Still using the ordered probit model when the parallel regression assumption for it is violated!

[Bui Thanh Yen Thao]

Hi All,
I face the problem: the parallel regression assumption for the ordered probit model is violated. It is usually advised that we should alternate other possible models: multinomial logit model, generalized ordered logit model. However, I would like to use the ordered oprobit for my quantitative model after some following explanation.

"The drawback of using the multinomial logit model is that it does not preserve the inherent ordering of the categories of dependent variable (6 values) and therefore does not incorporate this information when estimating the coefficients of the explanatory variables. This results in a loss in the efficiency of the estimators. While the generalized ordered logit model provides an alternative model that does preserve the ordering (e.g., it is a restricted version of the multinomial logit model), it is very sensitive to low frequency counts (e.g., small cell sizes). Thus, it is often necessary to combine the dependent variable categories that have low frequencies with adjacent categories in order for the estimation procedure to work. However, combining categories may also lead to a loss in information, especially if the underlying latent variable is multi-leveled or continuous. As a result, we have chosen to present the results from the ordered probit model. A larger sample size and fewer explanatory variables would have made the use of generalized models more feasible."

I wondered what happens or if there is any important mistake when we still keep the original model (oprobit). Could you please help me to understand this situation? Thanks.
I am a new-comer in this field, specially with ordered probit model. However, this is the problem for my thesis. Pls help me!
YT

[ben earnhart]

Does ordered logit really return different substantive findings than ordered probit? Can you show us the output from the two models? If your dependent variable is ordered, then multinomial logit is clearly inappropriate. But if ologit and oprobit are giving different results, then it does seem to take some strong argumentation why you're sticking with one over the other.

[Richard Williams]

I don't think Bui is saying that ologit and oprobit are giving substantively different results. If they were, that would be pretty unusual. Rather, he is saying that he has somehow tested the parallel regressions assumption of the model and found that it is violated.

How much of a problem this is is hard to say, at least without seeing more. Sometimes statistically significant violations can have substantively trivial consequences.

You might try some of the other approaches mentioned and see if they lead to different conclusions or if they would create more problems than they solve. As it now stands, you are just asserting the other methods would be worse than sticking with ordered probit, but at least in the one paragraph presented there is no proof of that.

[Bui Thanh Yen Thao]

Hi Richard Williams and Ben Earnhart,

Thank you for your guidance.
Yep, I would like to mention the violation of testing parallel regressions assumption.
I tried to run alternative model like: mlogit and gologit but the iteration can't stop.
So we can use oprobit, can't we?