Proportional odds assumption violated in ordered logistic regression, what to use ideally for the dependent, ordinal variable: gologit2?

[Isabel Hostettler]

Dear all,

I am trying to run a multivariable regression analysis with a dependent variable (mRS= modified Ranking Scale) being a ordered variable that, as often is the case, violates the proportional odds assumptions. I have tried several things such as combining different categories to even it out but the violation stays. I have come across the gologit2 command and wondered if that might be a good way to analyse the data. Unfortunately, our statistician does not have any experience and he recommended to either use the multinomial regression or to try to treat the variable as linear.
I have been through the different sections and links provided on the below address kindly provided by Richard William (https://www3.nd.edu/~rwilliam/gologit2/index.html) but am still not sure if gologit2 is suitable and am confused about a few steps.


I have attached a document with the original dependent variable as well as the 2 gologit2 calculations I have tried: one with autofit and one without autofit. The one with autofit specified indicates that antiHTN (antihypertensive medication) and Sex fulfil the pl assumption but Age doesn't. If I do not specify the autofit, I thought the model autofits it automatically but the OR turn out all different between the different groups compared to the autofit. Or is gologit2 without specifying autofit indeed running a ologit?
The output seems hard to interpret in my eyes.

One more question I have: how can I add categorical variables into the model such a Haptoglobin which as three categories (or more) instead of being binary like Sex and antiHTN? It doesn't accept the i. specification.

And my final question is: would you really recommend the gologit2 for this example or do you think something else would be more suitable and eventually easier for me to understand?

Happy to provide more information if helpful.

Thanks for your help!

[Richard Williams]

gologit2 without autofit is the same as gologit2 with the npl option, i.e. parallel lines constraints are not imposed for any variables.

gologit2 with the pl option produces results equivalent to ologit.

To help with understanding and interpretation, you may want to look at the following paper and handout:

https://www.dropbox.com/s/arkevwxlyf...t2016.pdf?dl=0

https://www3.nd.edu/~rwilliam/stats3/Margins05.pdf

If you haven't read it already, this was the first published paper on gologit2:

http://www.stata-journal.com/article...article=st0097

Fond as I am of gologit2, it doesn't gain you that much parsimony in this case, so you may want to go with the more widely known mlogit. But it is hopefully worth your while to better understand gologit2 regardless of what you do in this instance.

Finally, the current version of gologit2 does support factor variable notable notation. Make sure you have the latest version, available from SSC, not the old SJ version. You should have this version:

. which gologit2
c:\ado\plus\g\gologit2.ado
*! version 3.1.1 18oct2016  Richard Williams, [email protected]

[Isabel Hostettler]

Dear Richard,

thank you very much for your answer. I checked and I had the old version indeed. I have now downloaded the new version. Thanks for flagging this to me. I will continue reading about gologit2 as it seems very attractive and with great potential to me even though you are right and I think this time mlogit will be better. Just one more question: the new version supporting factor variable notable notation I might indeed be able to use it for my study?

Thanks for your great help!

[Richard Williams]

I don't know why you couldn't use gologit2. Whether or not it is the best choice is another matter.

[Isabel Hostettler]

Okay. Thanks a lot for your help. I'll have a go.

[Isabel Hostettler]

Dear Richard,

can I ask a further question? I have been through the material you advised. I just want to clarify that i understand everything correctly:
When I evalute the influence of ethnicity (coded as 0=white, 1=black and 2=asian) on the dependent variable mrs (a functional outcome score from 0-6) the proportional odds assumption are violated (looks like due to the black category).



When I now use gologit2 without specification I do nothin else then repeating the ologit, right?



If I understood the reading material correctly the autofit options lets stata evaluate the best model and see which variable meets and which variable violates the proportional-odds assumption.



What I dont understand is how here it says that the proportional-odds assumption is not violated so theoretically I could use the ologit but the ologit says the proportional-odds assumption is indeed violated. What is now correct and the right thing to do?

[Isabel Hostettler]

Unfortunately it hasn't sent the models. Attached you find the three different models.

[Richard Williams]

It is generally better just to post the output (using code tags; see pt 12 of the FAQ) than to include attachments.

Do NOT use the xi: prefix. It is not necessary and screws up some post-estimation commands. See

https://www3.nd.edu/~rwilliam/stats3/Margins01.pdf

I'm not totally sure what you think is a problem. But in any event, neither independent variable in your model seems to have any effect.

[Isabel Hostettler]

But does gologit2 recognise an independent categorical variable?
I have left out the xi and used following command:

gologit2 mrsFU Age Sex SBP DM HTN Hyperchol AF Statins GCSbin crICHvol crPHEvol IVH i.Hpgeno i.HpSNP

But doing so results in the program running endlessly or following error:

gologit2 mrsFU Age Sex SBP DM HTN Hyperchol AF Statins GCSbin crICHvol crPHEvol IVH i.Hpgeno i.HpSNP
factor-variable and time-series operators not allowed
r(101);


Hpgeno and HpSNP are categorical variables. If I also leave out the i. will these two variables still be recognised as categorical?

Apologies for using the attachment. When I add the stata output it doesnt look as nice as the ones of others and I still can't figure out how to insert my command and the outputnicely after reading pt 12. I hope it looks nicer now.

[Richard Williams]

Are you sure you are running the latest version of gologit2? Factor variables should be fine.

The troubleshooting FAQ has suggestions for convergence problems:

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

I would usually include the auto option when running gologit2. Also try starting with a simpler model and add variables gradually.

Your earlier simpler models did not produce significant results. If those had what you thought were your best variables, it may just be that there isn't much there.

[Isabel Hostettler]

Thanks a lot for all your help. I will have a closer look at it and starting with a simpler model. Will update here if there is a specific point or step when I run into problems.