Sequential Underspecified Instrument Selection for Cause-Effect Estimation
Authors: Elisabeth Ailer, Jason Hartford, Niki Kilbertus
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Since a real-world evaluation of our approach would require access to sequential randomized experimentation in a complex setting, we are restricted to simulation studies. We first illustrate the properties of our proposed (combined) causal effect estimators in the underspecified IV setting and then evaluate our full sequential instrument selection method. |
| Researcher Affiliation | Collaboration | 1Helmholtz AI, Helmholtz Munich, Munich, Germany 2School of Computation, Information and Technology, Technical University Munich, Munich, Germany 3MILA Quebec Artificial Intelligence Institute, Montreal, Quebec, Canada 4Recursion, Montreal Quebec Canada 5Munich Center for Machine Learning, Munich, Germany. Correspondence to: Elisabeth Ailer <elisabeth.ailer@helmholtzmunich.de>. |
| Pseudocode | Yes | Algorithm 1 Sequential selection of instrument sets |
| Open Source Code | Yes | Code. The implementation as well as experimental details are publicly available on Github: https://github.com/EAiler/underspecified-iv. |
| Open Datasets | No | The paper describes generating random scenarios for simulations but does not provide access information for a fixed, publicly available dataset in the conventional sense. It states: "The data generation process follows the model described in (Janzing & Sch olkopf, 2018)." |
| Dataset Splits | No | The paper describes generating random scenarios for simulations but does not specify fixed training, validation, or test dataset splits in the conventional sense. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers. |
| Experiment Setup | Yes | For the cost function we choose c(d) = log(d) for d NIV/exp and c(d) = otherwise, effectively limiting the maximum number of instruments per round to NIV/exp. We set dz = 30, did = 15 and use two different treatment dimensions dx = 50 and dx = 150. Further, we allow NIV/exp = 3 instruments per round with a total budget of T = 6 experimental rounds. |