Do you winsorize continuous variables before running a regression to predict values?

[Stan Hissink]

I am using panel data, consisting of many continuous firm-level. Now my sample is balanced and complete, I want to run several regressions to predict values (for earnings management proxies) following Roychowdhury (2006) in where many continuous variables are the dependent variables. I want to control for extreme outliers and therefore winsorize my continuous variables. After inspecting several variables, I concluded winsorizing is needed. What do you think? Do I need to winsorize before estimating the proxies or after? At this moment, I tend to winsorize the variables before predicting to control for potential errors.

Thanks in advance!

[Nick Cox]

Roychowdhury (2006)

Please read and act on what we say in the FAQ Advice about bare name (date) references.

Otherwise I would search the forum for discussions. Questions like this about from people apparently new to statistics on whether it is a good idea are pretty well matched by scepticism if not hostility from those not so new to statistics. It does appear that these questions all arise from people working from business or finance data. So, why do you think it is a good idea? What about generalized linear models or their relatives? What about transformations?

I'd bear in mind that winsorizing variables scrunches them to fall within a (high-dimensional) parallelipiped and (in all the versions I know of) uses univariate information only, a really strange idea as a prelude to modelling.