turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Normalize A Variable

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 07:00 PM

If a variable has value ranging from 5000 to -5000, how does normalizing it to the range of 1 and -1 affect the regression model result? Will this variable become less influential?

Accepted Solutions

Solution

07-07-2017
08:38 AM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-09-2014 08:17 AM

how does normalizing it to the range of 1 and -1 affect the regression model result? Will this variable become less influential?

If you are talking about ordinary least squares regression, then changing the units like this will not make the variable "less influential". If you are fitting the model with some other fitting technique (like for example partial least squares regression) then it is possible that this will make the variable "less influential".

At least, that's for my definition of "influential". Maybe you have a different definition? What do you mean by "influential"? Is there a standard mathematical description that you can provide for the meaning of "influential"? My definition would be the same sum of squares, in which case there should be no change to the sum of squares if you rescale an independent variable in ordinary least squares regression.

All Replies

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 08:35 PM

Hi,

Do you want to standardize variable using range?

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 08:41 PM

Yes, standarize/normlaize variable from [-5000,5000] to [-1,1] and want to know the implication for model?

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 08:57 PM

Okay, you can use this syntax to standardize variable. Independent variables in regression usually become in different units (kgs, meters etc.) so we convert them into comaptible units by stadardizing them to have mean equal to zero and standard deviation 1. Hope this helps.

proc stdize data=have out=want method=range;

var var;

run;

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 09:00 PM

This is not what I am looking for. I would like to know what the effect of normalizing a variable is on model building.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-08-2014 09:08 PM

Are you normalizing dependent variable or independent variables?

Solution

07-07-2017
08:38 AM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-09-2014 08:17 AM

how does normalizing it to the range of 1 and -1 affect the regression model result? Will this variable become less influential?

If you are talking about ordinary least squares regression, then changing the units like this will not make the variable "less influential". If you are fitting the model with some other fitting technique (like for example partial least squares regression) then it is possible that this will make the variable "less influential".

At least, that's for my definition of "influential". Maybe you have a different definition? What do you mean by "influential"? Is there a standard mathematical description that you can provide for the meaning of "influential"? My definition would be the same sum of squares, in which case there should be no change to the sum of squares if you rescale an independent variable in ordinary least squares regression.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-09-2014 08:45 AM

And to carry this further, are you simply rescaling, or are you in fact normalizing. In the first, you would just divide the values by 5000. In the second, since the values are already centered on zero (I assume), you would divide the values by the standard deviation. As Paige points out, this should have no effect on the "influence" of the variable, EXCEPT...

If this is a multiple regression, where there are several independent variables, and you normalize only one of them, it's pretty obvious that you are going to change the (partial) covariance between that variable and the dependent variable. If you normalize all of the variables, influence on the dependent variable is more a matter of changes in standard deviation units rather than on the original scale.

Short answer: Normalizing MAY (and probably WILL) change influence as measured by partial R squared, or by change in IC. Rescaling should NOT change influence.

Steve Denham

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-09-2014 08:52 AM

It would help the discussion if aha123 would state the actual model used, how many independent variables are there, also are there interactions or quadratic or other polynomial terms (or is it a nonlinear model), and it would further help to know if we are discussing ordinary least squares or some other model fitting technique.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

06-09-2014 09:23 AM

Good call--REG, PLS, ROBUSTREG, QUANTREG and MIXED may all give different results, as they use five different algorithms for optimizing the function. And that is just for a linear function.

Steve Denham