Date: Fri, 13 Jul 2007 05:42:15 -0400
Reply-To: Peter Flom <firstname.lastname@example.org>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject: Re: Regression Skewed data!
Content-Type: text/plain; charset=UTF-8
Pooch <sree.seetharam@GMAIL.COM> wrote
>I have a data where both my dependent and independent variables are
>left-skewed. What transformation do I use to normalize both of them
>and use them in regression?
First, ordinary least squares regression does not assume a particular distribution for either the dependent or independent variables, it assumes that the *residuals* are normally distributed. So, run the regular regression and look at them
If the residuals are not normally distributed, then transformations are *one* option. But not the only one, and perhaps not the best one. There are other regression models that may be better (e.g. robust regressions). Some of this is a matter of preference, but some is not. Should you transform? Well, does the transformation make substantive sense? Some people use the Box-Cox method to come up with a good transformation, but raising a variable to a fractional power usually does not lead to easy interpretation. Also, transforming to normalize the residuals may cause other problems (e.g. heteroskedasticity).
I suggest you write back to SAS-L and list
1) What your DV and IVs are
2) What you are trying to figure out
3) How many IVs, how many N, and so on
4) What the residuals from OLS regression look like
Then someone may be able to provide you with more guidance