Matching using GRF

[ukwarwicker]

trafficstars

Hi team,

I understand that one could use the GRF procedure as a matching technique to identify the weighted sample and then apply the weighting to a linear regression model (e.g., a difference-in-differences model). That is, the GRF procedure can be used to identify the weighted set of neighbors that share similar covariates with each treatment member. I was wondering if there is a way to separately identify/construct the weighted sample so that it can serves as input to a DID model. I saw that there was a recent function lm_forest that may be relevant, but couldn’t get it to work on my comp, presumably it is under development of some sort? Would appreciate any thoughts/suggestions on this, particularly relating to how one may be able to separately identify/construct the weighted sample, which then can be used as input to a linear regression model. Also, please provide guidance on how a balance test can be conducted on the weighted vs. nonweighted sample in this case. Thanks.

[erikcs]

Hi @ukwarwicker, a straight-forward way to apply GRF in the standard DiD setting is to use causal forest on the differenced outcomes: #1064. (If the setting is you don't observe each unit before/after treatment, https://arxiv.org/pdf/1905.11622.pdf provides an another approach using causal forest)