Technical Summary: Causal Inference with Categorical Unobserved Confounder via Mixture Learning

Problem Statement

Estimating causal effects from observational data is fundamentally challenged by unobserved confounding—latent variables that influence both treatment assignment and outcomes. Standard methods like regression or propensity score matching fail when these confounders are unmeasured, leading to biased estimates (e.g., Simpson's paradox).

Existing literature attempts to address this via two primary avenues, both of which face limitations:

1. Proximal Causal Inference: Uses proxy variables (negative controls) to identify causal effects. However, current frameworks (e.g., Miao et al., 2018; Tchetgen Tchetgen et al., 2020) require proxies to be rigidly classified into asymmetric roles: treatment-inducing proxies, outcome-inducing proxies, and common causes. In practice, finding variables that satisfy these specific, asymmetric conditions is often difficult.
2. Multiple Treatments (The "Deconfounder"): Leverages multiple conditionally independent treatments to infer latent structure (Wang and Blei, 2019). While conceptually promising, the identifiability conditions for this approach are debated (Ogburn et al., 2019), and existing estimation methods often rely on heuristic algorithms (e.g., Expectation-Maximization) that lack global convergence guarantees or finite-sample consistency.

This paper addresses the gap by establishing that causal effects are identifiable in both multi-proxy and multi-treatment settings when the unobserved confounder is categorical (taking a finite number of values). The authors propose a framework based on mixture learning and tensor decomposition to recover the latent confounding structure without requiring asymmetric proxy classification or heuristic estimation.

Methodology

The core insight is that when the latent confounder $U$ is categorical and there are at least three conditionally independent views (proxies or treatments), the identification problem reduces to recovering a finite mixture distribution. The authors utilize the method of moments via tensor decomposition to achieve this.

1. Multi-Proxy Setting

The framework assumes a latent confounder $U \in \{1, \dots, K\}$ and $p \ge 3$ proxy variables $Z_1, Z_2, Z_3$ that are conditionally independent given $U$. Unlike prior work, these proxies do not need to be classified into specific causal roles; they are treated symmetrically.

• Identification: Under assumptions of conditional independence and linear independence of the conditional measures, the parameters of the mixture distribution (mixing weights $\pi_u$ and component densities $P_{v,u}$) are strictly identifiable up to permutation (Allman et al., 2009).
• Estimation Procedure:
  1. Unsupervised Tensor Learning: Apply robust tensor decomposition (Anandkumar et al., 2014) to the joint distribution of the proxies to estimate the mixing weights $\hat{\pi}$ and component densities $\hat{f}_{v,u}$.
  2. Posterior Recovery: Compute the posterior probability of the latent state for each unit: $\hat{w}_u(Z) = P(U=u|Z)$.
  3. Treatment Model Estimation: Regress the treatment $A$ on features weighted by the recovered posteriors to estimate cluster-specific treatment parameters.
  4. Outcome Model Estimation: Update the posteriors to account for the treatment assignment ($P(U=u|A, Z)$) and regress the outcome $Y$ on features weighted by these updated posteriors to estimate cluster-specific outcome parameters.
  5. Causal Effect Calculation: Compute the Average Treatment Effect (ATE) and Conditional Average Treatment Effect (CATE) by integrating over the recovered latent structure.

2. Multi-Treatment Setting

The framework extends to scenarios with no proxies but multiple (at least three) conditionally independent treatments $A_1, A_2, A_3$.

• Identification: Similar to the proxy setting, the joint distribution of treatments allows for the unique identification of the latent mixture parameters.
• Estimation: A tensor decomposition is applied directly to the treatment variables to recover the latent prior and emission probabilities. These are then used to construct weighted features for a final regression on the outcome to estimate causal parameters.

3. Theoretical Guarantees

The paper provides non-asymptotic error bounds for the entire estimation pipeline.

• Finite-Sample Consistency: The authors derive error bounds for the tensor decomposition step (estimating posteriors) and propagate these errors through the subsequent regression stages.
• Convergence: The proposed algorithm achieves finite-sample consistency for the causal parameter (ATE), with an error bound of $O(\log(\delta) \cdot n^{-\frac{b}{2b+7d}})$, where $b$ is a smoothness parameter and $d$ is the dimension of the proxies.
• Global Convergence: Unlike EM algorithms, the tensor power method used ensures global convergence under mild non-degeneracy conditions.

Key Contributions

1. Relaxed Proxy Assumptions: The paper introduces a multi-proxy causal framework that does not require the rigid classification of proxies into negative control exposures or outcomes. It only requires a collection of three or more conditionally independent proxies.
2. Identifiability and Estimation: The authors establish the identifiability of model parameters and causal estimands (ATE, CATE) for both multi-proxy and multi-treatment settings under the assumption of a categorical latent confounder.
3. Algorithmic Stability: They propose an efficient estimation procedure based on tensor decomposition that offers global convergence guarantees, avoiding the local optima issues common in variational inference or EM.
4. Finite-Sample Theory: The paper provides rigorous finite-sample error bounds for the nonparametric estimation of latent posterior distributions and the subsequent causal parameters, culminating in a consistency guarantee for the ATE.
5. Non-Parametric Generalization: The identifiability results are extended to a broader non-parametric setting (Theorem 5.2), though the specific algorithmic implementation for the fully non-parametric case is noted as future work.

Experimental Results

The authors validate their method through simulations and real-world data analysis:

• Simulations: In both multi-proxy and multi-treatment settings, the proposed estimator demonstrates convergence to the true causal coefficients as sample size increases, performing well even with limited data.
• Real Data (BWGHT Dataset): The method was applied to estimate the effect of maternal smoking on infant birth weight. The unobserved confounder was identified as latent family socio-economic status.
  • Using three proxies (family income, father's education, mother's education), the tensor decomposition recovered a trivalent latent structure ($K=3$).
  • The resulting clusters aligned with epidemiological consensus (high-risk disadvantaged, discordant middle-tier, and highly advantaged protective classes).
  • The estimated causal effect of smoking was negative across all latent strata, reinforcing prior findings on smoking-induced fetal growth restriction.

Significance and Claims

The paper claims to establish a "concrete bridge between mixture learning and causal inference." By reframing the problem of unobserved categorical confounding as a latent mixture identification problem, the authors avoid the ill-posed inverse problems typical of integral equation approaches used in proximal causal inference.

The significance of this work lies in:

• Practicality: It removes the difficult requirement of finding asymmetric negative controls, making proximal causal inference more applicable to domains where such variables are hard to identify.
• Theoretical Rigor: It provides the first consistent estimation method for the multi-treatment setting with global convergence guarantees and finite-sample error bounds, addressing the theoretical gaps in the "Deconfounder" literature.
• Methodological Innovation: It demonstrates that tensor decomposition can be effectively leveraged to "unmix" latent confounder distributions, enabling consistent causal recovery in settings previously considered intractable without strong parametric assumptions or untestable identification conditions.

The authors remain modest regarding future directions, noting that while their current results treat proxy distributions non-parametrically, the estimation algorithm assumes parametric forms for the treatment and outcome models. Generalizing the framework to a fully non-parametric setting for estimation is identified as a natural extension for future research.