A Comparison of Logistic Regression Pseudo R2 Indices

Abstract

The present study examined seven distinct pseudo R² indices used in logistic regression, how values of these indices compared to values of ordinary least squares (OLS) R² obtained under similar conditions, and how values of these indices varied as a function of multicollinearity among predictors and base rate of the dependent variable. Monte Carlo simulation methods suggested that the Aldrich-Nelson pseudo R² index with Veall-Zimmermann correction resulted in values that most closely approximated the OLS R² values. Additionally, lower multicollinearity among predictors was associated with increased variability among the values resulting from the various indices. Changes in base rate had little effect on “corrected” versions of the indices, but did affect uncorrected versions of the indices.