Propensity Scores and Common Support

[Milos Drezga]

Hi, everybody!

I'm currently in the process of writing my MSc thesis in economics. I am studying the effect of an anti-corruption policy on corruption as perceived by the private sector in the Balkan countries. I decided to use propensity scores to reduce selection bias, as I am working with observational data. I generated the scores by running a probit regression, regressing my treatment on time-invariant characteristics of the firms in my sample. I plotted the density of my propensity scores using the following code:

probit tregion begin employ regi when
predict p
kdens p, ul(1) ll(0)
loc w=r(width)
kdens p if tregion==1, ul(1) ll(0) bw(`w') g(d1 x) nogr
kdens p if tregion==0, ul(1) ll(0) bw(`w') g(d0) at(x) nogr
line d0 d1 x, name(comp, replace)

As a result, I got the following graph:

What I would want to do now is create a restricted sample for which common support exists, as suggested by the graph. The question is: how? What command would I use to keep only those observations?

Thank you in advance,
Miloš

[Matt Warkentin]

Hi Milos,

To answer your question with a question, what would you consider as "enough" common support? Perhaps this is the root of what you meant with your question.

Just based on the eye-ball test it looks like you have at least some common support over the entire range of probabilities, albeit with a mass just below 0.4. So, a decision would need to be made to determine what is "enough" support. If you can specify this condition then what you want to achieve could be implemented somewhat trivially with -keep- or -drop-, or if you just want to set to missing for those observations that could be done too.

Unless you are asking how to determine what is enough support, is this your question? Not how to implement, but how to determine?

[Milos Drezga]

Hi Matt,

Thank you for your reply. Indeed, the graph seems to be suggesting that there is support across the whole range. I guess the question is both about what is enough and how to implement it. I understand that the simple -keep- or -drop- can be used, but I was wondering if there was a more sophisticated method that could use the information found in the above graph to generate the restricted sample.