Content area
Full Text
About the Authors:
Christopher Rentsch
Affiliation: Atlanta Veterans Affairs Medical Center, Decatur, Georgia, United States of America
Ionut Bebu
Affiliation: Infectious Disease Clinical Research Program, Uniformed Services University of the Health Sciences, Bethesda, Maryland, United States of America
Jodie L. Guest
Affiliations Atlanta Veterans Affairs Medical Center, Decatur, Georgia, United States of America, Emory University School of Medicine, Atlanta, Georgia, United States of America, Rollins School of Public Health, Emory University, Atlanta, Georgia, United States of America
David Rimland
Affiliations Atlanta Veterans Affairs Medical Center, Decatur, Georgia, United States of America, Emory University School of Medicine, Atlanta, Georgia, United States of America
Brian K. Agan
Affiliation: Infectious Disease Clinical Research Program, Uniformed Services University of the Health Sciences, Bethesda, Maryland, United States of America
Vincent Marconi
* E-mail: [email protected]
Affiliations Atlanta Veterans Affairs Medical Center, Decatur, Georgia, United States of America, Emory University School of Medicine, Atlanta, Georgia, United States of America, Rollins School of Public Health, Emory University, Atlanta, Georgia, United States of America
Introduction
The methods commonly used in selecting variables for multivariate models can be overshadowed with objective statistical tests while others, in contrast, grant several subjective decisions to the researcher. There are risks associated with methods that rely too heavily on either objectivity or subjectivity. Most notably for the latter is the concept of researcher bias, which is a process in which the researcher influences the results of an analysis by forcing certain variables into or out of regression models typically based upon associations found in previous studies or logic determined by causal pathways [1]. For example, statistical tools, such as stepwise regression and collinearity diagnostics, could suggest a variable be dropped from further analysis in a multivariate model, but the researcher may decide to ignore the suggestion and force the variable into the model. Researcher bias has the ability to impact the accuracy and precision of conclusions output from data analyses.
Conversely, some variable selection processes allow for very little, if any, influence on which variables are kept in or dropped from multivariate analyses. Such processes include using statistical significance tests (e.g., Wald) in a stepwise fashion [2], [3], in which variables are selected and/or deleted from an analysis using a pre-specified significance level (p-value). In essence, the researcher inputs...