在用2SLS解决内生性问题的时候，如果要考察的是一个treatment, (D), which is dummy varialbe that takes the value of 1 if receives the treatment. 它的instrument 是另一个或多个dummy variables, (Z1,Z2...).  用传统的2SLS来estimate很可能是biased的，最新的改进方案是用Wald estimator, 也叫LATE （Local Average Treatment Effect）. 这里我附件了一份参考文件，它后面有个例子，但是用STATA实现的，不是很懂。请问哪问高手知道Wald estimator,用SAS怎么实行的。跪拜，感谢.



那个例子用STATA是这样的：

clear
set seed 54687
set obs 20000
/* first, randomly assign the instrument - say half-half */
ge z = uniform()>.5
/* then, generate never-takers (d00), always-takers (d11) and compliers
(d01), independent of z */
ge d00=(_n<=5000)
ge d11=(_n>5000 & _n<=10000)
ge d01=(_n>10000)
/* observed outcomes: always zero for never-takers, always one for
always-takers, depends on the IV for compliers */
ge D=d11+z*d01
/* now give the three groups different LATE. Without loss of
generality, assume within group homogeneity. */
ge late=-1 if d00==1
replace late=0 if d11==1
replace late=1 if d01==1
/* next generate potential outcomes y0,y1 */
ge y0=0.25*invnorm(uniform())
ge y1=y0+late
/* actual outcome depends on treatment status */
ge y = D*y1+(1-D)*y0
/* the average treatment effect is simply the sample mean of late */
sum late
/* OLS doesn't give you ATE or LATE */
reg y D
/* IV gives you the LATE for the compliers */
ivreg y (D=z)
exit