cloglog and proportional odds assumption

[Hanqi Ge]

hello,everyone.it's the first time i write here.I hope you can understand my poor English because it is not my mother tongue.
I am doing a study of influence factors of people’s fertility intention and i use ordinal categorical variable "people’s fertility" as dependent variable.At the beginning,i try to fit a model with "ologit" and "oprobit" .However, the two link cannot pass the score test for the proportional odds assumption.Then i use the link "cloglog" but i find that Stata’s "oparallel" command only do for logit.
So I want to consult:
Whether i should test the proportional odds assumption when fit an ordinal regression model with the link "cloglog"?If should,how to do it in Stata？
Thankyou for your time and attention and thankyou in advance for your help!

[Richard Williams]

How did you estimate an ordinal model with cloglog link in Stata? I didn't think built-in commands could do it.

I think you can do what you want with the user-written gologit2 command, available from SSC:

webuse nhanes2f, clear
gologit2 health age i.black i.female, link(cloglog) pl store(m1)
gologit2 health age i.black i.female, link(cloglog) npl store(m2)
lrtest m1 m2, stats

However, you may want to use link(loglog) instead, as loglog and cloglog get parameterized differently by different programs.. Compare gologit2's results with whatever you have been getting with your current approaches.

Incidentally, you only refer to proportional odds when using link logit. A more general term is parallel lines or parallel regressions.

gologit2 also provides a means for selectively relaxing the parallel lines assumption for variables that violate it.

It may be that you want to stick with logit or probit and let gologit2 selectively relax restrictions. gologit2 works best when most variables meet pl and only a few don't. But, if several/most variables violate pl, you may be better off going with a better known method like mlogit.

For more on gologit2, see

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

[Hanqi Ge]

Dear Richard，
Thank you very much for your attention and help.
Kind regards,
Ge

[Hanqi Ge]

How did you estimate an ordinal model with cloglog link in Stata? I didn't think built-in commands could do it.

I think you can do what you want with the user-written gologit2 command, available from SSC:

webuse nhanes2f, clear
gologit2 health age i.black i.female, link(cloglog) pl store(m1)
gologit2 health age i.black i.female, link(cloglog) npl store(m2)
lrtest m1 m2, stats

However, you may want to use link(loglog) instead, as loglog and cloglog get parameterized differently by different programs.. Compare gologit2's results with whatever you have been getting with your current approaches.

Incidentally, you only refer to proportional odds when using link logit. A more general term is parallel lines or parallel regressions.

gologit2 also provides a means for selectively relaxing the parallel lines assumption for variables that violate it.

It may be that you want to stick with logit or probit and let gologit2 selectively relax restrictions. gologit2 works best when most variables meet pl and only a few don't. But, if several/most variables violate pl, you may be better off going with a better known method like mlogit.

For more on gologit2, see

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

Dear Richard
I have used your command"gologit2",but i recieved the messages"convergence not achieved" and "variance matrix is nonsymmetric or highly singular"under the parallel lines test.
My code is:

gologit2 intention i.gender i.marriage age education-year  i.residence i.children income i.insurance ,link(logit) autofit difficult

Thanks.

[Richard Williams]

It is hard to tell without having access to the data. Can you start by showing the frequencies for your dependent variable?

[Richard Williams]

The gologit2 troubleshooting FAQ is at

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

There is a section specifically on "gologit2 is very slow and/or does not converge/and/or produces implausible estimates."

Start with the basics though. It makes a big difference if N = 100 or N = 10,000. If N is small or individual categories have very small Ns, you probably have to make modifications to the model or the data. If gologit2 can't be made to work and you are not happy with ologit either, you may just want to use mlogit instead.