Inference for the Complier-Average Causal Effect From Longitudinal Data Subject to Noncompliance and Missing Data, With Application to a Job Training Assessment for the Unemployed

Abstract

Longitudinal studies involving human participants are often complicated by subjects who do not comply with their treatment assignment or do not provide complete data. A treatment effect of interest in the presence of noncompliance is the complier-average causal effect (CACE; Imbens and Rubin 1997a), which is the treatment effect for subjects who would comply regardless of the assigned treatment. Imbens and Rubin (1997a,b) proposed maximum likelihood and Bayesian inferential methods for CACE, which make explicit assumptions for causal inference in the presence of noncompliance and are more efficient than standard instrumental variable methods. A model for inference about the CACE based on this approach is developed which allows for the inclusion of baseline covariates and handles missing data in the repeated outcome measures. Our methods are applied to a randomized trial of a job training intervention for unemployed workers. Results suggest that the intervention trial significantly reduced depression for high-risk compliers up to six months postintervention but not for low-risk compliers.