Contextual Effects in Multilevel Models: A Discussion of Centering Strategies From a Graphical Models Perspective

• In Event: Methods for Causal Inference and Estimation of Treatment Effects

Abstract

In educational research, contextual effects are often of interest to researchers and policymakers. Conventionally, compositional multilevel models are used to assess contextual effects, where both the level-1 predictor and the corresponding cluster-level average are included in the regression model using the recommended centering strategies. Such conventional approaches, however, may fail to correctly capture the true data generating processes that involve contextual effects due to within-cluster interference. The current study examines the causal identification of contextual effects using causal interference graphs and outlines the scenarios when the conventional multilevel models and predictor centering strategies fall short for causally identifying contextual effects. Following the conceptual illustration and analytical derivation, simulation studies will be conducted to empirically demonstrate the theoretical results.

Authors

• Yi Feng, University of Maryland - College Park

• Peter M. Steiner, University of Maryland - College Park