Inconsistent ATT weights
OtterCuber opened this issue · 2 comments
In one of your references (Austin, 2011), ATT weights are calculated as 1 for the treated group and ps/(1-ps) for the control group. However, this doesn't appear to be how you implemented the ATT weighting estimator in get_w_from_ps().
Could you clarify how ATT is calculated and what reference you used? Thank you.
In get_w_from_ps(), I use the formula w_ATT = ps * w_ATE, where w_ATE are the usual ATE weights (A/ps + (1-A)/(1-ps), where A is treatment) and ps is the propensity score. You can see that these definitions are equivalent. This perspective on computing weights (i.e., starting with the ATE weights and then multiplying them by a function of the propensity score) is discussed in Li et a. (2018). You are welcome to verify that the weights are correctly computed yourself. If you find they are not, please let me know!
The way get_w_from_ps() is coded makes the code generalizable to multiple (i.e., more than 2) treatment levels, so it uses a propensity score matrix, where each unit gets a propensity score for each level of the treatment.
I can confirm that these two methods are equivalent. Thank you for your clarification!