Exploratory Approaches for Determination of Important Physical Attributes of Cigarettes in the Yields of Tar, Nicotine and CO: a Case Study

• Activity Number: 78 - Statistical Consulting Applications
• Type: Contributed
• Date/Time: Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Christopher Ellison* and Qian Li
• Companies: Food and Drug Administration and Food and Drug Administration
• Keywords: Variable importance ; Wold value ; commonality analysis ; parsimonious regression models

Abstract

Several statistical metrics exist to identify important contributors to dependent variables in regression models. Individual approaches may not be satisfactory when the contributors are correlated and may not provide information on relative importance of the contributors in explaining variability of dependent variables. The goal of this project is to demonstrate exploratory statistical methods that may collectively identify key regressors and provide assessment on their relative importance. We present a case study using data from 51 brands of cigarette which utilizes exploratory statistical methods to identify correlated physical attributes that are important in describing variability in yields of tar, nicotine, and carbon monoxide (TNCO) in smoke as collected by smoking machine. Attribute importance was assessed using parsimonious regression models, partial least squares Wold values, and commonality analyses. We explain how similarities and differences among these statistical methods may lead to identification of attribute importance. Our results suggest that a combined-assessment approach can reveal the relative importance of the attributes in the yields of TNCO.