Joint Models and Tests of Multiple Noncommensurate Outcomes

• Activity Number: 212
• Type: Invited
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Frank Yoon*+ and Stuart Lipsitz and Garrett A. Fitzmaurice and Nicholas Jon Horton and Sharon-Lise Theresa Normand
• Companies: Harvard Medical School and Brigham and Women's Hospital and Harvard School of Public Health and Smith College and Harvard Medical School
• Keywords: Multiple outcomes ; Clinical trial ; Observational study ; Testing ; Likelihood ; Quasi-likelihood

Abstract

Multiple outcomes in randomized and observational studies in psychiatry are often non-commensurate, for example, measured on different scales or constructs. Standard multiplicity adjustments can control for Type I error, though such procedures can be overly conservative when the outcomes are highly correlated. Recent literature demonstrates that joint tests can capitalize on the correlation among the outcomes and are more powerful than univariate procedures using Bonferroni adjustments. However, joint tests are little used in practice, perhaps, due in part, to the specification of a joint model for the non-commensurate outcomes. Additionally, software routines to estimate joint models have not been widely publicized despite their wide availability. This work presents an evaluation of likelihood and quasi-likelihood methods for jointly testing treatment effects in a simulation study. Applications to a clinical trial and an observational study of mental health care illustrate their benefits. Adoption of these methods will lead to more efficient psychiatric clinical trials.