Winsorization as robustness test

[Prathvajeeth Rajmohan]

Hi I am investigating the effect of IR derivative usage on firm value, where my focus variable is IR derivative and there also control variables included:

Code:
. regress firm value lnassets IRderivatives 10 bookleverage_w1 roa_w1 cratio_w1 rnd_rev_w1 div_yield_w1
this regression contains all winsorized variables, and I have found that the coefficient on Ir: 0.05, is positive and significant.

Can you use the same regression without the winsorized variables as a check of robustness?

I ran the same regression again without the winsorized variables and I found that while the magnitudes of all the controls changed, the ones that were signficant in the initial winsorized version were still significant and had the same sign, and the there that ir coefficient was now 0.1 and still positive and significant. Can I claim that the IR derivative coefficient results are robust to the effect of outliers. thanks

[Bruce Weaver]

Can you use the same regression without the winsorized variables as a check of robustness?

I think you have it the wrong way around. One would estimate the model with the original variables first, and then do something else (e.g., use Winsorized variables) as a check of the robustness of the model with original variables.

Not everyone is a fan of Winsorizing. I would consider using rreg to check for robustness instead. To read about it, type help rreg, then click [R] rreg to open the PDF documentation.

HTH.

[Nick Cox]

Whenever Winsorizing is at all tempting it is likely that some functional form other than Y = Xb may make (more) sense than just doing what you did otherwise but with and without Winsorizing.

I have to disagree with Bruce a little. rreg implements a flavour of robust regression that was at most a competitive method art in 1985. In 2017 there are better robust regression methods now available. Vincenzo Verardi is the leading Stata author here and a search for his name will find more details.

But we seem to be answering the same questions repeatedly here.

[Bruce Weaver]

Nick, thanks for the tip about Vincenzo Verardi. I found his Stata Journal article with Croux.