Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Identification and estimation of interference effects with contextual multilevel models
Authors: Yi Feng, Peter Steiner,
Presenting Author: Yi Feng*
Interference (spill-over) effects play an important role in the social, behavioral, and health sciences. While different types of causal interference effects can be clearly defined in terms of potential outcomes and interference graphs, their estimation remains a considerable challenge and generally employable analytic tools are still not available (Hong, 2015; Hudgens & Halloran, 2008; VanderWeele, 2015). To advance applied research on interference effects, we start by considering contextual multilevel models (MLM, mixed effects models) where the average treatment exposure of multiple subjects within groups (group-level means) is used as an additional predictor to assess contextual treatment effects (Enders, 2013). Using different data-generating interference models (compositional, direct interference, contagion; Ogburn & VanderWeele, 2014) we investigate whether contextual MLMs are able to identify average causal interference effects. When interference effects are caused by the groups’ mean exposure, the MLM estimator identifies the average interference effect. However, even when spill-over effects are caused by direct interference or via contagion, which is more likely in practice than compositional effects transmitted via group means, we show that the corresponding interference effects can be recovered from MLM estimates. Theoretical identification results using causal graphs will be presented and discussed, including extensions to heterogeneous interference effects.