351 – Contributed Poster Presentations: Section for Statistical Programmers and Analysts
Assessing the Impact of Multicollinearity in Various Regression Scenarios
Lisa W. Kay
Eastern Kentucky University
Daniel J. Mundfrom
Eastern Kentucky University
Michelle L. DePoy Smith
Eastern Kentucky University
Although it is well known that multicollinearity can impede one's ability to evaluate model predictors (Montgomery, Peck, & Vining, 2001; Pedhazur, 1982), it has been suggested that the presence of multicollinearity may not affect the accuracy of a prediction of the response variable, given a set of observations taken on the predictor variables (Kutner, Nachtsheim & Neter, 2004; Weiss, 2012). In a previous study, the authors explored a model's ability to make predictions under different scenarios by varying the number of predictors, the strength of the association between the predictor variables and the response variable, the sample size, and the level of multicollinearity (Mundfrom, Smith & Kay, 2016). Simulations in the study indicated that confidence intervals are wider in the presence of multicollinearity. Furthermore, differences in confidence interval width appeared to depend on degree of taintedness in multivariate normal data, sample size, and number of predictors in the model. The purpose of the present study was to determine which scenarios result in more appreciable effects and to examine alternative ways of measuring the impact of multicollinearity on predictions.