Randomized experiments are the gold standard for impact measurement. However, it is common that the treatment of interest cannot be directly randomized, like the usage of a feature, or participation in a program. Instead, nudges to use the feature or join the program are randomly assigned to induce the treatment of interest, but units can choose not to comply with the assignment, hindering simple analysis. This package, BIVA (Bayesian instrumental variable analysis), helps unlock flexible impact measurement for randomized experiments with noncompliance.
You can install the development version of biva from GitHub with:
# install.packages("devtools")
devtools::install_github("google/biva")Examples for how to apply the BIVA workflow can be found in the ‘vignettes’ folder.