Generalised propensity score for multiple treatments

[Padarn]

Hi all, really interesting work

I am trying to think about a problem in which we have multiple continuous treatments, so I'd like to estimate a multi-dimensional generalized propensity score (GPS).

The literature on GPS (https://www.math.mcgill.ca/dstephens/PSMMA/Articles/HIrano-Imbens-2004.pdf) doesn't really seem to cover this case, and it doesn't seem like a totally straightforward extension.

Does anyone have tips on

• Is literature I'm missing on this?
• Is this something DoWhy already has?
• If not, is it something of interest?

Thanks

[EgorKraevTransferwise]

EconML's docs seem to suggest some of their models support multiple continuous treatments

Some estimation models allow joint estimation of the effects of multiple treatments on outcome(s)

Not sure if tbe DoWhy wrapper for the EconML models supports that as well - hope you can try both raw EconML models and the DoWhy wrapper for them and tell us what you find :)

[Padarn]

Thanks, I'll report back!

[Padarn]

Coming back to somewhat report on my findings for anyone else interested.

I didn't use the wrapper, as EconML had most of what I wanted, but EconML doesn't directly support multi-dimensional generalized propensity scores exactly. However, the DML (double machine learning) approach, which they have good support for, does allow for multiple treatments and estimates a model to predict treatment from confounders and controls. This does not directly result in propensity scores, but I found the DML approach to work fairly well.