Timeline for Should I use a machine learning model to calculate propensity score?
Current License: CC BY-SA 4.0
18 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 3, 2023 at 18:46 | comment | added | usεr11852 | (+1) Great! Thank you. My work has moved me away from Causal Inference stuff but these seem like a bunch of really interesting papers. I will try to read them in Feb. | |
| Feb 3, 2023 at 17:34 | comment | added | Noah | @usεr11852, well, EB doesn't estimate propensity scores at all. It estimates weights directly, skipping the PS. So it's not that the PS are something different, it's that there aren't any involved at all. A bunch of methods have been developed recently that work similarly (skipping PS altogether). See Chattopadhyay et al. (2020) for information on this distinction and other examples of direct weighting methods. Li and Li (2021) compare these methods with PS methods in a simulation study. | |
| Feb 3, 2023 at 16:24 | comment | added | usεr11852 | I mean, sorry, that's what I mean, it balances the ones you want. Thanks, I will look at the Y&Y paper! In general, entropy balancing is interesting in the sense of the prop. scores aren't really probabilities of getting the treatment but rather weights to balance the potential confounders. | |
| Feb 3, 2023 at 15:13 | comment | added | Noah | @usεr11852 It doesn't, it only balances the moments you specifically request, and requesting too many makes the problem infeasible. A new development tries to address this problem by penalizing higher order terms: Yu & Yang (2022) | |
| Feb 3, 2023 at 13:09 | comment | added | usεr11852 | From what I recall entropy balancing tries to explictly match higher moments too. | |
| Mar 2, 2022 at 16:16 | comment | added | Noah | @PlasticMan Both good points :) | |
| Mar 2, 2022 at 15:44 | comment | added | riccardo-df | Thanks for the references, I will look into them. Anyway, the main point of my comment was about how to define covariate balance (not just comparing first moments), and to points to some theoretical results about the propensity score (at the population level). | |
| Mar 2, 2022 at 14:39 | comment | added | Noah | @PlasticMan regarding checking balance on the propensity score, Ho et al. (2007) and Stuart et al. (2013) argue otherwise. Balance on a bad propensity score won't yield covariate balance, but the only way to assess whether a propensity score is good is to see if it achieves covariate balance. So covariate balance is primal and balance on the propensity score is incidental. | |
| Mar 2, 2022 at 9:31 | comment | added | riccardo-df | Moreover, Imbens and Runib (2015, chapter 14) shows that covariates have the same distribution across groups iff the average propensity score (at population level) is the same across treatment arms. So, investigation of the (estimated) propensity score may be useful in order to check for the balance (clearly assuming that our estimated function is at least a very good approximation of the true score). | |
| Mar 2, 2022 at 9:29 | comment | added | riccardo-df | Wonderful answer. I want just to make a comment about the following statement: "I'm defining balance as the case where the means of every term in the outcome model are the same between the treatment groups." Covariates balance should be achieved for the whole distribution of the variables, not just for the first moment. Clearly, it may be infeasible chekcing all the univariate distributions, or the entire multivariate distribution, and so we are satisfied with comparisons of first and second moments. But ideally we would like that $X_j$ are distributed identically in both groups. | |
| Dec 14, 2019 at 0:11 | history | edited | Noah | CC BY-SA 4.0 |
deleted 1 character in body
|
| Aug 8, 2019 at 17:49 | history | edited | Noah | CC BY-SA 4.0 |
edited body
|
| Aug 8, 2019 at 17:12 | comment | added | usεr11852 | Credit where is due. This is a proper answer. (+1) Great to see some newer references. | |
| Aug 8, 2019 at 16:51 | comment | added | EdM | +1 particularly for a superb list of references | |
| Aug 8, 2019 at 16:22 | history | edited | Noah | CC BY-SA 4.0 |
added 249 characters in body
|
| Aug 8, 2019 at 9:25 | comment | added | lsfischer | Thank you for the clear and complete explanation on both paths of propensity scores! | |
| Aug 8, 2019 at 9:24 | vote | accept | lsfischer | ||
| Aug 7, 2019 at 19:54 | history | answered | Noah | CC BY-SA 4.0 |