Primary Submission Category: Bayesian Causal Inference
Bayesian inference for direct and spillover effects for two-stage randomized experiments affected by noncompliance
Authors: Claudia Mastrogiacomo, Laura Forastiere, Joshua Warren,
Presenting Author: Claudia Mastrogiacomo*
Interference occurs when one individual’s outcome depends on the treatments of other individuals. Two-stage randomized experiments, where clusters are first randomly assigned to a certain proportion of treated individuals (dosage) and individuals are then randomly assigned to a treatment, have been used to estimate direct and spillover effects in the presence of interference. However, estimation of causal effects in two-stage randomized trials can be complicated by noncompliance since, in a setting with interference, an individual’s potential treatment uptake and potential outcomes may be influenced by the treatment assignment and uptake of others.
Under the principal stratification framework, we define principal strata based on potential treatment uptake and on whether and at which dosage an individual would switch their compliance behavior. We develop a Bayesian inference method to impute compliance types and estimate direct and spillover effects within strata, modeling the outcomes as a Gaussian process with the covariance matrix depending on the switching thresholds. We apply our novel method to data from a two-stage randomized experiment conducted in villages in rural Honduras to estimate direct and spillover effects of a maternal and child health intervention.