I have a bunch of independent variables which are skewed and have negative and zero values. I am seeing a lot of suggestions of using cube root as a transformation.
What would be the harm in using $\text{sign}(x)\log(1+|x|)$ instead?
$\begingroup$
$\endgroup$
5
-
4$\begingroup$ Why are you transforming these independent variables? $\endgroup$– Glen_bCommented Jun 8, 2016 at 17:28
-
$\begingroup$ Because the distributions are very skewed and outliers prone. I'm Afraid that it'll affect the estimates. $\endgroup$– IamNotLegendCommented Jun 8, 2016 at 17:46
-
$\begingroup$ What do you intend to do with these variables (after a possible transformation)? $\endgroup$– Joris BierkensCommented Jun 8, 2016 at 20:17
-
$\begingroup$ Use it as a predictor in a logistic regression $\endgroup$– IamNotLegendCommented Jun 8, 2016 at 20:30
-
1$\begingroup$ Why do you think skewness or outliers are a problem for these variables? $\endgroup$– Glen_bCommented Jun 9, 2016 at 1:34
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
One reason to avoid such a transformation is that it will make the interpretation of the regression coefficient very difficult.
Moreover, there is no requirement for independent variables to be normally distributed, and as a rule you should avoid doing so unless there are substantive reasons for it, such as a known nonlinear relationship, to deal with heteroscedasticity, or to help interpretation
answered Jun 8, 2016 at 21:32