weighting data and winsoring

[natalia malancu]

hi guys!

quick beginner question.
i am estimating earnings and weighting my estimates. i would like to winsorize the earnings distribution - top and bottom 1% set respectively to the value of the 1st and 99th percentile
should i:

a. winsorize the distribution as it it and then perform the regression with weights on?

sum wg,de
gen wins=wg
replace wins=r(p1) if wg<=r(p1)
replace wins=r(p99) if wg>=r(p99)
svy: regress wins x y z t

or
b. winsorize the weighted distribution and then perform the regression with weights on?

svy: mean wg
gen wins=wg
replace wins=r(p1) if wg<=r(p1)
replace wins=r(p99) if wg>=r(p99)
svy: regress wins x y z t

thanks,
natalia

[Phil Bromiley]

Since winsorizing has no particular statistical properties, I don't think your question has a good answer. By the way, there is an automatic winsorizing routine in Stata.
If it makes much difference which one you do, I'd be very concerned.

[Nick Cox]

Phil:

an automatic winsorizing routine in Stata

Which is? I count only user-written contributions, and I've often regretted writing the one I did.

[natalia malancu]

I did not imply it has statistical properties. I am doing it more as a check - not a believer of its mystical powers.
It matters in as far as the value at which you winsorize under svy is different then the value at which you do it outside svy. My intuition is since the analysis is conducted under svyset then the winsorizing of the dependent should account for that. maybe i am wrong?!