What You Observe Determines How You Identify Causal Effects: Evaluating Causal Models across Observational Views

2026-09-29

AI Summary

This study addresses the limitation that performance evaluations of existing causal foundation models are confounded by observational perspectives, lacking a fair comparison benchmark. We propose CausalIDView, a multi-perspective benchmark that enables controlled experiments by fixing the structural causal model and target estimand while varying only the observational perspective. Furthermore, it adopts a modular approach that decouples predictive tabular foundation models from specific identification mechanisms, facilitating equitable comparisons across different identification strategies. Our experiments reveal significant instability in model rankings across perspectives, with no single model consistently outperforming others. Additionally, we demonstrate that this modular approach surpasses several existing causal foundation models in performance, establishing a new paradigm for the robustness evaluation of causal models.

Abstract

Causal foundation models (CFMs) pre-trained on data generated from various structural causal models (SCMs) have been proposed for estimating causal effects from observational data. However, differences in pre-training environments and evaluation protocols make it difficult to assess how their performance depends on the information available for causal identification. To enable controlled comparisons, we introduce CausalIDView, a multi-view benchmark that holds fixed SCM realization and target estimand while varying only the observational view available to the estimator. Each observational view corresponds to a distinct identification regime under the benchmark's maintained causal assumptions. Across these matched views, no CFM consistently performs best and model rankings vary substantially. Under controlled structural changes, CFMs exhibit model-specific failures to maintain stable estimates when true effects are unchanged and to track genuine effect changes. We also examine whether combining explicit identification with strong predictive estimation is effective. A modular approach that pairs a predictive tabular foundation model with regime-specific identification procedures is competitive with CFMs and outperforms several of them. These findings motivate cross-regime comparisons to assess the empirical value of CFMs.

Problem

• causal foundation models
• observational views
• causal identification
• benchmark evaluation
• structural causal models

Innovation

• Causal foundation models
• Multi-view benchmark
• Structural causal models
• Causal identification
• Modular approach