Ordered logit regression with fixed effects

[Khalid Noor]

Hello,

I would like to test the following model:

Dependent Variable: ordinal variable (firms' credit ratings. This ranges between 1 and 22 where 22 is AAA rating).
The main independent variable of interest:continuous variable.


But I have two questions:
First: Can I (or should I) treat my dependent variable as continuous and perform an OLS since there are 22 categories?

Second question: How do I include industry- and year-fixed effects in ologit model? I have 48 industries and 30 years, is it okay to code it as follows:

ologit Dependentvar independentvars i.industry i.year, vce(cluster firmid)

I looked up -feologit- function but it takes forever to run and then returns:
1,223 (group size) take 1,216 (# positives) combinations results in numeric overflow; computations cannot proceed


Your help is appreciated!

[Maxence Morlet]

ssc install feologit

by Baetschmann et al. (2015)

[Fei Wang]

First: Can I (or should I) treat my dependent variable as continuous and perform an OLS since there are 22 categories?

I would do both to examine the robustness of the results.

Second question: How do I include industry- and year-fixed effects in ologit model? I have 48 industries and 30 years, is it okay to code it as follows:

Yes, your code is correct.

[Richard Williams]

With 22 categories, I would hope/expect linear regression is ok, at least if the categories seem equally spaced. It will certainly make your life easier if you don’t have to use ordinal models in this situation.

But, consider using xthybrid, available from ssc. There are some Stata Journal articles about it. Allison discusses hybrid models at

https://statisticalhorizons.com/prob...hybrid-method/

[Khalid Noor]

ssc install feologit

by Baetschmann et al. (2015)

I did try -feologit- but did not work. It works when I group(firmid) but does not work when I group(industry) and returns the following:

1,223 (group size) take 1,216 (# positives) combinations results in numeric overflow; computations cannot proceed


I have around 1400 distinct firm id and 48 distinct industry id

[Khalid Noor]

With 22 categories, I would hope/expect linear regression is ok, at least if the categories seem equally spaced. It will certainly make your life easier if you don’t have to use ordinal models in this situation.

But, consider using xthybrid, available from ssc. There are some Stata Journal articles about it. Allison discusses hybrid models at

https://statisticalhorizons.com/prob...hybrid-method/

I had the same intuition that OLS should be ok. I performed these three models and got similar results but do not know which of these has fewer problems and thus should be reported:

reghdfe Dependentvar independentvars, absorb(industry year) vce(cluster firmid)
ologit Dependentvar independentvars i.industry i.year, vce(cluster firmid)
ologit Dependentvar independentvars Mean_independentvars i.year, vce(cluster firmid)

where Mean_independentvars is the time averages by industry for the time varying variables:

egen Mean_independentvars = mean(independentvars), by(industry)

Thanks for suggesting xthybrid. I will need to read about it as I have never used it.