QuantileComparator

This is the code for the paper "Conditional Quantile Comparator: A Quantile Alternative to CATE" by Josh Givens, Henry Reeve, Song Liu, and Katarzyna Reluga. The paper is published in proceedings for NeurIPS 2024 and can be found at [Placeholder].

As a quick summary, this method creates a function to transform values from one conditional distirbution to a value of the equivalent quantile in another conditional distribution.

Folder contents

CDTE

This contains the code for the CDTE method presented in Kallus and Oprescu 2023. The paper and original code for this can be found at: https://proceedings.mlr.press/v206/kallus23a.html

Code

This contains all the code for the method. All estimators are defined in nonparamcdf.py with the kernels themselves implemented in kernel.py which is adapted from https://github.com/wittawatj/kernel-gof/ the original licence for this code can be found in the file as well. utils.py contains general helper functions for the code.

Experiments

This contains notebooks for all the experiments in the paper. ColonExample.ipynb contains the code for the colon cancer example, EmploymentExample.ipynb contains the code for the employment example, and SimulatedExperiment.ipynb contains the code for all the simulated examples. All experimental results are saved in the Test_Results folder.

Plots

This contains all the plots for the paper.

RealData

The contains the colon data and employment data used in the paper as well as the R code to retrieve and process the colon data.

Datasets

Employment data

The Employment data is taken from https://www.journals.uchicago.edu/doi/suppl/10.1086/687522/suppl_file/12062data.zip, the supplementary material for https://www.journals.uchicago.edu/doi/10.1086/687522

Colon Cancer Data

The colon cancer data is found in the survival package in R. The data is loaded with the following code:

library(survival)
data(colon)

Code to convert this data from long into wide format and then save to csv can be found in RealData/data_read.R.

Code Usage

The main classes are dr_learner, pseudo_ipw, and separate_learner which can all be used to estimate h at given points using the get_single_h method, evaulate at all y_1 step points using the get_all_hs method and estimate $g^*$ using the predict method. Each must be initially fit using the fit method with the data used to estimate the function.

There are also kernel_cdf and exact_cdf methods for estimating the cdf with cdf for giving evaluation of the cdf at specified $y,x$ points and getallcdfs for evaluating the cdf at all $y$ points for specified $x$ points. Both have to be fit using the fit method with the data used to estimate the cdf.

Finally there is the kernel_regressor class to perform standard regression with the fit method and evaluate the regression at specified points with the predict method.