10 Apr
2009
10 Apr
'09
12:17 p.m.
Dear Miya and Patrick,
Quick question on 3.c
I am not sure how to understand " alter the propensity score model in any
way you see fit"
indeed, do we have to change the *type* of distance measures used ( and see
if it improves the imbalance) or improve the specific type we chose in 3.a.
if it is the latter we have to do, I do not see how we can do that without
knowing the data set and the name of the variables. Don't we need a bit of
theory to play with the IV in the propensity score model?
thanks
ps: also, can we use matchit for 3.b?
Charlotte
10 Apr
10 Apr
12:53 p.m.
see below.
On Fri, Apr 10, 2009 at 12:17 PM, charlotte cavaille <
charlotte.cavaille at gmail.com> wrote:
Dear Miya and Patrick,
Quick question on 3.c
I am not sure how to understand " alter the propensity score model in any
way you see fit"
indeed, do we have to change the *type* of distance measures used ( and
see if it improves the imbalance) or improve the specific type we chose in
3.a.
either is acceptable, assuming i correctly understand the distinction you
are making. basically, the goal is to improve the propensity score
specification to increase balance. whatever changes you make, just be sure
that the "distance" you are estimating is still summarized by a propensity
score -- i.e. Pr(T|X).
if it is the latter we have to do, I do not see how we
can do that without
knowing the data set and the name of the variables. Don't we need a bit of
theory to play with the IV in the propensity score model?
the goal of matching is to increase balance. in the absence of having any
theory, you can still play around with the model and see whether or not
balance improves.
thanks
ps: also, can we use matchit for 3.b?
yes, but you should match treatment on the propensity score estimated in 3a
(not the covariates). this applies for when you repeat 3a and 3b in 3c as
well.
Charlotte
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
Miya Woolfalk
Ph.D. Student
Harvard University
Government and Social Policy
1:01 p.m.
Thanks Miya!
I just don't get the emphasis on "propensity score estimated" vs.
"covariates". Indeed, once we are doing *nearest.matching*, matchit needs a
*distance* and this distance is by default the one produced by a logit
regression (default setting in matchit), so how could we anyways match on
covariates if we are doing nearest matching???? (am I getting this right?)
is there a way to actually disenable the "distance" argument?
2009/4/10 Miya Woolfalk <woolfalk at fas.harvard.edu>
see below.
On Fri, Apr 10, 2009 at 12:17 PM, charlotte cavaille <
charlotte.cavaille at gmail.com> wrote:
Dear Miya and Patrick,
Quick question on 3.c
I am not sure how to understand " alter the propensity score model in any
way you see fit"
indeed, do we have to change the *type* of distance measures used ( and
see if it improves the imbalance) or improve the specific type we chose in
3.a.
either is acceptable, assuming i correctly understand the distinction you
are making. basically, the goal is to improve the propensity score
specification to increase balance. whatever changes you make, just be sure
that the "distance" you are estimating is still summarized by a propensity
score -- i.e. Pr(T|X).
if it is the latter we have to do, I do not see
how we can do that without
knowing the data set and the name of the variables. Don't we need a bit of
theory to play with the IV in the propensity score model?
the goal of matching is to increase balance. in the absence of having any
theory, you can still play around with the model and see whether or not
balance improves.
thanks
ps: also, can we use matchit for 3.b?
yes, but you should match treatment on the propensity score estimated in 3a
(not the covariates). this applies for when you repeat 3a and 3b in 3c as
well.
Charlotte
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
Miya Woolfalk
Ph.D. Student
Harvard University
Government and Social Policy
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
1:20 p.m.
see below.
On Fri, Apr 10, 2009 at 1:01 PM, charlotte cavaille <
charlotte.cavaille at gmail.com> wrote:
Thanks Miya!
I just don't get the emphasis on "propensity score estimated" vs.
"covariates". Indeed, once we are doing *nearest.matching*, matchit needs
a *distance* and this distance is by default the one produced by a logit
regression (default setting in matchit), so how could we anyways match on
covariates if we are doing nearest matching???? (am I getting this right?)
think about what matchit() is doing when you feed it a formula containing
covariates vs a formula containing propensity scores and leave all of the
other arugments the same.
is there a way to actually disenable the "distance" argument?
you can't disable it but you can specify your own distances by feeding it a
vector of user specified distances. but note, disabling the distance
argument is not required to answer problem 3 (though you can certainly play
around with this option if you like)
2009/4/10 Miya Woolfalk <woolfalk at fas.harvard.edu>
see below.
--
Miya Woolfalk
Ph.D. Student
Harvard University
Government and Social Policy
On Fri, Apr 10, 2009 at 12:17 PM, charlotte cavaille <
charlotte.cavaille at gmail.com> wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
Dear Miya and Patrick,
Quick question on 3.c
I am not sure how to understand " alter the propensity score model in any
way you see fit"
indeed, do we have to change the *type* of distance measures used ( and
see if it improves the imbalance) or improve the specific type we chose in
3.a.
either is acceptable, assuming i correctly understand the distinction you
are making. basically, the goal is to improve the propensity score
specification to increase balance. whatever changes you make, just be sure
that the "distance" you are estimating is still summarized by a propensity
score -- i.e. Pr(T|X).
if it is the latter we have to do, I do not see
how we can do that
without knowing the data set and the name of the variables. Don't we need a
bit of theory to play with the IV in the propensity score model?
the goal of matching is to increase balance. in the absence of having any
theory, you can still play around with the model and see whether or not
balance improves.
thanks
ps: also, can we use matchit for 3.b?
yes, but you should match treatment on the propensity score estimated in
3a (not the covariates). this applies for when you repeat 3a and 3b in 3c
as well.
Charlotte
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
Miya Woolfalk
Ph.D. Student
Harvard University
Government and Social Policy
_______________________________________________
gov2001-l mailing list
gov2001-l at lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
5515
days inactive
5515
days old
gov2001@lists.gking.harvard.edu
3 comments
2 participants
participants (2)
-
charlotte.cavaille＠gmail.com -
woolfalk＠fas.harvard.edu