I am familiar with Inverse Propensity Weighting (IPW) for the estimation of causal effects, and recently, I came across the 2016 paper by Chernozhukov et al. on Double/Debiased Machine Learning. From what I can see, both DML and IPW make similar assumptions over the causal structural equations: is there any reason why I would choose one approach over the other when estimating the ATE or CATE?

  • 3
    $\begingroup$ Mostly a rule of cool situation. (Someone will make the case about non-linearities too. Maybe we get a mention of Causal Forests and a historical reference to doubly robust methods.) (+1 cause I am curious for any other half-reasonable points - oh and welcome to CV.SE) $\endgroup$
    – usεr11852
    Commented Mar 24, 2022 at 1:40
  • 1
    $\begingroup$ Check out this related question. In short: DML uses a doubly-robust estimator; IPW is singly robust except for a few specific methods. The causal identification assumptions are the same; they differ in their ability to remove confounding by the observed variables. $\endgroup$
    – Noah
    Commented Mar 24, 2022 at 3:58
  • 1
    $\begingroup$ Look up AIPW vs IPW. Double ML for the ATE is pretty much just the decades-old AIPW estimator combined with cross-fitting of nuisance estimators (which goes back decades as well). This writeup looks good: scholar.harvard.edu/files/aglynn/files/AIPW.pdf $\endgroup$
    – Lars
    Commented Apr 15, 2022 at 2:47


Your Answer

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