[Paper Review] Copula-based Sensitivity Analysis for Multi-Treatment Causal Inference with Unobserved Confounding

Jiajing Zheng, Alexander D’Amour+1 more|arXiv (Cornell University)|Feb 18, 2021

Advanced Causal Inference Techniques50 references4 citations

TL;DR

This paper proposes a copula-based sensitivity analysis framework for multi-treatment causal inference under unobserved confounding, using Gaussian copulas to model confounder-outcome relationships and bound causal effects even when point identification is impossible. The method provides robust, calibrated bounds on treatment effects that remain informative under plausible latent variable models.

ABSTRACT

Recent work has focused on the potential and pitfalls of causal identification in observational studies with multiple simultaneous treatments. Building on previous work, we show that even if the conditional distribution of unmeasured confounders given treatments were known exactly, the causal effects would not in general be identifiable, although they may be partially identified. Given these results, we propose a sensitivity analysis method for characterizing the effects of potential unmeasured confounding, tailored to the multiple treatment setting, that can be used to characterize a range of causal effects that are compatible with the observed data. Our method is based on a copula factorization of the joint distribution of outcomes, treatments, and confounders, and can be layered on top of arbitrary observed data models. We propose a practical implementation of this approach making use of the Gaussian copula, and establish conditions under which causal effects can be bounded. We also describe approaches for reasoning about effects, including calibrating sensitivity parameters, quantifying robustness of effect estimates, and selecting models that are most consistent with prior hypotheses.

Motivation & Objective

• To address the challenge of causal inference in multi-treatment settings where unobserved confounders prevent point identification.
• To develop a sensitivity analysis framework that quantifies the range of causal effects compatible with observed data under plausible assumptions about latent confounders.
• To provide practical tools for calibrating sensitivity parameters, assessing robustness, and selecting models consistent with prior hypotheses.
• To demonstrate that even without full identification, latent variable models can sharpen causal conclusions via bounded effect estimation.

Proposed method

• Uses a copula factorization to model the joint distribution of outcomes, treatments, and unmeasured confounders, decoupling the confounder-outcome dependence from the treatment model.
• Applies Gaussian copulas to characterize the confounder-outcome relationship, enabling flexible and tractable sensitivity analysis without altering model fit.
• Imposes a latent factor model on treatments to represent shared unobserved confounding, with the number of factors determined by data-driven criteria.
• Derives bounds on causal effects under the assumption that the latent variable model for treatments is identifiable, ensuring finite bounds even when effects are otherwise unbounded.
• Calibrates sensitivity parameters using prior knowledge or negative control exposures to assess robustness of effect estimates.
• Employs Bayesian additive regression trees (BART) for outcome modeling and posterior inference, enabling uncertainty quantification in the presence of confounding.

Experimental results

Research questions

• RQ1 Can causal effects be meaningfully bounded in multi-treatment settings when unobserved confounding prevents point identification?
• RQ2 How can sensitivity analysis be adapted to leverage latent variable structures in multi-treatment causal inference?
• RQ3 What role do copulas play in modeling confounder-outcome dependence without restricting the outcome model?
• RQ4 How can sensitivity parameters be meaningfully calibrated and interpreted in the context of shared unobserved confounders?
• RQ5 To what extent do latent variable models improve the precision and robustness of causal effect estimates in the absence of full identification?

Key findings

• Causal effects remain unbounded under unrestricted sensitivity models, but become bounded when the latent variable model for treatments is identifiable.
• The Gaussian copula specification ensures that bounds on treatment effects are finite and interpretable, even when full identification is unattainable.
• In a gene expression analysis, Igfbp2 showed a significant negative causal effect on mouse weight only at expression levels above the 75th quantile, and this effect was robust to confounding explaining up to 100% of residual outcome variance.
• For expression levels below the 75th quantile, no significant causal effect was detected, even under no confounding, indicating no robust effect at lower expression levels.
• The method successfully bounds causal effects across multiple treatments and quantifies robustness, with results showing that high Igfbp2 expression reduces mouse weight regardless of confounding strength.
• The framework enables practical sensitivity analysis that is both statistically rigorous and interpretable, with code and an R package publicly available for replication and application.