Hello,
I just read a study that recommends the application of either propensity or matching methods using random forests, and then the use of raking iterations methods as a second stage to establish the weights of the subjects of a survey. According to the authors, it would be more effective to use a combination of these methods to reduce bias than the application of any single method on its own.
Can someone guide me through how we can combine these methods together with a concrete example using R in order to weighting the survey sample with the least bias possible?
The MatchIt package supports propensity value matching methods. I am not sure that it will do this based on random forest, but it seems straightforward. A propensity score is just a likelihood for being “enrolled” in the “intervention”. With a regression, you use “enrollment” as the dependent variable, all the antecedent variables as independent variables, then you leave out the study’s dependent variable. This basically uses the antecedents to predict who is most likely to enroll in the study/receive the treatment.
Random Forest is a fancier way of modeling this than normally used, but the formula is the same.:
Enrollment ~ all the antecedents.
thank you so much for referring to this article. yeah I am familiar with this package. quite useful. but it does not include raking. I found out a package called pewmethods made by the pew research center. quite effective when applying raking to estimate the weights of the subjects with the use of a benchmark data set that is already available.