Winsorizing or dealing with outliers

[Fimi Karimi]

Hi everyone,

I am working a panel data. My variables are social economic. 2 of my variables, Religious_Perc and NonReligious_Perc are percentage of non religious and religious populations in each country-year base on a survey. Some of countries have a high percentage of religious population, 99 and 100%, which is not reasonable.
Do you have any suggestion for dealing with these outliers? I think they may make my results biased.
Thanks for your time.

[Nick Cox]

If the upper quartile is 99.6, then values of 99 or 100 may be implausible to you, but they certainly aren't uncommon or way beyond the rest of the data.

What is acceptable? 98? 97?

I wouldn't on this evidence call those outliers at all.

[Nick Cox]

It may well be that either variable (or even both) looks too problematic to use. That's a quite different decision from changing some values and not others.

If non-religious PLUS religious sum to 100% where both are given, one is redundant. Otherwise why is one ever missing when the other isn't?

Winsorizing is often mentioned on Statalist. I'll repeat a challenge often given: Please give a reference in a good text or paper explaining why it's a good idea for your analysis, and a better idea than others.