables. However, the data is cross-country and I want to weight by country population using aweight, which oprobit does allow but gologit2 does not allow. Since the variables that fail the proportional odds assumption are just controls and not the variables of interest (as revealed when using the autofit option on gologit2), would it be wrong to use oprobit and ignore the prop. odds assumption in this case? Will the estimates for the variables of interest be reliable?

Secondly, when running oprobit or gologit2 for some subsamples, Stata sometimes gives messages such as “Note: 11 observations completely determined. Standard errors questionable.” or “WARNING! 11 in-sample cases have an outcome with a predicted probability that is less than 0.” What does this mean, and how concerned should I be by this?

Thank you!

[Mike Lacy]

My perspective on the proportional odds assumption is that, with a sample large enough to get good performance from a ordinal response model, it's likely to be large enough that the model will fail a hypothesis of this assumption, even if the departure is not substantively large. Rather than look at the hypothesis test, I like to compare predicted probabilities, perhaps with some ordinal pseudo-R2. (See my -ssc describe r2o- program and related literature if the latter is of interest.) My experience is that hypothesis tests here can commonly lead to model choices that are not justifiable on the grounds of parsimony.

For a sophisticated ordinal response model with relaxed assumptions, and which accepts weights, take a look at -gencrm-:
net describe st0546, from(http://www.stata-journal.com/software/sj18-4)

I have not tried this but the authors are in my view on the leading edge of work about such models.

[Leonardo Guizzetti]

Just a note about continuation ratio models which are employed by -gencrm- (as well as a few other packages): these are most suitable when an individual experiencing the ordinal outcome may proceed sequentially from one outcome to the adjacent category, always going in a forwards (increasing) or backwards (decreasing) direction, and with implicit assumption that one cannot skip outcome levels nor reverse directions. With enough model flexibility (unconstrained or proportionally constrained parameters) the continuation ratio models can also yield negative predicted probabilities for being in a certain outcome levels, which would of course be non-sense. A disappointing downsize of the -gencrm- command is that it does not allow for the use of post-estimation commands like -predict- or -margins- which means you will need to manually program those yourself.

Btw, two of the -gencrm- authors, Fullerton and Xu have a clear and excellent book on Ordered Regression Models.

[Richard Williams]

A few things:

1. oprobit does NOT allow aweights (unless the documentation is wrong). So I don't see why weighting capabilities would be a factor in your choice of models. The Stata manual and other Stata sources explain why the use of aweights is often incorrect or nonsensical.

2. I'd be very surprised if oprobit gave you messages like “WARNING! 11 in-sample cases have an outcome with a predicted probability that is less than 0.”. gologit2 can though. See the troubleshooting FAQ at https://www3.nd.edu/~rwilliam/gologit2/tsfaq.html .

3. Like Mike says, with enough cases any violation of proportional odds will be significant. You might set autofit(.01) or autofit(.001).

4. Some suggested readings on gologit2:

Williams, Richard. 2006. "Generalized Ordered Logit/ Partial Proportional Odds Models for Ordinal Dependent Variables." The Stata Journal 6(1):58-82. The published article is available for free at http://www.stata-journal.com/article...article=st0097.

Williams, Richard. 2016. 2016. "Understanding and interpreting generalized ordered logit models." The Journal of Mathematical Sociology, 40:1, 7-20. http://www.tandfonline.com/doi/full/...X.2015.1112384.

Williams, R. A., & Quiroz, C. (2020). Ordinal Regression Models. In P. Atkinson, S. Delamont, A. Cernat, J.W. Sakshaug, & R.A. Williams (Eds.), SAGE Research Methods Foundations. doi: 10.4135/9781526421036885901. https://methods.sagepub.com/Foundati...ression-models

5. Like Leonardo says, continuation models are very good sometimes but only with variables that can change in one direction, e.g. you can increase your education but you can't decrease it.