Hi there,
I need to know who matched with who using full method. I
can get the match.matrix when I use nearest, optimal, or other
methods. But when I set the method as "full", the match.matrix I got was
null. Do you know how to get that matched matrix?
All your help will be greatly appreciated.
Best,
Xin
Hi there,
I need to know who matched with who using full method. I can get the match.matrix when I use nearest neighbor, optimal, or other methods. But when I set the method as "full", the match.matrix I got is null. Do you know how to get the matched matrix?
Dear MatchIt Team,
I am using the MatchIt package to implement genetic matching. As a result, the
achieved balance on all covariates is better compared to other matching methods
except for the distance measure (derived through logistic regression,using the
command: distance ="logit"). I could fix that issue using the Matching- Package
(J.S. Sehkon) by defining, on which variables I would like to achieve balance on
via the "BalanceMatrix" option. I included the distance measure in that matrix.
Is there an option to define such a matrix with MatchIt? Otherwise, the balance
on the distance measure is not satisfying after matching with matchit(stand.
mean difference > 0.15). What I need is a command to tell the genetic algorithm
to achieve balance on all covariates AND the combined distance measure.
Thank you for your help in advance and kind regards,
Aleks
Aleksander Kocaj <a.kocaj(a)iqb.hu-berlin.de> wrote:
> I wonder, if there is any possibility to obtain the Abadie - Imbens SE, t- statistics & p- value of the ATT via the summary command in Zelig.
On Sat, Oct 13, 2012 at 6:30 PM, Kosuke Imai <kimai(a)princeton.edu> wrote:
> Unfortunately, that option is not currently available in Zelig.
An underlying philosophy in MatchIt is that the first job of analysing
observational data is to get the design right. Once you have a good
design, you go about analysing that cleaned-up design asif that data
came from an experiment. This is a simple and attractive perspective.
I happened to discuss this with Alberto Abadie a few weeks ago and his
argument lay in two parts. The design that comes out of any matching
process involves a statistical estimate of which are the matched
units, which units are deleted, etc. In general, we can't ignore that
dimension. Specifically, he pointed out a paper coauthored by him
where the bootstrap doesn't work well in these situations owing to
non-differentiability of the likelihood function across choices of
different permutations. (Abadie and Imbens in their Econometrica 2008
paper show that classical bootstrap schemes fail to provide correct
inference for K-nearest neighbour (KNN) matching estimators of
average causal effects.)
In short, he said (a) You can't ignore the origins of your design and
proceed on analysis and (b) You can't just bootstrap your way to
happiness.
I am not an expert in this field and I am aware the literature is
moving fast. E.g. http://ftp.iza.org/dp5361.pdf in 2010 seem to have
figured out how to do bootstrap
inference works for a 1:K nearest neighbour matching estimator.
Elizabeth Stuart's beautiful survey article had pointed out that oddly
enough, in many situations, taking the sampling perspective on the
matching process gives you improved precision.
As an applied researcher I am delighted to be in the world of Matchit,
where painstaking care is applied in the design stage. I have no doubt
in my mind that this is a superior world when compared with the old
econometrics. I am curious what the wise men on this list think about
these questions. And, if a journal referee insists that I should do
alternative inference procedures such as Abadie/Imbens, what would one
do?
--
Ajay Shah
ajayshah(a)mayin.org
http://www.mayin.org/ajayshahhttp://ajayshahblog.blogspot.com
Dear MatchIt- Team,
I am currently using the "MatchIt" package in R to do some nonparametric
preprocessing of my data (school achievement data of students in
Germany). Afterwards I want to estimate the ATT of a certain
intervention with the matched data (using Zelig). After reading several
of your papers (which was really a pleasure!) I wonder, if there is any
possibility to obtain the Abadie - Imbens SE, t- statistics & p- value
of the ATT via the summary command in Zelig. By default, only the point
estimate, its standard error and the 95- confidence interval is printed.
Do you know if and how this could be achieved?
Thank you in advance and kind regards,
Aleksander Kocaj
--
Dipl. -Psych. Aleksander Kocaj
Humboldt-Universität zu Berlin
Forschungsdatenzentrum
Research Data Centre
Institut zur Qualitätsentwicklung im Bildungswesen
Institute for Educational Quality Improvement
POSTANSCHRIFT: Unter den Linden 6, 10099 Berlin
GERMANY
Dienstsitz: Hannoversche Straße 19, 10115 Berlin
Tel: +49 (30) 2093-46507
eMail:fdz@iqb.hu-berlin.de
Web:www.iqb.hu-berlin.de\FDZ
Hello MatchIt community,
I am relatively new to MatchIt and I am now working on a causal inference study with MatchIt.
I my study, I would like to use my own distance measure for matching.
I have read the manual which mentions that MatchIt allows users to use a vector of user-made distance measure as the input.
Can anyone give me an example for this?
Given a data set with 1000 items, I have computed a distance matrix (1000*1000). And what should I do in the next step?
Thanks you very much!
Best,
Gang