Estimating Causal Effects using a Multi-task Deep Ensemble
Authors: Ziyang Jiang, Zhuoran Hou, Yiling Liu, Yiman Ren, Keyu Li, David Carlson
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method across various types of datasets and tasks and find that CMDE outperforms state-of-the-art methods on a majority of these tasks. |
| Researcher Affiliation | Academia | 1Department of Civil and Environmental Engineering, Duke University, Durham, NC, USA 2Department of Biostatistics and Bioinformatics, Duke University School of Medicine, Durham, NC, USA 3Program in Computational Biology and Bioinformatics, Duke University School of Medicine, Durham, NC, USA 4Department of Economics, University of Michigan Ross School of Business, Ann Arbor, MI, USA 5Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA 6Department of Computer Science, Duke University, Durham, NC, USA. |
| Pseudocode | No | The paper describes the architecture of its baselearner (Figure 1) and provides mathematical formulations, but it does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code to replicate all experiments is available at: https://github.com/jzy95310/ICK/tree/main/experiments/causal_inference |
| Open Datasets | Yes | All datasets used in our experiments are available at https://github.com/jzy95310/ICK/tree/main/data and are released under the MIT license. For ACIC, Twins, Jobs, and STAR, the original datasets have an open-access license and are publicly available. For COVID-19 experiment, the original dataset is available at: https://github.com/ieee8023/covid-chestxray-dataset |
| Dataset Splits | Yes | For the Twins dataset... we average over 10 train-validation-test splits with a ratio of 56:24:20. ...For the Jobs dataset... we average over 10 train-validation-test splits with a ratio of 56:24:20. ...The train-validation-test split ratio is again set to be 40:20:40. |
| Hardware Specification | No | The paper mentions "Due to time and memory constraints, we pick a small subset..." but does not specify any hardware details like GPU models, CPU types, or memory used for running the experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., Python version, specific library versions like PyTorch or TensorFlow). |
| Experiment Setup | Yes | For our experiment, we set σ2 x = 9 and σ2 ξ = 0.0025 and sample N = 3000 data points. The CMDE model consists of 10 estimators where f H, f T , and f HT in each estimator are single-hidden-layer neural networks with Re LU activation and 2048 units in the hidden layer. We set the initial values of αH, αT , and αHT to be αH = 0, αT = 0, and αHT = 1. All weight and bias parameters in f H, f T , and f HT are independently drawn from a normal distribution N(0, σ2 w I) a priori and σ2 w = 0.1. |