Speaker : Song Zhang, PhD
Partially overlapping research units in a pre-post study lead to a mixed nature of observed data: paired outcomes from those who contribute complete pairs of observations and independent outcomes from those who contribute incomplete (pre only or post only) observations. It is frequently encountered by practitioners for various reasons. For example, substantial missing data in the pre and post measurements effectively leads to partially overlapping units. It also occurs when researchers conduct random survey in a small community before and after a radio campaign. We examine the limitations of the traditional estimator of the intervention effect. A new hybrid estimator is proposed which we theoretically prove to be more efficient than the traditional estimator under practical conditions. We further derived a closed-form sample size formula to help researchers determine how many subjects need to be enrolled in such studies. Simulation and a real application example are presented.
Dr. Song Zhang is an assistant professor at the Department of Clinical Sciences, UT Southwestern Medical Center. His research interests include Bayesian hierarchical modeling (with applications in mssing data imputation, joint inference of longitudinal and survival data, microarray data analysis incorporating biological knowledge, risk prediction based on electronic medical records, item response theory, and disease mapping) and clinical trial design (accounting for clustered randomization, longitudinal measurements, historical control arms, missing data, and financial constraints).