Efficiently Exploiting Structure for Causal Inference in the Presence of Network Interference
Abstract
Causal inference has risen to the forefront in importance within data science, as its focus is to answer the deeper what if and why questions that enable us to predict outcomes under interventions that change the status quo of our society. Consider a new public health policy or proposal for social welfare reform. Before deciding to accept or reject a newly proposed policy, we would like to estimate the potential benefit or risks of adopting the new policy. Unfortunately the classical literature has heavily relied on an assumption that an individual’s outcome only depends on their own assignment to treatment or control, which is referred to as the Stable Unit Treatment Value Assumption (SUTVA). In complex network settings, this assumption is clearly violated, as the network can mediate information flow amongst individuals. Network interference refers to the phenomenon wherein treating individual A may in turn affect individual B’s outcomes. Under network interference, classically used estimators and randomized designs can result in highly misleading and biased estimates for the treatment effect. The proposed work will develop new methods to both efficiently exploit structure in the potential outcomes as well as utilize additional richness of data and flexibility available in the experiment design, most notably the ability to implement an experiment over time and collect intermediate measurements. What has been overlooked in the development of causal inference methods under network interference is that in many applications we have access to data over time, and the randomized experiment itself can be implemented gradually over time, staggering the rollout of the treatment across the treatment group. The results of this research will build a practical theory for causal inference under network interference which can be applied to any network whether it is known or unknown. The solutions can be combined with the rich literature in constructing randomized designs to build a coherent theory alongside practical insights and guides for optimal experimental design and estimation in the presence of network interference. The limited recent work in network interference either have weak results due to imposing no structure on the model, or have fragile results due to relying on strong structural assumptions. We propose a new hierarchy of potential outcomes models which relies on representing the outcomes as a polynomial in the treatment vector. The degree of the polynomial is the natural measure of complexity, and our results extrapolate from simple linear models with degree 1 to fully general nonparametric models when the degree is sufficiently large. We develop new methods that extend classical concepts in causal inference to models with polynomial network interference, providing solutions that are flexibly parameterized by the model complexity. In Thrust 1, we will develop new estimators for the total treatment effect under staggered rollout experiment designs. The key concept we will use is that the additional measurements enable us to reduce estimation to polynomial extrapolation. The proposed research involves designing optimal estimators under a variety of model classes, and extending this framework to time-varying models. In Thrust 2, we will develop techniques for optimal experimental design in the context of structured network interference that exploits additional available information from the network structure, observed covariates, or intermediate measurements. This will involve studying graph matching heuristic for covariate balancing, and optimally combining cluster randomized designs with structure-aware estimators to exploit both graph and model structure. In Thrust 3, we will develop techniques for estimating causal effects from observational studies under network interference.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 29, 2024
- Source ID
- FA95502310301
Entities
People
- Christina Yu
Organizations
- Air Force Office of Scientific Research
- Cornell University
- United States Air Force