Seminars 2007— Abstracts

Friday April 20

Speaker: Liana Jacobi, University of Melbourne

Title: Analysis of Treatment Response Data from Eligibility Designs

Abstract: We develop two approaches for isolating the effect of a treatment on an outcome of interest in settings where subjects are randomized into a treatment or a control arm, but where the actual intake of the treatment is not necessarily the same as the assignment, except in the control arm. We suppose that the lack of compliance with assignment in the treatment arm is potentially due to observed and unobserved confounders. In the first approach that we develop, which stems from Sommer and Zeger (1991), the unobserved confounders are modeled by a discrete indicator variable that represents subject-type, defined in terms of their potential intake in the face of each possible treatment assignment. In the second approach, confounding is modeled through a continuous variable without reference to subject-type. Both approaches are formulated in Bayesian terms and implemented by tuned MCMC methods. Further comparisons of the approaches are possible via marginal likelihoods and Bayes factors. The techniques are illustrated in detail with simulated data and with real data where the question is to find the effect of job-training for the unemployed on subsequent depression scores.