|Date: ||Thu, 12 Dec 2002 17:03:07 -0500|
|Reply-To: ||Richard Ristow <email@example.com>|
|Sender: ||"SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>|
|From: ||Richard Ristow <firstname.lastname@example.org>|
|Subject: ||Re: log transform|
|Content-Type: ||text/plain; charset="us-ascii"; format=flowed|
At 04:15 PM 12/12/2002 -0500, Jessica L. Kenty wrote:
>I have very skewed data. I want to use log transformation to "fix"
>it. I realize I can't have any 0's (need to add +1 to all
>values). But, my variable has negative & positive values. Do I need
>to adjust all values to be above zero (i.e. -85000 adjust by adding
>85001 to = 1)?
Well, you'd have to do something like that; but I think you should look
at your data and give some thought to what a 'fix' would mean.
What does the distribution look like? Is it skewed away from zero in
If you log-transform, say, income (I'm looking at your institutional
affiliation in your sig), you're making an implicit judgement (with, I
believe, some support in psychology) that the perceived size of a
change in income is the proportional change: that a doubling of your
income or mine, and Bill Gates's, would affect us all about equally.
If, say, you're looking at net worth skewed in both directions, it
might work better to take the log of the absolute value in each direction.
Alternatively (following Tukey and others), you can accept that any
transformation is fundamentally arbitrary, and use non-parametric
statistical methods on your data. Unfortunately, in an economic project
you may be committed to multiple regression, which doesn't have
non-parametric analogs that I know of.
Sorry for more questions than answers. You sound, though, like you have
a situation where the proper statistics depend on what you're
measuring, and what you regard as meaningful about it.
>Jessica L. Kenty-Drane
>Assets & Educational Inequality Project
>Department of Sociology and Anthropology