Covariate Adjustment Cannot Hurt: Treatment Effect Estimation under Interference with Low-Order Outcome Interactions

1. Problem Statement

The paper addresses the challenge of estimating the Total Treatment Effect (TTE) in randomized experiments where the Stable Unit Treatment Value Assumption (SUTVA) is violated due to interference (i.e., one unit's treatment affects another's outcome).

• Context: In many real-world settings (e.g., social networks, public health interventions), units are connected, and outcomes depend on the treatment assignments of neighbors.
• The Gap: While covariate adjustment (e.g., regression) is a standard technique to reduce variance in SUTVA settings (Lin, 2013), its application under interference is non-trivial. Direct regression adjustment can inadvertently increase variance because it fails to account for the complex covariance structures induced by network interference.
• Goal: Develop a class of covariate-adjusted estimators for TTE under interference that guarantees a "no-harm" property: the asymptotic variance of the adjusted estimator is no larger than that of the unadjusted estimator.

2. Methodology

2.1. Model and Setup

• Interference Model: The authors adopt the Neighborhood Interference Model combined with a Low-Order Interaction Outcome Model (Cortez-Rodriguez et al., 2023).
  • A unit's outcome depends only on its own treatment and the treatments of its neighbors in a known interference network.
  • The potential outcome is modeled as a polynomial of degree $\beta$ (low-order interactions) of the treatment assignments of neighbors.
• Baseline Estimator: The paper builds upon the Structured Neighborhood Interference Polynomial Estimator (SNIPE), denoted as $\hat{\tau}_{unadj}$, which is unbiased for TTE under the low-order interaction model.

2.2. General Covariate-Adjusted Estimator

The authors propose a general class of estimators parameterized by a vector $\theta$:
$$\hat{\tau}(\theta) = \frac{1}{n} \sum_{i=1}^n \omega_i (Y_i - \theta^\top X_i)$$
where:

• $Y_i$ is the observed outcome.
• $X_i$ is the covariate vector (can include network-derived features).
• $\omega_i$ are weights derived from the SNIPE estimator, which are mean-zero ($E[\omega_i]=0$).
• This formulation resembles the Control Variates method. Since $E[\omega_i]=0$, the estimator remains unbiased for any fixed $\theta$.

2.3. Proposed Estimators for $\theta$

The paper investigates two specific strategies for selecting $\theta$:

1. Regression-Based Adjustment ($\hat{\theta}_{Reg}$):

   • Minimizes the approximate variance assuming no interference (weighted least squares).
   • Limitation: The paper demonstrates via a toy example that in the presence of interference, minimizing marginal variance can increase the covariance between units, leading to a net increase in total variance compared to the unadjusted estimator.

2. Variance-Improvement-Maximized (VIM) Adjustment ($\hat{\theta}_{VIM}$):

   • Core Innovation: Instead of minimizing the total variance directly (which is hard to estimate unbiasedly under interference), the authors maximize the estimated variance reduction relative to the unadjusted estimator.
   • They derive an unbiased estimator for the difference in variance, $\Delta(\theta) = \text{Var}(\hat{\tau}(0)) - \text{Var}(\hat{\tau}(\theta))$.
   • $\hat{\theta}_{VIM}$ is chosen to maximize this estimated difference.
   • Key Feature: This approach explicitly incorporates the interference network structure (via cross-unit interaction terms) into the optimization, ensuring that the reduction in variance outweighs any potential increase in covariance.

2.4. Variance Estimation

• The authors develop a conservative variance estimator for the adjusted estimators.
• Unlike existing conservative estimators (e.g., Cortez-Rodriguez et al., 2023) which rely on worst-case bounds and are often overly conservative (leading to wide confidence intervals), the proposed estimator leverages the low-order interaction structure.
• It uses unbiased estimators for second-order terms where possible and applies Cauchy-Schwarz bounds only where necessary, resulting in significantly tighter confidence intervals in finite samples.

3. Key Contributions

1. No-Harm Guarantee: The paper proves that under sparsity conditions on the interference network, the VIM-based estimator ($\hat{\tau}(\hat{\theta}_{VIM})$) is asymptotically unbiased and its asymptotic variance is no larger than the unadjusted SNIPE estimator. This parallels Lin's (2013) result in SUTVA settings but extends it to interference.
2. Optimality: Within the class of estimators parameterized by $\theta$, the VIM estimator is shown to be asymptotically optimal in terms of Mean Squared Error (MSE).
3. Robustness to Covariate Dependence: The framework allows covariates to be arbitrarily dependent on each other and even on the interference network itself, provided they are independent of the treatment assignment.
4. Improved Inference: The proposed variance estimator is asymptotically conservative but empirically much less conservative than previous methods, yielding tighter confidence intervals.

4. Theoretical Results

• Consistency: Both $\hat{\theta}_{Reg}$ and $\hat{\theta}_{VIM}$ are consistent under standard regularity assumptions (boundedness, sparsity of the network, invertibility of design matrices).
• Asymptotic Normality: The adjusted estimators satisfy a Central Limit Theorem (CLT), allowing for valid large-sample inference.
• Variance Bound: Theorem 1 establishes that $n[\text{Var}(\hat{\tau}_{unadj}) - \text{Var}(\hat{\tau}(\hat{\theta}_{VIM}))] \to \theta^{*\top} V_X \theta^* \geq 0$, confirming the no-harm property.

5. Empirical Results (Simulation Study)

The authors conducted extensive simulations across various network models (Erdős–Rényi and Soft Random Geometric Graphs) and interaction orders ($\beta=1, 2$).

• Bias: All proposed estimators ($\hat{\tau}_{unadj}$, $\hat{\tau}(\hat{\theta}_{Reg})$, $\hat{\tau}(\hat{\theta}_{VIM})$) remained unbiased, whereas standard methods like Difference-in-Means (DM) and Lin's estimator (designed for SUTVA) exhibited significant bias under interference.
• Efficiency (MSE):
  • $\hat{\tau}(\hat{\theta}_{VIM})$ consistently outperformed the unadjusted estimator in terms of MSE.
  • In scenarios with strong interference, $\hat{\tau}(\hat{\theta}_{Reg})$ sometimes performed worse than the unadjusted estimator (confirming the theoretical risk of variance inflation), whereas $\hat{\tau}(\hat{\theta}_{VIM})$ never did.
• Confidence Intervals: The proposed variance estimator produced confidence intervals that were substantially narrower (often by orders of magnitude) than those from the conservative estimator of Cortez-Rodriguez et al. (2023), while maintaining valid coverage.

6. Significance

This paper resolves a critical open question in causal inference under interference: Can we safely use covariates to improve precision without risking variance inflation?

• Practical Impact: It provides a "safe" methodology for researchers to incorporate rich covariate information (including network features) into treatment effect estimation, leading to more precise estimates and narrower confidence intervals in real-world networked experiments.
• Theoretical Advancement: It extends the "covariate adjustment cannot hurt" paradigm from the SUTVA world to the interference world, offering a principled framework (VIM) that accounts for the complex dependencies introduced by network interference.
• Generalizability: The authors note that the underlying logic (optimizing variance reduction relative to a baseline) can be applied to other baseline estimators and causal estimands beyond TTE, such as direct and indirect effects.