Simulation-Extrapolatino for Estimating Principal Causal Effect Surfaces

Abstract

Amid the “big data” revolution, background information on participants is becoming ever more available for experimental researchers to predict treatment effect heterogeneity, including heterogeneity on some intermediate variable collected post-treatment. At the same time, the recently developed principal stratification framework allows researchers to assess heterogeneity on an intermediate variable in a manner that maintains causal interpretations. This paper details the shortcomings of two-stage least squares and imputation methods as viable estimators if used to assess treatment effect heterogeneity when the intermediate variable is continuous and traditional assumptions are not tenable. Results from an alternative estimator that relies on simulation-extrapolation is evaluated to inform future research.