[Paper Review] Using Exposure Mappings as Side Information in Experiments with Interference

David Choi|arXiv (Cornell University)|Jun 28, 2018

Advanced Causal Inference Techniques22 references4 citations

TL;DR

This paper proposes robust causal inference methods for experiments with interference by using exposure mappings as side information, enabling valid confidence intervals even under severe exposure model misspecification. It introduces a nonnegative treatment effect assumption for conservative inference and a weaker 'contrasts attributable to treatment' estimand that requires no structural assumptions on interference.

ABSTRACT

Exposure mappings are widely used to model potential outcomes in the presence of interference, where each unit's outcome may depend not only on its own treatment, but also on the treatment of other units as well. However, in practice these models may be only a crude proxy for social dynamics. In this work, we give estimands and estimators that are robust to the misspecification of an exposure model. In the first part, we require the treatment effect to be nonnegative (or "monotone") in both direct effects and spillovers. In the second part, we consider a weaker estimand ("contrasts attributable to treatment") which makes no restrictions on the interference at all.

Motivation & Objective

• Address the challenge of interference in randomized experiments where units' outcomes depend on others' treatments, violating standard ignorability assumptions.
• Overcome the limitation of existing exposure mapping methods that require correct model specification, which may be unrealistic in practice.
• Develop inference procedures that remain valid even when exposure models are misspecified, ensuring robustness in real-world networked experiments.
• Introduce a new estimand, 'contrasts attributable to treatment' (CAT), that avoids structural assumptions on interference and enables inference from randomization alone.
• Provide a framework for conservative inference under minimal assumptions, offering a fallback for settings where modeling peer effects is highly uncertain or contentious.

Proposed method

• Propose a new estimand, τ^ZCAT, based on contrasts between potential outcomes under treatment and control, which is robust to exposure model misspecification.
• Use randomization-based inference to derive confidence intervals for τ^ZCAT without requiring correct exposure model specification.
• Introduce a nonnegativity assumption on treatment effects (direct and spillover) to ensure conservative inference under mild regularity conditions.
• Leverage the largest eigenvalue of a transformed design matrix to bound the variance of the test statistic, enabling finite-sample confidence regions.
• Apply the method to both simulated and real-world data (e.g., Facebook voting experiment) to demonstrate empirical validity.
• Use Bernoulli randomization to assign treatments and derive asymptotic concentration bounds for the estimand under weak dependence assumptions.

Experimental results

Research questions

• RQ1 How can we perform valid causal inference in network experiments when the exposure model is misspecified?
• RQ2 What estimand can be used to estimate treatment effects under interference without assuming a specific interference mechanism?
• RQ3 Can we construct conservative confidence intervals for treatment effects using only randomization and exposure mappings, without relying on correct exposure model specification?
• RQ4 What are the finite-sample properties of the proposed estimand under general interference structures?
• RQ5 How does the proposed method compare to traditional exposure model-based approaches in terms of robustness and conservativeness?

Key findings

• The proposed estimand τ^ZCAT yields a 95% confidence interval of [2.06%, 2.26%] for the Facebook voting experiment, indicating a significant treatment effect with no assumptions on interference.
• Under the nonnegativity assumption, the method produces conservative confidence intervals that remain valid even if the exposure model is misspecified.
• The method ensures that the true treatment effect lies within the confidence interval with probability converging to at least 1−α, even under complex interference patterns.
• The largest eigenvalue of the design matrix plays a key role in bounding the variance of the test statistic, enabling finite-sample inference.
• The CAT estimand allows inference without modeling the full interference mechanism, making it suitable for settings with unknown or heterogeneous peer effects.
• The method provides a fallback inference strategy when subjective modeling assumptions (e.g., exposure models) are controversial or unverifiable.