problem with result variable generated using xtile

[Mariana LO]

Hello,

I am trying to generate an income quantile variable using xtile. For some reason, the new variable has less levels with actual values than what I specify, and one level is all missing.

For example, if a generate the new variablw specifying n(5), I get 4 levels with values and one with all missing values. The original variable is a continous variable with lots of zeros.

sum Y_total

Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
Y_total | 61665 4868.114 17333.08 0 199998


Here is an example of what is happening with my new variables:

. xtile Y_quart = Y_total, nq(4)
. tab Y_quart
4 quantiles |
of Y_total | Freq. Percent Cum.
------------+-----------------------------------
1 | 33,457 54.26 54.26
3 | 13,489 21.87 76.13
4 | 14,719 23.87 100.00
------------+-----------------------------------
Total | 61,665 100.00


The only way it seems to not-do this is if I set n(3), which generates:

xtile Y_quart2 = Y_total, nq(3)
. tab Y_quart2
3 quantiles |
of Y_total | Freq. Percent Cum.
------------+-----------------------------------
1 | 33,457 54.26 54.26
2 | 9,253 15.01 69.26
3 | 18,955 30.74 100.00
------------+-----------------------------------
Total | 61,665 100.00

Do you have any ideas what is causing this? I am using Stata SE 11.2

THANKS!!!

[Phil Clayton]

This is most likely correct behaviour due to all of the zeroes. See the technical note in the PDF documentation for xtile, eg page 8 of http://www.stata.com/manuals13/dpctile.pdf

Simple example:

clear
set obs 100
gen x=0
replace x=_n in 51/100
xtile q=x, n(4)
tab q

[Nick Cox]

This is a common complaint with xtile, but the constraint that the same values must get put into the same quantile band can bite very hard.

In general using a different convention about the inequalities may give a better answer, but it sounds as if that's unlikely in this case. There is some detailed discussion within http://www.stata-journal.com/article...article=pr0054

[Mariana LO]

Thanks for the replies.