A Statistical Method for Analyzing Risk Difference in Trials with a Three-Level Paired Design

Abstract

This thesis is motivated from an animal trial for a wearable external cardiac defibrillator. Each pig in the trial will be treated with two devices (test vs. control) after induced to experience ventricular fibrillation multiple times. A set of different defibrillation waveforms will be tested and are randomized in blocks. The efficacy of two waveforms, defined as the probability of shock success, will be compared in absolute difference. We propose a statistical method, applying t-test on a discrete variable transformed from the original binary outcome, to study the (average) risk difference in trials with a similar three-level paired design, which can properly adjust for the random effects for pig (subject) and block, as well as for the random slope of treatment for pig, which is also known as the “Subject – Varying Treatment Effect”. We develop two naive data generating procedures for the simulation study, and test the performance of both procedures. The results of a pilot study are used to inform design parameters used to simulate data. The performance of our proposed method is evaluated through a set of simulation studies.