Dear Stata users,
I am working with xtabond2, the one step difference GMM i.e. main equation in differences and instrumenting with second lag. I am trying to replicate the estimator myself. When I look at the H matrix in post estimation e(H), it's a diagonal with 2s and -1s on the off diagonals as specified in Roodman (2009). However, the size of the matrix is T by T where the number of observations in the panel is NT (the panel is treated as balanced by substituting 0s for missing in Z, X, Y). The dimensions of Z are TN by (T-1)*(number of endogenous variables). How is H made conformable in order to obtain the first step estimator beta1 = (X'Z(Z'HZ)^-1Z'X)^(-1)X'Z(Z'HZ)^-1Z'Y ? I rep the H matrix (N,N) so as to make it TN by TN but the resulting matrix is singular. When I take A1 from post estimation I am able to get beta1 from above. So ultimately I wonder how is A1 found?
Thank you!
I am working with xtabond2, the one step difference GMM i.e. main equation in differences and instrumenting with second lag. I am trying to replicate the estimator myself. When I look at the H matrix in post estimation e(H), it's a diagonal with 2s and -1s on the off diagonals as specified in Roodman (2009). However, the size of the matrix is T by T where the number of observations in the panel is NT (the panel is treated as balanced by substituting 0s for missing in Z, X, Y). The dimensions of Z are TN by (T-1)*(number of endogenous variables). How is H made conformable in order to obtain the first step estimator beta1 = (X'Z(Z'HZ)^-1Z'X)^(-1)X'Z(Z'HZ)^-1Z'Y ? I rep the H matrix (N,N) so as to make it TN by TN but the resulting matrix is singular. When I take A1 from post estimation I am able to get beta1 from above. So ultimately I wonder how is A1 found?
Thank you!
