Erik Sverdrup

Interesting, policytree takes around 15 sec for the two first ones and 11 minutes for the last.

Yes, we should consider handling repeated Xj in policytree, your package makes it clear it's very worthwhile. I'm not sure when I will get time to have a closer look...

Yes, you can probably work out empirical N/n heuristics to switch between implementations during runtime if you really want to keep tinkering! (In these things there are practically always areas...

Great, maybe you could put some timings on your GH README for reference? (considering I guess you'll keep making changes you'll always have the latest timings in your repo, we...

a) When things are getting more polished, we could also add an option to policytree that would dispatch to your package depending on input type? I.e. something like: ``` if...

Sounds good! Yes, we should add multithreading to policytree then as well, else it looks would look weird. Some practical considerations to keep in mind is guaranteeing same output for...

a) sounds good! Different compilers on different OS's sometimes produce different result. Also, different OS users may face different installation hassles, I tried installing the package on stanford's linux HPC...

PS: I think your package in its current from already seem very useful. The algo suggestion above could be more of a longer term project if you are interested in...

Just a heads up that this will be an annoying wrapper to make, since it involves a lot of tedious calculations from the wrapper side (sorry about that). It will...

Hi Robert, the idea behind MCF is to deliver an efficiency gain in estimates in the setting where you think it’s reasonable to believe that the treatment effects across arms...