When to Winsorize the Panel Data Variables of Non-Financial Firms

[Ammar Hussain]

Dear Statalist Community,

I have panel data on more than 200 variables, all from balance sheets, income statements, CFS, CSR reporting, board member counts, etc.

So, my question is: what is the best stage to winsorize the variables? Do I winsorize the variables in their current raw form, e.g., total assets, net income, audit fees, expenditures, liabilities, etc., OR first develop required proxies like ROA (net/total assets), discretionary accruals using the Jones model, natural log of total assets as a proxy of size, etc., and then winsorize the ROA, discretionary accruals, and size, etc.? Or do I need to winsorize both?

Your valuable feedback will really help me a lot. Thanking you in Advance.

Best Regards:
Ammar Hussain (Final Year PhD Candidate)

[Nick Cox]

Your question in turn raises many more questions.

I first wrote a command winsor in 1998 which was later posted to and remains on SSC. The Statalist context surrounding posting of that code is lost and I don't remember any details myself. It may have been that someone wanted code to Winsorize and it seemed fun to write the code. Otherwise my interest then was essentially in one application, Winsorizing a variable before taking a mean, as a robust or resistant complement to other summary measures. I mention this because what follows may seem surprising given that small historical fact. (Others have since gone further in offering Stata code.)

We get questions on Winsorizing here on Statalist that are more like yours than they are like the application just mentioned. Often the idea is, or seems to be, essentially to Winsorize everything as a prelude to some modelling exercise! From time to time I challenge questioners to provide authoritative references explaining why wholesale Winsorizing is (a) a good approach and (b) better than alternatives. I don't recollect any replies to that challenge.

What is a good approach goes hand in hand with what are the goals of any project and particularly the motive for Winsorizing.

Any use of Winsorizing has to be based on choosing how much to Winsorize, whether the same for every variable or different for different variables. That shouldn't be chosen arbitrarily, for example just following some previous study that used 5% or whatever.

The use of Winsorizing envisaged is (it seems) one variable at a time. That would be naive if not analytically dangerous. It is easy to identify definite or possible outliers on one variable that make perfect sense in the light of other variables. It is only a little more difficult to see that outliers could be outliers in some space that aren't discernible in lower-dimensional subspaces.

I don't routinely use the kind of data discussed in #1 but the kinds of environmental data I do know more about are not vastly different.

Winsorizing may (is designed to!) pull in outliers brutally but it can't itself cope with nonlinearities in particular.

For that and other reasons, it seems to me that carefully chosen transformations of some variables and use of generalized linear models (or similar) with appropriate link functions are strongly competitive with Winsorizing as you describe it.

[Nick Cox]

A crucial detail for panel data is that outliers may make perfect sense given previous or following values. It is not just a matter of where they lie within a distribution.

[M Crain]

Financial ratios tend to be very skewed distributions that will likely lead you to use of medians rather than means even if you remove extreme values.

Offhand, consider logs of accounting measures. But that will not help very much for earnings measures that may have a lot of negative values (unless first transforming the observation values). If you have not done already, look at the accounting literature.