A Bayesian Approach for Estimating Causal Effects from Observational Data

Johan Pensar, Topi Talvitie, Antti Hyttinen, Mikko Koivisto

Introduction

Estimating causal effects from observational data is a fundamental challenge in science. When the underlying causal structure—represented as a Directed Acyclic Graph (DAG)—is known, techniques like back-door adjustment provide a clear path to calculating the effect of an intervention. However, in most real-world scenarios, the true DAG is unknown.

Existing frequentist methods, such as IDA (Iterative Causal Discovery), typically attempt to identify a single "best" graph or a set of equivalent graphs (a CPDAG) to estimate effects. This approach often fails to account for structural uncertainty; when data is noisy or sparse, many different graphs may fit the data equally well while implying vastly different causal conclusions. This paper introduces BIDA, a fully Bayesian framework that integrates structural uncertainty directly into the causal estimate.

Methodology: The BIDA Framework

BIDA treats the causal effect not as a single point estimate, but as a posterior distribution. This is achieved through Bayesian Model Averaging (BMA), where the causal effect is averaged across all possible DAGs, weighted by their posterior probability given the data.

The $O(3^d d)$ Algorithm

Summing over all possible DAGs is computationally prohibitive, as the number of DAGs grows super-exponentially with the number of variables $d$. The authors overcome this by proving that for linear Gaussian models, the causal effect of a variable $X$ on $Y$ depends only on the parent set of $X$. By focusing on the posterior distribution of parent sets rather than full graphs, they developed a dynamic programming algorithm with $O(3^d d)$ complexity. This allows for exact Bayesian inference in systems with up to approximately 20 variables.

The Mixture Posterior

A key insight of the Bayesian approach is the resulting "mixture distribution" of the causal effect. Part of the probability mass is concentrated at exactly zero—representing the probability that $X$ is not an ancestor of $Y$—while the remainder forms a continuous distribution representing the magnitude of the effect if one exists.

Posterior distributions of causal effects across different cases and sample sizes. As $n$ increases, the distribution concentrates around the true effect (dotted line).

Empirical Evaluation

Synthetic Benchmarks

The method was tested on synthetic datasets with 20 nodes and varying sample sizes ($n=50, 200, 800$). BIDA consistently achieved lower Mean Squared Error (MSE) than various IDA benchmarks. The performance gap was most pronounced at small sample sizes ($n=50$), where frequentist methods struggle to identify a stable graph structure, whereas BIDA's averaging approach effectively manages the high structural uncertainty.

Comparison of Mean Squared Error (MSE) for BIDA and various IDA benchmarks across different sample sizes. BIDA shows significantly lower error, especially in low-data regimes.

Real-World Validation: Sachs Dataset

To validate BIDA on real-world data, the authors applied it to the Sachs flow cytometry dataset, a gold standard in causal discovery involving 11 protein measurements. BIDA identified 13 high-confidence causal pairs, 7 of which matched established biological consensus. Other identified pairs were supported by modern methods like Invariant Causal Prediction (ICP), demonstrating BIDA's reliability on noisy biological signaling data.

Conclusion and Future Work

BIDA represents a significant step forward in causal inference by providing a rigorous way to handle structural uncertainty. While the $O(3^d)$ complexity is a breakthrough for exact inference, it remains limited to systems of roughly 20–25 variables. Future research will focus on scaling the method using approximate inference (such as MCMC), relaxing the assumption of causal sufficiency to account for latent confounders, and extending the framework to non-linear relationships.

Summary updated

Abstract

We present a novel Bayesian method for the challenging task of estimating causal effects from passively observed data when the underlying causal DAG structure is unknown. To rigorously capture the inherent uncertainty associated with the estimate, our method builds a Bayesian posterior distribution of the linear causal effect, by integrating Bayesian linear regression and averaging over DAGs. For computing the exact posterior for all cause-effect variable pairs, we give an algorithm that runs in time O(3d d) for d variables, being feasible up to 20 variables. We also give a variant that computes the posterior probabilities of all pairwise ancestor relations within the same time complexity, significantly improving the fastest previous algorithm. In simulations, our Bayesian method outperforms previous methods in estimation accuracy, especially for small sample sizes. We further show that our method for effect estimation is well-adapted for detecting strong causal effects markedly deviating from zero, while our variant for computing posteriors of ancestor relations is the method of choice for detecting the mere existence of a causal relation. Finally, we apply our method on observational flow cytometry data, detecting several causal relations that concur with previous findings from experimental data.
BibTeX
@inproceedings{pensar2020bayesian,
  author = {Johan Pensar and Topi Talvitie and Antti Hyttinen and Mikko Koivisto},
  title = {{A Bayesian Approach for Estimating Causal Effects from Observational Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.5988},
}
Figure