A Comparison of Semiparametric Approaches to Model Nonlinear Outcome Trajectories in the Presence of Nonignorable Dropout

• Activity Number: 358 - Contributed Poster Presentations: Section on Statistics in Epidemiology
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Andrew Hammes* and Samantha MaWhinney and Nichole E Carlson and Peter DeWitt and Jeri Forster
• Companies: University of Colorado-Biostatistics and University of Colorado Anschutz Medical Campus and University of Colorado Anschutz and University of Colorado - Biostatistics and University of Colorado - Biostatistics
• Keywords: B-Splines; Nonignorable dropout; nonlinear; semi-parametric models; mixture models; CD4 T-Cell trajectories

Abstract

Dropout is a common problem in longitudinal cohort studies. If the probability of dropout depends on unobserved outcomes, it is considered missing not at random and is therefore nonignorable. We have previously considered a varying-coefficient mixture model using natural B-splines assuming a continuous dropout distribution with a linear outcome trajectory. We extend our existing method to allow a nonlinear trajectory over time using B-splines to model both time and dropout time. Given a lack of a rectangular basis, a tensor product between these B-splines is not applicable. We considered two approaches: (1) a nonlinear B-spline transformation of both time and dropout time, and (2) adding a B-spline transformation of the interaction between time and dropout time to the model. Simulation studies, consistent with nonlinear CD4+ T-cell count trajectories in untreated HIV-infected patients identified during early infection, were used to evaluate method performance. The interaction model performed best based on mean squared error. This model had a stable, reasonable fit; although it was not as flexible as may be desired.