It seems I can only get the fixed portion when using margins after running xtmelogit model with a cross-classification structure. Does anyone know whether STATA can calculate the estimate probabilities that take both fixed and random effects into consideration?
until recently i only knew one way of doing what you are after in Stata, that uis by the post-estimation command
gllapred prob, mu marg
used after estimating your model with the gllamm suite. But the latest versions of the me commands may do the trick as well.
Could you give me an example of using gllamm? I am aware of this command but never used it in real life. Let's say my model is xtmelogit read c.grade##c.grade race || _all: R.school:, || student:, mle covariance and after running the model I would want to get the predicted probabilities (that take both fix and random effects into consideration) for each of the racial groups. I know for sure that predict, mu marginal won't work after running the model using xtmelogit. Thanks!
Take a look at this http://biostat.jhsph.edu/~fdominic/teaching/bio656/software/gllamm.manual.pdf . I have never estimated cross-classified models with gllamm but the suite is quite flexible and I think it does accommodate more complex specificationas for the random part. From the title of your post I thought you were after gee probabilities (i.e. obtained by integrating out the distribution of the random intercepts). In fact you seem to (also/instead?) want subject-specific predicted probabilities. So the code I gave you above refers to the former, but gllamm can also give you the latter simply by removing the "marg" option.
Thanks Luca. I will take a look at the manual. So one thing you are sure about is that using xtmelogit along with the margins command won't work in my case, right? I tried an example listed on this website http://stats.idre.ucla.edu/stata/faq/how-can-i-compute-predictive-margins-for-xtmelogit-with-random-effects/ But since in their example, they are not using a cross-classification model, and I found myself in a desperate situation where after running the xtmelogit model, predict re*, reffects produces me nothing.
gllamm stands for Generalized Linear Latent and Mixed Models so it is designed to handle non-linear models. It is also designed for hierarchical structures. However, the cross-classified model can be fitted by defining a pseudo top-level unit. The trick is explained in chapter 8 of Rabe-Hesketh, S. and Skrondal, A. (2005) Multilevel and Longitudinal Modeling using Stata. College Station, TX: Stata Press. If you can't find any simpler solution then try and get hold of this reference.
Bottomline is: yes you can solve your problem in stata using gllamm.
- Badges
- Science topic