Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: RD and binary outcomes


From   D-Ta <altruist81@gmx.de>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: RD and binary outcomes
Date   Tue, 18 Oct 2011 18:45:20 +0200

I am totally on your side. I just used the flexible parametric model as a comparison. And, as mentioned, the standard errors of the -rd- estimator are oftentimes even smaller when the outcome is continous. Also, the point estimates are largely in line. What I do not understand, however, is why -rd- point estimates differ considerable from the parametric counterparts and standard errors are so much larger when the outcome is binary.

This is not about publications (at least not at this stage), I just want to understand the properties of the -rd- procedure when the outcome of interest is binary. Is -rd- adequate in such (binary) settings or shouldnt one rather use a local logit version of -rd- ?

Darjusch

Am 18.10.2011 17:23, schrieb Austin Nichols:
D-Ta<altruist81@gmx.de>:
It is true that -rd- uses only linear models, but this is by design.
I have seen many papers using a parametric model, but it is not always
clear that the parametric model is justified.  The strength of the RD
design is that it relies on so few assumptions; using a parametric
model adds many more assumptions, and it is not always clear what they
are.  Think about what IV model you would run, using assignment to
treatment (one side or the other of the cutoff) as an instrument for
treatment received (for a linear model this is easy, otherwise not so
much); how would you weight (i.e. what kernel and bandwidth choice
would you make)?  I would want some simulation evidence for any
application of a parametric model to an RD design, using data built to
look like the problem at hand, before I trusted the results.  Smaller
standard errors are often desirable not because they are closer to the
truth, but because they lead to higher publication probabilities!  In
the case of RD, the large standard errors often reflect true
uncertainty or variability of estimates, not simply the sample loss
that accompanies focusing on the discontinuity.  But I am sympathetic;
looking only near the discontinuity increases internal validity at the
cost of sample and leads many apparently strong effects to look
statistically insignificant.  Start with a good simulation, and make
sure you put in some wiggles in the conditional mean functions (under
treatment and no treatment).

On Tue, Oct 18, 2011 at 9:41 AM, D-Ta<altruist81@gmx.de>  wrote:
Dear Stata-Users,

I am currently working with the RD command provided by  Austin Nichols. I
investigate continous as well as binary outcome variables. I use cubic and
quartic parametric models as benchmarks to compare the RD results against.

In the case of continous variables the RD command does fine (i.e. very close
to the parametric model with even smaller standard errors). However, when
looking at binary outcomes, the RD procedure produces much wider standard
errors and the point estimates differ a lot from the parametric (probit)
specifications.

I assume that this is due to the underlying local linear model used by RD.
(Perhaps it would be better to use a local logit(?))

Am I right that -in its current format- the RD command is not the right tool
to use in case of binary outcome variables?

Many thanks

Darjusch

*
*   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/

*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index