Discovering Heterogeneous Causal Effects in Relational Data

Shishir AdhikariAAAI 2024 · OpenReview

Abstract

Causal inference in relational data should account for the non-IID nature of the data and the interference phenomenon, which occurs when a unit's outcome is influenced by the treatments or outcomes of others. Existing solutions to causal inference under interference consider either homogeneous influence from peers or specific heterogeneous influence contexts (e.g., local neighborhood structure). This thesis investigates causal reasoning in relational data and the automated discovery of heterogeneous causal effects under arbitrary heterogeneous peer influence contexts and effect modification.

Introduction

Standard causal inference often relies on the "Independent and Identically Distributed" (IID) assumption, which posits that one unit's treatment does not affect another's outcome. However, in interconnected environments—such as social media platforms, healthcare systems, and public policy domains—this assumption frequently fails due to interference. For example, a vaccination campaign does not just protect the individual; it alters the infection risk for their entire social circle.

This research addresses the challenge of measuring causal effects in these complex environments. By moving beyond the simplistic assumption that peer influence is uniform (homogeneous), this work develops frameworks to discover heterogeneous causal effects—identifying how the impact of a treatment varies based on individual traits, relationship types, and network structures.

The Problem: Beyond Homogeneous Influence

Most existing methods for handling interference assume that every peer’s impact is equal, often reducing influence to a simple count of treated neighbors. In reality, influence is highly contextual. It depends on "ego" traits (e.g., an individual's age), peer traits (e.g., their political affiliation), and the nature of the relationship (e.g., the strength or duration of a friendship). Ignoring this diversity leads to biased results and a failure to identify which subpopulations are most affected by an intervention. This research seeks to automate the discovery of these complex, varying causal effects within relational data.

Methodological Contributions

The research presents a progression of four major methodological innovations designed to capture the nuances of relational causality:

  1. Quasi-Experimental Design for Social Data: A framework for testing causal hypotheses in "flattened" relational data. This approach was used to analyze how policy changes, such as cannabis legalization, affect public opinion over time by aggregating user interactions.
  2. Ensemble Causal Structure Learning: To avoid the risks of relying on a single model with untestable assumptions, this method uses an ensemble of algorithms. It identifies "effect modifiers"—factors that trigger or inhibit outcomes, such as repeat emergency room visits—across different subpopulations, providing a measure of confidence in the discovered causes.
  3. Network Structural Causal Model (NSCM) & NAGG: The core innovation of this work is the NSCM, which extends traditional causal models to account for network-specific complexities like homophily (the tendency to associate with similar people) and latent variables. This is paired with the Network Abstract Ground Graph (NAGG), a tool that allows researchers to reason about network-wide interventions similarly to how Structural Causal Models work for independent data.
  4. Representation Learning: The research utilizes Graph Neural Networks (GNNs) to learn high-dimensional representations of relational contexts. This allows for the estimation of individual effects even when peer influence is highly varied and complex.

Key Insights and Results

Peer Influence as a Variable Context

A central insight of this work is that peer influence is not a monolith. By treating relationships and network properties as Relational Random Variables, the researcher can capture arbitrary contexts that modify how a treatment works. The evaluation of the NSCM framework demonstrated that accounting for this heterogeneity significantly reduces bias in estimating individual treatment effects compared to models that assume homogeneous influence.

Robustness through Ensembles

In practical discovery tasks, such as healthcare, relying on a single causal model is risky. The ensemble approach identified specific preventive factors for ER visits, providing a more robust map of vulnerable subpopulations than traditional single-model methods.

Real-World Policy Impact

The application of these methods to Twitter data successfully characterized how attitudes toward cannabis shifted differently across various states and timeframes following legalization, demonstrating the framework's utility for large-scale social science research.

Limitations and Future Work

While this research advances the field of relational causal inference, several challenges remain. Estimating effects in "large counterfactual neighborhoods"—where a single person may be connected to thousands of others—remains computationally intensive. Future work aims to map complex network representations to human-understandable concepts like "echo chambers" or "polarity." Additionally, the author plans to extend the NSCM/NAGG framework to discover subpopulations with varying "total effects" (the sum of direct and peer effects) in retrospective observational data, which is often noisier than experimental data.

Summary updated

Abstract

Causal inference in relational data should account for the non-IID nature of the data and the interference phenomenon, which occurs when a unit's outcome is influenced by the treatments or outcomes of others. Existing solutions to causal inference under interference consider either homogeneous influence from peers or specific heterogeneous influence contexts (e.g., local neighborhood structure). This thesis investigates causal reasoning in relational data and the automated discovery of heterogeneous causal effects under arbitrary heterogeneous peer influence contexts and effect modification.
BibTeX
@inproceedings{adhikari2024discovering,
  author = {Shishir Adhikari},
  title = {{Discovering Heterogeneous Causal Effects in Relational Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=kixaDtBkH0},
}
Figure