Technical Summary: Causal Inference for Sequential Settings under Interference and Latent Confounding

Problem Statement

This work addresses the challenge of causal inference in sequential, observational settings where the Stable Unit Treatment Value Assumption (SUTVA) is violated due to outcome interference and the presence of latent confounding.

In many real-world scenarios, such as the spread of infectious diseases or the adoption of vaccines, the outcome of a unit (e.g., an individual or county) depends not only on its own treatment and characteristics but also on the outcomes, treatments, and characteristics of other units in the population. Furthermore, these relationships are often observed over time, introducing temporal dependencies. Existing literature typically handles interference or sequential settings separately, or assumes confounders are observed. This paper tackles the complex setting where:

1. Interference: Outcomes are interdependent across units (spatial dependence).
2. Sequentiality: Outcomes evolve over time steps with dependencies between adjacent time steps.
3. Latent Confounding: Unobserved variables influence both treatments and outcomes, modeled with a low-rank structure similar to panel data factor models.
4. Data Scarcity: The available data is a single sample (one realization) from this high-dimensional joint distribution of $N$ units over $T$ time steps.

The goal is to estimate causal quantities of interest, specifically the Generalized Treatment Effect (GTE), which compares average outcomes under two arbitrary intervention patterns, despite having only one observed trajectory.

Methodology

Model Formulation

The authors model the system as a Markovian process over $T$ time steps for $N$ units. The binary outcome $x_i^{(t)} \in \{-1, 1\}$ for unit $i$ at time $t$ is governed by an Ising model conditional on the previous time step's outcomes, the current interventions, and latent confounders.

The probability distribution is defined as:
$$p(x | z) = \prod_{t=1}^T p(x^{(t)} | z^{(t)}, x^{(t-1)})$$
where the conditional probability is proportional to:
$$\exp\left( \sum_{i=1}^N \alpha_i^{(t)} x_i^{(t)} + \beta \sum_{i=1}^N x_i^{(t)} z_i^{(t)} + \xi \sum_{i \neq j} \gamma_{ij} x_i^{(t)} x_j^{(t)} + \eta \sum_{i=1}^N x_i^{(t)} x_i^{(t-1)} \right)$$

Key parameters include:

• $\alpha_i^{(t)}$: Latent confounders (external field).
• $\beta$: Direct effect of treatment.
• $\xi$: Strength of spatial interference (interaction matrix $\Gamma$).
• $\eta$: Strength of temporal dependence.

Key Assumptions:

1. Low-Rank Latent Structure: The matrix of latent confounders $A = \{\alpha_i^{(t)}\}$ has a rank $k \ll N, T$, expressible as $A = UV^\top$.
2. Bounded Interaction: The interaction strength satisfies Dobrushin's uniqueness condition ($\|\xi^* \Gamma\|_\infty < 1$), ensuring correlation decay and efficient inference.
3. Excitability of Interventions: The intervention matrix $Z$ is not perfectly collinear with the latent confounders, ensuring identifiability of the treatment effect $\beta$.

Learning Algorithm: Sequential MPLE

To learn the parameters $\theta = (A, \beta, \xi, \eta)$ from a single sample, the authors propose a Sequential Maximum Pseudo-Likelihood Estimation (MPLE) approach.

• Objective: Minimize the negative log pseudo-likelihood:
  $$\hat{\theta} \in \arg\min_{\theta \in \Theta} -\sum_{t=1}^T \sum_{i=1}^N \log p_\theta(x_i^{(t)} | x_{-i}^{(t)}, z_i^{(t)}, x_i^{(t-1)})$$
• Optimization: The objective is convex in $\theta$ if $A$ were fixed, but the low-rank constraint on $A$ makes the problem non-convex. The authors parameterize $A = UV^\top$ and use alternating minimization (fixing $U$ to update $V$, and vice versa) with regularization.
• Inference: Once parameters are estimated, causal quantities (GTE) are estimated by sampling from the learned model using Sequential Gibbs Sampling. Due to the Dobrushin condition, the sampler mixes rapidly, allowing for efficient computation of counterfactual expectations.

