How do I test the parallel lines assumption when doing ordinal regression in JMP?

[#1]

It can be done in SAS but I can't find that this is available in JMP and it is not clear to me how to edit the ordinal regression script in JMP to get the results I need. 

[#2]

You could fit the ordinal logistic model and the nominal logistic model and compare their respective AICc values. If the additional slope parameters in the nominal logistic model are unnecessary, then the AICc should be higher. You could also determine the significance of the additional slope parameters in the nominal logistic model through the regression analysis.

[#3]

How do I do a nominal logistic model when the dependent variable is a three level categorical variable?

David Herrington, MD, MHS

Dalton McMichael Chair in Cardiovascular Medicine

Professor, Section on Cardiovascular Medicine

Vice-Chair for Research, Dept. of Internal Medicine

Department of Internal Medicine

Medical Center Boulevard \ Winston-Salem, NC 27157

p 336.716.4950 \ f 336.716.9188 \ pager 336.716.6770+pin 9249

dherring@wakehealth.edu \ WakeHealth.edu

[#4]

odd… Just tried again and this time “nominal logistic” option was available. Perhaps I missed it earlier. Sorry. - DH

David Herrington, MD, MHS

Dalton McMichael Chair in Cardiovascular Medicine

Professor, Section on Cardiovascular Medicine

Vice-Chair for Research, Dept. of Internal Medicine

Department of Internal Medicine

Medical Center Boulevard \ Winston-Salem, NC 27157

p 336.716.4950 \ f 336.716.9188 \ pager 336.716.6770+pin 9249

dherring@wakehealth.edu \ WakeHealth.edu

[#5]

If I'm understanding you correctly, if the AIC is higher for the MNL than it is for the OL then the parallel line assumption is not violated and the OL is correct model?

Why is that? 

Thanks

[#6]

The AICc is a useful criterion for model selection. It measures the reduction in bias (-2L) to avoid under-fitting but penalizes model complexity (2k(1 + (k=+1)/(n-k-1)) to avoid over-fitting and the resulting increase in variance. The model with the lowest AICc is the best model. You can compare two or more models using AICc. The one with lowest AICc is still the best model but how much support does the second best model have from the data? If the difference in AICc is less than 4, then the second best model has some support. If the difference in AICc is between 4 and 10, then the support is considerably less for the second best model. If the difference is more than 10, then there is essentially no support for the second best model.

This criteria could be used to compare the ordinal logistic regression model to the more complex nominal logistic model to determine which is best and if the additional complexity is worthwhile. This comparison is valid as long as the data sample is the same for both models.

Note: AICc is not a statistical hypothesis test and therefore cannot be used for inference. That is, it is meaningless to say that a model or a term in the model is significant or not based on the AICc or any other such criterion.