Random forests (RF) are a very simple and flexible nonparametric method for estimating the propensity score. They automatically allow for interactions, non-linear effects, can handle high-dimensional features, etc. RF tends to work well with default values too.
Currently weightit supports bart and gbm which also have the above benefits. I'm finding that the gbm method is too slow for large datasets. bart still works reasonably fast for large datasets but is more of a black box to me compared to RF. In any case, might be worth adding RF as an alternative approach?
EDIT: In the meantime, here is a slow workaround to use randomForest:
weightit(
...
method = 'super',
SL.library = c("SL.randomForest")
)
Random forests (RF) are a very simple and flexible nonparametric method for estimating the propensity score. They automatically allow for interactions, non-linear effects, can handle high-dimensional features, etc. RF tends to work well with default values too.
Currently
weightitsupportsbartandgbmwhich also have the above benefits. I'm finding that thegbmmethod is too slow for large datasets.bartstill works reasonably fast for large datasets but is more of a black box to me compared to RF. In any case, might be worth adding RF as an alternative approach?EDIT: In the meantime, here is a slow workaround to use randomForest: