Deep Learning Methods for Proximal Inference via Maximum Moment Restriction
Authors: Benjamin Kompa, David Bellamy, Tom Kolokotrones, james m robins, Andrew Beam
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method achieves state of the art performance on two well-established proximal inference benchmarks. Finally, we provide theoretical consistency guarantees for our method. Section 6 provides details on the experiments conducted. |
| Researcher Affiliation | Academia | Benjamin Kompa Department of Biomedical Informatics Harvard Medical School benjamin_kompa@hms.harvard.edu David R. Bellamy* Department of Epidemiology, CAUSALab Harvard School of Public Health david_bellamy@g.harvard.edu Thomas Kolokotrones Department of Epidemiology Harvard School of Public Health thomas_kolokotrones@hms.harvard.edu James M. Robins Department of Epidemiology, CAUSALab Harvard School of Public Health robins@hsph.harvard.edu Andrew L. Beam Department of Epidemiology, CAUSALab Harvard School of Public Health andrew_beam@hms.harvard.edu |
| Pseudocode | No | The paper describes the method using mathematical equations and textual explanations, but it does not include a distinct pseudocode block or algorithm outlining the steps in a structured, code-like format. |
| Open Source Code | Yes | The code to reproduce our experiments can be accessed on Git Hub.2 [2]https://github.com/beamlab-hsph/Neural-Moment-Matching-Regression |
| Open Datasets | Yes | Hartford et al. [21] introduced a data generating process for studying instrumental variable regression, and Xu et al. [7] adapted it to the proximal setting. The second benchmark uses the d Sprite dataset from Matthey et al. [27]... |
| Dataset Splits | No | Each method was trained on simulated datasets with sample sizes of 1000, 5000, 10,000, and 50,000. To assess the performance of each method, we evaluated a at 10 equally-spaced intervals between 10 and 30. We compared each method s estimated potential outcomes, ˆE[Y a], against estimates of the truth, E[Y a], obtained from Monte Carlo simulations (10,000 replicates) of the data generating process for each a. The evaluation metric is the causal mean squared error (c MSE) across the 10 evaluation points of a: 1 10 i=1(E[Y ai] ˆE[Y ai])2. For MMR-based methods, predictions are computed using a heldout dataset, DW with 1,000 draws from W so ˆE[Y ai] = |DW| 1 P|DW| j ˆh(ai, wj), i.e. a sample average of the estimated bridge function over W. |
| Hardware Specification | Yes | Experiments were conducted in Py Torch 1.9.0 (Python 3.9.7), using an A100 40GB or Titan X 12GB GPU and CUDA version 11.2. |
| Software Dependencies | Yes | Experiments were conducted in Py Torch 1.9.0 (Python 3.9.7), using an A100 40GB or Titan X 12GB GPU and CUDA version 11.2. |
| Experiment Setup | Yes | Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix B and C as well as our code. We choose k to be an RBF kernel (see Appendix B). |