topic Re: Percentiles from proc surveymeans differ slightly from proc means in SAS Procedures

Percentiles from proc surveymeans differ slightly from proc means

Monicapn — Tue, 18 Dec 2018 17:15:22 GMT

Hi,

I am getting slightly different values for percentiles of a continuous variable when I use proc surveymeans vs. proc means. I meant to calculate crude geometric means for my variable (Serum_conc) and also get percentiles of its distribution, so I did not enter any sampling weights, stratum, cluster or finite population correction into proc surveymeans. The reason to use proc surveymeans is to calculate survey-method adjusted GMs and compare them to the crude estimates, so I started with the crude GM calculations and also included percentiles. I expected proc surveymeans to calculate identical percentiles for the same variable as proc means, but the values differ slightly in the decimals -- do the two procedures use different default calculations for percentiles? My code is below.

PROC MEANS DATA=datafile n nmiss min p25 p50 p75 p90 p95 max;
var Serum_conc;
RUN;

PROC SURVEYMEANS DATA=datafile GEOMEAN GMCLM NMISS percentile=(25,50,75,90,95) max;
VAR Serum_conc;
ODS OUTPUT Statistics=dataset1 GeometricMeans=dataset2; *Specify ODS table names as output datasets;
RUN;
*Print temporary datasets/tables;
PROC PRINT data=dataset1;
RUN;
PROC PRINT data=dataset2;
RUN;

Thank you in advance.

Re: Percentiles from proc surveymeans differ slightly from proc means

ballardw — Tue, 18 Dec 2018 17:58:15 GMT

The documentation for Surveymeans includes:

PROC SURVEYMEANS uses Woodruff’s method (Dorfman and Valliant 1993; Särndal, Swensson, and Wretman 1992; Francisco and Fuller 1991) to estimate the variances of quantiles. For more details, see the section Quantiles.

So I suspect that percentiles are calculated somewhat differently to account for the additional things Surveymeans has to address.

Re: Percentiles from proc surveymeans differ slightly from proc means

Monicapn — Tue, 18 Dec 2018 18:52:46 GMT

Thank you. I had found that in the SAS documentation for each procedure -- the formula from proc means is pretty straight forward and the formula from proc surveymeans uses probability. We decided to go ahead and use the proc surveymeans procedure for the percentiles, since we are using it to estimate the geometric means, for consistency.