Interpretable Matching for Causal Inference

Efficient implementations of the algorithms in the Almost-Matching-Exactly framework for interpretable matching in causal inference. These algorithms match units via a learned, weighted Hamming distance that determines which covariates are more important to match on. For more information and examples, see the Almost-Matching-Exactly website.

FLAME (Fast, Large-scale, Almost Matching Exactly)

The FLAME algorithm provides fast and large-scale matching approach to causal inference. FLAME creates matches that include as many covariates as possible, and iteratively drops covariates that are successively less useful for predicting outcomes based on matching quality. Currently the FLAME package applies to categorical data, and provides two approaches for implementation - bit vectors and database management systems (e.g., PostgreSQL, SQLite). For data that has been preprocessed and fits in memory, bit vectors should be applied. For extremely large data that does not fit in memory, database systems should be applied.

For more details about the FLAME algorithm, please refer to the paper: FLAME: A Fast Large-scale Almost Matching Exactly Approach to Causal Inference and FLAME: An R Package for a Fast Large-scale Almost Matching Exactly Approach to Causal Inference


# Current version on GitHub

Required data format

FLAME package requires input data to have specific format. First, input data should be a R Data Frame. Second, all covariates in the input data should be categorical covariates, represented by factor R data type. If there are continuous covariates, please consider regrouping. Third, input data columns should contain (1) covariates in factor data type, (2) outcome variable in numeric data type, and (3) variable specifying a unit is treated or control (treated = 1, control = 0) in factor data type. Lastly, though there are no requirements for input data column names, the column order should follow [covariates, outcome, treated]. Below is an example of input data with n units and m covariates.

x1 x2 ... xm − 1 xm outcome treated
R data type factor factor factor factor factor numeric factor
unit 1 0 1 ... 1 2 3.8 1
unit 2 1 0 ... 1 0 1.36 0
unit 3 0 1 ... 0 1 -7.25 0
... ... ... ... ... ... ... ...
unit n 0 1 ... 1 0 20 1

Holdout training set should also follow the same format.

Other requirements

  1. FLAME requires installation of python, specifically with at least python 2.7 version. If your computer system does not have python 2.7, install from here.

  2. For database systems implementation, FLAME package provides two versions - SQLite and PostgreSQL. PostgreSQL requires installation of external database system but it is faster. SQLite does not require external database system but is slower. If your computer does not have PostgreSQL installed, install from here. For connecting and setup of PostgreSQL server, please refer to the tutorial

  3. For database systems implementation, please name the database connection as db.

Some examples

Apply the FLAME Algorithm to Synthetic Data

Get Detailed Information After Matching

Match Quality


FLAME 1.0.0

This is a new submission to CRAN.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.


2.1.0 by Vittorio Orlandi, 5 months ago,

Report a bug at

Browse source code at

Authors: Vittorio Orlandi [aut, cre] , Sudeepa Roy [aut] , Cynthia Rudin [aut] , Alexander Volfovsky [aut]

Documentation:   PDF Manual  

MIT + file LICENSE license

Imports glmnet, gmp

Suggests knitr, mice, rmarkdown, testthat, xgboost

See at CRAN