Propensity Score Matching and Interaction Terms

[Philipp Schmitt]

Hi all!

I am currenlty analyzing a cross section dataset with propensity score matching and got the following problem:

As soon as I include interactions and quadratic variables like age^2 to estimate my propensity scores, the resulting ATT is really far off the former results. Is it a problem to include interactions etc. in psmatch and teffects?

Thanks in advance.

Best

Philipp

[FernandoRios]

Hi Phillipp
That is not a problem.
Often, adding squares helps with balancing, which in turn gives you a better match. However, if you had a problem with overlapping, and now you add squared terms, then the problems you describe may appear.
One thing I would do is check how the summary statistics compare before and after matching before you try adding those squared terms.
Fernando

[Philipp Schmitt]

Hi Fernado,

thanks!
I'm not 100% sure but I think overlapping looks rather good and balancing is good aswell... Moreover, all observations are on support.

Are you sure you mean summary statistics? Why should that change?

I just dont get why these 2 variables change my result so much as they are just an interaction and potency. Could this be a indication that I dont have enough characteristic variables to perform a proper PS matching?