Disentangled Instrumental Variables for Causal Inference with Networked Observational Data

TL;DR

This work tackles causal effect estimation in networked observational data plagued by unobserved confounding. It introduces DisIV, a disentangled instrumental variable framework that uses network homogeneity as an inductive bias to isolate individual-specific latent IVs through structural disentanglement, enforcing exogeneity via orthogonality and exclusion constraints. The method employs an asymmetric inference–generation architecture with a confounder proxy learned from graph structure and a latent IV learned from observation features, optimized through a two-stage procedure that separates IV recovery from outcome modeling. Empirical results on semi-synthetic BlogCatalog and Flickr datasets show that DisIV consistently outperforms state-of-the-art baselines, with ablation and disentanglement analyses confirming the necessity and validity of the latent IVs for unbiased ITE and ATE estimation. The approach offers a scalable, principled path to causal inference in networked settings where environmental confounding and neighbor-induced endogeneity are entangled with individual-specific signals.

Abstract

Instrumental variables (IVs) are crucial for addressing unobservable confounders, yet their stringent exogeneity assumptions pose significant challenges in networked data. Existing methods typically rely on modelling neighbour information when recovering IVs, thereby inevitably mixing shared environment-induced endogenous correlations and individual-specific exogenous variation, leading the resulting IVs to inherit dependence on unobserved confounders and to violate exogeneity. To overcome this challenge, we propose $\underline{Dis}$entangled $\underline{I}$nstrumental $\underline{V}$ariables (DisIV) framework, a novel method for causal inference based on networked observational data with latent confounders. DisIV exploits network homogeneity as an inductive bias and employs a structural disentanglement mechanism to extract individual-specific components that serve as latent IVs. The causal validity of the extracted IVs is constrained through explicit orthogonality and exclusion conditions. Extensive semi-synthetic experiments on real-world datasets demonstrate that DisIV consistently outperforms state-of-the-art baselines in causal effect estimation under network-induced confounding.

Figures (4)

• Figure 1: Conceptual diagram illustrating the discovery of latent IVs in networked data. Within the observed feature $X$, neighbour information is entangled with individual-specific information. Our method identifies the individual-specificity information as a latent IV. By structurally disentangling $X$, this latent IV can be extracted to eliminate bias introduced by unobserved confounder $U$ and neighbour information.
• Figure 2: Overall architecture of DisIV. First, DisIV extracts a confounder proxy $\mathbf{e}$ from neighbour information. Subsequently, during the inference phase, the encoder extracts a latent representation $\mathbf{z}$ as a candidate IV; in the generation phase, we condition on $\mathbf{e}$ and $\mathbf{z}$ to reconstruct the observed data. This asymmetric design compels $\mathbf{z}$ to capture solely residual information (i.e., individual specificity) that cannot be explained by $\mathbf{e}$. Furthermore, the constraint term is introduced to further purify $\mathbf{z}$, ensuring it satisfies the definition of an IV.
• Figure 3: Quantitative analysis of latent factor validity on the BC dataset. The GT-IV denotes $\mathbf{Z}_{true}$, and the GT-Conf denotes $\mathbf{C}_{net}$.
• Figure 4: Quantitative analysis of latent factor validity on the Flickr dataset. The GT-IV denotes $\mathbf{Z}_{true}$, and the GT-Conf denotes $\mathbf{C}_{net}$.

Theorems & Definitions (1)

• Definition 1.1: Valid Instrumental Variable