re: Re: st: Balance for PSM

["Ariel Linden" <ariel.linden@gmail.com>]

Let me add to Jorge's advice here.

You provide no rationale for running several different matching algorithms.
What are you hoping to achieve? 

As a general rule, researchers should give thought to the research design
and evaluation techniques they intend to employ (a priori). It doesn't
appear that you've done that. 

Different matching algorithms will very likely elicit different results,
depending on many factors, such as sample sizes actually used (i.e., number
of matches vs using entire dataset with weights), choice of estimator (ATE,
ATT, or ATC), and when using kernel weights, the choice of kernel and
bandwidths. All of these issues (and several others) will likely lead you to
different conclusions. So you would be expected to think about these issues
ahead of time, rather than reviewing the results after the fact and choosing
the approach that gave you the result you liked the best...

I suggest you read the following excellent papers:

Stuart, E.A. (2010) Matching methods for causal inference: a review and a
look forward. Statistical Science, 25(1), 1?21.

Caliendo, M. Kopeinig, S. (2008) Some practical guidance for the
implementation of propensity score matching. Journal of Economic Surveys,
22, 31-72.


I hope this helps,

Ariel   


Date: Mon, 2 Dec 2013 16:26:59 -0500
From: =?ISO-8859-1?Q?Jorge_Eduardo_P=E9rez_P=E9rez?= <jorge_perez@brown.edu>
Subject: Re: st: Balance for PSM

Carlos,

You are not supposed to send attachments to Statalist. I did not open it.

You are also supposed to say that psmatch2 is an user written command from
SSC.

Having said that, you may want to rethink your problem. Do you think
that informality is as good as randomly assigned to workers after
controlling for the limited set of covariates you have? I think not:
you lack quite few controls. Are workers within some industries more
likely to be informal than others? Are workers in different cities
more likely to be informal than others? I could go on and on, but this
is the Stata list, not the economics one.

Your results show that your covariates are not balanced in either your
unmatched or matched sample, with the exception of gender which seems
balanced according to the t-test (which has it's own problems, see
http://imai.princeton.edu/research/files/matchse.pdf) . So you need to
redefine your model before estimating ATE or ATT before proceeding
with the matching.

It seems that what you ran was a nearest neighbour matching. Radius
matching can be more computationally demanding, but before buying a
new computer I would change the propensity score specification, make
sure I have balance, and then start obtaining matching estimates. And
before doing that, I would think about whether propensity score
matching is the right tool to use.

Regards, Jorge Pérez.
- --------------------------------------------
Jorge Eduardo Pérez Pérez
Graduate Student
Department of Economics
Brown University


On Mon, Dec 2, 2013 at 3:27 PM, Carlos Tendilla González
<carlos.tendilla@imss.gob.mx> wrote:
> Hi,
>
> I am using Stata 13. I am doing a study about Informality and its effect
on Wage. The data base contains information about employees and their work
status, and also some personal characteristics  (age, sex, state, civil
status and others).
>
> I have to perform the Propensity Score Matching for NN, Startification,
Radius and Kernel Matching. I started doing a PS Match using psmatch2.ado,
and the results I had were (also available in attached):
>
> . pstest familiar casado hombre edad edad2 escolaridad escolar2 edadsexo,
raw t(totalformal)
> . probit totalformal familiar casado hombre edad edad2 escolaridad
escolar2 edadsexo
> . predict double ps
> . psmatch2 totalformal, outcome (lsalhora) pscore(ps) ate
> . pstest familiar casado hombre edad edad2 escolaridad escolar2 edadsexo,
both
>
>
----------------------------------------------------------------------------
--
>                 Unmatched |       Mean               %reduct |     t-test
>     Variable      Matched | Treated Control    %bias  |bias| |    t
p>|t|
>
--------------------------+----------------------------------+--------------
--
>      familiar   Unmatched | .47932   .29533     38.5         |  59.46
0.000
>                  Matched  | .47932   .48352     -0.9    97.7 | -61.65
0.000
>                           |                                  |
>        casado   Unmatched |   .545   .37322     35.0         |  54.35
0.000
>                  Matched  |   .545   .54642     -0.3    99.2 | -55.63
0.000
>                           |                                  |
>        hombre   Unmatched |  .6161   .62242     -1.3         |  -2.03
0.043
>                  Matched  |  .6161   .62591     -2.0   -55.0 |  -0.86
0.390
>                           |                                  |
>          edad   Unmatched | 35.085   31.907     26.9         |  42.43
0.000
>                  Matched  | 35.085   34.781      2.6    90.4 | -38.79
0.000
>                           |                                  |
>         edad2   Unmatched | 1348.2   1179.4     19.4         |  30.44
0.000
>                  Matched  | 1348.2   1322.7      2.9    84.9 | -27.09
0.000
>                           |                                  |
>   escolaridad   Unmatched | 11.209   7.9337     57.0         |  88.51
0.000
>                  Matched  | 11.209   11.156      0.9    98.4 | -95.41
0.000
>                           |                                  |
>      escolar2   Unmatched | 159.96   94.585     14.8         |  22.94
0.000
>                  Matched  | 159.96   156.09      0.9    94.1 | -27.02
0.000
>                           |                                  |
>      edadsexo   Unmatched |  21.85   19.616     11.9         |  18.39
0.000
>                  Matched  |  21.85   22.111     -1.4    88.3 | -18.50
0.000
>                           |                                  |
>
----------------------------------------------------------------------------
--



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/