Dichotomizing Non-Normal Continuous Data While Retaining Statistical Precision for Informing a Commensurate Prior

• Activity Number: 297
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Byron Gajewski*+ and C. Shane Reese and John A. Colombo and Susan E. Carlson
• Companies: University of Kansas Medical Center and Brigham Young University and University of Kansas and University of Kansas Medical Center
• Keywords: dichotomization ; statistical precision ; mixture of normal distributions ; CI

Abstract

Dichotomizing continuous data is a huge problem because of the reduction in precision. However, clinicians prefer a threshold for clinical and policy decision making. Peacock et al (2012) addressed this by using a normal distributional approach. However, care must be taken when directly applying their distributional approach when the clinical thresholds are in the tails of the distribution because a deviation from normality is more crucial in the extremes (Peacock, et al, 2012, p. 14). Our application involves clinical thresholds (< 1.5kg or < 2.5 kg) for low birth weight. It is very well known that birth weight is non-normal and thresholds are at the tail. We propose an extension that analyzes non-normal data using a mixture of normal distributions to replace dichotomizing alone. We demonstrate our proposition on birth weight data and simulation and show better performance of the mixture approach. We conclude with an explanation of how this approach is used to develop a commensurate prior used in a model for predicting public health impact of treatment for reducing low birth weight from a clinical trial.