Rethinking Nonlinear Instrumental Variable Models through Prediction Validity
Authors: Chunxiao Li, Cynthia Rudin, Tyler H. McCormick
JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 5, we apply our two-stage and one-stage methods to simulated datasets. The simulation results that accord with theorems in Section 4 verify the theoretical results of our two stage method. We also show that our two stage method can efficiently identify non-instruments. In Section 6, we apply our two stage method to a real-world dataset on climate policy perspectives of voters who live near a wind-energy project. We show that our two stage method with more flexible model constructions does often outperform the traditional two-stage method in terms of more accurate estimation on the true causal effect. |
| Researcher Affiliation | Academia | Chunxiao Li EMAIL Department of Statistical Science Duke University Durham, NC 27708, USA; Cynthia Rudin EMAIL Departments of Computer Science, Electrical and Computer Engineering, Statistical Science, Mathematics and Biostatistics & Bioinformatics Duke University Durham, NC 27708, USA; Tyler H. Mc Cormick EMAIL Department of Statistics and Department of Sociology University of Washington Seattle, WA 98195-4322, USA |
| Pseudocode | Yes | Appendix A. Pseudocode for the Two-Stage and One-Stage ML-IV Methods |
| Open Source Code | No | The paper mentions using "tensorflow, cvxopt and cvxpy in python for our two stage method and tensorflow for our one stage method." However, it does not explicitly state that the authors are releasing their own code for the methodology described in the paper, nor does it provide a link to a repository. |
| Open Datasets | Yes | In this section, we tested our two-stage and one-stage methods on data from the paper entitled Electoral Backlash Against Climate Policy: A Natural Experiment on Retrospective Voting and Local Resistance to Public Policy by Stokes (2016). |
| Dataset Splits | Yes | In the modeling process, 10-fold cross validation was used to ensure the stability of our results. |
| Hardware Specification | No | The paper mentions the use of software tools like "tensorflow, cvxopt and cvxpy" but does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | In our experiments we used tools such as tensorflow, cvxopt and cvxpy in python for our two stage method and tensorflow for our one stage method. While specific software names are mentioned, no version numbers are provided for Python, TensorFlow, cvxopt, or cvxpy. |
| Experiment Setup | Yes | Second, we built four different models using different input features and used RMSE on the outcome as the metric to compare prediction performance. In the modeling process, 10-fold cross validation was used to ensure the stability of our results. ... Note that we have normalized all input variables when pre-processing, which means the results shown in the table are standardized. |