Hi,
How do you match when your treatment is modeled to interact with one of your
covariates? For instance, if you have a regression of Y on treatment*X1 and
X2, X3, X4, how do we match?
Thanks,
Miya
Hi...
When using Match and replace=T, multiple controls are matched to single
treatment observations in my data. And then, when taking the subset of
matched data for the parametric post-matching analysis, extracting it into a
new object using $index.treat and $index.control, I end up with duplicate
treatment observations because of the replace=T matching procedure. Should I
keep the duplicate observations for analysis or drop them? It would seem
strange to get rid of them since it's what enforces balance in the matched
data.
Thanks,
Andy
Hi,
Running match using biasadjustment=T and replace=T, I get better balance
than when running MatchIt using "nearest". Are there other parameters for
MatchIt that can improve the balance in a similar fashion to Match with its
biasadjust and replace?
Andy
Hey everyone,
I've posted a revised answer key to problem set 7. The previous version
of the answer key stated that the stochastic component of a negative
binomail is,
Y| \eta \sim Negative Binomial(\theta \lambda_i)
\eta \sim 1/\theta Gamma(\theta)
But the *correct* stochastic component is (notice the change from \theta
to \eta in the Y|\eta portion)
Y| \eta \sim Negative Binomial(\eta \lambda_i)
\eta \sim 1/\theta Gamma(\theta)
I hope this did not cause any problems--
Cheers,
Justin
Patrick,
you have at least two options: you can either let matchit() estimate the
propensity score itself, or estimate the propensity score separately and
then give it to Match() (part of the Matching library) to match on.
There probably is a way to also get matchit() to accept propensity
scores but I'm not aware of it. If anyone knows how to do this please
let us know. In any case, you can simply use the Match() function instead.
Holger
Patrick Lam wrote:
> Hi Holger/Justin,
>
> I had a question that Miya hinted on in her earlier email. So the
> question is whether to use the covariates or the pscores in MatchIt.
> The response was that we should estimate the pscores first, and then
> match on them using MatchIt, and not to enter the covariates in
> MatchIt. So I interpreted it as something like this.
>
> matches <- matchit(D~pscores, method="nearest", data=newdata...)
>
> However, it seems like MatchIt does not work that way. In the
> documentation for MatchIt, we are supposed to enter a formula, which
> MatchIt then takes to estimate distance measures and match on those
> distance measures. So the command should look like this.
>
> matches2 <- matchit(D~X1+X2+X3+X4+X5, method="nearest", data=data,
> distance="logit")
>
> With this, MatchIt will calculate pscores using the logit link, and then
> match on the pscores. In the first way of doing it, matchit will run a
> logit regression of D on our manually calculated pscores, and then match
> on the resulting probabilities, which are not pscores.
>
> I've done it both ways, and the matched datasets are identical.
> However, this could be due to the fact that there is a direct
> relationship between pscores and the subsequent probabilities produced
> by a logit of D on pscores. It seems like the first way of doing it is
> not actually correct, but the answers just happen to be the same. If
> you take the second way of doing it and then call the distance
> (matches2$distance), the resulting output is exactly the same as our
> manually calculated pscores.
>
> To sum up, it seems as if we should be inputing the actual covariates in
> MatchIt in order to match on pscores. Otherwise, we are matching on
> some other probabilities which result from a logit of D on pscores. Am
> I interpreting everything correctly?
>
> Thanks
>
> Patrick
--
Holger Lutz Kern
Graduate Student
Department of Government
Cornell University
Institute for Quantitative Social Science
Harvard University
1737 Cambridge Street N350
Cambridge, MA 02138
www.people.cornell.edu/pages/hlk23
Hi,
When I use "Match" to generate the matched pairs and estimate ATT, the ATT
value is different each time I run "Match" anew although the number of
successful matches remains the same.
Secondly, it is still unclear to me how Match estimates ATT to begin with.
It doesn't say in the help file how $est is arrived at.
Thanks you!
Andy
> match.out <- Match(Y=y, Tr=Tr, X=pscore, estimand="ATT", M=1,
> BiasAdjust=F, replace=F, sample=T)
> summary(match.out)
Estimate... 293.31
SE......... 7.9789
T-stat..... 36.760
p.val...... < 2.22e-16
Original number of observations.............. 1000 Original number of
treated obs............... 255 Matched number of
observations............... 255 Matched number of observations
(unweighted). 255
> match.out <- Match(Y=y, Tr=Tr, X=pscore, estimand="ATT", M=1,
> BiasAdjust=F, replace=F, sample=T)
> summary(match.out)
Estimate... 296.92
SE......... 7.9923
T-stat..... 37.15
p.val...... < 2.22e-16
Original number of observations.............. 1000 Original number of
treated obs............... 255 Matched number of
observations............... 255 Matched number of observations
(unweighted). 255
Hi Gang, Holger, Justin and I would like to invite everyone over my house
for lunch in Brookline on Sunday May 6th around noon. More info to come.
I hope you can make it... Gary
hi all,
please read
GARY KING, JAMES HONAKER, ANNE JOSEPH, KENNETH SCHEVE. 2001. Analyzing
Incomplete Political Science Data: An Alternative Algorithm for Multiple
Imputation. American Political Science Review 95 (1) for next week.
http://gking.harvard.edu/files/evil.pdf
Holger
--
Holger Lutz Kern
Graduate Student
Department of Government
Cornell University
Institute for Quantitative Social Science
Harvard University
1737 Cambridge Street N350
Cambridge, MA 02138
www.people.cornell.edu/pages/hlk23
Hi all,
I have to move my office hours from Wednesday to tomorrow (Monday), 12
to 2, in the basement computer lab.
Holger
--
Holger Lutz Kern
Graduate Student
Department of Government
Cornell University
Institute for Quantitative Social Science
Harvard University
1737 Cambridge Street N350
Cambridge, MA 02138
www.people.cornell.edu/pages/hlk23
hi all,
please read for next week:
Kosuke Imai, Gary King, and Elizabeth Stuart. ``Misunderstandings among
Experimentalists and Observationalists: Balance Test Fallacies in Causal
Inference.''
http://gking.harvard.edu/files/abs/matchse-abs.shtml
If you are interested in matching, we suggest you also take a look at
Alexis Diamond and Jasjeet S. Sekhon. Genetic Matching for Estimating
Causal Effects: A General Multivariate Matching Method for Achieving
Balance in Observational Studies.
http://sekhon.berkeley.edu/papers/GenMatch.pdf
before section on Thursday. We'll talk about genetic matching, among
other things.
cheers,
Holger
--
Holger Lutz Kern
Graduate Student
Department of Government
Cornell University
Institute for Quantitative Social Science
Harvard University
1737 Cambridge Street N350
Cambridge, MA 02138
www.people.cornell.edu/pages/hlk23