proc stdize producing wrong value for percentiles

Question

I have a dataset like so

data test;
    do i = 1 to 100;
    x1 = ceil(ranuni(0) * 100);
    x2 = floor(ranuni(0) * 1600);
    x3 = ceil(ranuni(0) * 1500);
    x4 = ceil(ranuni(0) * 1100);
    x5 = floor(ranuni(0) * 10);
    output;
    end;
run;

data test_2;
    set test;

    if mod(x1,3) = 0 then x1 = .;
    if mod(x2,13) = 0 then x2 = .;
    if mod(x3,7) = 0 then x3 = .;
    if mod(x4,6) = 0 then x4 = .;
    if mod(x5,2) = 0 then x5 = .;
    drop i;
run;

I plan to calculate a number of percentiles including two non-standard percentiles (2.5th and 97.5th). I do this using proc stdize as below

PROC STDIZE 
    DATA=test_2
    OUT=_NULL_
    NOMISS 
    PCTLMTD=ORD_STAT
    pctldef=3
    OUTSTAT=STDLONGPCTLS
    pctlpts=(2.5 5 25 50 75 95 97.5);
    VAR _NUMERIC_;
RUN;

Comparing to proc means

DATA TEST_MEANS;
    SET TEST_2;
    IF NOT MISSING(X1);
    IF NOT MISSING(X2);
    IF NOT MISSING(X3);
    IF NOT MISSING(X4);
    IF NOT MISSING(X5);
RUN;

PROC MEANS 
    DATA=TEST_MEANS NOPRINT; 
    VAR _NUMERIC_;
    OUTPUT OUT=MEANSWIDEPCTLS P5= P25= P50= P75= P95= / AUTONAME;
RUN;

However, something to do with how SAS labels missing values as -inf, when I compare the results above, to the results produced in excel and proc means, they aren't aligned, can someone confirm which would be correct?

This type of things occurs when you have a lot of ties. How many ties are in your data and how do they impact the results? — Reeza, Nov 18 '18 at 19:28
And percentiles don't have a 'common' definition, so there isn't a right, there's a what's appropriate for you data and which one you want to use. Pick one and make it clear what definition you used. — Reeza, Nov 18 '18 at 19:29

score 1 · Accepted Answer · answered Nov 18 '18 at 16:11

1

You are using pctldef=3 in PROC STDIZE but the default definition for PROC MEANS, and that is 5. I tested your code with PCTLDEF=3 using PROC MEANS and get matching results.

answered Nov 18 '18 at 16:11

data _null_

8,534
12
14

Thanks for the clarification, what concerns me is the difference between the options (over 20 units in some cases). How can I trust this output and how can i distinguish which is most accurate? – 78282219 Nov 18 '18 at 16:54
1

@78282219 Just look at the definition of the algorithms used by the different options of the PCTLDEF= option and pick the one that you want to use. – Tom Nov 19 '18 at 04:12
I find that pctldef =5 replicates proc univariate and I am content – 78282219 Nov 19 '18 at 15:24

proc stdize producing wrong value for percentiles

1 Answers1