I want to run a diff-in-diff model. To choose an appropriate control group, I use a nearest-neighbor matching model based on several determinants of the outcome variable that I study.

I was wondering: is it possible to directly match my observations using the values of the outcome variable before the treatment instead of simply using the determinants? It seems like a perfect way to obtain a parallel trend for the outcome variable for the two groups in the pre-treatment period. But most papers I read avoid matching directly on the outcome variable in the pre-treatment period, there must be an explanation.


1 Answer 1


Try Propensity Score Matching, a method to match observations by the likelihood of being treated (instead of observing the outcome variable)

  • $\begingroup$ Thanks a lot, I will look at it. But just out of curiosity: do you know what would be wrong in matching directly the values of the outcome variable before the treatment date? Thanks again $\endgroup$ Commented Dec 4, 2018 at 12:47
  • $\begingroup$ The whole reason you do a diff-in-diff analysis is that observations usually do not match in the outcome variable (and, of course, to get rid of time trends). If you had observations that have similar (or identical) determinants AND a similar outcome, you could just take the difference after the treatment. Importantly, a treatment is only meaningful when being applied randomly (i.e., the probability of being treated does not depend on your determinants). Otherwise you could make no (causal?) inference whether the treatment is driving your effect or one of the determinants. $\endgroup$
    – D. Beck
    Commented Dec 5, 2018 at 13:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.