21 Apr
2006
21 Apr
'06
11:31 p.m.
Hi,
I had another look over GenMatch. It looks to me as though it is nothing more
than a stochastic beam search for Euclidean distance metrics. Surely I must be
missing something.
Geoff
22 Apr
22 Apr
3:46 a.m.
Hi Geoff,
I hadn't before heard the term "stochastic beam search", but from what
I've gleaned on the web, Genetic algorithms work in a similar way. At a
given point in the algorithm, they have a set of possible solutions that
they are modifying by combining them, mutating them, and terminating some.
The analogy is to a gene pool that is improving through similar processes.
The innovation here is that genetic alogrithms are better at searching
across non-smooth surfaces as I understand it, because they are less prone
to getting caught in local maxima.
On this topic, I would see the paper by Jas Sekhon and Alexis Diamond:
http://sekhon.berkeley.edu/papers/GenMatch.pdf
Best,
Dan
----
Ph.D. Student
Department of Government
Harvard University
Tutor, Currier House
dhopkins(a)fas.harvard.edu
http://www.danhopkins.org
On Sat, 22 Apr 2006 ghumphr(a)fas.harvard.edu wrote:
Hi,
I had another look over GenMatch. It looks to me as though it is nothing more
than a stochastic beam search for Euclidean distance metrics. Surely I must be
missing something.
Geoff
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
11:56 a.m.
New subject: [gov2001-l] Posttreatment bias and matching
For our paper, Amy and I have a binary treatment variable that indicates
an increase in the diversity index of at least 10% from 1970 to 2000.
What we'll be testing via matching will be the influence of "treatment"
(rapid increase in diversity) on growth from 1970 to 2000 (dep. var).
Should I only balance on conditions prior to 1969 to avoid posttreatment
bias since the treatment is growth in diversity from 1970 to 2000? Or do
we want to balance on everything (even conditions in the middle of this
time period) and only worry about post treatment bias in the regressions
we run once the treatment and control group have been made?
Thanks!
Suzanna
12:11 p.m.
New subject: [gov2001-l] Re: Posttreatment bias and matching
A good question. I would recommend matching on only those covariates that
are definitely pre-treatment--including, if you have them, change
variables BEFORE 1970, to capture the differing trajectories. I'd deal
with other, potentially confounding treatments (e.g. other changes in the
1970-2000 period) at the stage of parametric adjustment, though keep in
mind the twin dangers of omitted variable bias and post-treatment bias.
Best,
Dan
----
Ph.D. Student
Department of Government
Harvard University
Tutor, Currier House
dhopkins(a)fas.harvard.edu
http://www.danhopkins.org
On Sat, 22 Apr 2006, Suzanna Chapman wrote:
For our paper, Amy and I have a binary treatment
variable that indicates
an increase in the diversity index of at least 10% from 1970 to 2000.
What we'll be testing via matching will be the influence of "treatment"
(rapid increase in diversity) on growth from 1970 to 2000 (dep. var).
Should I only balance on conditions prior to 1969 to avoid posttreatment
bias since the treatment is growth in diversity from 1970 to 2000? Or do
we want to balance on everything (even conditions in the middle of this
time period) and only worry about post treatment bias in the regressions
we run once the treatment and control group have been made?
Thanks!
Suzanna
2:27 p.m.
New subject: [gov2001-l] Summary() in Matching
Hi everyone,
When we use the summary() command to check balance, we are getting means
of the treated, means of the control, mean difference, and QQ plot summary
stats - but we do not have standardized biases or standard deviations for
the treated and control reported. The matchit.pdf document online says we
are supposed to have these results reported in the summary command - is
anyone else getting this information?
If not - we want to calculate standardized biases - and that requires the
sd of the treated group. Would it be equivalent to just manually
calculate the sd. for each variable in the treated subsample?
Thanks!
3:32 p.m.
New subject: [gov2001-l] Summary() in Matching
Hi Suzanna,
Check out the code we presented in section to see how we calculated the
standardized bias--and yes, you could do it manually as below. And yes, I
don't *think* the most recent version of MatchIt returns the standardized
bias--it didn't in the code we created for section. Also, keep in mind
you want to evaluate balance across all the key covariates, not just those
that you matched on.
Best,
Dan
On Sat, 22 Apr 2006, Suzanna Chapman wrote:
Hi everyone,
When we use the summary() command to check balance, we are getting means
of the treated, means of the control, mean difference, and QQ plot summary
stats - but we do not have standardized biases or standard deviations for
the treated and control reported. The matchit.pdf document online says we
are supposed to have these results reported in the summary command - is
anyone else getting this information?
If not - we want to calculate standardized biases - and that requires the
sd of the treated group. Would it be equivalent to just manually
calculate the sd. for each variable in the treated subsample?
Thanks!
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
3:56 p.m.
New subject: [gov2001-l] Summary() in Matching
is there code to isolate the meandifference column in summary out for use
in calculating standardized biases? right now we are manually entering
it:
0.582/sd(TR$variable)
but it would be easier if there was someway in code to indicate which row
and column of the summary to link to - do you know?
thanks!
On Sat, 22 Apr 2006, Dan Hopkins wrote:
Hi Suzanna,
Check out the code we presented in section to see how we calculated the
standardized bias--and yes, you could do it manually as below. And yes, I
don't *think* the most recent version of MatchIt returns the standardized
bias--it didn't in the code we created for section. Also, keep in mind
you want to evaluate balance across all the key covariates, not just those
that you matched on.
