Abstract
The estimation of causal effects from non-experimental data is a fundamental problem in many fields of science. One of the main obstacles concerns confounding by observed or latent covariates, an issue often addressed by adjusting for a set of variables. This paper introduces a data-driven framework for selecting adjustment sets using conditional independence tests, providing a rigorous way to determine if a treatment variable has a causal effect on an outcome without full knowledge of the underlying causal graph.
Introduction
Measuring causal effects typically requires Randomized Controlled Trials (RCTs) to eliminate the influence of confounders. However, in many real-world scenarios, RCTs are impractical or unethical. Researchers must instead rely on observational data and "adjust" for a set of covariates ($W$). Selecting the correct set is critical: adjusting for too few variables leaves bias, while adjusting for the wrong ones—such as "colliders"—can introduce new biases. This paper addresses this challenge by proposing two elementary rules for data-driven covariate selection that are proven to be both sound and complete.
Methods
The authors utilize Directed Acyclic Graphs (DAGs) and Pearl’s back-door criterion, assuming observed covariates are background conditions not influenced by the treatment ($x$) or outcome ($y$). They introduce two core inference rules based on conditional independence ($⊥\!\!⊥$):
- Rule R1 (Finding an Effect): If a covariate $w$ and a set $Z$ exist such that $w$ is dependent on $y$ given $Z$ ($w \not\u22a5\!\!⊥ y | Z$) but becomes independent once you also condition on $x$ ($w \u22a5\!\!⊥ y | \{x, Z\}$), then $x$ has a causal effect on $y$, and $Z$ is a valid adjustment set.
- Rule R2 (Finding No Effect): If $x$ and $y$ are independent given some set $Z$ ($x \u22a5\!\!⊥ y | Z$), or if there is a $w$ such that $w$ is dependent on $x$ but independent of $y$ given $Z$ ($w \not\u22a5\!\!⊥ x | Z$ and $w \u22a5\!\!⊥ y | Z$), then $x$ has no causal effect on $y$.
If neither rule is triggered, the causal effect is deemed indeterminate ('?').
Experimental Evaluation
The approach was evaluated using synthetic datasets generated from linear Gaussian models across four distinct tasks, illustrated in the causal structures above. These tasks represent scenarios with varying presence of effects and availability of adjustment sets. The simulations included 10 observed covariates and 5 latent variables to create complex confounding patterns, tested across sample sizes ranging from 100 to 10,000, as well as an "independence oracle."
Results and Discussion
The proposed rules proved to be highly reliable. In simulations, the rules were conservative; when they predicted an effect or no effect, they were almost always correct, especially as sample sizes increased.
Compared to the Fast Causal Inference (FCI) algorithm, these specialized rules performed significantly better at identifying indeterminate cases. For instance, in scenarios where an effect exists but no valid adjustment set is observable (Task #2), the rules correctly returned '?' most of the time, whereas FCI often erroneously predicted a zero effect.
Key Insights
- The "Proxy" Intuition: Rule R1 essentially looks for a "test" variable $w$ that sends a signal to $y$ only through $x$. If $x$ blocks that signal, $x$ must be on the path to $y$.
- Avoiding the "Adjust for Everything" Trap: The paper demonstrates that blindly adjusting for all available variables can be disastrous if the causal structure is unfavorable. Their rules provide a surgical alternative.
Limitations and Future Work
The framework relies on the assumption that $x$ and $y$ do not cause any of the covariates in $W$ (the "pre-treatment" assumption). If $W$ contains mediators, the rules would require modification. Additionally, while the theory is nonparametric, the practical tests used in the paper assume linear Gaussian relationships. Future research could explore non-linear, kernel-based independence tests and methods to handle the $O(2^{|W|})$ computational complexity of searching through all possible subsets in high-dimensional data.
Summary updated