Date Posted: 08/19/2020
Traditional methods for doing driver analysis (using correlation analysis or regression analysis) suffer from well-known flaw, particularly from the fact that collinearity degrades both of them badly. After showing the evidence for not doing simple correlation or regression analysis, we show how each of them can be done better. And after considering the red herring that is factor analysis, we’ll go on to cover three modern methods (averaging over orderings, Johnson’s relative importance weights and random forests) that enable an analyst to conduct driver analysis in ways that accommodate collinearity. In addition, we’ll provide advice for how to run these new models free software, including R code for all methods we discuss and other free web-based tools for those who don’t use R.