A0833
Title: Stochastic noncompliance and endogenous confounding in evaluating a multi-phase treatment: Multi-site IV as a solution
Authors: Guanglei Hong - University of Chicago (United States) [presenting]
Abstract: The average cumulative effect of a multi-phase treatment sequence is hard to identify due to stochastic noncompliance and endogenous confounding. Despite the initial randomization of treatment assignment, individual responses to the phase-1 treatment may predict non-compliant behaviors in the subsequent phase, thereby confounding the effect of the phase-2 treatment on the outcome. Principal stratification resorts to a deterministic framework, often a mismatch with reality. In contrast, non-compliant behaviors are allowed to be influenced by stochastic random events. Extending the instrumental variable (IV) method to a multi-site randomized trial for evaluating a multi-phase treatment sequence, the approach requires neither sequential ignorability nor exclusion restriction. The key is to obtain the conditional distribution of the potential intermediate outcome under the counterfactual phase-1 treatment condition as a function of not just baseline covariates but also the observed intermediate outcome under the actual phase-1 treatment condition for each individual. The cumulative treatment effect is further allowed to depend on these potential/counterfactual intermediate outcome values. Reanalyzing the well-known Project STAR data, a multi-site randomized trial for studying class size reduction, the average impact of receiving two years of instruction is evaluated in a small-size class as opposed to a regular-size class on student achievement.
