Hello,
I am struggling with how to winsorize my variables. I know that winsorizing different variables at different levels is uncommon. However, in my opinion, some of the variables I am using do not need winsorizing (for example, age and log firm size (because age is a natural count and because the log serves to mitigate outliers)). Can I choose not to winsorize these variables, and at what base should I decide whether or not to winsorize a certain variable? Should I use a test? What is your view on this?
Kind regards,
Timea De Wispelaere
No don't, M estimators work much better See the attached, Best, David Booth
Here is an R program that uses variable selection as well. If you don' want variable selection replace hqreg by rlm which is robust only . If you have questions, ask.
Best, David Booth An intro to R is here:
http://sgpwe.izt.uam.mx/files/users/uami/gma/R_for_dummies.pdf
Instead of Winsorizing, you may use robust estimators for your analysis.
But if you really want to winsorize variables, you can do this in a multivariate manner as well.
Univariate: symmetrize your variables with a transformation (e.g. log for income, etc.), estimate a threshold using robust estimators (like median + 3 * mad). Then all values of a variable are shinked to this threshold, for example. When you do this for all variables, you have different data-dependent threshold on each variable.
Multivariate: estimate tollerance ellipses and shrink the values to the boarder of the tolerance ellipse (this is a bit tricky).
Hope this helps.
you seem to be assuming (implicitly) that your data is uniform. extreme values and outliers are important, study them before deciding to DELETE them [winsorising = deleting]...II investigated such problems in the attached paper...
There are literatures that suggest robust regression method is preferable to winsorization.