The Stata code to implement the method is:
“ivregress 2sls Y C (X = D DW) [k]”
Y is the outcome
X is the vector of endogenous variables. In the paper these are dummies, but it works with any set of variables
C=(1,W',Z,DZ,ZW',DZW)'. If this is not clear, see details in the Estimation section in the paper.
D is the indicator of crossing the threshold, 1(Z>=0)
DW is the product of the threshold indicator D and the covariates W
k is the kernel weights. We recommend the triangular kernel: k_i=(1-|Z_i/h|)1(|Z_i|<=h), where h is the bandwidth.
We do not have a theory for the optimal bandwidth choice yet, but we are working on this. We recommend reporting results for several bandwidths, though the method is fairly robust to bandwidth changes in our simulations.
