Panel Data: Conducting N time-series regressions and averaging coefficients

#1
Hi,
I stumbled upon a special regression method in a published paper (A+ journal) that I hadn't heard before. In dealing with a panel data set, the authors perform a time-series regression for each single entity. As they explain, "The reported coefficient values are averages of the resulting N regression estimates for the coefficient on each variable. Associated t-statistics are calculated from the cross-sectional variation over the Nj estimates for each coefficient by dividing each reported coefficient value by the standard deviation of the Nj estimates and scaling by the square-root of N."

I know of pooled regressions (with and without clustered errors), fixed, random, between effects and dynamic panel regressions, but I haven't heard of that method which seems odd to me. Apparently, they don't even give it a name.

If anybody recognizes this method or just would like to comment on it, I'd be happy about any input. I'm particularly interested how it compares to pooled or fixed effect regressions, if the t-statistics can be viewed as unbiased, etc.
In case this method is quite common, I'd be happy to hear how it is implemented in standard statistic packages like Stata, eviews, or others.

Thanks,
Steve
 
Last edited:
#2
Hi Steve,

I have never heard of this method, but I am having a problem which seems to be similar to yours, from data structure point of view. So any insights you find can be valuable to me as well.
 

OKar

New Member
#3
To me it sounds like the Mean Group Estimator (MGE or just MG), first introduced by Pesaran & Smith in 1995 in the Journal of Econometrics. There is a user created package in Stata for doing this called "xtmg". Its helpfile explains the procedure briefly. I have only used this with dynamic panels (where lagged dependent variables are used as independent variables), but it should work with any kinds of panels. The idea is that this model allows heterogeneity in coefficients between panel groups. It calculates regressions for each panel group separately and then takes the average of the estimated coefficients. You may also want to check out it's "random effects counterpart", the random coefficients model (xtrc in Stata).
 
#4
hi NN_stat did you find the solution to this problem?? i am stuck in the same problem, any help would be highly appreciated
thanking you in anticipation


Hi Steve,

I have never heard of this method, but I am having a problem which seems to be similar to yours, from data structure point of view. So any insights you find can be valuable to me as well.