The famous paper Dorie,2017 shows that BART performs dramatically well in causal inference. In my replication, MSE in BART can be 40% lower than MSE in other machine learning methods.

But all machine learning methods just regress $Y(0) \sim X$ on control group, get a predictor $f_0(X)$, and use it to predict $Y(0)$ on treatment group, similarly regress $Y(1) \sim X$ on treatment group, get a predictor $f_1(X)$, and use it to predict $Y(1)$ on control group, finally use $f_1(x) - y(0)$ (on control group) or $y(1) - f_0(x)$ (on treatment group) to estimate causal effects. It seems like that the accuracy of causal estimation depends on the accuracy of regression $Y \sim X$, so the great performance of BART in causal inference implies that BART also performs great in the general regression problem.

But BART is not very famous in the general regression problem, for example, the citations of Breiman's random forest paper are 50 times more than the citations of Chipman's BART paper. So, is BART just ignored in machine learning, or is BART particularly accurate in causal inference? If BART is particularly accurate, WHY?

  • 2
    $\begingroup$ Can you please point to where BART is "40% lower than MSE in other machine learning methods"? To quote the authors: "... it is fair to say that most of the methods in this competition performed reasonably well with both (the absolute value of) bias and RMSE (...) with respect to the outcome measure." That paper does not claim that BART is amazing; it consistently outperforms linear models, aside that? No clear wins. I think the authors advocate that a main factor for the success of a causal inference method is to focus on conditional means rather than the true data generative process. $\endgroup$
    – usεr11852
    Commented Jun 6, 2020 at 22:08
  • 1
    $\begingroup$ Thanks for your comment:) I mean in my replication the MSE of BART is 40% lower than MSE in other methods. And in the homepage of the competition jenniferhill7.wixsite.com/acic-2016/competition you can see that BART performs better than all other methods in the black box part. $\endgroup$ Commented Jun 7, 2020 at 3:05
  • $\begingroup$ Can you provide more details, perhaps some reproducible, on your replication of the work? There are programmatic elements as well as statistical fundamentals at play in the question. $\endgroup$ Commented Sep 29, 2022 at 13:27
  • $\begingroup$ BART = Bayesian additive Regression tree in this context $\endgroup$
    – Ggjj11
    Commented Sep 8, 2023 at 18:11


Your Answer

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

Browse other questions tagged or ask your own question.