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

Jiajing Zheng, Alexander D’Amour, Alexander Franks

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.

Introduction

In an ideal causal inference scenario, researchers observe every variable that influences both the treatment and the outcome. However, in real-world observational studies, "unobserved confounding" is the rule rather than the exception. When key confounders are missing, the causal effect is partially identified: even with an infinite amount of data, we can only narrow the effect down to a range of possible values, known as the "ignorance region," rather than a single point estimate.

While Bayesian methods are naturally equipped to handle uncertainty, they face a unique challenge in partially identified models. A researcher might select a seemingly "neutral" prior, only to find that the resulting posterior concentrates on a specific value. This concentration often occurs not because the data supports that value, but because the prior accidentally imposed strong, invisible assumptions on the unidentified portions of the model. This paper introduces a "transparent parameterization" framework to bridge the gap between the mathematical rigor of causal inference and the practical intuition of scientific researchers.

The Framework: Transparent Parameterization

To address the sensitivity of Bayesian results to prior specifications, the authors propose splitting model parameters into two distinct categories:

  1. Identifiable Parameters ($\phi$): These are the components that the data can directly inform, such as the observed correlations between treatments and outcomes.
  2. Unidentified or Sensitivity Parameters ($\lambda$): These represent the relationships the data cannot reveal, such as the strength of the link between unobserved confounders and the outcome.

By separating these parameters, researchers can conduct a formal sensitivity analysis. This allows them to observe how the causal range shifts as they vary their assumptions about the unobserved variables. The framework incorporates several key techniques to refine these estimates:

  • Partial R-Squared ($R^2$): This allows researchers to reason about the fraction of outcome variance explained by unobserved confounders.
  • Negative Controls: By identifying treatments known to have zero causal effect, researchers can "anchor" the model, significantly shrinking the ignorance region.
  • Sparsity (Horseshoe Priors): In high-dimensional settings with many potential causes, the model assumes that only a few treatments are likely to have a significant impact.

The Multi-Treatment "Blessing"

While unobserved confounding is typically viewed as a hurdle, the presence of multiple treatments provides a unique leverage point. The dependencies between different treatments can be used to "learn" the underlying structure of unobserved confounders. The authors demonstrate that while we cannot see the confounders directly, the collective behavior of multiple treatments reveals the latent dimensions that influence the system.

However, the authors also warn of the "Uniform Prior Trap." They prove that a uniform prior on the direction of confounding—which might seem objective—actually acts as a very strong assumption. As the number of confounders increases, this prior tends to force the estimated bias toward zero, potentially leading to overconfidence in a "naive" result that ignores confounding altogether.

Experimental Validation

Simulated Data

The framework was first tested using a synthetic environment with 1,000 observations, 10 treatments, and 2 unobserved confounders. This controlled setting allowed the authors to compare the "ground truth" causal effects against various prior specifications, confirming that the transparent parameterization accurately reflected the bounds of the ignorance region.

Mouse Obesity Study

The authors applied their method to a classic benchmark dataset consisting of 227 mice. In this study, the "treatments" are the expression levels of 37 different genes, and the "outcome" is body weight.

Key findings from this analysis include:

  • Robustness of Igfbp2: The gene Igfbp2 was found to have a significant effect on weight that remained robust even under the assumption that unobserved confounders explained up to 50% of the residual variance.
  • Evaluating Assumptions: The authors found that strict "negative control" assumptions used in previous literature actually resulted in a significantly worse model fit. This highlights the framework's ability to flag when a researcher's prior intuition contradicts the observed data.

Conclusion and Future Directions

This research provides a vital tool for making Bayesian causal inference more interpretable and honest. By distinguishing between what is learned from data and what is assumed via priors, researchers can provide more reliable intervals for causal effects.

Currently, the framework relies on Gaussian (normal) distributions and assumes linear relationships between treatments, confounders, and outcomes. Future work may explore relaxing these assumptions to accommodate non-linear interactions and non-Gaussian data. Additionally, further research into automating the selection of the "latent rank" (the number of unobserved confounders) would further enhance the model's utility in complex biological and social science applications.

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