Estimation in Additive Exposure Models

• Activity Number: 80 - Graphical Models and Causal Inference
• Type: Contributed
• Date/Time: Sunday, July 28, 2019 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Kelly Kung* and Daniel L Sussman
• Companies: Boston University and Boston University
• Keywords: causal inference; exposure; additivity; interference ; network; treatment effect

Abstract

In network settings, it can be difficult to distinguish between direct and indirect treatment effects. However, exposure models capture the relevant treatment components that affect outcomes through a unit's exposure level, enabling estimation of these effects. In this paper, we present monotonic atomic estimators that form an affine space of unbiased estimators for causal effects in additive exposure models. These estimators are atomic because they require a minimal set of exposure levels, and monotonicity ensures they form a basis. Additive exposure models isolate treatment components relevant to each effect and increase the support of unbiased estimators to more treatment allocations. We use these monotonic atomic estimators to derive minimum integrated variance linear unbiased estimators (MIVLUE) of the interference effect, and we study these estimators in a social network data.