Bayesian Inference and Partial Identification in Multi-Treatment Causal Inference with Unobserved Confounding

Jiajing Zheng, Alexander D'Amour, Alexander Franks

In the field of causal inference, a primary challenge is the presence of unobserved confounding—factors that influence both the treatment and the outcome but are not recorded in the dataset. This paper, presented at AISTATS 2022, introduces a Bayesian framework designed to navigate the "ignorance region," the range of possible causal effects consistent with observed data when confounding is unmeasured. By employing a "transparent parameterization," the authors demonstrate how to separate empirical data signals from prior scientific assumptions, such as sparsity or negative controls.

The Challenge of Unobserved Confounding

In an ideal scenario, researchers measure every variable influencing both treatments and outcomes. However, real-world data is often incomplete. When unobserved confounders exist, the causal effect is not uniquely identified; even with infinite data, the effect cannot be narrowed to a single value. Instead, we obtain a partial identification region—a set of values that are all equally plausible given the observed data.

While Bayesian methods are naturally suited for expressing this uncertainty through posterior distributions, they are highly sensitive to the choice of prior in partially identified settings. A prior that appears "uninformative" in one coordinate system may inadvertently impose strong, hidden assumptions in another. This research aims to make these assumptions explicit and provide tools to refine the uncertainty region using valid scientific knowledge.

Methodology: Transparent Parameterization

The core of the proposed approach is a "transparent parameterization" strategy that divides model parameters into two distinct groups:

  1. Identified parameters ($\phi$): Quantities that can be estimated directly from the data, such as the observed correlation between treatments and outcomes.
  2. Unidentified parameters ($\lambda$): Sensitivity parameters representing missing information about unobserved confounders.

Key methodological innovations include the use of Partial $R^2$ Parameterization, where researchers reason about the fraction of variance explained by confounding ($R^2_{Y \sim U|T}$) rather than raw coefficients. Additionally, the authors utilize regularized horseshoe priors to encode sparsity assumptions—the idea that most treatments likely have no effect—which helps collapse the uncertainty region toward a single point.

Comparison of different prior parameterizations (Gamma, Null controls, R2, and Sparsity) and their impact on the identification region and variance explained.

Key Insights: The "Uniform Prior" Trap

A significant finding of the paper (Proposition 1) is that seemingly neutral uniform priors on the direction of confounding can act as strong assumptions that confounding bias is zero. In high-dimensional spaces, "random" vectors tend toward orthogonality, which can accidentally wash out bias in the model. The authors show that while confounders are unobserved, assumptions about them are testable. For instance, if a "negative control" (a treatment assumed to have zero effect) is highly correlated with the outcome in a way the model cannot explain, the model fit will suffer, allowing for data-driven validation of causal assumptions.

Empirical Results: Mice Obesity Study

The framework was applied to a real-world dataset involving gene expression levels (37 genes as "treatments") and body weight (outcome) for 227 mice. This benchmark dataset is often used to test multi-treatment causal inference methods.

A standout result was the analysis of the gene Igfbp2. Its negative effect on mouse weight remained robust even in scenarios where unobserved confounders were assumed to explain a significant portion of the outcome variance. This sensitivity analysis allows practitioners to define the bounds of an effect: for example, stating that a gene's effect is significant unless unobserved factors explain more than 50% of the variance.

Sensitivity analysis for the Igfbp2 gene, showing the estimated causal effect across different quantiles and varying levels of confounding variance (R²).

Conclusion and Future Work

The authors provide a robust toolkit for handling the inherent ambiguity of unobserved confounding. While the current method assumes a latent factor model for treatment relationships and relies on linear treatment-confounder relationships, it sets the stage for more complex extensions. Future work may involve formal Bayesian model averaging over the rank of the confounder matrix and expanding the framework to fully non-linear, copula-based models.

Summary updated

Abstract

In causal estimation problems, the parameter of interest is often only partially identified, implying that the parameter cannot be recovered exactly, even with infinite data. Here, we study Bayesian inference for partially identified treatment effects in multi-treatment causal inference problems with unobserved confounding. In principle, inferring the partially identified treatment effects is natural under the Bayesian paradigm, but the results can be highly sensitive to parameterization and prior specification, often in surprising ways. It is thus essential to understand which aspects of the conclusions about treatment effects are driven entirely by the prior specification. We use a so-called transparent parameterization to contextualize the effects of more interpretable scientifically motivated prior specifications on the multiple effects. We demonstrate our analysis in an example quantifying the effects of gene expression levels on mouse obesity.
BibTeX
@inproceedings{zheng2022bayesian,
  author = {Jiajing Zheng and Alexander D'Amour and Alexander Franks},
  title = {{Bayesian Inference and Partial Identification in Multi-Treatment Causal Inference with Unobserved Confounding}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
}
Figure