-----Urspr?ngliche Nachricht-----
Von: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]
Gesendet: Sunday, December 05, 2004 8:28 PM
An: gov2001-l(a)lists.fas.harvard.edu
Betreff: Re: [gov2001-l] Problem 2c
Jens, I disagree.
If by endogenaity, you mean that there is an *un*observed
covariate that determines whether the unit actually recieves
treatment if assigned, then we don't have that problem here
because we actually observe the characteristics that matter.
Thus, retirees are more likely to be home and more likely to
vote and this we control for this by matching on age. (And
similarly for the other X.) 2(c) is just to show that there
is
lack of covariate balance. Reciept of treatment *isn't*
endogenous and there is no need for IVE in this model.
If you still don't believe me, come to OH, but please don't
confuse your classmates.
Yours,
Olivia Lau
----- Original Message -----
From: "Jens Hainmueller" <jens_hainmueller(a)ksg05.harvard.edu>
To: <gov2001-l(a)lists.fas.harvard.edu>
Sent: Sunday, December 05, 2004 8:07 PM
Subject: AW: [gov2001-l] Problem 2c
Thanks Olivia. One thing is still unclear to me.
The point of checking for covariate balance is to
see if
the
group that recieved treatment differs from the group
assigned
control. We don't really care about the group assigned to
treatment that didn't recieve the treatment except in so
far
as
it changes p(treat = 1 & comply = 1|X) compared to p(treat
=
0|X) because we are looking for the average treatment
effect
of
Y(1) - Y(0), where Y(1) is the potential outcome under
treatment
and Y(0) is the potential outcome under control. Neither
of
these are observed for the group assigned treatment that
did
not
comply.
I don't understand why we do not care about the group
assigned
to the
treatment that didn't received the treatment, because in
our
case receiving
the treatment is *endogenous*, right? Imagine that those
with
little time
are less likely to pick up the phone and those are also
less
likely to vote.
Since we do not control for time availability, doesn't this
completely
confound the ATE estimate using your method (due to
selection
to treatment)?
The only thing that is random is *assignment* to treatment.
So
it may make
sense to check covariates balance between those assigned
and
those not
assigned. Receiving the treatment is endogenous, however,
and
thus checking
the covariates balance does not help much because it
doesn't
tell us whether
those that received the treatment are systematically
different
than those
that did not in some crucial unobservable characteristic.
Your estimate is based on the assumption of selection on
observables, i.e.
that Y1, Y0 are orthogonal to D|X. (with D being the
treatment). This
assumption of conditional random assignment says that you
believe that you
control for all confounding Xs, conditional upon which D is
random. This
seems totally unrealistic to me in our case, and even if
this
assumption
were to hold we would not need a randomized experiment. As
far
as I
understand the point of randomization is to overcome the
problem of bias
originating from unobservable characteristics.
In our case, the only way I can think of to get around this
endogneiety bias
(and make the advantages of randomization work), is to use
assignment (which
is random) as an instrument (Z) for receiving the treatment
(D). Use a Wald
Estimator: Local ATE
={E[Y|Z=1]-E[Y|Z=0]}/{E[D|Z=1]-E[D|Z=0]}
the
denominator here is proportion of people that take up
treatment if assigned
minus proportion of people that take up treatment if not
assigned. If we
assume monotonicity (no defiers) this last term is zero and
thus the LATE
equals:cov(Y,Z)/cov(D,Z) or simply E[Y1-Y0|D1>D0]] the
Local
Average
Treatment Effect for the compliers.
Jens
> -----Urspr?ngliche Nachricht-----
> Von: gov2001-l-bounces(a)lists.fas.harvard.edu
> [mailto:gov2001-l-bounces@lists.fas.harvard.edu]
> Gesendet: Sunday, December 05, 2004 7:28 PM
> An: gov2001-l(a)lists.fas.harvard.edu
> Betreff: Re: [gov2001-l] Problem 2c
>
>
> Hi, Jens. Good questions. See below.
>
> ----- Original Message -----
> From: "Jens Hainmueller"
> <jens_hainmueller(a)ksg05.harvard.edu>
> To: <gov2001-l(a)lists.fas.harvard.edu>
> Sent: Sunday, December 05, 2004 5:29 PM
> Subject: [gov2001-l] Problem 2c
>
>
> > Two questions:
> >
> > 1. For problem 2c (covariates balance), what do you
> > consider
> > the control
> > group here? Those that were not assigned to any
> > treatment
> > (as
> > in 1 and 2a)
> > OR those that were not assigned to any treatment plus
> > those
> > that were
> > assigned to the phone treatment but didn't take it up
> > (did
> > not
> > answer the
> > phone).
>
> Take the control group to be the group that was not
> targeted
> to
> recieve any of the three treatments. By using this group
> as
> the
> control group for all 3 treatments, we are effectively
> running 3
> separate experiments. You can't include the group that
> was
> assigned to treatment (treat =1) that didn't recieve the
> treatment (comply = 0) in the control group because comply
> is
> post-treatment (assignment). Then assignment to treatment
> would
> be confounded.
>
> > 2. It seems to me that the more appropriate comparison
> > would
> > be to first
> > check whether those that were assigned to the treatment
> > and
> > did take it up
> > differ systematically from those that were assigned to
> > the
> > treatment but did
> > not take it up. This would get us at least some sense to
> > which
> > degree
> > selection to treatment might be an issue (at least for
> > the
> > observable
> > characteristics recorded in the covariates). No?
>
The point of checking for covariate balance is to
see if
the
group that recieved treatment differs from the group
assigned
control. We don't really care about the group assigned to
treatment that didn't recieve the treatment except in so
far
as
it changes p(treat = 1 & comply = 1|X) compared to p(treat
=
0|X) because we are looking for the average treatment
effect
of
Y(1) - Y(0), where Y(1) is the potential outcome under
treatment
and Y(0) is the potential outcome under control. Neither
of
these are observed for the group assigned treatment that
did
not
comply.
>
> If you were curious about why some people complied and
> others
> didn't, you could check to see how p(treat = 1 & comply =
> 0|X)
> differs from p(treat = 1 & comply = 1|X), but it wouldn't
> affect
> the final estimate for the ATE and ATT.
>
> Hope this makes sense! Ask more questions...
>
> Olivia
>
> > Thanks
> > Jens
> >
> >
> > _______________________________________________
> > gov2001-l mailing list
> > gov2001-l(a)lists.fas.harvard.edu
> >
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >
>
> _______________________________________________
> gov2001-l mailing list
> gov2001-l(a)lists.fas.harvard.edu
>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
>
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu