Re: st: propensity score matching in 2 stages(if match is not found inside region then match outside)

[Adam Olszewski <adam.olszewski@gmail.com>]

I would echo Ariel's expert comments. Additionally, if you find the
issues of common support daunting, you can use procedures such as
optimal matching (-net describe optmatch-) or coarsened exact matching
(-ssc describe cem-) which are alternative approaches that will match
all your cases.
But you should consider what your primary question is: the effect of
natural disaster for children in specific districts (in which case you
would exclude cases from other districts), or the effect of the
disaster on any child in your sample (in which case you could use the
districts as a variable in your PS model and they will get
approximately balanced, though not necessarily in a 1:1
correspondence).
Additionally, to be clear, the 0.25 (or, 0.2 as in the quoted paper)
standard deviation is "supposed" to be a pooled SD of the LOGIT of the
propensity score. [Austin PC. Optimal caliper widths for
propensity-score matching when estimating differences in means and
differences in proportions in observational studies. Pharm Stat 2011;
10: 150-161.]
Best,
Adam Olszewski

On Wed, Nov 28, 2012 at 12:08 PM, Ariel Linden, DrPH
<ariel.linden@gmail.com> wrote:
> Hi Auditya,
>
> First off, you mention both the 0.25 caliper and common support in the same
> breath. These are mutually exclusive issues and effect the analysis
> separately. More specifically (as related to your question):
>
> Calipers: setting the caliper at 0.25 is a rule of thumb, but by no means is
> it written in stone. You should try various caliper settings (including NO
> caliper at all and setting the algorithm to nearest-neighbor), and see if
> that increases sample size while preserving covariate balance. Similarly,
> you could forego the matching and use propensity score weighting instead.
> That would ensure you use the full set of non-treated (those not-impacted
> but the natural disaster).
>
> Common support: this is a more straight-forward issue. Generally speaking,
> you should always run your analyses using only those units within common
> support. This is to ensure that we are not extrapolating beyond the data.
> That said, you should always check the covariate balance with and without
> common support, since it is quite possible that if the propensity score is
> incorrectly estimated, you could have good covariate balance for units
> beyond common support (or vice-versa).
>
> I would start with these items before moving on to a different control group
> scenario. I am also not enamored with the mix-and-match approach you
> describe, using controls from one area and then supplementing them with
> controls from another area. You are running the risk of increasing
> unobserved bias. If you want to include different sets of controls, I would
> suggest running the matching strategy twice, once using the "local"
> controls, and once using the "distant" controls. If the results are similar,
> you may have more confidence in your treatment effect estimates.
>
> I hope this helps
>
> Ariel
>
>
>
> Date: Tue, 27 Nov 2012 19:15:50 -0600
> From: Auditya Shamsuddin <auditya46@gmail.com>
> Subject: st: propensity score matching in 2 stages(if match is not found
> inside region then match outside)
>
> Fellow Stata-users
> I am a relatively new STATA user.
> I am using psmatch2(version 4.0.4 10nov2010 E. Leuven, B. Sianesi) to
> find out impact of natural disaster on child labor. Based on the
> propensity to be affected by natural disaster, I would like to match
> children within same district(same as county in US) i.e., a natural
> disaster affected child will be matched with another child from the
> same district who has similar likelihood of being affected by the
> natural disaster but who is not. To ensure good quality match within
> the district, I specify the caliper to be of 0.25 standard deviation
> of propensity score. But problem with this condition is that 35 out of
> my 105 treatment observations stay out of common support and, this
> cannot be used for estimating average treatment effect on the treated.
> So, I would like to match these 35 observations, which did not find
> close  match within district, with control observations from outside
> the district, if close matches exist within 0.25 standard deviation of
> propensity score. If a treatment fails to find a match within the
> specified caliper even outside the region/district, I will discard
> that from the analysis. In brief
> i) First stage: matching has to be done within district.
> ii) In Second stage the unmatched treatments remaining from the first
> stage  will be tried to match with control observations from outside
> the region/district.
> How can I  use psmatch2 for this? Or, any other package in Stata? I
> really lack knowledge on writing program in STATA. Can anyone save me
> with this, specifically, how to write code to execute my model. I
> appreciate your thought and instruction in advance.
> SIncerely
> Auditya
>
>
> *
> *   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/
*
*   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/