Changing reference category affects estimates for multinomial model

[emilybennitt]

Good day,

I am fitting a multinomial model to a dataset, after going through the LEMMA course. My response variable has three categories of behaviour: rest, graze and walk. I need to identify parameters that can distinguish between those categories, but none of them is an obvious reference category. I ran the model with Rest as the reference, then again with Graze as the reference, since that should give me all possible combinations. However, I have found substantial differences in the parameter estimates between the two models with different reference categories, as well as between the chi-squared estimate for the overall models. I had to clear the first model, then build the second one again because I could not merely change the reference category.

Please could you advise me whether I am going wrong somewhere?

Many thanks
Emily

[ChrisCharlton]

The IGLS estimation can sometimes struggle with multinomial models, so you might want to also run them using MCMC to see if you get the same behaviour there. If you can send any example worksheets (you can find my email on http://www.bristol.ac.uk/cmm/team/) then we can look into this further.