Power in the Case of Extreme Bimodal Distribution: Analysis of Delirium/Coma-Free Days in the ICU

• Activity Number: 360
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Jennifer L. Thompson*+ and Ayumi Shintani
• Companies: Vanderbilt University and Vanderbilt University
• Keywords: bootstrap ; proportional odds ; nonparametric ; regression ; simulation ; bimodal

Abstract

When data are complex, statistical methods considered most correct may produce results which are not easily interpreted by clinicians, making choice of analysis difficult. We consider data measuring acute brain dysfunction among critically ill patients. Severity of dysfunction was quantified as number of days alive and without delirium or coma over a 14-day study period; this variable often creates clusters at both minimum and maximum values, making normalization of data challenging. Proportional odds models are useful, but results can be difficult to express to clinicians. We present a simulation study to compare the power and relative benefits of the Wilcoxon rank sum test, proportional odds logistic regression, and bootstrapped linear regression - all robust to distribution assumptions - in detecting differences in outcomes between two treatment groups, using normal and bimodal data.