Noncompliance in Randomized Experiments: A Stochastic Perspective

The purpose of this dissertation was to estimate the average effect (ATE) of choosing a treatment in a randomized experiment with noncompliance. This required modeling noncompliance as a hybrid choice model or an integrated choice and latent variable model (ICLV). The identification of ATE required an econometric model of noncompliance choice under inherent uncertainty (ex-ante choice) as an ICLV model. The situated expectancy value theory (SEVT) provided a model for the cognitive process underlying individual noncompliance behavior. SEVT provided a plausible model of vocational education and training choice (VET choice) among at-risk youth. SEVT helped in identifying the pretreatment covariates – observed and unobserved – which constitute a sufficient information set to fully describe the noncompliance behavior. These are also a sufficient set of confounders which can be adjusted with to obtain unbiased ATE of a job-training program (Job Corps) participation on the risk of violent crime victimization among at-risk youth. I estimated a four percent reduction in the risk of violent crime victimization by participating in the Job Corps program.