26 Apr
2011
26 Apr
'11
5:48 p.m.
Hi everyone,
Here is a question about instrumental variables. It might be useful
for everyone. Say you get lucky and come up with a bunch of valid
instruments. When running your IV regressions, should you use the
instruments separately, or should you throw them all in one
regression?
I am not sure which approach is more theoretically justifiable. I've
seen the first reported in economics papers frequently, but the Zelig
option where you can stick in a bunch of instruments at once calls out
at me.
Best,
Erin
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
gov2001 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
26 Apr
26 Apr
7:58 p.m.
Erin,
Maybe others have different thoughts but I am deeply skeptical that
you could ever have multiple valid instruments for any question in
political science. Its a very specific set of assumptions. Further,
when you have multiple IV's you have to assume they don't interact (I
think). So the assumptions become further complicated. In general it
seems like you would want to focus on in defending one instrument
really well rather than spreading yourself thin.
If you are in some crazy world where multiple instruments could work,
I think showing that its robust across instruments in separate
regressions is compelling. These estimators are very sensitive to the
choice of instrument typically.
The Zelig option comes from the underlying package on which the
implementation is based (systemfit). This is just by way of saying
that you can never be sure of the reason an option is there! It might
only be a good idea for other fields.
That being said, this is just my take. I am sure you can find people
with a less grumpy worldview on the topic :)
Brandon
On Tue, Apr 26, 2011 at 5:48 PM, Erin Baggott <ebaggott(a)fas.harvard.edu> wrote:
Hi everyone,
Here is a question about instrumental variables. It might be useful
for everyone. Say you get lucky and come up with a bunch of valid
instruments. When running your IV regressions, should you use the
instruments separately, or should you throw them all in one
regression?
I am not sure which approach is more theoretically justifiable. I've
seen the first reported in economics papers frequently, but the Zelig
option where you can stick in a bunch of instruments at once calls out
at me.
Best,
Erin
_______________________________________________
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 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
9:08 p.m.
i agree with Brandon, grump included. the direct answer to your question,
Erin, is that you can include all the valid instruments, which works all
conditional on the model.
Gary
--
*Gary King* - Albert J. Weatherhead III University Professor - Director,
IQSS - Harvard University
GKing.Harvard.edu <http://gking.harvard.edu/> - King(a)Harvard.edu -
@kinggary<http://twitter.com/kinggary>- 617-500-7570 - Asst 495-9271 -
Fax 812-8581
On Tue, Apr 26, 2011 at 7:58 PM, Brandon Stewart
<brandonmstewart(a)gmail.com>wrote;wrote:
Erin,
Maybe others have different thoughts but I am deeply skeptical that
you could ever have multiple valid instruments for any question in
political science. Its a very specific set of assumptions. Further,
when you have multiple IV's you have to assume they don't interact (I
think). So the assumptions become further complicated. In general it
seems like you would want to focus on in defending one instrument
really well rather than spreading yourself thin.
If you are in some crazy world where multiple instruments could work,
I think showing that its robust across instruments in separate
regressions is compelling. These estimators are very sensitive to the
choice of instrument typically.
The Zelig option comes from the underlying package on which the
implementation is based (systemfit). This is just by way of saying
that you can never be sure of the reason an option is there! It might
only be a good idea for other fields.
That being said, this is just my take. I am sure you can find people
with a less grumpy worldview on the topic :)
Brandon
On Tue, Apr 26, 2011 at 5:48 PM, Erin Baggott <ebaggott(a)fas.harvard.edu>
wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
Hi everyone,
Here is a question about instrumental variables. It might be useful
for everyone. Say you get lucky and come up with a bunch of valid
instruments. When running your IV regressions, should you use the
instruments separately, or should you throw them all in one
regression?
I am not sure which approach is more theoretically justifiable. I've
seen the first reported in economics papers frequently, but the Zelig
option where you can stick in a bunch of instruments at once calls out
at me.
Best,
Erin
_______________________________________________
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
27 Apr
27 Apr
1:43 p.m.
New subject: calculating standard errors for multiple imputation
Hi all,
Can anyone tell me how to calculate appropriate multiple imputation standard errors for
the marginal effect of an independent variable that is part of an interaction term?
That is, the basic model is:
Y = beta1*x1 + beta2*x2 + beta3*x1*x2.
The standard errors for each of the three coefficients can be calculated with the usual MI
adjustment but to calculate the standard error of the marginal effect you need to deal
somehow with the covariance term between x1 and x2.
Does anyone know whether the covariances can be adjusted in a similar way to what happens
with the standard errors or do we have to calculate the standard error of the marginal
effect for each dataset/regression and then perform the SE adjustment?
Cheers,
James
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
gov2001 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
1:46 p.m.
New subject: calculating standard errors for multiple imputation
You can definitely do it using the covariance but its waaay easier to
just simulate. Simulate 1000 times from each dataset, combine the
simulations and then take the SE as normal.
