Falsification before Extrapolation in Causal Effect Estimation

Authors: Zeshan M Hussain, Michael Oberst, Ming-Chieh Shih, David Sontag

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the properties of our approach on semi-synthetic and real world datasets, and show that it compares favorably to standard meta-analysis techniques. ... Our conclusions are supported by semi-synthetic experiments, based on the IHDP dataset, as well as real-world experiments, based on clinical trial and observational data from the Women s Health Initiative (WHI), that demonstrate various characteristics of our meta-algorithm.
Researcher Affiliation Academia Zeshan Hussain MIT CSAIL & IMES Cambridge, MA zeshanmh@mit.edu Michael Oberst MIT CSAIL & IMES Cambridge, MA moberst@mit.edu Ming-Chieh Shih National Dong Hwa University Hualien, Taiwan mcshih@gms.ndhu.edu.tw David Sontag MIT CSAIL & IMES Cambridge, MA dsontag@csail.mit.edu
Pseudocode Yes Algorithm 1 Extrapolated Pessimistic Confidence Sets
Open Source Code Yes For code and instructions on data access, please visit: https://github.com/clinicalml/rct-obs-extrapolation
Open Datasets Yes We generate semi-synthetic RCTs and observational datasets with covariates from the Infant Health and Development Program (IHDP)... as well as real-world experiments, based on clinical trial and observational data from the Women s Health Initiative (WHI).
Dataset Splits Yes To construct the subgroups, we consider all pairs of a selected set of binary covariates... We treat two of the subgroups as validation subgroups and two as extrapolated subgroups.
Hardware Specification No The paper does not mention any specific hardware (GPU, CPU, or cloud instance types) used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes Data generation parameters include K, r, mc, mb, cz, and the significance level . By default, we set K = 5, r = 10, mc = 4, mb = 3, cz = (0, 0, 2, 4, 6), and = 0.05. The full hyperparameter search is provided in Appendix F, and details of hyperparameter tuning can be found in Appendix C.