Random dummy variable with restrictions

[Felix Stips]

I have a panel dataset and want to create a random dummy variable with mean 0.2 that marks a random subpart of the dataset. This would work with some code like this

set seed 2803
gen id = _n
gen random = runiform()
sort random
local cut = round((_N / 10) * 2)
gen sample = _n <= `cut'
drop random

However, I want to impose the restriction that each individual i and each time period t is marked at least once by this dummy, i.e. is present in the sample. Any ideas on how to approach this?

Note: I tried to use the gsample command by B. Jann, but it does not seem to work for this

clear
use http://www.stata-press.com/data/r16/grunfeld.dta
distinct company
distinct time
xtset company year
gsample 20, wor cluster(company year) percent alt
distinct company
distinct time

clear
use http://www.stata-press.com/data/r16/grunfeld.dta
distinct company
distinct time
xtset company year
gsample 20, wor strat(company year) percent alt
distinct company
distinct time

[Felix Stips]

Okay, so first of all using the ssc command randomtag by Robert Picard makes this easier as the data is kept in place. Secondly, one approach to do this is to pick one observation from each group and then generate the remaining observations randomly until we reach target sample size.

set seed 12345
use http://www.stata-press.com/data/r16/grunfeld.dta

local percent = 20
local samplesize = round((_N / 100 ) * `percent')

egen company1 = tag(company)
qui count if company1 == 1
local n1 = r(N)

egen time1 = tag(time)
qui count if time1 == 1
local n2 = r(N)

local n3 = `samplesize' - `n1' - `n2'

randomtag if company1 == 0 & time1 == 0, count(`n3') gen(sample)
replace sample = 1 if time1 == 1 | company1 == 1
drop time1 company1