gologit2 with autofit is too slow. Is there any alternative?

[Santiago Valdivieso]

I'm running a regression which has life satisfaction (measured in a 1-10 scale) as a dependent variable and a relatively big set of consumption variables (amount spent on amusement, cultural good, music-related goods, etc) as the regressors. Including the controls (demographics, education, etc.), I have almost 50 independent variables. I thought about using an ordered probit model, but I suspect that the parallel trends assumption is violated. Therefore, I want to use the partial proportional odds model, which uses a different equation for every category of the dependent variable *only* when the IV violates the parallel lines assumption. For this purpose, I used the gologit2 command with the autofit option. However, after 7 hours runnnig, STATA is still processing the regression. This is very unpractical because I have many other models that I have to run and I can't spend a day in each one. Therefore, I was wondering if there is a way to make it faster, or an alternative command that fits with my objective. If there is nothing to do, do you think it would be a good idea to run the regression in an Amazon's EC2 cloud computer?

[Richard Williams]

The gologit2 troubleshooting FAQ has several suggestions:

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

I strongly suggest your model is too huge for gologit2 to handle (there are up to 500 parameters to estimate) and needs to be simplified (i.e. variables dropped), but maybe one of the suggestions will work for you.

Incidentally, with a 10 pt scale, it may not be all that bad just to use OLS regression.

[Sebastian Rodriguez-Raza]

Hi Richard Williams,
Thank you for your reply. I am currently working with Santiago Valdivieso. We are reding your paper "Understanding and interpreting generalized ordered logit models". We would like to know how to calculate the overall p-value that you report in Table 2, Section 4. For the variables that do not meet the proportional odds assumption, is there an specific option in the gologit2 command to perform the test or you do it with a different command (In this case which command do you use).
Thanks in advance.

[Richard Williams]

Hi Sebastian. You can do it with plain old -test- commands. For example,

use https://www3.nd.edu/~rwilliam/statafiles/ordwarm2, clear
gologit2 warm yr89 male white age ed, auto(.01)
test yr89
test male
test white
test age
test ed

If proportional odds is met you'll be testing whether the one coefficient for a variable significantly differs from zero. If proportional odds is violated, there will (in this case) be four coefficients for each variable and the test will be of whether any of them significantly differ from zero.

Incidentally, you could do the same thing after mlogit.