Propensity Score Matching with DiD regression

[Nikita Shukla]

Hi everyone,

I had a really quick question about doing propensity score matching with panel data. I have values for pre- and post-treatment and I want to use propensity score matching before I perform the DiD regression. Do I perform matching with the households before they've been put through the treatment? I would also appreciate any advice on what kind of variables I should be using and if there's a minimum/perfect amount of variables that should be used to ensure they're matched properly?

Thank you!

[Jeff Wooldridge]

Is this panel data? Two periods or many?

[Jared Greathouse]

I would advise a reading of this paper too. Athey and co detail the differences between the unconfoundedness literature, synthetic controls and related estimators. It's very technical in nature, but it's instructive.

Either way, you'll match on pre-intervention trends.

[Maxence Morlet]

You may want to check this paper out: https://gking.harvard.edu/files/gkin...900011_rev.pdf

[Nikita Shukla]

Is this panel data? Two periods or many?

yes, it is panel data with only two periods ( 0 = pre-treatment, 1 = post-treatment)

[Nikita Shukla]

Thank you everyone for linking the papers! I will read through them

[Jeff Wooldridge]

I would recommend you use matching after creating the change in the outcome variable. The usual DiD with two-period panel data is to regress the change in y on a constant, the treatment indicator, treat, and control variables, x. But you can match on the control variables and use regression, too. I actually discussed this in my recent short course on DiD methods.

Here's how it would work:

xtset id time
gen cy = D.y
teffects psmatch (cy) (treat x1 ... xK), atet

The x variables should be determined prior to the intervention to be sure they are "good" controls. You can apply any other teffects estimator on cy, too.