Approximate Causal Effect Identification under Weak Confounding
Authors: Ziwei Jiang, Lai Wei, Murat Kocaoglu
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we conduct synthetic and real data simulations to compare our bounds with the bounds obtained by the existing work that cannot incorporate such entropy constraints and show that our bounds are tighter for the setting with weak confounders. We conduct experiments using both simulated and real-world data to test our method and demonstrate that our bound is tighter than existing approaches that are unable to incorporate entropy constraints. We demonstrated our method with simulated and real-world datasets in this section. First, we show the behavior of the bounds with randomly sampled distributions P(X, Y ). We change the entropy constraint θ from 1 to 0 for each sampled distribution. We also experiment with the full distribution P(X, Y, Z) where Z is the low entropy confounder and X, Y in high dimensions. We show the experimental results with the real-world dataset Adult (Dua & Graff, 2017). Since our algorithm works for discrete random variables with binary treatment or outcome, we take a subset of features in the graph and modify some features by discretizing continuous variables or combining states with very low probabilities. And finally, we experiment with our method in the finite sample setting and compare two optimization problem formulations. |
| Researcher Affiliation | Academia | 1Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, USA. Correspondence to: Ziwei Jiang <jiang622@purdue.edu>. |
| Pseudocode | No | No pseudocode or algorithm block found. |
| Open Source Code | No | No explicit statement about open-source code release or link to repository found. |
| Open Datasets | Yes | For the INSURANCE dataset, we aim to estimate the causal effect of Car Cost on the expected claim of the Property Cost. We consider the variable Accident as an unobserved variable with known entropy. The Car Cost and Property Cost claim are confounded through the cost of the other car in the accident as shown in Figure 6b. The results in Table 3 indicate narrow bounds on the causal effect when the entropy of the confounder is small. Therefore, we can have confidence in predicting the expected claim based on car cost, even in the presence of the confounding variable. For the ADULT Dataset (Dua & Graff, 2017), we take a subset of variables from the dataset with the causal graph as shown in Figure 6a. In this experiment, we treat age as a protected feature, which may not be accessible from the dataset, and only the entropy of age is known. We experiment with the ASIA dataset (Lauritzen & Spiegelhalter, 1988). |
| Dataset Splits | No | The paper describes data generation and sampling for evaluating bounds but does not specify training/validation/test splits for a learnable model. |
| Hardware Specification | No | No specific hardware details (GPU/CPU models, memory, etc.) are mentioned for running experiments. |
| Software Dependencies | No | We use the CVXPY package to solve the problem and formulate the constraint according to the Disciplined Convex Programming rules. (CVXPY is mentioned, but no version number is provided). |
| Experiment Setup | No | The paper describes parameters for data generation and evaluation of bounds (e.g., 'sampling P(x) from 0.01 to 0.8 and P(y|x) from 0 to 1', 'use {10, 102, 103, 104} samples from each distribution'), but it does not provide specific experimental setup details like hyperparameters (learning rate, batch size) for a machine learning model. |