Abstract
Methods for learning causal relationships among a set of foreground variables $X$ based on signals from a (potentially much larger) set of background variables $Z$, which are known non-descendants of $X$. The confounder blanket learner (CBL) uses sparse regression techniques to simultaneously perform many conditional independence tests, with complementary pairs stability selection to guarantee finite sample error control. CBL is sound and complete with respect to a so-called "lazy oracle", and works with both linear and nonlinear systems. For details, see Watson & Silva (2022) doi:10.48550/arXiv.2205.05715.
Introduction
Causal discovery—the task of inferring cause-and-effect relationships from observational data—is a cornerstone of modern statistical inference. However, the presence of unobserved confounders often makes this task notoriously difficult. The R package cbl (Causal Discovery under a Confounder Blanket) provides a robust, theoretically grounded framework for tackling this challenge by leveraging background information to clear the "causal fog."
In many research scenarios, we have a set of "foreground" variables ($X$) whose causal structure we want to map out. We often also have access to a much larger set of "background" variables ($Z$)—such as baseline characteristics, genetic markers, or environmental factors—that are known non-descendants of the foreground variables. The challenge is that these background variables can act as confounders, creating spurious correlations between variables in $X$. Standard causal discovery algorithms often struggle when the number of potential confounders is large. The cbl package addresses this by using the background variables as a "blanket" to perform conditional independence tests, allowing researchers to distinguish between true causal links and mere associations.
The Confounder Blanket Learner (CBL) Algorithm
The core of the package is the Confounder Blanket Learner (CBL) algorithm. Its approach is built on three primary pillars:
- Conditional Independence via Supervised Learning: Instead of traditional statistical tests, CBL treats independence testing as a feature selection problem. If variable $X_i$ is a useful predictor for $X_j$ even after accounting for background variables $Z$, it suggests a potential causal link.
- Sparse Regression & Boosting: The package supports both linear models (via the Lasso) and nonlinear models (via Gradient Boosting/LightGBM). This flexibility allows it to handle complex, real-world data structures.
- Stability Selection: To prevent false positives (errors of commission), CBL employs "complementary pairs stability selection." This technique involves running the model on many subsamples of the data to ensure that the discovered causal links are consistent and not just artifacts of a specific sample.
The primary interface for this is the cbl() function, which accepts foreground and background matrices as inputs. The parameters allow for fine-tuning the selection method and the stability selection threshold.
Implementation and Subroutines
CBL is designed to be computationally tractable even when the background set $Z$ is very large. Unlike some algorithms that try to condition on every possible subset of variables, CBL focuses on the intersection of known non-descendants. This "minimal conditioning" strategy is supported by a feature selection subroutine that maps features and outcomes to a bit vector indicating which features are selected.
Theoretical Foundations and Results
CBL is theoretically "sound and complete" with respect to a "lazy oracle"—a hypothetical entity that only answers independence queries about variable pairs conditioned on their shared non-descendants. This provides a strong mathematical foundation for the algorithm's reliability. Furthermore, by using stability selection, the algorithm provides a bound on the probability of making a wrong causal claim, ensuring finite sample error control.
The output of the cbl function is an ancestrality matrix. This matrix provides nuanced information about the inferred relationships:
- 1: $X_i$ is an ancestor of $X_j$ (a clear causal path).
- 0.5: $X_i$ is a possible ancestor or descendant (the direction is ambiguous).
- 0: The variables are causally disconnected.
- NA: The relationship is unknown.
Conclusion
The cbl package offers a powerful tool for causal discovery in high-dimensional settings. Its most significant requirement is that background variables $Z$ must be known non-descendants of the foreground variables $X$. While the use of stability selection increases computational cost, the package supports parallel processing to mitigate this, making it suitable for modern research applications where background data is plentiful.
Summary updated