Conditional Outcome Equivalence: A Quantile Alternative to CATE

Authors: Josh Givens, Henry Reeve, Song Liu, Katarzyna Reluga

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our theory in numerical simulations which show that our method produces more accurate estimates than baselines. Finally, we apply our methodology to a study on the effect of employment incentives on earnings across different age groups.
Researcher Affiliation Academia Josh Givens University of Bristol josh.givens@bristol.ac.uk Henry W J Reeve University of Bristol henry.reeve@bristol.ac.uk Song Liu University of Bristol song.liu@bristol.ac.uk Katarzyna Reluga University of Bristol katarzyna.reluga@bristol.ac.uk
Pseudocode Yes Algorithm 1 DR estimation procedure for the CCDF contrasting function h", "Algorithm 2 DR estimation procedure for the CQC g
Open Source Code Yes Code implementation can be found at: github.com/joshgivens/Conditional Outcome Equivalence", "Code to implement our approach alongside Jupyter notebooks running our numerical experiments can be found in the supplementary materials.
Open Datasets Yes We use a dataset on an employment programme which has been studied in various prior works [4, 5, 26].", "This dataset was originally introduced in Laurie et al. [19] and can be found in the survival" package in R and loaded with the line data(colon, package="survival").
Dataset Splits No In our experiments half the samples are used to estimate the propensity score and CCDFs and the other half are used to regress against the pseudo-outcome.
Hardware Specification Yes Each experiment took no longer than 1 hour to run on a single 4 core CPU with 8GB of RAM.
Software Dependencies No We can use the Pool Adjacent Violators Algorithm (PAVA) which performs isotonic projection and is implemented in the Isotonic Regression class of sci-kit learn in Python [25].
Experiment Setup Yes In our experiments half the samples are used to estimate the propensity score and CCDFs and the other half are used to regress against the pseudo-outcome. In our first experiment, we let 2n = 1000 and vary γ in [0, 10]. In our second experiment, we let γ = 6 and 2n vary in [200, 5000].