Webcast Lecture Series

Navigating the Complexities of Multicollinearity in Regression Analysis

This presentation will discuss the inferential risks associated with multicollinearity among independent variables in ordinary least squares regression, and will offer methods for identifying and addressing these risks. Specifically, I will illustrate how multicollinearity can result in a phenomenon I call “beta polarization,” where the coefficients of correlated independent variables take on opposite signs and show statistical significance, even though one of the signs appears counter-intuitive and theoretically implausible. Such counter-intuitive results are usually type 1 errors (false positives) without causal validity; they are not likely to be surprising findings worthy of new theorizing.

Throughout this presentation, I will debunk four myths that have hindered serious discussions about multicollinearity in empirical research.

1. Misleading Claim of Unbiasedness: While it is technically true that multicollinearity does not bias coefficients, this is only a helpful insight in unrealistic and idealized scenarios. Generally, this claim is very misleading.

2. Misguided Faith in VIF Scores: Many research studies incorrectly dismiss multicollinearity concerns based on low VIF scores. However, low VIF scores do not actually eliminate these concerns.

3. Misidentification of Correlated Variables as Suppressor Variables: Variables affected by beta polarization due to multicollinearity are sometimes wrongly identified as suppressor variables. Such erroneous identification suggests that the beta polarized type 1 errors are legitimate causal effects that were “suppressed” when the correlated variable was excluded.

4. Sample Size Misconception: Contrary to the belief that larger sample sizes can mitigate multicollinearity, the frequency of multicollinearity-induced type 1 errors actually increases with larger sample sizes

Arturs Kalnins is Professor of Management and Entrepreneurship and Professor of Economics (by courtesy) at the Tippie College of Business, University of Iowa. His main research interests are business geography and research methods. His work has appeared in journals such as Strategic Management Journal, Organizational Research Methods, Academy of Management Journal, Academy of Management Review, Management Science, Marketing Science, and the RAND Journal of Economics.