I have a mix of continuous and binary data. promotion here is binary so 1= yes 0=no. i standardize (mean 0 and s.d. 1 ) the continuous variables so they are same unit.

I get the following output:

enter image description here

is this the correct interpretation?

  • a one unit s.d increase in temperature increase the s.d in products sold by 1030 therefore it has a large affect ; if we compare this to costs, costs has a greater coeff than temperature and therefore is more important
  • for binary : if we run a promotion we can expect an increase in 5727 units sold.
  • since coeff is greater for promotion than any others it is much more important feature : this one i struggle with, since the data for binary is not standardized and the coefficient has a different meaning is it right to compare relative importance between binary and continuous -- if not how can i RANK the predictors against one another?

You shouldn't compare continuous and binary features without a transformation. In this situation (because promotion seems to be important) I would make the continuous variables into binary variables by assigning a 1 or 0 based on the range, for example:

Temperature between 0-32: 1 or 0 Temperature between 33-70: 1 or 0 Temperature between 70-90: 1 or 0 Temperature above 90: 1 or 0

Now you have four new binary features that represent temperature. You can now run a regression model and safely compare the coefficients with "promotion." Do the same for costs.

Caution: don't add too many binary features! For regression you should try and keep the number of coefficients under 10. If you add too many coefficients it will artificially inflate your R^2 value.

  • $\begingroup$ -1 There is no issue mixing continuous and binary variables in a regression. ANCOVA with two groups is one example. Further, binning data like this loses information. It says that 32 and 33 are less alike than 70 and 33. $\endgroup$ – Dave Feb 9 at 23:56
  • $\begingroup$ Dave my mistake on mixing the variables, thanks for the correction. Binning can be helpful in this case, but you are right there is a trade-off. The ranges I picked are pretty arbitrary I leave it up to Maths12 to find the best ones. $\endgroup$ – bstrain Feb 10 at 1:59
  • $\begingroup$ How do you figure that the binning helps in this case? $\endgroup$ – Dave Feb 10 at 2:09
  • $\begingroup$ If you bin and run regression you can learn if temp's effect changes over its range. $\endgroup$ – bstrain Feb 10 at 2:14
  • $\begingroup$ Robert Long and others on here have addressed that on here and how splines are superior if you’re concerned that temperature should have a nonlinear term. $\endgroup$ – Dave Feb 10 at 2:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.