Hi,
I was wondering if I could use coarsened exact matching (CEM) to
create the following type of blocks/strata for a randomized
experiment:
1st stage: block 50 schools into pairs (25 pairs total), randomization
would take place in each pair
2nd stage: Take the students in each of the above schools (say there
are 60 students in each school) and stratify them into triples as
well, randomization would take place within each triple (assigned to
treatment A, treatment B, and control).
I have read for example that the cem command in Stata can create
blocks for a randomized experiment, but I am not sure if it can create
matched pairs and triples as in the above situation or not, since CEM
may create different size blocks/strata according to the variables and
degree of coarsening one chooses, rather than creating pairs/triples
per say -- is there a way to actually create pairs/triples using CEM
or do I have to additionally rely on another matching method (say
using some distance measure) to create pairs/triples? And is further
creating pairs/triples (beyond the blocks from cem) necessarily
helpful in reducing bias and increasing power?
Thank you so much in advance for your help,
Prashant
-
cem Mailing List, served by Harvard-MIT Data Center
Send messages: cem(a)lists.gking.harvard.edu
[un]subscribe Options: http://lists.gking.harvard.edu/?info=cem
More information on cem: http://gking.harvard.edu/cem
Hello,
I am currently working on a paper where I use one-shot survey data from 2006
and 2008 (repeated cross-sections, exact same variables) and, although they
are supposed to be random samples drawn from the same population, there is
considerable multivariate imbalance on demographic variables. Thus, I used
CEM on the pooled sample (with year as "treatment") to make the samples
comparable:
UNWEIGHTED MEANS
year education male age urban
2006 8.572 0.493 37.611 0.792
2008 8.269 0.495 40.841 0.692
Multivariate L1 distance: .34852375
CEM WEIGHED MEANS
year education male age urban
2006 8.204 0.492 40.278 0.700
2008 8.256 0.492 40.696 0.700
Multivariate L1 distance: 2.355e-15
My question is whether it makes sense to run separate regressions for 2006
and 2008 using the respective CEM weights obtained from the pooled sample.
Hope you can help. Thanks!
Reynaldo T. Rojo Mendoza
Ph.D. Student
Department of Political Science
University of Pittsburgh