Diagnosing and Dealing with Multicollinearity

Abstract

The purpose of this article was to increase nurse researchers' awareness of the effects of collinear data in developing theoretical models for nursing practice. Collinear data distort the true value of the estimates generated from ordinary least-squares analysis. Theoretical models developed to provide the underpinnings of nursing practice need not be abandoned, however, because they fail to produce consistent estimates over repeated applications. It is also important to realize that multicollinearity is a data problem, not a problem associated with misspecification of a theorectical model. An investigator must first be aware of the problem, and then it is possible to develop an educated solution based on the degree of multicollinearity, theoretical considerations, and sources of error associated with alternative, biased, least-square regression techniques. Decisions based on theoretical and statistical considerations will further the development of theory-based nursing practice.

Factors Associated with Underscreening for Cervical Cancer among Women in Canada

“…Based on this cut-off, location of residence was not retained in the logistic regression analysis. A variance inflation factor (VIF) of greater than 10 was used to identify possible multicollinearity among independent variables (Schroeder, 1990). Using this technique, location of birth was excluded from further analyses due to possible multicollinearity with other variables.…”

Failure to reduce C-reactive protein levels more than 25% in the last 24 hours before intensive care unit discharge predicts higher in-hospital mortality: A cohort study

“…The backward elimination method was performed using P b .05 as statistical significant, and variables with P b .10 were retained in the final multivariate model after the likelihood ratio analysis. Single collinearity was evaluated with the Pearson correlation among the independent variable, and multicollinearity was evaluated with the variance inflation factor [22]. The odds ratio (OR) and corresponding 95% CI for each variable were computed.…”

Detecting the number of buildings in a single high-resolution SAR image

“…The value is between 0 and 1, where 0 corresponds to a circle region and 1 to a line segment; -Compactness -the ratio of the square of the perimeter to the area of the patch. If we directly use all of these features for modelling, there will be a multicollinearity problem [Schroeder, 1990] because some features may be correlated with each other. In the presence of serious multicollinearity, regression estimates are unstable, and the possibility of overfitting the data increases.…”