In polynomial regression, it is recommended to center predictor input variables to break multi colinear relationships of x to x^2.

From Wikipedia: The underlying monomials can be highly correlated "For example, x and x2 have correlation around 0.97 when x is uniformly distributed on the interval (0, 1). "

When a variable x is between -1 and 1, x^2 makes the magnitude smaller while when x is outside of that range, x^2 makes x's magnitude larger.

Making the variable into an integer variable could change the behavior.


df$x=round((df$x - mean(df$))*100)

Any opinions on the scale especially in regards to interval [-1,1] vs [-100,100]

It is common to normalize predictors subtracting the mean and dividing by the standard deviation when doing inference analysis but this question pertains to regression prediction.

Asking a similar question in regards to natural log, a variable that has a range (0,1] has a dramatically different transformed value than [1,100].

log(seq(0.1,1,.1)) #mostly negative
log(seq(0.1,1,.1)*100) #rather positive

If the predictor variable in the case of log happened to be sometimes less than 1 and others greater than 1, that could make the transformation act a little "wild". Would it be best to transform the variable to be within (0,1] or [1,] but not both?


Your Answer

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

Browse other questions tagged or ask your own question.