www.cmm.bristol.ac.uk/forum

Hi everyone,
I got a question when I used MLwiN to estimate a multilevel multinomial logit model. After I finished the model estimation, I tried to calculate ICC (intr-class correlation). I tried to find out the level-one (individual-level) residuals variance. The estimation results show that level-one residuals variance seems equal to 1. I am not sure if I read the results right or not because for multinomial logit model, the level-one residuals is distributed as Gumbel distribution with variance π2/3. If the level-one residuals variance equal to 1 and mean equal to 0, it should be a multinomial probit model.
So, my question is:
1. what's the assumption of the level-one residuals distribution for multilevel multinomial logit model in MLwiN??
2. How to find out or calculate the level-one residuals variance for multilevel multinomial logit model in MLwiN??

Hope someone can answer my questions. Thank you very much.

Chien-Pang

In the formulation that MLwiN uses the variance of one that you see is actually a multiplier for the variance (if you turn on extra-multinomial variation then this will be estimated instead of being fixed to one). For details of the formulation see section 4.4 of Harvey Goldstein's book (Multilevel Statistical Models, 4th Edition, http://onlinelibrary.wiley.com/book/10. ... 0470973394).