A python implementation of the dual bounds framework for inference on partially identified estimands (and the solutions to optimization problems more generally). See https://dualbounds.readthedocs.io/en/latest/ for detailed documentation and tutorials.
To install dualbounds
, just use pip:
pip install dualbounds
Documentation and tutorials are available at https://dualbounds.readthedocs.io/en/latest/.
Given a response vector y
, binary treatment vector W
, covariate matrix X
, and an (optional) propensity score vector pis
, dualbounds allows analysts to perform inference on partially identified estimands of the form E[f(Y(0), Y(1), X)], for any choice of f. For example, the code below shows how to perform inference in P(Y(0) < Y(1)), the proportion of individuals who benefit from the treatment. Dual bounds can wrap on top of any machine learning model to provide provably valid confidence intervals in randomized experiments.
import dualbounds as db
# Generate synthetic data from a heavy-tailed linear model
data = db.gen_data.gen_regression_data(
n=900, # Num. datapoints
p=30, # Num. covariates
r2=0.95, # population R^2
tau=3, # average treatment effect
interactions=True, # ensures treatment effect is heterogenous
eps_dist='laplace', # heavy-tailed residuals
sample_seed=123, # random seed
)
# Initialize dual bounds object
dbnd = db.generic.DualBounds(
f=lambda y0, y1, x: y0 < y1, # defines the estimand
X=data['X'], # n x p covariate matrix
W=data['W'], # n-length treatment vector
y=data['y'], # n-length outcome vector
pis=data['pis'], # n-length propensity scores (optional)
Y_model='ridge', # description of model for Y | X, W
)
# Perform inference
dbnd.compute_dual_bounds(
nfolds=5, # number of cross-fitting folds
alpha=0.05 # nominal level
)
If you use dualbounds
in an academic publication, please consider citing Ji, Lei, and Spector (2023). The bibtex is below:
@misc{ji2023modelagnostic,
title={Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects},
author={Wenlong Ji and Lihua Lei and Asher Spector},
year={2023},
eprint={2310.08115},
archivePrefix={arXiv},
primaryClass={econ.EM}
}