Time series cross section forecasting with R I have a (I suspect) simple question. I have time series cross section data on voting behaviour in the Council of the European Union (the monthly number of yes, no and abstentions for each member state from 1999 to 2007). So basically the variables are counts, thus a Poisson/negative binomial regression would be appropriate, possibly with lagged dependent variables on the right hand side to control for time dependencies. I have seen papers with people using such negative binomial models to forecast, for instance the number of monthly legislative acts adopted in the future, and I have three questions in this regard:


*

*How can i run a negative binomial regression on panel data without making any inferential mistakes?

*How can I use a negative binomial model with lags to forecast future values of the dependent variable.

*Can this be done in R?
Thomas
 A: After a bit of research, I can give a partial answer. In his book  Wooldridge discusses Poisson and negative binomial regressions for cross-section and panel data. But for regression with lagged variables he only discusses Poisson regression. Maybe negative binomial is discussed in the new edition. The main conclusion is that for random effects Poisson regression with lagged random variable can be estimated by mixed effects Poisson regression model. The detailed description can be found here. The mixed effects Poisson regression in R can be estimated with glmer from package lme4. To adapt it to work with panel data, you will need to create lagged variable explicitly. Then your estimation command should look something like this:
glmer(y~lagy+exo+(1|Country),data,family=quasipoisson)

You should also look  into gplm package suggested by @dickoa. But be sure to check, whether it supports lagged variables. Yves Croissant, the creator of gplm and plm packages writes wonderful code, but unfortunately in my personal experience, the code is not tested enough, so bugs crop up more frequently than in standard R packages. 
A: May be you can take a look at the pglm package (from the same author of plm), use the family negbin.  You can also try from a Bayesian point of view the MCMCglmm package.
