This study examines the performance of eight methods of predictor importance under varied correlational and distributional conditions. The proportion of times a method correctly identified the dominant predictor was recorded. Results indicated that the new methods of importance proposed by Budescu (1993) and Johnson (2000) outperformed commonly used importance methods.
Whittaker, Tiffany A.; Fouladi, Rachel T.; and Williams, Natasha J.
"Determining Predictor Importance In Multiple Regression Under Varied Correlational And Distributional Conditions,"
Journal of Modern Applied Statistical Methods:
2, Article 44.
Available at: http://digitalcommons.wayne.edu/jmasm/vol1/iss2/44