Elizabeth Ogburn, Ilya Shpitser, and Eric J. Tchetgen Tchetgen
Abstract
This note provides a rigorous theoretical critique of the "deconfounder" method proposed by Wang and Blei (2019). The deconfounder is intended to control for unmeasured confounding by modeling the dependence between multiple causes using a latent variable $Z$. We demonstrate that this method is theoretically unsound, relying on two primary logical errors: (1) the assumption that any variable rendering causes independent is a valid substitute for a confounder, and (2) the conflation of pairwise independence with joint independence. Through synthetic counterexamples, we show that the deconfounder fails to identify causal effects without additional, often unrealistic, parametric assumptions.
Introduction
In the field of causal inference, the "gold standard" for estimating the effect of a treatment is the measurement of all variables that influence both the treatment and the outcome—known as confounders. In "The Blessings of Multiple Causes" (2019), Wang and Blei proposed a provocative alternative: when multiple causes are present (e.g., several different medical treatments), one can use a factor model to identify a latent variable $Z$ that renders these causes independent. They claimed this $Z$ could serve as a "substitute confounder," allowing for valid causal inference even in the presence of unmeasured confounding.
This note serves as a theoretical "reality check" for the deconfounder claim. We argue that the existence of such a variable $Z$ does not guarantee the control of confounding and that the mathematical foundations of the deconfounder are built on flawed logical leaps.
The Logical Failures of the Deconfounder
Our critique identifies two fundamental errors in the original deconfounder theory:
1. The Independence Trap
The deconfounder theory assumes that if causes are independent conditional on some unobserved confounder $U$, and we find a different variable $Z$ that also renders the causes independent, then $Z$ must be a valid substitute for $U$. We show this is a non-sequitur. In probability theory, two different variables can both satisfy an independence condition while being entirely unrelated to one another. Consequently, finding a $Z$ that makes causes independent does not mean $Z$ captures the confounding information necessary to block the bias between the causes and the outcome.
2. Pairwise vs. Joint Independence
A significant oversight in the deconfounder argument is the distinction between pairwise and joint independence. The deconfounder method often relies on showing that causes $A_1$ and $A_2$ are independent given $Z$. However, for causal identification, one requires the causes to be collectively independent of the outcome given $Z$. We demonstrate that pairwise independence between causes does not imply the joint independence required to satisfy the back-door criterion or other identification strategies.
The Deterministic Dilemma
In response to early critiques, Wang and Blei (2020) suggested that $Z$ could be treated as a deterministic function of the causes as the number of causes grows to infinity. We argue that this creates a "logical impossibility." If $Z$ is a deterministic function of the causes, the requirement of "positivity" (or overlap) is violated. Specifically, the conditional expectation of an outcome given the causes and a deterministic function of those same causes becomes mathematically undefined or redundant, rendering the calculation of causal effects impossible without restrictive parametric assumptions.
Theoretical Counterexamples
To prove these points, we constructed two elegant counterexamples using Bernoulli trials:
- Counterexample 1: We define a scenario where causes are independent given a latent variable $Z$, but $Z$ is a constant (a "null" variable). In this case, $Z$ satisfies the deconfounder's requirements but provides zero control over confounding, leaving the causal estimate biased.
- Counterexample 2: We demonstrate a case where even if a latent variable is constructed exactly as the deconfounder prescribes, the joint dependence between the causes and the outcome persists. This confirms that the deconfounder fails to block the necessary paths to identify a true causal effect.
Conclusion
While factor models are valuable tools for dimensionality reduction and pattern recognition, they do not inherently solve the problem of unmeasured confounding. The deconfounder's failure stems from an attempt to derive causal conclusions from purely statistical dependencies without sufficient structural assumptions. We conclude that valid multi-cause causal inference requires moving away from "universal" deconfounders and toward models that reflect specific, well-understood mechanisms of dependence, such as those found in genetics or clustered data environments.
Summary updated