Models for a Longitudinal Quantitative Outcome with Multivariate Normal (MVN) Distribution Censored by Therapeutic Intervention

• Activity Number: 469
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Wanjie Sun*+ and Michael D. Larsen and John M. Lachin
• Companies: The George Washington University and The George Washington University and The George Washington University
• Keywords: Multiple imputation ; Monte Carlo EM ; Censored Normal Regression ; Truncated MVN

Abstract

In longitudinal clinical trials, epidemiologic, or genetic studies, a quantitative outcome may be altered by the administration of a non-randomized, non-trial intervention during follow up. The re- sulting effect of the non-trial intervention may seriously bias the study results, including treatment or exposure effects or associations. Current methods to address this issue including multilevel models (White et al 2001) or multiple imputation (MI) (Cook 1997, Cook 2006), are either restricted to a specific longitudinal data structure or are valid only under special circumstances. We propose two new methods for general longitudinal data - a modified Expectation-Maximization (EM)-type model and a modified Monte Carlo EM-MI model. These combine Monte Carlo EM (Wei and Tanner 1990) and MI, and are extensions of Censored Normal Regression (Tobin et al 2005) to longitudinal data. These replace the intractable calculation of a multi-dimensionally truncated MVN posterior distribution with a simplified but sufficiently accurate approximation. Simulation shows that the two proposed methods have the least biased treatment effect estimate in a majority of simulated scenarios amongst six methods applied.