ICC for negative binomial multilevel model

[Dirk Enzmann]

Is there a way to calculate an intraclass correlation of an intercept only negative binomial multilevel model with two levels? I am looking for a working example.

Searching for an answer I found answers to two related questions. The first discusses this question for lm4 using R (not Stata) (see: http://stats.stackexchange.com/quest...xed-model-in-l ), the other gives a negative answer (see: http://www.stata.com/statalist/archi.../msg01195.html ), but maybe meanwhile a positve answer (or an answer that explains why it would make no sense to try to calculate an ICC in this situation) is possible?

[Clyde Schechter]

I think it makes no sense to calculate an ICC following a negative binomial 2-level model. The ICC is defined as var_u/(var_u + var_e), where var_u is the variance of the intercepts at the higher level and var_e is the variance of the residuals at the bottom level. After linear regression, var_e is estimated from the data; its existence is a presumption of the model. After logit or probit estimation, var_e is a constant attribute of the logistic or normal distribution ((pi^2)/3 or 1, respectively) that characterizes the latent variable. In the case of the Poisson or negative binomial regression, there is no latent-variable formulation of the model, and there is no variance to estimate from the data because it is an assumption of these models that the variance of the error term is a function of the predicted value (e.g. in Poisson variance = mean). Since there is no such thing as var_e for these models, there is nothing from which to calculate ICC.

[Dirk Enzmann]

Thanks!

[vahid yazdi]

Dear Clyde schechter
Hi, I also am working on a two-level (students negative binomial regression model in Stata software. I selected an outcome variable (a count variable related to behavior of students). this variable is a variable with over-dispersion (no equality of variance and mean). I have a question that how can I calculate the variance of bottom level (e.g individual level in case of students within schools) in final model. The output of Stata give a var(cons) for school level which is statistically significant. However it does not show as for individual level. I want to know that how much the variance of outcome variable is explained by individual level (i.e students). please guide me more.

[Steve Samuels]

Welcome to Statalist, Yshid. Be sure to read the FAQ here (especially FAQ 12) and extra advice here. In this case, you've asked a question on a thread that has been closed by the original poster because his question was answered. Your question is unrelated (has nothing to do with ICC), so start a new topic on the main forum page (+ Topic button).