PS matching and Difference-in-Differences

[Simone Chinetti]

Dear all,

I am facing some troubles in understanding which STATA command is best at estimating propensity score and then use it to estimate a difference-in-differences model.

In the setting I have to study, there is a policy that firms can access to if they request it. So my sample is composed of firms who obtained access to the policy in different years (staggered treatment). However, I do not observe firms that asked for gaining access to policy but failed to. Hence, in order to assess the impact of the policy I need to construct a control group, as the best way to do it is to match treated firms with the overall sample of existing firms who did not benefit from the treatment, based on pre-treatment characteristics for instance.

In STATA there are several ways, according to my understanding and infos I found online, to estimate propensity score so as to construct the control group. The best one methods should be:

• teffects, which however directly estimate ATE or ATET; but it is not very clear to me if I can only estimate PS and then apply it to my setting (difference-in-differences)
• pscore (algorithm released by Becker and Ichino)
• psmatch2

As far as I understood, the preferable approach should be that one based on teffects (by Abadie) given that it computes properly SE. However, as I already pointed it's not very clear how to estimate only the PS in order to use it to estimate the ATET based on different model rather than that which is embedded in the program.

Does anyone has any clue on how to overcome this issue ? Does anyone has any suggestion to help me ?

Thank you very much in advance.

[Simone Chinetti]

No one has any clue ?

[Tim Kohler]

Hi Simone.

I trust you got an answer already, considering this post is nearly two years old. In case you have not received an answer, or maybe someone else on here is looking for a similar answer, see the below.

I believe you can incorporate PSM into DID using one of the the following three ways:

(1) Estimate the propensity scores using either "teffects pscore" or psmatch2, then simply control for the propensity score in the DID regression as a covariate.
(2) After estimating the propensity scores as above in (1) using psmatch2, estimate the DID regression as normal, but weight the regression using the frequency weights generated from the psmatch2 process (this restricts the sample to the matched sample and weights accordingly).
(3) Use the command "diff", which has an option to generate propensity scores simultaneously and account for them in your DID regression.

I am, however, not sure which option (1) - (3) is preferred, and whether this varies by context.

[Haoyu Wang]

Hi Simone.

I trust you got an answer already, considering this post is nearly two years old. In case you have not received an answer, or maybe someone else on here is looking for a similar answer, see the below.

I believe you can incorporate PSM into DID using one of the the following three ways:

(1) Estimate the propensity scores using either "teffects pscore" or psmatch2, then simply control for the propensity score in the DID regression as a covariate.
(2) After estimating the propensity scores as above in (1) using psmatch2, estimate the DID regression as normal, but weight the regression using the frequency weights generated from the psmatch2 process (this restricts the sample to the matched sample and weights accordingly).
(3) Use the command "diff", which has an option to generate propensity scores simultaneously and account for them in your DID regression.

I am, however, not sure which option (1) - (3) is preferred, and whether this varies by context.

Hi TIm,

I like your method to solve the problem! It's simple but useful. And I take a great interests in the option(1) and (2).

Could you please give me any clues of them that which paper mentioned them or use them?

[Justin Niakamal]

You could look into the user-written diff (from SCC). diff includes Propensity Score Matching diff-in-diff. See the recommended references in the help file and the Stata Journal article for implementation.

Hope this helps.

[Haoyu Wang]

You could look into the user-written diff (from SCC). diff includes Propensity Score Matching diff-in-diff. See the recommended references in the help file and the Stata Journal article for implementation.

Hope this helps.

Right. I have used command diff with option kernel, but it provide me different results by using same command and same dataset. According to this, I turn to psmatch2 for help. I don't think diff is robust in PSM-DID. May be I'm worng. I'll find out the reference.

Thank you for your help.