Brandon
On Wed, Apr 27, 2011 at 1:43 PM, James Conran <jkconran(a)mit.edu> wrote:
Hi all,
Can anyone tell me how to calculate appropriate multiple imputation standard errors for
the marginal effect of an independent variable that is part of an interaction term?
That is, the basic model is:
Y = beta1*x1 + beta2*x2 + beta3*x1*x2.
The standard errors for each of the three coefficients can be calculated with the usual
MI adjustment but to calculate the standard error of the marginal effect you need to deal
somehow with the covariance term between x1 and x2.
Does anyone know whether the covariances can be adjusted in a similar way to what happens
with the standard errors or do we have to calculate the standard error of the marginal
effect for each dataset/regression and then perform the SE adjustment?
Cheers,
James
_______________________________________________
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 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
1:58 p.m.
New subject: calculating standard errors for multiple imputation
As an addendum: would simulation be appropriate if it was a panel dataset? My inclination
is that it would not be, since the rows aren't interchangeable
Jeremy
________________________________________
From: gov2001-l-bounces(a)lists.fas.harvard.edu [gov2001-l-bounces(a)lists.fas.harvard.edu] On
Behalf Of Brandon Stewart [brandonmstewart(a)gmail.com]
Sent: Wednesday, April 27, 2011 1:46 PM
To: Class List for Gov 2001/E-2001
Subject: Re: [gov2001] calculating standard errors for multiple imputation
You can definitely do it using the covariance but its waaay easier to
just simulate. Simulate 1000 times from each dataset, combine the
simulations and then take the SE as normal.
Brandon
On Wed, Apr 27, 2011 at 1:43 PM, James Conran <jkconran(a)mit.edu> wrote:
Hi all,
Can anyone tell me how to calculate appropriate multiple imputation standard errors for
the marginal effect of an independent variable that is part of an interaction term?
That is, the basic model is:
Y = beta1*x1 + beta2*x2 + beta3*x1*x2.
The standard errors for each of the three coefficients can be calculated with the usual
MI adjustment but to calculate the standard error of the marginal effect you need to deal
somehow with the covariance term between x1 and x2.
Does anyone know whether the covariances can be adjusted in a similar way to what happens
with the standard errors or do we have to calculate the standard error of the marginal
effect for each dataset/regression and then perform the SE adjustment?
Cheers,
James
_______________________________________________
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 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
2:03 p.m.
New subject: calculating standard errors for multiple imputation
Jeremy,
(sorry this has to be really quick as I am running out the door)
Simulation is always appropriate given the model. The only new
assumption simulation is introducing is that the # of simulations is
high enough to asymptotically approach the correct value and you can
just set the #. Once you have a vector of means and variance
covariance matrix you have already assumed any conditional
exchangeability of observations.
Brandon
On Wed, Apr 27, 2011 at 1:58 PM, Jeremy Ferwerda <ferwerda(a)mit.edu> wrote:
As an addendum: would simulation be appropriate if it
was a panel dataset? My inclination is that it would not be, since the rows aren't
interchangeable
Jeremy
________________________________________
From: gov2001-l-bounces(a)lists.fas.harvard.edu [gov2001-l-bounces(a)lists.fas.harvard.edu]
On Behalf Of Brandon Stewart [brandonmstewart(a)gmail.com]
Sent: Wednesday, April 27, 2011 1:46 PM
To: Class List for Gov 2001/E-2001
Subject: Re: [gov2001] calculating standard errors for multiple imputation
You can definitely do it using the covariance but its waaay easier to
just simulate. Simulate 1000 times from each dataset, combine the
simulations and then take the SE as normal.
Brandon
On Wed, Apr 27, 2011 at 1:43 PM, James Conran <jkconran(a)mit.edu> wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
gov2001 mailing list served by Harvard-MIT Data Center
List Address: gov2001(a)lists.gking.harvard.edu
Subscribe/Unsubscribe: http://lists.gking.harvard.edu/?info=gov2001
Hi all,
Can anyone tell me how to calculate appropriate multiple imputation standard errors for
the marginal effect of an independent variable that is part of an interaction term?
That is, the basic model is:
Y = beta1*x1 + beta2*x2 + beta3*x1*x2.
The standard errors for each of the three coefficients can be calculated with the usual
MI adjustment but to calculate the standard error of the marginal effect you need to deal
somehow with the covariance term between x1 and x2.
Does anyone know whether the covariances can be adjusted in a similar way to what happens
with the standard errors or do we have to calculate the standard error of the marginal
effect for each dataset/regression and then perform the SE adjustment?
Cheers,
James
_______________________________________________
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
4756
days inactive
4757
days old
gov2001@lists.gking.harvard.edu
6 comments
5 participants
participants (5)
-
Brandon Stewart -
Erin Baggott -
Gary King -
James Conran -
Jeremy Ferwerda