Finding Regions of Heterogeneity in Decision-Making via Expected Conditional Covariance
Authors: Justin Lim, Christina X Ji, Michael Oberst, Saul Blecker, Leora Horwitz, David Sontag
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In a semi-synthetic experiment, we show that our algorithm recovers the correct region of heterogeneity accurately compared to baselines. Finally, we apply our algorithm to real-world healthcare datasets, recovering variation that aligns with existing clinical knowledge. |
| Researcher Affiliation | Collaboration | Justin Lim MIT CSAIL and IMES Cambridge, MA justinl@mit.edu Christina X Ji MIT CSAIL and IMES Cambridge, MA cji@mit.edu Michael Oberst MIT CSAIL and IMES Cambridge, MA moberst@mit.edu Saul Blecker NYU Langone New York, NY saul.blecker@nyulangone.org Leora Horwitz NYU Langone New York, NY leora.horwitz@nyulangone.org David Sontag MIT CSAIL and IMES Cambridge, MA dsontag@csail.mit.edu. This work was supported in part by Independence Blue Cross |
| Pseudocode | Yes | Algorithm 1 Identifying regions with variation |
| Open Source Code | Yes | Our code is available at https://github.com/clinicalml/finding-decision-heterogeneity-regions. |
| Open Datasets | Yes | Dataset: We use publicly available data from Lin et al. (2020), who ask participants on Amazon s Mechanical Turk platform to make recidivism predictions based on information present in the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) dataset for Broward County, FL (Dressel and Farid, 2018). |
| Dataset Splits | Yes | After requiring at least 4 patients per agent, 3,576 patients and 176 group practices are included. This filter ensures each group practice has at least 1 sample in the training and validation sets and at least 2 samples in the test set. |
| Hardware Specification | No | The paper mentions 'CPU with 32 cores, 256GB of RAM' in Appendix B.1 and 'All experiments ran on CPUs with 32 cores and 256GB of RAM' in the main text of the appendix, but it does not specify exact CPU models (e.g., Intel Xeon E5-2630, AMD Ryzen) or GPU models, if any were used. |
| Software Dependencies | No | The paper mentions using PyTorch and Scikit-Learn but does not provide specific version numbers for these or any other software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | In this experiment, we choose β = 0.25 as input to our algorithm. Adam optimizer with learning rate 0.001. |