3

A question that frequently comes up while presenting the findings of a predictive model to a business audience (with non-statistical background) is: which variable/predictor is most important in explaining the phenomenon being modeled? Or they ask for the set of predictors included in the final model to be listed as per their importance.

Now, what measure of "importance" should I be using?:

  1. I could use p-values for the predictors as a rough measure of importance, though am not sure if that will be accurate

  2. Coefficient values could be another option to compare relative importance, but given that different variables may be on different scales it won't be an apple-to-apple comparison

  3. I tend to prefer the the list of important variables that the Random Forest algorithm generates (for e.g. using the VarImp function in the randomForest package in R)

Are there any better variable importance measures that I can look at?

mkt
  • 18,245
  • 11
  • 73
  • 172
Dataminer
  • 375
  • A quick point worth mentioning. A lot of models insist that you center and scale the variables for precisely the reason you mention in point 2. What do you mean by accurate p-values? – Sid May 27 '15 at 05:30
  • note that there are two uses of the word "explaining": one is based on mere covariance and without a proper design, this would not have the second meaning: causal explanation of changes observed in the dependent variable. Another complication: some variables on the predictor side can be actually influenced while others are themselves outcomes. – jank May 27 '15 at 12:53

2 Answers2

1

You could generate the entire Lasso path, and the variables can be listed to be important in the order in which they enter the fit.

rakesh a
  • 11
  • 1
0

I could use p-values for the predictors as a rough measure of importance, though am not sure if that will be accurate.

It won't. P values say nothing about how important a variable is. A better measure would be the effect size.

Coefficient values could be another option to compare relative importance, but given that different variables may be on different scales it won't be an apple-to-apple comparison

Better. You could always standardize covariates to have mean 0 and standard deviation 1. That puts things on the same scale.

I tend to prefer the the list of important variables that the Random Forest algorithm generates (for e.g. using the VarImp function in the randomForest package in R)

This is fine, but there are more interpretable models than a random forest.

If prediction is your main concern and not inference, I think a penalized model (a la LASSO or Ridge Regression) is the way to go. You can examine what variables are most important by tuning the penalty parameter and seeing what drops out of the model last.

Demetri Pananos
  • 36,121