Marcello and Torsten Santavirta--
On a related note, if you are working from the example code in
http://pped.org/stata/ciwod.pdf, it reads:
ssc install gsample, replace
gsample 100, wor gen(pick)
levelsof z if pick==1, loc(p)
foreach val of local p {
cap drop newz
g newz=z-'val'
bootstrap r(d), reps(1000): discont y znew
bootstrap r(d), reps(1000): discont xt znew
}
but the example given by the poster has
levelsof x2co if pick==1, loc(p)
and
g newz=x2co-'val'
which makes no sense--the idea is to pick a random Z to be the new Z0,
where we don't expect to find a discontinuity, and see what proportion
of the time we would reject the null of no discontinuity. For this,
we need to define a new assignment variable newz based on a possible
choice of Z, not on X.
I should also point out that if you want to put standard error bars
around every point, you will need to modify the program to save the
local linear regression estimates at every point (to bootstrap) or add
an se() option to the -lpoly- command, to use analytic SEs. If you
don't want to put standard error bars around every point, you can
modify the program to compute the local linear regression estimates at
only the cutoff point, which will cut down execution time considerably
(esp with 1000 reps on your bootstrap).
In the example, where the cutoff is at 0, the current code in -discont- reads
loc step=r(max)/50
local N=50+floor(-r(min)/(r(max)/50))
cap set obs 'N'
qui g 'z'=(_n-1)*r(max)/50 in 1/51
qui replace 'z'=-(_n-50)*r(max)/50 in 51/'N'
(creating a z variable with 50 steps above the cutoff and an
appropriate number of steps of equal size below the cutoff, where the
local `step' is superfluous and left over from a more elaborate
version of the program) but those 5 lines could be usefully rewritten
as
qui g 'z'=0 in 1
in a parallel version of the program designed not to graph the local
linear regressions, but just for bootstrapping the jump, as was
pointed out to me last week by Vince Wiggins. The program -rd- on SSC
should also be modified to either (1) give SEs at every Z value, or
(2) only bootstrap the jump at the cutoff. I will try to modify it in
the next few days.
On Dec 6, 2007 11:16 AM, Austin Nichols <[email protected]> wrote:
> Marcello--
> After installing -moremata-, close Stata and open Stata again. Then
> Mata should be able to find everything you need.
>
>
> On Dec 6, 2007 11:02 AM, Marcello Pagano <[email protected]> wrote:
> > Hello,
> > I would like to use gsample sampling command for cross validation of my
> > regression discontinuity as suggested bu Austin Nichols in "Causal
> > inference with observational data".
> > The idea is to randomly choose placebo cut off points and test
> > continuity of the forcing variable and the outcome variable. Now Nichols
> > suggests generating random samples from the data. When using the gsample
> > command stata suggests that mm_sample() is required asks me to install
> > moremata. Having installed moremata the same problems occurs again. What
> > could be the problem. I'm using stata 10.0 and my code looks as follows:
> >
> > ssc install gsample, replace
> > gsample 100, wor gen(pick)
> > levelsof x2co if pick==1, loc(p)
> > foreach val of local p {
> > cap drop newz
> > g newz=x2co-'val'
> > bootstrap r(d), reps(1000): discont y znew
> > bootstrap r(d), reps(1000): discont x2co znew
}
> >
> > Does anyone have an idea what goes wrong with using gsample?
> > Thank You in advance,
> > Torsten Santavirta
> > Helsinki School of Economics
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
