Counterfactual (Non-)identifiability of Learned SCMs
Abstract
Recent advances in probabilistic generative modeling have motivated learning Structural Causal Models (SCM) from observational datasets using deep conditional generative models, also known as Deep Structural Causal Model (DSCM). If successful, DSCM s can be utilized for causal estimation tasks, e.g., for answering counterfactual queries ([Pawlowski et al.(2020)Pawlowski, Coelho de Castro, and Glocker]). In this work, we warn practitioners about non-identifiability of counterfactual inference from observational data, even in the absence of unobserved confounding and assuming known causal structure. We prove counterfactual identifiability of monotonic generation mechanisms with single dimensional exogenous variables. For general generation mechanisms with multi-dimensional exogenous variables, we provide an impossibility result for counterfactual identifiability, motivating the need for parametric assumptions. As a practical approach, we propose a method for estimating worst-case errors of learned DSCM s’ counterfactual predictions. The size of this error can be an essential metric for deciding whether or not DSCM s are a viable approach for counterfactual inference in a specific pro
中文速览
用深度生成模型来学习结构因果模型（Structural Causal Model, SCM），再用它回答"反事实"问题（比如"如果当初不吃这药，病人会怎样？"），是近年来一个很热门的方向——但这套方法在理论上到底可不可信，此前几乎没人认真追问过。这篇论文系统地研究了从观测数据中学习深度结构因果模型（Deep SCM, DSCM）时反事实推断的可识别性（identifiability）：对于生成机制满足单调性且外生变量是一维的情形，作者证明了反事实是可识别的；而对于更一般的多维外生变量情形，则给出了不可识别的反例，说明仅凭观测数据无法唯一确定反事实结果。为了让这一问题具有实际操作价值，作者进一步提出了一种估计学到的DSCM在反事实预测上"最坏情况误差"的计算方法，让从业者可以用这个误差上界来判断DSCM在具体问题中是否足够可靠。实验表明，该方法在可识别的SCM上给出了接近零的误差界，在不可识别的合成SCM上也提供了有意义的误差范围，为实际应用中能否信任深度因果模型的反事实输出提供了有据可查的量化依据。
原文 arXiv:2301.09031;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2301.09031v1