The FLAME package offers an efficient implementation of the Fast, Large-Scale, Almost Matching Exactly algorithm, described in detail here. FLAME allows for interpretable matching within observational settings in causal inference. It does so by matching units via a learned, weighted Hamming distance that determines which covariates are more important to match on. The package will soon be updated to include other algorithms in the Almost Matching Exactly framework, such as DAME, FLAME-Networks, and FLAME-IV.
We can generate some toy data for matching by using the gen_data function included in the package:
matching_data <- gen_data()
holdout_data <- gen_data()
We can then match the units in matching_data, by using holdout_data to learn a weighted Hamming distance between units by calling the main function in the package:
FLAME_out <- FLAME(matching_data, holdout_data)
The matched data, which can be accessed via FLAME_out$data contains asterisks for the covariates on which units were not matched. The groups of units matched together are stored in FLAME_out$MGs, the corresponding conditional average treatment effect (CATE) estimates in FLAME_out$CATEs, the covariates values units were matched on in FLAME_out$matched_on, and the covariates at every iteration of FLAME by FLAME_out$matching_covs.
For a more in-depth discussion of the package, please see the accompanying vignette.
FLAME can be installed directly from CRAN via running install.packages('FLAME') in R. This will also install some tidyverse dependencies, gmp for FLAME's efficient bit-vectors implementation, mice for handling missing data, and glmnet and xgboost for outcome prediction.
Alternatively, you can install the package from the author's GitHub, here via devtools::install_github('https://github.com/vittorioorlandi/FLAME').