Variability in the Complier Average Causal Effect of E-Assist on a Binary Outcome by a Shared Random Effects Model

• Activity Number: 480 - Causal Inference and Optimal Decision-Making
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Xinxin Sun* and Yongyun Shin
• Companies: Virginia Commonwealth University and Virginia Commonwealth University
• Keywords: one-sided compliance; missing at random; adaptive Gauss Hermite quadrature; the EM algorithm; Newton Raphson; maximum likelihood

Abstract

e-Assist is a decision aid program to help eligible patients complete colorectal cancer screening(CRCS); they are randomized to e-Assist or control within physicians(PH). However, the effect of e-Assist may vary randomly over heterogeneous PH and compliance C to treatment assignment is imperfect. Because those assigned to control cannot access e-Assist, one assigned to e-Assist is an "e-Assist complier" if the patient takes the assignment or "never taker" otherwise; one assigned to control, however, could be a "control complier" if she would have taken e-Assist or "never taker" otherwise under the alternative assignment. Assuming C missing at random, we jointly model CRCS and C for the random, PH-specific CRCS rates of the C groups; the difference between e-Assist and control compliers is a complier average causal effect. We compare the means, variances and covariances of random rates to learn who compliers are. We integrate random effects out by adaptive Gauss Hermite quadrature(AGHQ) to compute the likelihood. To deal with the intensive computation by AGHQ and highly collinear random effects, we estimate the joint mixed model with shared random effects by maximum likelihood.