The parallel trends assumption for a difference-in-differences analysis does not require the level of the trends to be similar (just parallel). However, in a case where we want the levels to be similar, can we match on the pre-treatment outcomes? Is there a reason not to do this? Should we just match on pre-treatment covariates to try to reduce the gap between the levels of the treated and control groups before the treatment?

  • $\begingroup$ Have you considered using a Synthetic Cohort method? $\endgroup$
    – dimitriy
    Aug 26, 2020 at 3:29

2 Answers 2


Combining difference-in-differences (DiD) and conditioning on pre-treatment outcomes is a practice that is being used in the applied literature (at least in economics). My understanding of the literature is that the conditions under which this is good practice are not completely clear. This paper, which I find relevant for this question, argues that one should be careful about this combination. Conditioning on pre-treatment outcomes does not necessarily (in theory) reduce the bias of the DiD estimator. In simulations, the author actually finds that conditioning on pre-treatment could increase the bias. A safer practice, according to the paper, is to increase the set of pre-treatment covariates (rather than outcomes) one controls for, to improve the validity of the parallel trend assumption.

  • $\begingroup$ My understanding is that matching on pre-treatment outcomes is problematic if their distributions in the intervention and control groups vary significantly. For example, matching on pre-treatment medical costs of $y_t$=1,000 in the intervention cohort and $y_c$=1,000 in the control cohort ignores huge regression-to-the-mean effects if $y_t \sim{N(\mu = 2,000,\sigma=100)}$ and $y_c \sim{N(\mu = 500,\sigma=100)}$, leading to a biased estimate of $ATE$ or $ATT$. $\endgroup$
    – RobertF
    Apr 6, 2022 at 21:37
  • $\begingroup$ interesting comment @RobertF, but a bit cryptic so far, would be great to expand on it as a full answer? Thanks! $\endgroup$
    – Matifou
    Oct 21, 2022 at 7:16

Take a look at this wonderful paper that helped me understand whether or not to do so in my setting. Essentially, matching on pre-treatment outcomes means that you move from a world in which the main identifying assumption is one of parallel trends, to an unconfoundedness world. If you don't believe that parallel trends hold, that might be a better bet. If you have an exogenous shock, it could induce bias (They cite a lot of papers that might be more or less applicable depending on your set-up).



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.