9 Feb
2012
9 Feb
'12
8:41 a.m.
Hi,
I have a hopefully quick question: Are there any prohibitions/issues
with imputing missing values for a treatment variable in a quasi-exp
or exp design (for example somehow treatment status wasn't reported in
an observational study or failed to be recorded in an experimental
study)? What about the dependent variable?
Thank you very much!
Prashant
11 Feb
11 Feb
9:27 a.m.
there's 2 perspectives on this: one yes, and one no. Over time, some of
the same people have taken each of the two positions!
If we think of the data as drawn from a given joint distribution, then its
best to impute all the variables together. If instead, we're focused on a
causal inference then some think that its best to not impute the outcome
variable with the others. The first is more efficient; the latter is not
biased, but its conceivable that a researcher could fool themselves by
tweaking the imputation model to (perhaps inadvertently) favor themselves.
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 Thu, Feb 9, 2012 at 8:41 AM, Prashant <presearchwork(a)gmail.com> wrote:
Hi,
I have a hopefully quick question: Are there any prohibitions/issues
with imputing missing values for a treatment variable in a quasi-exp
or exp design (for example somehow treatment status wasn't reported in
an observational study or failed to be recorded in an experimental
study)? What about the dependent variable?
Thank you very much!
Prashant
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=amelia
More info about Amelia: http://gking.harvard.edu/amelia
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
https://lists.gking.harvard.edu/mailman/listinfo/amelia
14 Feb
14 Feb
2:49 a.m.
Many thanks Prof. King for the thoughtful response.
A brief follow-up: I was consulting with someone who was looking at
the effects of teacher characteristics (say teacher ranking) on
student achievement by running a student-fixed effects model (where
one essentially takes first differences between student scores in two
different subjects -- say math and English, and uses the variation in
math teacher and English teachers' ranking as the treatment). The
person was using cross-sectional survey data which had some missing
responses for teacher rank. So based on Professor King's response, my
understanding is that imputing values for teacher rank (say using just
the teacher and classroom level data at a single-level for
simplicity), the treatment variable, may lead to imprecision in the
effect estimates, but not bias (as long as the MAR assumption behind
using multiple imputation is satisfied). Is this correct?
best wishes,
Prashant
On Sat, Feb 11, 2012 at 10:27 PM, Gary King <king(a)harvard.edu> wrote:
there's 2 perspectives on this: one yes, and one
no. Over time, some of
the same people have taken each of the two positions!
If we think of the data as drawn from a given joint distribution, then its
best to impute all the variables together. If instead, we're focused on a
causal inference then some think that its best to not impute the outcome
variable with the others. The first is more efficient; the latter is not
biased, but its conceivable that a researcher could fool themselves by
tweaking the imputation model to (perhaps inadvertently) favor themselves.
Gary
--
Gary King - Albert J. Weatherhead III University Professor - Director, IQSS
- Harvard University
GKing.Harvard.edu - King(a)Harvard.edu - @kinggary - 617-500-7570 - Asst
495-9271 - Fax 812-8581
On Thu, Feb 9, 2012 at 8:41 AM, Prashant <presearchwork(a)gmail.com> wrote:
Hi,
I have a hopefully quick question: Are there any prohibitions/issues
with imputing missing values for a treatment variable in a quasi-exp
or exp design (for example somehow treatment status wasn't reported in
an observational study or failed to be recorded in an experimental
study)? What about the dependent variable?
Thank you very much!
Prashant
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=amelia
More info about Amelia: http://gking.harvard.edu/amelia
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
https://lists.gking.harvard.edu/mailman/listinfo/amelia
7:34 a.m.
exactly. missing data that fits MAR won't bias anything.
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, Feb 14, 2012 at 2:49 AM, Prashant <presearchwork(a)gmail.com> wrote:
Many thanks Prof. King for the thoughtful response.
A brief follow-up: I was consulting with someone who was looking at
the effects of teacher characteristics (say teacher ranking) on
student achievement by running a student-fixed effects model (where
one essentially takes first differences between student scores in two
different subjects -- say math and English, and uses the variation in
math teacher and English teachers' ranking as the treatment). The
person was using cross-sectional survey data which had some missing
responses for teacher rank. So based on Professor King's response, my
understanding is that imputing values for teacher rank (say using just
the teacher and classroom level data at a single-level for
simplicity), the treatment variable, may lead to imprecision in the
effect estimates, but not bias (as long as the MAR assumption behind
using multiple imputation is satisfied). Is this correct?
best wishes,
Prashant
On Sat, Feb 11, 2012 at 10:27 PM, Gary King <king(a)harvard.edu> wrote:
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=amelia
More info about Amelia: http://gking.harvard.edu/amelia
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
https://lists.gking.harvard.edu/mailman/listinfo/amelia
there's 2 perspectives on this: one yes,
and one no. Over time, some
of
the same people have taken each of the two
positions!
If we think of the data as drawn from a given joint distribution, then
its
best to impute all the variables together. If
instead, we're focused
on a
causal inference then some think that its best to
not impute the outcome
variable with the others. The first is more efficient; the latter is not
biased, but its conceivable that a researcher could fool themselves by
tweaking the imputation model to (perhaps inadvertently) favor
themselves.
Gary
--
Gary King - Albert J. Weatherhead III University Professor - Director,
IQSS
- Harvard University
GKing.Harvard.edu - King(a)Harvard.edu - @kinggary - 617-500-7570 - Asst
495-9271 - Fax 812-8581
On Thu, Feb 9, 2012 at 8:41 AM, Prashant <presearchwork(a)gmail.com>
wrote:
Hi,
I have a hopefully quick question: Are there any prohibitions/issues
with imputing missing values for a treatment variable in a quasi-exp
or exp design (for example somehow treatment status wasn't reported in
an observational study or failed to be recorded in an experimental
study)? What about the dependent variable?
Thank you very much!
Prashant
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=amelia
More info about Amelia: http://gking.harvard.edu/amelia
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
https://lists.gking.harvard.edu/mailman/listinfo/amelia
4464
days inactive
4469
days old
amelia@lists.gking.harvard.edu
3 comments
2 participants
participants (2)
-
Gary King -
Prashant