Propensity score matching

[Alberto Alvarez]

Dear Experts,

Can you please help me with the following issue.

I have a sample where some observations experience a shock, let's say "t". So, what I want to do is to create a control group given certain characteristics (x1, x2, x3, etc.) which will represent a treated sample. I run a probit model:

probit t x1 x2 x3 x4 i.industry i.year, vce(cluster id)

then "predict prop_score"

what I get: for each firm in the treated sample I get a firm with more or less similar characteristics. Finally, what I want to do is to run diff-in-diff given treated and control sample, say:

y= t post t*post x1 x2
where post is 1 if observation falls in one year after the treatment ( zero otherwise)

QUESTION:
1. is it right what I am doing?
2. How an I create a control sample which will have more matches, say for one firm (treated group) 2/3 control firms?

Please, help me with this issue

[Clyde Schechter]

You can download -psmatch2- from SSC to get a very popular user-written program that calculates propensity scores and matching.

If you find it doesn't do exactly what you need in terms of matching, there have been numerous posts on this Forum about how to generically pick matched controls for cases. For example, http://www.statalist.org/forums/foru...gory-sex-state. (And there are many others--use the search function to find more.)

[Alberto Alvarez]

Thank you for your response! I used the following code:

psmatch2 t x1 x2, out(y)

and it weights each observation in accordance with the matching score. However, my question is how can I do diff-in-diff in this case guaranteeing that the observations that are used in regression are from treated and control group (from psmatch2 command)?

Please, help me with this question.

[Clyde Schechter]

After you match your cases to your controls, assign a match-group id (actually, I think that is part of the output of -psmatch2-) to each case and its associated controls. Drop all the observations that are not matched. Then -xtset- the data using that match-group id as the panel variable. The diff-in-diff analysis then uses a panel estimator (-xtreg-, or -xtlogit-, or -xtpoisson-, etc. depending on the type of outcome variable) and includes an interaction between t and post.

[Dimitriy V. Masterov]

You might find my answer to a similar CV question relevant.