Best,
Dan
On Sat, 22 Apr 2006, Suzanna Chapman wrote:
Hi everyone,
When we use the summary() command to check balance, we are getting means
of the treated, means of the control, mean difference, and QQ plot summary
stats - but we do not have standardized biases or standard deviations for
the treated and control reported. The matchit.pdf document online says we
are supposed to have these results reported in the summary command - is
anyone else getting this information?
If not - we want to calculate standardized biases - and that requires the
sd of the treated group. Would it be equivalent to just manually
calculate the sd. for each variable in the treated subsample?
Thanks!
_______________________________________________
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
7:16 p.m.
New subject: [gov2001-l] Summary() in Matching
Typically, the best way to proceed with this is to figure out what objects
exist from which you might extract the column. So, if "mout" is the
output from matching, run:
names(mout)
and
names(summary(mout))
and look for the relevant object.
Best,
Dan
----
Ph.D. Student
Department of Government
Harvard University
Tutor, Currier House
dhopkins(a)fas.harvard.edu
http://www.danhopkins.org
On Sat, 22 Apr 2006, Suzanna Chapman wrote:
is there code to isolate the meandifference column in
summary out for use
in calculating standardized biases? right now we are manually entering
it:
0.582/sd(TR$variable)
but it would be easier if there was someway in code to indicate which row
and column of the summary to link to - do you know?
thanks!
On Sat, 22 Apr 2006, Dan Hopkins wrote:
Hi Suzanna,
Check out the code we presented in section to see how we calculated the
standardized bias--and yes, you could do it manually as below. And yes, I
don't *think* the most recent version of MatchIt returns the standardized
bias--it didn't in the code we created for section. Also, keep in mind
you want to evaluate balance across all the key covariates, not just those
that you matched on.
Best,
Dan
On Sat, 22 Apr 2006, Suzanna Chapman wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
Hi everyone,
When we use the summary() command to check balance, we are getting means
of the treated, means of the control, mean difference, and QQ plot summary
stats - but we do not have standardized biases or standard deviations for
the treated and control reported. The matchit.pdf document online says we
are supposed to have these results reported in the summary command - is
anyone else getting this information?
If not - we want to calculate standardized biases - and that requires the
sd of the treated group. Would it be equivalent to just manually
calculate the sd. for each variable in the treated subsample?
Thanks!
_______________________________________________
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
4:35 p.m.
New subject: [gov2001-l] Summary() in Matching
One more question, sorry for filling your inboxes!
We made a dummy variable for our treatment variable using the code you
suggested dan:
treatment<-1*(abs(data[,1]-data[,2])>=.1)
So this is our treatment variable, and it works well in the matchit() to
have (treatment ~ x1 + x2)
However, because this dummy is not a column in the dataframe, I can't use
the code from section to run the balance tests... specifically I can't
create the original balance "TR.pre" and "CON.pre" groups. Is there a
way
to attach the treatment variable to the dataframe?
Thanks again!
4:56 p.m.
New subject: [gov2001-l] Summary() in Matching
Susana,
I think that the following should work:
new.data <- cbind(data.frame, treatment)
Good luck,
Elena
----- Original Message -----
From: Suzanna Chapman
To: gov2001-l(a)lists.fas.harvard.edu
Sent: Saturday, April 22, 2006 8:35 PM
Subject: Re: [gov2001-l] Summary() in Matching
One more question, sorry for filling your inboxes!
We made a dummy variable for our treatment variable using the code you
suggested dan:
treatment<-1*(abs(data[,1]-data[,2])>=.1)
So this is our treatment variable, and it works well in the matchit() to
have (treatment ~ x1 + x2)
However, because this dummy is not a column in the dataframe, I can't use
the code from section to run the balance tests... specifically I can't
create the original balance "TR.pre" and "CON.pre" groups. Is there
a way
to attach the treatment variable to the dataframe?
Thanks again!
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
7:14 p.m.
New subject: [gov2001-l] Summary() in Matching
And you might also look at the as.data.frame() function... Yet another
approach is simply to add a variable to an existing dataframe, as follows:
dataframe$treatment<-1*(abs(data[,1]-data[,2])>=.1)
Best,
Dan
----
Ph.D. Student
Department of Government
Harvard University
Tutor, Currier House
dhopkins(a)fas.harvard.edu
http://www.danhopkins.org
On Sat, 22 Apr 2006, Elena Llaudet wrote:
Susana,
I think that the following should work:
new.data <- cbind(data.frame, treatment)
Good luck,
Elena
----- Original Message -----
From: Suzanna Chapman
To: gov2001-l(a)lists.fas.harvard.edu
Sent: Saturday, April 22, 2006 8:35 PM
Subject: Re: [gov2001-l] Summary() in Matching
One more question, sorry for filling your inboxes!
We made a dummy variable for our treatment variable using the code you
suggested dan:
treatment<-1*(abs(data[,1]-data[,2])>=.1)
So this is our treatment variable, and it works well in the matchit() to
have (treatment ~ x1 + x2)
However, because this dummy is not a column in the dataframe, I can't use
the code from section to run the balance tests... specifically I can't
create the original balance "TR.pre" and "CON.pre" groups. Is
there a way
to attach the treatment variable to the dataframe?
Thanks again!
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
6599
days inactive
6599
days old
gov2001@lists.gking.harvard.edu
10 comments
4 participants
participants (4)
-
dhopkins＠fas.harvard.edu -
ellaudet＠fas.harvard.edu -
ghumphr＠fas.harvard.edu -
schapman＠fas.harvard.edu