Key Contributions

1. Unified Model: The paper introduces a model that simultaneously captures sequential dynamics, network interference (outcome interference), and low-rank latent confounding.
2. Computationally Efficient Estimation: It provides a computationally tractable method (Sequential MPLE) to learn high-dimensional model parameters from a single observation of the joint distribution.
3. Non-Asymptotic Guarantees:
   • Parameter Estimation: Theorem 1 establishes non-asymptotic consistency for parameter estimation. The error bound scales as $O\left(\frac{k(N+T)\log T}{T\|\Gamma\|_F^2}\right)$, showing that the effective sample size is driven by the interaction strength and the number of time steps.
   • Causal Estimation: Theorem 2 and Corollary 2 show that the error in estimating the Generalized Treatment Effect (GTE) is bounded by the parameter estimation error. This relies on correlation decay properties under Dobrushin's condition, ensuring that local estimation errors do not propagate spatially or temporally to corrupt the global causal estimate.
4. Empirical Validation: The method is validated on synthetic data and a real-world case study involving COVID-19 vaccination rates and death rates across US counties.

Results

Synthetic Experiments

In synthetic settings with $N=500$ and $T=50$, the proposed method ($\hat{\theta}$) significantly outperforms baselines that ignore interference ($\hat{\theta}_{\xi=0}$) or latent confounding ($\hat{\theta}_{A=0}$).

• Parameter Recovery: The method accurately recovers the treatment effect $\beta$, interaction strength $\xi$, and temporal strength $\eta$.
• Causal Accuracy: The estimated GTE was improved by approximately 92% in absolute error compared to the no-interference baseline and 91% compared to the no-latent-confounding baseline.
• Robustness: The method successfully recovered parameters even when the theoretical Dobrushin condition was slightly violated ($|\xi^*| + |\eta^*| > 1$), suggesting practical robustness.

Real-World Case Study: COVID-19

The authors applied the model to county-level data in the US (March 2020 – May 2022) to estimate the causal effect of vaccination on death rates.

• Setup: Outcomes were binary indicators of high death rates; interventions were binary indicators of vaccination rates exceeding 30%.
• Hybrid Validation: To address potential violations of identifiability assumptions in the real data, "hybrid experiments" were conducted where the real intervention pattern was used, but outcomes were generated from the model. The method successfully identified parameters in this setting.
• Test Set Recovery: The model was trained on a subset of data and tested on a held-out test set constructed to respect spatio-temporal dependencies. The model achieved low absolute error (0.015 on the test set), comparable to the hybrid case, demonstrating its expressiveness.
• Causal Findings:
  • Modeling interference yielded a larger estimated causal effect of vaccination compared to the no-interference model.
  • Under the "no interventions" counterfactual, the model predicted worse outcomes than observed initially, while the "all interventions" counterfactual predicted better outcomes, converging as vaccination rates rose.
  • The interference model captured the "spill-over" effects where vaccination in one area benefits others, a nuance missed by standard models.

Significance and Claims

The paper claims to offer a provable framework for causal inference in a setting where three major complexities coexist: network interference, intertemporal dependencies, and latent confounding.

• Theoretical Contribution: It extends the literature on learning high-dimensional distributions from single samples (previously limited to non-sequential or non-confounded settings) to the sequential, latent-variable case. It provides the first non-asymptotic guarantees for causal estimation in this specific, highly complex regime.
• Practical Utility: By demonstrating efficacy on a real-world public health problem (COVID-19), the authors argue that their method can inform policy decisions by providing more accurate estimates of treatment effects in the presence of interference and unobserved confounders.
• Limitations: The authors modestly acknowledge that their theoretical guarantees rely on Dobrushin's uniqueness condition, which may not hold in all real-world scenarios with strong interference. They note that while their method performed well in experiments, understanding performance in regimes violating this condition remains an open area.

In summary, this work bridges the gap between theoretical causal inference and complex, real-world sequential data, providing a method that is both computationally efficient and theoretically grounded for estimating causal effects under interference and latent confounding.