www.cmm.bristol.ac.uk/forum

Dear all,

I have a longitudinal dataset of firms and apply a 3-level logistic regression with time as level 1, firms as level 2 and regions as level 3 using Bayesian methods with the runmlwin command in Stata. I now wonder how I can account for possible endogeneity issues in my model that might arise from correlations between level 2 firm characteristics and omitted regional variables at level 3. For example, I have a variable on the R&D expenditures of firms. However, the decision of firms on how much money they spend on R&D is likely to be influenced by regional factors that I cannot observe. Is there any established procedure how I can deal with these kind of endogeneity issues? And if so, how can I implement this with the runmlwin command?

I am happy about any help!

Best,
Thore

I asked George about this and his reply was as follows:

I guess you could apply ideas from fixed versus random effects to panel data which are motivated by endogeneity concerns.

So you could enter region as fixed-effects dummy variables instead of as random effects.

The effects of the firm covariates will then only be pure within region comparisons.

Thank you very much for the response! From a practical point of view, would that mean to change the model to a 2-level model with time as level 1 and firms as level 2 and then adding i.region as an additional explanatory variable to the fixed part of the model?

That would be my interpretation of George's suggest as well.