I am running the following nonparametric regression:
Code:
set seed 12345
npregress kernel spending age income i.year i.fips, kernel(gaussian)
margins if age >= 55 & age < 65
margins if age >= 64 & age < 85
But every time I run the model, I get a different average derivative for age and a different mean for spending and therefore a different observed margin from the margins command. I am not interested in bootstrapping the standard errors, but I thought having the seed set would handle this, but is there another way? The values are similar, but not the same.

Is this a bandwidth selection issue? Because by using cross-validation it's different each time?