Abstract
Unobserved confounding is a central barrier to drawing causal inferences from observational data. Several authors have recently proposed that this barrier can be overcome in the case where one attempts to estimate the effects of multiple causes simultaneously. This paper provides a rigorous "reality check" for this "multi-cause" causal inference hype. Through analytical counterexamples and a new impossibility theorem, we demonstrate that observing multiple causes is not a "silver bullet"; without strict parametric assumptions or extra data like proxy variables, causal effects remain fundamentally unidentified.
Introduction
In classical causal inference, unobserved confounding represents a fundamental challenge. When a hidden variable $U$ influences both a treatment $A$ and an outcome $Y$, it becomes impossible to distinguish the direct causal effect of $A$ on $Y$ from the spurious correlation induced by $U$.
Recently, a "string of work"—most notably the "Deconfounder" framework—suggested that this barrier might be bypassed when studying multiple causes simultaneously. The intuition is that if multiple causes are conditionally independent given a hidden confounder, factor models can "discover" $U$, allowing for adjustment. This paper critically evaluates this claim, asking whether causal effects can truly be identified from observational data alone simply by increasing the number of causes.
The Multi-Cause Framework
The multi-cause setting assumes a structure where an unobserved confounder $U$ affects a set of causes $A^{(1)}, \dots, A^{(m)}$, which in turn affect an outcome $Y$. The confounder $U$ may also have a direct effect on $Y$.
The core idea of recent proposals is that as the number of causes $m$ grows, $U$ can be reconstructed from the observed causes. However, this paper identifies several fundamental theoretical hurdles to this approach.
Theoretical Challenges to Identification
1. Nonparametric Impossibility
We present Theorem 1, a nonparametric impossibility theorem. It states that without specific, often untestable parametric assumptions about the underlying distributions, a single causal effect cannot be pinpointed. Factorization of the causes is not unique; different, equally valid factor models can lead to wildly different, even opposite, causal conclusions.
2. The Positivity Trap
A second challenge arises in the "large-$m$" limit. While having many causes might allow for a more accurate estimation of the hidden confounder $U$, perfect reconstruction of $U$ from $A$ leads to a violation of the "positivity" assumption. If $U$ is a deterministic function of the causes, we cannot observe the outcome for different cause values while holding $U$ constant, making counterfactual reasoning impossible.
Key Results: The Ignorance Region and Regularization Sensitivity
Our analysis reveals that instead of a single identifiable effect, there exists an "ignorance region"—a range of causal effects that are all equally compatible with the observed data.
As shown in the figure above (left), the true causal effect (green squares) lies within a shaded "ignorance region" bounded by the minimum and maximum possible correlations between $Y$ and $U$ given $A$. The observed conditional probability $P(Y|A)$ (purple circles) can be significantly biased.
Furthermore, we demonstrate that in many practical implementations, the reported "causal effect" is highly sensitive to the choice of regularizer. The figure (right) shows how the estimated probability $P(Y=1 | do(A))$ varies dramatically based on the log odds ratio targeted by a penalty, whereas a model incorporating proxy variables remains stable and close to the truth. This suggests that what is often perceived as a "discovery" from data is actually a byproduct of arbitrary mathematical choices made to prevent overfitting.
A Promising Alternative: Proxy Variables
The paper concludes on a constructive note. While multi-cause data alone is insufficient for identification, the problem becomes solvable with the addition of "proxy variables"—variables related to the confounder $U$ but not directly to the outcome $Y$. Integrating these "negative controls" into machine learning pipelines offers a mathematically sound path toward reliable causal identification.
Conclusion
Multi-cause models are currently "oversold" as a general solution to unobserved confounding. Researchers should shift from seeking point identification to performing sensitivity analyses that map out the "ignorance region." Future work should focus on robustly incorporating proxy variables to move beyond the limitations of purely observational multi-cause data.
Summary updated