help with transformation of variables in multiple regression

Moderators: statman, Analyst Techy, andris, Fierce, GerineL, Smash

Emily

help with transformation of variables in multiple regression

Postby Emily » Thu Jul 10, 2008 1:03 pm

I am conducting a mutliple regression with 6 IVs. Most of the IVs range from -2000 to 2000 (reaction time (RT) data). The direction of the scores is important as a negative score signifies faster responses to a particular group of words (group A) and a positive score indicates faster responses to a different set of words (group B). The IVs are calculated as mean RT to group A words - mean RT to group B words.

One of my variables was slightly positively skewed, so I performed a square root transformation adding a constant to make negative values positive. Now this has created a positive correlation with the DV when before there was a negative correlation. So do I need to back transform it to reinstate the negative scores? I have the same problem with another variable that I transformed using reflect and square root. Do I now need to reflect the scores back? And if so, how do I do this?

I would be really grateful of any advice you can give me on this.
statman
Administrator
Posts: 2723
Joined: Tue Jun 12, 2007 12:08 pm
Location: Florida, USA

Postby statman » Thu Jul 10, 2008 2:38 pm

So I am guessing that you over wrote the original variable? As expected, no a good idea. Anyway, a SQRT transform on a negative value - Will this work?

I suggest you attempt to get your original data, start over and ignore the transformation (if you scan the forum you will see I am not a supporter of transforms, there are other options in most cases.
uncouth_kitten
Posts: 2
Joined: Thu Aug 07, 2008 1:43 pm

Postby uncouth_kitten » Fri Aug 08, 2008 9:48 am

Thank you for your reply. I waited ages and ages for the activation email for registration but never received it, so I have registered with another username, but I am the above poster, Emily.

I didn't actually overwrite the original variable, I always make sure I keep all raw data. In the end I windsorised more of the extreme values, as I had been very conservative to begin with, and this seemed to solve the skewness to some degree.

Many thanks for your help though. I actually have another question now which I have just posted! :oops:
statman
Administrator
Posts: 2723
Joined: Tue Jun 12, 2007 12:08 pm
Location: Florida, USA

Postby statman » Fri Aug 08, 2008 1:13 pm

Good

Topic closed

Who is online

Users browsing this forum: No registered users and 1 guest

cron