Causal Effect Identification in a Sub-Population with Latent Variables

Authors: Amir Mohammad Abouei, Ehsan Mokhtarian, Negar Kiyavash, Matthias Grossglauser

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct a numerical experiment to demonstrate the significance of the s-ID problem and assess the output of Algorithm 1. and In this paper, we studied the S-ID problem in the presence of latent variables and provided a sufficient graphical condition to determine whether a causal effect is S-ID. Consequently, we proposed a sound algorithm for S-ID.
Researcher Affiliation Academia Amir Mohammad Abouei 1, Ehsan Mokhtarian 1, Negar Kiyavash 2, Matthias Grossglauser 1 1School of Computer and Communication Sciences, EPFL 2College of Management of Technology, EPFL
Pseudocode Yes Algorithm 1 Computing P S X(Y) from P S(V) and Algorithm 2 Reduction from S-Recoverability to S-ID
Open Source Code No As discussed in the previous question, our numerical experiment consists of a few simple steps. One can generate data according to the defined causal model in our experiment and reproduce the results. Hence, we have not provided the code.
Open Datasets No We utilize synthetic data corresponding to a causal model defined in Appendix D. Hence, the data generation process is straightforward. No concrete access information for a publicly available or open dataset is provided as it's synthetic and generated directly by the experiment description.
Dataset Splits No No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning was found. The paper generates synthetic data and then filters it based on a condition, but this is not a train/validation/test split.
Hardware Specification No In Appendix D, we note that our numerical experiment does not need specific computational power. This indicates that no specific hardware details are necessary or provided for reproducibility.
Software Dependencies No The paper describes mathematical operations and refers to a 'simple plug-in estimator' but does not specify any particular software libraries, packages, or their version numbers necessary for replication beyond basic computational tools.
Experiment Setup Yes Consider the following structural causal model (SCM) with the causal graph depicted in Figure 7. U1 Bern(0.5) U2 Bern(0.5) X = U1 εx, εx Bern(0.2) Y = X U2 Z = U1 εz, εz Bern(0.2) ... We then computed S for each generated sample and collected the samples where S = 1, resulting in 1469 available samples from our target sub-population. Recall that our goal was to estimate PX=0(Y = 1|S = 1).