Re: st: Regression Discontinuity Duration Analysis

[Austin Nichols <austinnichols@gmail.com>]

Jen Zhen <jenzhen99@gmail.com> :

You can simply restrict to small ranges of the forcing variable in the
vicinity of the cutoff.  You can also restrict attention to a range of
the forcing (assignment) variable Z by generating weights that decay
linearly as you move away from the cutoff, and include a dummy for Z
above the cutoff Z0, and the interaction with Z.  In the linear
regression case, that approach should be equivalent to a local linear
regression on either side of the cutoff, since the coef on the dummy
for Z>Z0 is the estimate of the jump at Z0.  If you take that
approach, make sure you try different ranges (bandwidths) to ensure
your estimates are not sensitive to bandwidth choice.  You may also
want to bootstrap, though you will have to do some more work to
properly -stset- etc. in that case.

Probably there are issues due to the nonlinearity of the link
function, i.e. E(y|X) is not linear in X, but I am guessing that would
change the interpretation of your results rather than change the basic
approach.  Someone who uses -stcox- might have ideas, or know of some
discussion of this in the literature.

Here is a somewhat silly example:

webuse nhefs, clear
svyset psu2 [pw=swgt2], strata(strata2)
stset age_lung_cancer if age_lung_cancer<. [pw=swgt2], fail(lung_cancer)
svy: stcox former_smoker smoker male urban1 rural age_nhanesi
g w5=max(0,5-abs(age_nhanesi-65))
line w5 age_nhanesi, sort
g new=swgt2*w5
svyset psu2 [pw=new], strata(strata2)
stset age_lung_cancer if age_lung_cancer<. [pw=new], fail(lung_cancer)
g age=(age_nhanesi-65)
g above=(age_nhanesi>=65)
g ageXabove=(age_nhanesi>=65)*age
svy: stcox former_smoker smoker male urban1 rural age above ageXabove


On Thu, Jul 29, 2010 at 9:21 AM, Jen Zhen <jenzhen99@gmail.com> wrote:
> Dear Listers,
>
> I would like to estimate a Regression Discontinuity equation, where
> the outcome is a duration.
> So it seems that I can't simply use Austin Nichols' -rd- command, but
> should better use some version of a Cox Proportional Hazard model,
> using -stcox-.
>
> So right now I am estimating a Cox model, using as regressors just the
> dummy for being above the threshold and the forcing variable, but this
> seems somewhat restrictive in that I have to assume that the effect of
> the forcing variable on the duration outcome is linear. So I am
> wondering how I could best incorporate a less parametric control
> function into the -stcox- command?
>
> Thank you for your suggestions and best regards,
> JZ
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/