Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


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

Re: st: Difference-in-Differences and Panel Data - In search of an adequate regression


From   Sjoerd van Bekkum <[email protected]>
To   "[email protected]" <[email protected]>
Subject   Re: st: Difference-in-Differences and Panel Data - In search of an adequate regression
Date   Thu, 3 Jan 2013 18:32:05 +0100

Austin, I fully agree. Nobody would contest clustering (my first point
merely points out an even more severe issue in Barbara's original
post). My model contains group fixed effects, not individual ones, but
your point (I think) that one cannot include both is valid since a
group FE model would be nested within an individual FE model.

However, I think neither of our points is exactly what Barbara means,
judging from her latest reply (in new thread but copied below):

Barbara, if you have, say, N*T observations for N individuals, T time
periods, and J treatment groups (J <= N), clustering over id indicates
N independent observations. However, if you suspect that individuals
within the groups are correlated, you should cluster over groups
rather than id. This indicates only J independent observations.

---

**** Barbara's latest message: ****

Dear Sjoerd,

thanks a lot for your kind reply.
I got the basic assumptions of the DiD and my model contains (of
course) the treatment group and time dummies as well as the
interaction term. I was too brief on that, sorry!
I am just too confused about the clustering: With panel data, I surely
cluster for the individuals,  let's say their id:

reg y   x   d1   d2    d1d2, cluster(id)

this accounts for the correlation between the residuals that arise
because the observations within one individuals are likely to be
non-independent.

but how do i account for common group errors, i.e. errors that might
affect the whole treatment group?  Am I done clustering for the
individuals? Or am I just terribly wrong?

I would greatly appreciate another kind answer.

Best from Berlin,

Barbara


On 3 January 2013 16:13, Austin Nichols <[email protected]> wrote:
>
> Sjoerd van Bekkum <[email protected]>:
> OP asked about FE, which presumably are collinear with the treatment
> dummy and its interaction. You did not include fixed effects in your
> model.  As I noted, the cluster-robust SE *do* make sense, but the FE
> probably not (unless some FE not collinear with the treatment are
> meant).
>
> On Wed, Jan 2, 2013 at 10:09 PM, Sjoerd van Bekkum <[email protected]>
> wrote:
> > Maybe my post was too brief. I think what Barbara wants to do (and
> > what I meant in my previous post) is, assuming two groups and 2
> > (pre-/post) periods:
> >
> > y = a + b*D(treatment=1) + c*D(post=1) + d*D(treatment=1)*D(post=1) + e,
> >
> > where D are indicator variables. As mentioned in the paper I cited
> > above, this leads to the following groups:
> >
> > E[y|treated, post] =a+b+c+d
> > E[y|treated, pre] = a+b
> > E[y|not treated,post] = a+c
> > E[y|not treated, pre] = a
> >
> > with the dif-in-dif captured by
> >
> > DID = {E[y|treated, post]-E[y|treated, pre]} - {E[y|not
> > treated,post]-E[y|not treated, pre]}
> >       = {a+b+c+d - (a+b)} - {a+c - a}
> >       = d
> >
> > with cluster-robust errors, as Austin mentioned. I don't see any
> > collinearity problems here.
> >
> >
> > On 3 January 2013 02:03, Austin Nichols <[email protected]> wrote:
> >>
> >> Barbara Engels <[email protected]> :
> >> You should certainly use cluster-robust SE to account for repeated
> >> observations, but how could you include FE and a dummy for treatment
> >> group? With a post dummy, and a treatment dummy, and the interaction,
> >> there would be a severe perfect collinearity problem.
> >>
> >> On Wed, Jan 2, 2013 at 3:49 PM, Barbara Engels <[email protected]>
> >> wrote:
> >> > Dear Stata people,
> >> >
> >> > I am currently working on a difference-in-differences model in its
> >> > simplest form  - treatment and control group, pre- and
> >> > post-intervention
> >> > period.
> >> > However, I got a large panel data set and I wonder what is the best
> >> > way
> >> > to estimate the DID in Stata to account for flaws like serial
> >> > correlation.
> >> > Should I go for a simple
> >> >
> >> > reg y x  incl. interaction term, ROBUST
> >> >
> >> > Or should I apply clustering?
> >> > Or even xtreg with fe?
> >> >
> >> > Any help is greatly appreciated.
> >> >
> >> > Thanks a lot, happy 2013!
> >> >
> >> > Barbara
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index