Friday, November 20, 2020 from Noon to 12:50pm
Yongyun Shin, Ph.D.
Departmet of Biostatistics
Virginia Commonwealth University
Joint with Xinxin Sun and Jennifer E. Lafata
Housed within a patient portal, e-Assist is a decision support program that assists patients with a physician recommendation for completing colorectal cancer screening (CRCS). Patients are randomly assigned to e-Assist intervention or usual care within primary care practices in a multi-site randomized trial. Two complications make the causal analysis of e-Assist nonstandard: The causal effects of e-Assist on CRCS (1 if completed, 0 otherwise) may vary over heterogeneous practices, and compliance to treatment assignment is imperfect. Because a control patient assigned to usual care is forced to accept the assignment without access to e-Assist, the patient could be either “complier” if he or she would have taken e-Assist or “noncomplier” otherwise under the alternative assignment to e-Assist. Because the alternative assignment is hypothetical or counterfactual, the compliance is unobservable or missing. A treatment patient assigned to e-Assist, however, is an observed “complier” if the patient takes the assigned treatment or “noncomplier” otherwise. Consequently, a noncomplier is a “never taker” who cannot be induced to take e-Assist no matter what. We will estimate the discrete joint distribution of CRCS and compliance, assumed missing at random, by maximum likelihood via the Newton Raphson method and adaptive Gauss Hermite quadrature; find the practice-specific complier average causal effect (CACE) and never-taker mean CRCS of every practice, and their means, variances and covariance; and compare them with the mean and variance of practice-specific intent-to-treat (ITT) effects.