Darrin Aase posted on Wednesday, August 05, 2009 - 8:44 am
I've searched the discussion board and not found any posts on this in a couple of years. First, is there an example of testing the proporitonal odds assumption in MPlus for an ordinal outcome, specifically with longitudinal data? Second, is there yet a way to estimate a non-proportional odds model?
1. As a series of binary regressions (see Brant's paper)
2. The full model involves this Say U =0,1,2 –> create dummy out of the three categories D1 D2 if you have U on X you will have to change the model to U on X, XD1, XD2 If you have q X variables you will have instead 3q X variables. Before proceeding further you should make sure that you understand the intricacies of the full model and why it doesn't guarantee P>0 as virtually all other models do.
Thank you for your response & guide. I already read the materials (2 of them) and understand on the concept of Brant test, also already run simple ordered logistic and look into the result of Brant test.
But is it means that MPLUS not able to generate the coefficients for partial proportional simultaneously? is it generate a series of binary logistic is the only option? Im not so clear on the second option given above.
because actually im planning to fit a multilevel model for ordinal response as my main model. is it possible to do this using MPLUS?
First to clarify - there is an error in my earlier post. Method 2 above doesn't work. Here is what you can do in Mplus.
Method 1: using the binary regressions. The method works ok and would produce consistent estimates. Here is a sample code that assumes 1 covariate and a categorical variable U that has 3 categories: 0,1,2
variable: names are x u; usevar are x u1 u2; categorical=u1 u2; define: if (u==0) then u1=0 else u1=1; if (u<=1) then u2=0 else u2=1; model: u1-u2 on x; analysis: link=logit; estimator=ml; data: file=1.dat;
Method 2: The full non-proportional model is possible in Mplus using the constraint command User's Guide example 5.23 features that option. The sample code for the non-proportional model would be as follows
variable: names are x u; categorical=u; constraint=x; model: u on x; [u$2] (t2); model constraints: new(a b); t2=a-b*x; analysis: link=logit; estimator=ml; data: file=1.dat;
Method 1 can be used for two-level models but Method 2 can not be used for two-level models because the constraint option is not available for two-level models. You can use Method 2 with type=complex to account for non-independence of the observations across clusters.