Partial proportional logit model with gologit2: interpretation of inconsistent p-values

[Edem Kpewou]

Hi,

Thank you all for your support. It is my first post here.

Please I am using gologit2 to perform PPOM on a 3 level ordered dependent variable that is measuring nutrition behaviour (nutbecat). The predictor variables are location (rural =0 urban =1), age (continuous variable), gender (male= 0 female = 1, ,) education level (Primary =0, Secondary=1 and above=2 and No education=3), wealth index (continuous) , marital status (not married =0, married = 1) and employment status (Not employed= 0, Employed =1). I fitted a POM using ologit command in STATA 14.0 Proportional assumption checks on these variables using Brant test showed that location (p=0.000) and employment status (p=0.025) failed the test.

I hence performed a PPOM with gologit2 using autofit.

Please I have to issues that In need help and education with.

1. In the Brant test as mentioned earlier, two variables; location and employment were said to fail that test. However in the gologit2 output the only unconstrained variable is location. Please what does this mean and why is it so?

2. From the PPOM results, location, my primary predictor was not statistically significant and it had a positive coefficient (coeff.=0.3038553; p-value = 0.159) at the Poor nutrition behaviour versus fair nutrition behaviour and good nutrition behaviour level. But this changed to a significant p-value with a negative coefficient (coeff. = -0.795017 p-value = 0.000) in the behaviour and fair nutrition behaviour versus good nutrition behaviour level. Even as I have a fair idea why things are this way especially with the coefficients, I don't know much about the p-values. Another issue here is how to adequately interpret results for location.

Thank you again for your support.

[Bruce Weaver]

Hello Edem Kpewou. Welcome to Statalist. For the benefit of other readers, the command you are asking about is a user-written* command by Richard Williams. More info about it can be found by executing this command:

ssc describe gologit2

Have you seen this page on Dr. Williams' website? Some of the links there may be helpful to you.

* I think we are encouraged to use some other term now, but I can't remember what it is. Community-contributed, perhaps?

[Richard Williams]

Gologit2 and Brant do their testing a little differently so it is not unusual for their conclusions to differ a bit. Brant tests restrictions on variables one at a time. Gologit2 does consecutive tests — first the constraints on the biggest violator are freed up. Then, given those relaxed constraints, you test whether any other variables significantly violate the constraints. And so on until no remaining variables violate assumptions. My experience is that gologit2 tends to report fewer violators than Brant does.

Also, with gologit2, I usually suggest that autofit(.01) or something even stricter be used. With so many variables being tested, some apparent violations could just be due to chance alone. If I had my life to live over I would have made .01 the autofit default when I released the program in 2006.

I will add a few more things about interpretation later.

[Edem Kpewou]

Hello Edem Kpewou. Welcome to Statalist. For the benefit of other readers, the command you are asking about is a user-written* command by Richard Williams. More info about it can be found by executing this command:

ssc describe gologit2

Have you seen this page on Dr. Williams' website? Some of the links there may be helpful to you.

* I think we are encouraged to use some other term now, but I can't remember what it is. Community-contributed, perhaps?

Thanks Prof Weaver I very much appreciate your kind response. I have seen the page you're referring to and I must admit that most of the things I know about generalized logistic regression are learnt from there and I am grateful. I have taken my time to read his (Dr Williams) articles and those of others and they have been very educative. The issues I have posted about are quite still grey to me hence the request for further enlightenment from this forum.

[Edem Kpewou]

Gologit2 and Brant do their testing a little differently so it is not unusual for their conclusions to differ a bit. Brant tests restrictions on variables one at a time. Gologit2 does consecutive tests — first the constraints on the biggest violator are freed up. Then, given those relaxed constraints, you test whether any other variables significantly violate the constraints. And so on until no remaining variables violate assumptions. My experience is that gologit2 tends to report fewer violators than Brant does.

Also, with gologit2, I usually suggest that autofit(.01) or something even stricter be used. With so many variables being tested, some apparent violations could just be due to chance alone. If I had my life to live over I would have made .01 the autofit default when I released the program in 2006.

I will add a few more things about interpretation later.

Thanks Dr Williams for your response. It makes things better already and I am grateful. I am honored. I will be looking up to when you comment on the interpretation bit. Kind regards.

[Richard Williams]

Edem Kpewou Sorry for not getting back to you sooner.

Probably the most helpful sources:

https://www3.nd.edu/~rwilliam/xsoc73994/Margins05.pdf -- Describes how margins and related techniques can help in interpreting multiple outcome models, such as mlogot, ologit, and gologit

https://www.tandfonline.com/doi/full...X.2015.1112384 -- 2016 Paper entitled "Understanding and interpreting generalized ordered logit models"

https://www.stata-journal.com/articl...article=st0097 -- 2006 paper that introduced gologit2


Also potentially helpful:

https://methods.sagepub.com/foundati...ression-models -- Overview of various Ordinal regression models, including, briefly, gologit

https://methods.sagepub.com/foundati...dent-variables -- Ordinal Independent variables


Not directly relevant but possibly useful for related analyses:

https://www.sciencedirect.com/scienc...132?via%3Dihub -- "Comparing logit & probit coefficients between nested models" I want this to eventually become one of my most cited papers ever but it has a ways to go!

[Edem Kpewou]

Edem Kpewou Sorry for not getting back to you sooner.

Probably the most helpful sources:

https://www3.nd.edu/~rwilliam/xsoc73994/Margins05.pdf -- Describes how margins and related techniques can help in interpreting multiple outcome models, such as mlogot, ologit, and gologit

https://www.tandfonline.com/doi/full...X.2015.1112384 -- 2016 Paper entitled "Understanding and interpreting generalized ordered logit models"

https://www.stata-journal.com/articl...article=st0097 -- 2006 paper that introduced gologit2


Also potentially helpful:

https://methods.sagepub.com/foundati...ression-models -- Overview of various Ordinal regression models, including, briefly, gologit

https://methods.sagepub.com/foundati...dent-variables -- Ordinal Independent variables


Not directly relevant but possibly useful for related analyses:

https://www.sciencedirect.com/scienc...132?via%3Dihub -- "Comparing logit & probit coefficients between nested models" I want this to eventually become one of my most cited papers ever but it has a ways to go!

Thanks Dr. Williams. These are very helpful and I appreciate.