Considering a panel dataset where N (units) ~50 and T (months) ~300. I assume this is considered small N Large T, is that right?
My understanding is panel data models such as fixed- and random-effects models are more suitable for large N small T datasets (will produce unbiased coefficients asymptotically with large N given a correctly specified model). How strong is this assumption? In other words, if I use FE/RE models for the type of dataset I mentioned above, would I get wrong estimates/model?
If yes, what are the alternatives? I read some posts suggesting -xtpcse- (which I am not familiar with), others suggesting SUR, and some mentioned that dynamic panel models (such as -xtabond-) can be a solution.
Also, my understanding is that FE/RE models work according to the assumption that the sample (N) is random. If the data consists of N units that represent the entire population, therefore, they are not random (say 20 stations at each we measure a dependent continuous variable and we would like to model that wrt some independent variables some of which are time-variant variables), how do we go about this?
I suspect that I will have to deal with heteroskedasticity and endogeneity issues. So, a tool/model that can accommodate instrumental variables and robust/cluster standard error estimation might be required.
Any suggestions/directions would be very helpful. Thanks!
I posted a similar post here as well and I will try to share useful replies.
My understanding is panel data models such as fixed- and random-effects models are more suitable for large N small T datasets (will produce unbiased coefficients asymptotically with large N given a correctly specified model). How strong is this assumption? In other words, if I use FE/RE models for the type of dataset I mentioned above, would I get wrong estimates/model?
If yes, what are the alternatives? I read some posts suggesting -xtpcse- (which I am not familiar with), others suggesting SUR, and some mentioned that dynamic panel models (such as -xtabond-) can be a solution.
Also, my understanding is that FE/RE models work according to the assumption that the sample (N) is random. If the data consists of N units that represent the entire population, therefore, they are not random (say 20 stations at each we measure a dependent continuous variable and we would like to model that wrt some independent variables some of which are time-variant variables), how do we go about this?
I suspect that I will have to deal with heteroskedasticity and endogeneity issues. So, a tool/model that can accommodate instrumental variables and robust/cluster standard error estimation might be required.
Any suggestions/directions would be very helpful. Thanks!
I posted a similar post here as well and I will try to share useful replies.
Comment